text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'Let $G$ be a connected quasi-split reductive group over ${\mathbb{R}}$, and more generally, a quasi-split $K$-group over ${\mathbb{R}}$. Arthur had obtained the formal formula for the spectral side of the stable local trace formula, by using formal substitute of Langlands parameters. In this paper, we construct the spectral side of the stable trace formula and endoscopy trace formula directly for quasi-split $K$-groups over ${\mathbb{R}}$, by incorporating the works of Shelstad. In particular we give the explicit expression for the spectral side of the stable local trace formula, in terms of Langlands parameters.'
address:
-
- 'Institute of Mathematics, Academia Sinica, Taipei'
- 'Department of Mathematics, National University of Singapore'
author:
- 'Chung Pang Mok, Zhifeng Peng\*'
title: The spectral side of stable local trace formula for real groups
---
[^1]
Introduction {#sect: introduction}
============
In this paper, which is a sequel to [@P], we give the explicit formula for spectral side of stable local trace formula of a connected quasi-split reductive group over ${\mathbb{R}}$, and more generally, a quasi-split $K$-group over ${\mathbb{R}}$.
In general, the local trace formula is an identity, one side which is called the geometric side, is constructed by the semisimple orbital integrals; the other side, which is called the spectral side, is constructed in terms of tempered characters. Arthur [@A6] has obtained the stabilization of the geometric side, and consequently obtained the formal formula for the spectral side of the stable trace formula. However, the stable distributions and the coefficients that occurred in the formal formula for the spectral side, are not explicit.
By combining with Shelstad’s works [@S1; @S2; @S3], we will directly stabilize the spectral side of the local trace formula, which in particular give the explicit formula for the spectral side of the stable local trace formula, in terms of Langlands parameters.\
In more details, let $G$ be a quasi-split $K$-group over ${\mathbb{R}}$, a notion for which we refer to Section 1 of [@A4] (where it is called [*multiple groups*]{}) or section 2.2 of [@P], and $f$ a test function on $G({\mathbb{R}})$ with central character $\zeta$. The endoscopic decomposition of the spectral side of the invariant local trace formula, as obtained in [@A6], takes the following form: $$I^G_{{\operatorname{disc}}}(f)=\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}(G)}\iota(G,G^{\prime})\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}})$$ where $\mathcal{E}_{{\operatorname{ell}}}(G)$ is the set of $\widehat{G}$-equivalence classes of elliptic endoscopic data, and $f^{G^{\prime}}$ is the Langlands-Shelstad transfer of $f$ to $G^{\prime}$ [@S1]. One has the formal formula: $$\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}})=\int_{\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)}s^{G^{\prime}}(\phi^{\prime})f^{G^{\prime}}(\phi^{\prime})d\phi^{\prime}$$ with $\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$ being defined only in terms of formal substitute of Langlands parameters $\phi^{\prime}$ of $G^{\prime}$, and the coefficients $s^{G^{\prime}}(\phi^{\prime})$ are not explicit. Our task is to show that $\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$ can be taken as actual Langlands parameters, and also to give explicit formula for the coefficients in terms of Langlands parameters; [*c.f.*]{} Section 4 for the definition of these terms.\
In the global context of automorphic representations, the method of [@A1; @A9] is based on the comparison of the spectral and endoscopic objects for the global trace formula for $G$. This could be expressed schematically in terms of (conjectural) Langlands parameters: $$(M,\phi_{M})\longrightarrow(\phi,s)\longleftarrow (G^{\prime},\phi^{\prime})$$
In the archimedean local setting, the Langlands parametrization is available for general $G$. This allows us to adapt the comparison process to the setting of local trace formula:
$$(\tau)\longrightarrow(\phi,s)\longleftarrow (G^{\prime},\phi^{\prime})$$ in order to give the explicit construction of the spectral side of stable local trace formula.
We now give more details for the comparison process. Firstly recall the classification theory of tempered representations, due to Harish-Chandra. The tempered representations can be classified by triplets $\tau=(M,\pi,r)$, where $M$ is the Levi subgroup, $\pi\in \Pi_{2}(M)$ is square integrable modulo the split centre of $M$, and $r\in R_{\pi}$, the representation theoretic $R$-group of $\pi$, which is a finite abelian elementary $2$-group. For a test function $f= f_1 \times \bar{f}_2$, with $f_1,f_2 \in \mathcal{H}(G({\mathbb{R}}),\zeta)$, the Hecke space with central character $\zeta$, the spectral side of the invariant local trace formula for $G$ takes the following form [@A2]: $$I^G_{{\operatorname{disc}}}(f)=\int_{T_{{\operatorname{disc}}}(G,\zeta)}i^{G}(\tau) f_{G}(\tau)|R_{\pi}|^{-1}d{\tau}.$$
where $f_{G}(\tau)= \Theta(\tau,f_1) \Theta(\tau^{\vee},\bar{f}_2)$, and the coefficient $$i^{G}(\tau)=|W^{\circ}_{\pi}|^{-1}\sum_{w \in W_{\pi}(r)_{{\operatorname{reg}}}}\varepsilon_{\pi}(w)|\det(w-1)_{\mathfrak{a}^{G}_{M}}|^{-1},$$ encodes combinatorial data from Weyl groups that is relevant to the comparison of global trace formulas [@A1; @A9]; [*c.f.*]{} Section 3 for the definition of these terms.\
The first step in the process of stabilization is to express the coefficient $i^G(\tau)$ in terms of data defined in terms of Langlands parameters for $G$, and more precisely, defined in terms of the data $(\phi,s)$ above, an important point being that, by the works of Knapp-Zuckerman [@KZ] and Shelstad [@S1], the representation theoretic $R$-group is canonically isomorphic to the endoscopic $R$-group defined in terms of Langlands parameters; [*c.f.*]{} Section 4. $$i^{G}(\tau)=|W^{\circ}_{\phi}|^{-1}\sum_{w\in W_{\phi}(x)_{{\operatorname{reg}}}}s^{\circ}_{\phi}(w)|\det(w-1)_{\mathfrak{a}^{G}_{M}}|^{-1}.$$ where $x$ is the image of $s$ in $\mathcal{S}_{\phi} = \pi_0(\overline{S}_{\phi})$. In turn, this can be expressed in terms of the constants $\sigma(\overline{S}_{\phi,s}^{\circ})$ as defined in [@A1]: $$i^G(\tau) = i_{\phi}(x)= \sum_{s\in \mathcal{E}_{\phi,{\operatorname{ell}}}(x)}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ}).$$
Following [@A1], one defines the following subsets $\Phi_{s-{\operatorname{disc}}}(G,\zeta)\subset \Phi_{{\operatorname{disc}}}(G,\zeta)$ of the set $\Phi(G,\zeta)$ of Langlands parameters for $G$ (with central character $\zeta$), as: $$\Phi_{s-{\operatorname{disc}}}(G,\zeta)=\{\phi\in\Phi(G,\zeta):Z(\overline{S}_{\phi}^{\circ})<\infty\},$$ $$\Phi_{{\operatorname{disc}}}(G,\zeta)=\{\phi\in\Phi(G,\zeta):Z(\overline{S}_{\phi})<\infty\}.$$ One has the fact that the constants $\sigma(\overline{S}_{\phi,s}^{\circ})$ vanish if $Z(\overline{S}_{\phi,s}^{\circ})$ is not finite; in particular that the constants $\sigma(\overline{S}_{\phi,s}^{\circ})$ vanish if $\phi \notin \Phi_{s-{\operatorname{disc}}}(G,\zeta)$.
Having defined the basic endoscopic and stable objects on the spectral side, we can define the spectral transfer factors, in the spirit of Section 5 of [@A3], by combining the classification of tempered representations and Shelstad’s definition of spectral transfer factors [@S2; @S3]. For instance, in the case where $\phi \in \Phi(G,\zeta)$ is elliptic, we define, for $\tau =(M,\pi,r)$ and $s \in \overline{S}_{\phi}$ ([*c.f.*]{} Section 5): $$\Delta(\tau,\phi^{s})=\sum_{\chi\in\widehat{R}_{\phi}} \chi(r) \Delta(\Pi^{\chi},\phi^s).$$
We then have the spectral transfer: $$\Theta(\tau,f)=\sum_{x \in\mathcal{S}_{\phi}}\Delta(\tau,\phi^{x})f^{\prime}(\phi,x).$$
We then obtain the first main theorem of the paper in Section 6:
If $f=f_{1}\times\bar{f}_{2}, f_{i}\in \mathcal{H}(G({\mathbb{R}}),\zeta), i=1,2,$ then $$I^G_{{\operatorname{disc}}}(f) =\int_{\Phi_{{\operatorname{disc}}}(G,\zeta)}\sum_{s\in \mathcal{E}_{\phi,{\operatorname{ell}}}}|\mathcal{S}_{\phi}|^{-1}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ}) f^{\prime}_{1}(\phi,s)\overline{f^{\prime}_{2}(\phi,s)}d\phi .$$
To obtain the explicit formula for the spectral side of the endoscopy local trace formula and the stable local trace formula, we use the arguments of Chapter 4 of [@A9]; one analyzes the coefficients by using the bijective correspondence: $$\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta) \longleftrightarrow \widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta) ,$$ between the set of $\widehat{G}$-conjugacy classes of $X_{{\operatorname{disc}}}(G,\zeta)=\{(\phi,s):\phi\in\Phi_{{\operatorname{disc}}}(G,\zeta), s\in S_{\phi,{\operatorname{ell}}}\}$, and the set of $\widehat{G}$-conjugacy classes of $Y_{{\operatorname{disc}}}(G,\zeta) = \{(G^{\prime},\phi^{\prime}):G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}(G),\phi^{\prime}\in\Phi_{s-{\operatorname{disc}}}(G^{\prime}, \zeta) \}$ (see Section 7 for details).
We then obtain the following in Section 7:
If $f=f_{1}\times \bar{f}_{2}$, $f_{i}\in \mathcal{H}(G({\mathbb{R}}),\zeta),i=1,2$. Then we have $$I^G_{{\operatorname{disc}}}(f)=\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}(G)}\iota(G,G^{\prime}) \widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}})$$ where $$\iota(G,G^{\prime})=|{\operatorname{Out}}_{G}(G^{\prime})|^{-1}|\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}|^{-1},$$ $$\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}})=\int_{\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)}|\mathcal{S}_{\phi^{\prime}}|^{-1}\sigma(\overline{S}_{\phi^{\prime}}^{\circ})f_{1}^{G^{\prime}}(\phi^{\prime})\overline{f_{2}^{G^{\prime}}(\phi^{\prime})}d\phi^{\prime}.$$
The explicit formula for the spectral side of the stable trace formula will then be obtained in Section 8.
Here is the summary of the contents of the paper. After introducing preliminaries and notations in Section 2, we recall some formulation of the invariant local trace formula of Arthur in Section 3. Then in Section 4 we introduce the basic objects that occur in the spectral side of the endoscopic local trace formula and the stable trace formula. We will study the properties of the spectral transfer factors in Section 5. We then obtain the main theorem on stabilization of the local trace formula in Section 6. The explicit formula for the spectral side of the endoscopic and stable local trace formula are then obtained in Section 7 and 8.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to thank the referee for the careful reading of the manuscript. For this work, Zhifeng Peng had been supported by the National Natural Science Foundation of the People’s Republic of China, Grant No.11601503.
Preliminaries and notation {#preliminaries and notation}
==========================
Throughout the paper $G$ is a $K$-group over ${\mathbb{R}}$, which will be assumed to be quasi-split from Section 4 onwards. The center of $G$ is noted as $Z(G)$, while $Z$ stands for a fixed central induced torus in $G$ over ${\mathbb{R}}$, and $\zeta$ is a character on $Z({\mathbb{R}})$. Let $$\mathfrak{a}_{G}={\operatorname{Hom}}(X(G)_{{\mathbb{R}}},{\mathbb{R}})$$ be the real vector space dual of the module $X(G)_{{\mathbb{R}}}$ of ${\mathbb{R}}$ rational characters on $G$. There is a canonical homomorphism $$H_{G}:G({\mathbb{R}})\rightarrow \mathfrak{a}_{G}$$ defined by $$e^{\langle H_{G}(x),\chi\rangle}=|\chi(x)|, x\in G({\mathbb{R}}),\chi\in X(G)_{{\mathbb{R}}},$$ where $|\cdot|$ is the absolutely valuation on ${\mathbb{R}}$. Let $A_{G}$ be the split component of the center of $G$, then $$\mathfrak{a}_{G}=H_{G}(G({\mathbb{R}}))=H_{G}(A_{G})$$ and $\mathfrak{a}_G$ is the Lie algebra of $A_G$. It is convenient to fix a Haar measure on $\mathfrak{a}_{G}$. This determines a dual Haar measure on the real vector space $i\mathfrak{a}^{\ast}_{G}$. It also determines a unique Haar measure on $A_{G}({\mathbb{R}})$. We denote $\mathfrak{a}^{\ast}_{G,Z}$ by the subspace of linear forms on $\mathfrak{a}_{G}$ that are trivial on the image of $\mathfrak{a}_{Z}$ in $\mathfrak{a}_{G}$.
We recall some settings as in [@A2]. Denote by $\mathcal{L}^{G}(M)$ the finite set of Levi subgroups of $G$ which contain a given Levi subgroup $M$. We let $M_{0}$ be a fixed Levi component of some minimal parabolic subgroup of $G$. Put $\mathcal{L} := \mathcal{L}^G(M_0)$. For $M \in \mathcal{L}$, denote by $\mathcal{P}^G(M)$ the set of parabolic subgroups of $G$ having $M$ as Levi component. Define $\Pi_{2}(M({\mathbb{R}}))$ to be the set of (equivalence classes of) representations that are square integrable modulo the split center of $M$, and $\Pi_{{\operatorname{temp}}}(M({\mathbb{R}}))$ to be the set of (equivalence classes of) tempered representations.
For $M\in \mathcal{L}^{G}(M_{0})$ and $\pi \in \Pi_{2}(M({\mathbb{R}}))$, denote by $\Pi_{\pi}(G({\mathbb{R}}))$ the set of irreducible constituents of the induced representation $I_{P}(\pi)$, which is a finite subset of $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}))$ and independent of the parabolic subgroup $P$. The sets $\Pi_{\pi}(G({\mathbb{R}}))$ exhaust $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}))$. The classification of the representations in $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}))$ is reduced to classifying the representations in the finite sets $\Pi_{\pi}(G({\mathbb{R}}))$, and to determining the intersection of any two such sets. The second question is answered by Harish-Chandra’s work.
Write $W^{G}_{0}$ for the Weyl group of the pair $(G,A_{M_0})$; for $w \in W_0^G$, we generally write $\tilde{w}$ for any representative of $w$ in $K$; here $K$ is a (fixed) maximal compact subgroup $G({\mathbb{R}})$ that is in good position relative to $M_0({\mathbb{R}})$. If $M\in \mathcal{L}^{G}(M_{0})$ and $\pi \in \Pi_{2}(M({\mathbb{R}}))$, $wM=\tilde{w}M\tilde{w}^{-1}$ is another Levi subgroup, and $$(w \pi)(m^{\prime})=\pi(\tilde{w}^{-1}m^{\prime}\tilde{w}), \quad m^{\prime}\in(wM)({\mathbb{R}})$$ is a representation in $\Pi_{2}((wM)({\mathbb{R}}))$. We obtain an action $$(M,\pi)\rightarrow(wM, w\pi), w\in W^{G}_{0},$$ of $W^{G}_{0}$ on the set of pairs $(M,\pi), M \in\mathcal{L}^{G}(M_{0}),\pi \in\Pi_{2}(M({\mathbb{R}})).$
As stated in Proposition 1.1 of [@A2], one has the following: If $(M,\pi)$ and $(M^{\prime},\pi^{\prime})$ are any two pairs, and $(M^{\prime},\pi^{\prime})$ equals $(wM,w\pi)$ for an element $w\in W^{G}_{0},$ the subsets $\Pi_{\pi}(G({\mathbb{R}}))$ and $\Pi_{\pi^{\prime}}(G({\mathbb{R}}))$ of $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}))$ are identical. Conversely, if the sets $\Pi_{\pi}(G({\mathbb{R}}))$ and $\Pi_{\pi^{\prime}}(G({\mathbb{R}}))$ have a representation in common, there is an element $w\in W^{G}_{0}$ such that $(M^{\prime},\pi^{\prime})=(wM,w\pi)$.
In this paper we shall fix the central data $(Z,\zeta)$. Thus $\mathcal{H}(G({\mathbb{R}}),\zeta)$ is the Hecke space of smooth functions with compact support $f$ on $G({\mathbb{R}})$ that are left and right finite under the maximal compact subgroup $K$, and such that $f(zx) = \zeta(z)^{-1} f(x)$ for $z \in Z({\mathbb{R}})$ and $x \in G({\mathbb{R}})$. Similarly we define $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}),\zeta)$ to be the subset of $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}))$ consisting of those representations whose character on $Z({\mathbb{R}})$ is equal to $\zeta$. Similarly one defines the subsets $\Pi_2(G({\mathbb{R}}),\zeta) := \Pi_2(G({\mathbb{R}})) \cap \Pi_{{\operatorname{temp}}}(G({\mathbb{R}}),\zeta)$ and $\Pi_{\pi}(G({\mathbb{R}}),\zeta) := \Pi_{\pi}(G({\mathbb{R}})) \cap \Pi_{{\operatorname{temp}}}(G({\mathbb{R}}),\zeta)$ etc.
We can understand the finite set $\Pi_{\pi}(G({\mathbb{R}}),\zeta)$ by the representation theoretic $R$-group [@KZ]. More generally, we can parameterize the characters of the tempered representations $\Pi_{{\operatorname{temp}}}(G({\mathbb{R}}),\zeta)$ by the virtual characters of the sets $T(G,\zeta)$, where $T(G,\zeta)$ is the set of $G$-equivalence classes of the sets $\widetilde{T}(G,\zeta)$, with $\widetilde{T}(G,\zeta)$ being defined as $\widetilde{T}(G,\zeta)=\{\tau=(M,\pi,r): M\in \mathcal{L}^G(M_{0}),\pi\in \Pi_{2}(M({\mathbb{R}}),\zeta),r\in R_{\pi} \}$. Here $\Pi_{2}(M({\mathbb{R}}),\zeta)$ is as before the subset of $\Pi_2(M({\mathbb{R}}))$ consisting of those $\pi$ whose character on $Z({\mathbb{R}})$ is equal to $\zeta$. Finally the representation theoretic $R$-group $R_{\pi}$ of $\pi$ is defined as the quotient of $W_{\pi}$ by $W^{\circ}_{\pi}$, where $$W_{\pi}=\{w\in W(\mathfrak{a}_{M}):w\pi\cong\pi\}$$ is the stabilizer of $\pi$ in the Weyl group of $\mathfrak{a}_{M}$, and $W^{\circ}_{\pi}$ is the subgroup of elements $w$ in $W_{\pi}$ such that the operator $R(w,\pi)$ is a scalar ([*c.f.*]{} below). It is known that $W^{\circ}_{\pi}$ is a normal subgroup of $W_{\pi}$. The group $W_{\pi}^{\circ}$ is the Weyl group of a root system, composed of scalar multiples of those reduced roots $\alpha$ of $(G,A_{M})$ for which the reflection $w_{\alpha}$ belongs to $W^{\circ}_{\pi}$. These roots divide the vector space $\mathfrak{a}_{M}$ into chambers. By fixing such a chamber $\mathfrak{a}_{\pi}$, we can identify $R_{\pi}$ with the subgroup of elements in $W_{\pi}$ that preserve $\mathfrak{a}_{\pi}$.
The operator $$R(w,\pi)=A(\pi_{w})R_{\tilde{w}^{-1}P\tilde{w}|P}(\pi), \quad w\in W_{\pi}, \pi\in \Pi_{2}(M), P\in \mathcal{P}^G(M)$$ is an intertwining operator from $I_{P}(\pi)$ to itself. Here to define $A(\pi_w)$, first note that $\pi$ can be extended to a representation of the group $M^{\prime}({\mathbb{R}})$ generated by $M({\mathbb{R}})$ and $\tilde{w}$. We denote $\pi_{w}$ by such an extension, then the intertwining operator $$A(\pi_{w}): I_{\tilde{w}^{-1}P\tilde{w}}(\pi)\longrightarrow I_{P}(\pi), \pi\in W_{\pi}$$ between $I_{\tilde{w}^{-1}P\tilde{w}}(\pi)$ and $I_{P}(\pi)$ is defined by setting $$(A(\pi_{w})\phi^{\prime})(x)=\pi_{w}(\tilde{w})\phi^{\prime}(\tilde{w}^{-1}x), \phi^{\prime}\in I_{\tilde{w}^{-1}P\tilde{w}}(\pi).$$ Finally the intertwining operator $R_{\tilde{w}^{-1}P\tilde{w}|P}(\pi)$, and more generally $$R_{Q|P}(\pi)=r_{Q|P}(\pi)^{-1}J_{Q|P}(\pi): I_{P}(\pi)\longrightarrow I_{Q}(\pi),\mbox{ for }P,Q\in \mathcal{P}^G(M),$$ is the normalized intertwining operator between the induced representations $I_{P}(\pi)$ and $I_{Q}(\pi)$, with $r_{Q|P}(\pi)$ being the normalizing factors; see the discussion on p. 85 of [@A2]. In addition, since we are in the archimedean case, the operators $R(w,\pi)$ can be normalized so that $R(w,\pi)$ is the identity for $w \in W_{\pi}^{\circ}$ (hence $R(r,\pi)$ is well-defined for $r \in R_{\pi}$), and such that the map $r \mapsto R(r,\pi)$ is a homomorphism on $R_{\pi}$, [*c.f.*]{} p. 86 of [@A2]. We will always work with such a normalization in what follows.
Local trace formula
===================
In this section we recall the formalism of the local trace formula of Arthur [@A2], specifically in the archimedean case. Consider a test function $f=f_{1}\times \bar{f}_{2}$, with $f_{i}\in \mathcal{H}(G({\mathbb{R}}),\zeta)$ being in the Hecke space of test functions. Following Arthur, it is customary to think of the components $f_1,f_2$ of $f$ as being indexed by the two elements set $V=\{\infty_1,\infty_2\}$ (being regarded as two archimedean places). The local trace formula is given by the identity $I^G_{{\operatorname{disc}}}(f) =I^G(f)$, [*c.f.*]{} Theorem 4.2 of [@A2] and Proposition 6.1 of [@A6]. The geometric side of the local trace formula is given by: $$I^G(f)=\sum_{M\in\mathcal{L}}|W^{M}_{0}||W^{G}_{0}|^{-1}(-1)^{{\operatorname{dim}}(A_{M}/ A_{G})}\int_{\Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)}I^G_{M}(\gamma,f) d\gamma$$ defined in terms of the invariant distributions $I^G_{M}(\gamma,f)$ ([*c.f. loc. cit.*]{}) For the definition of $\Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)$, firstly, for a fixed basis $\Gamma(M,\zeta)$ of the space of invariant distributions $\mathcal{D}(M,\zeta)$ on $M({\mathbb{R}})$ introduced in Section 1 of [@A5] (which in particular are $\zeta$-equivariant under translation by $Z({\mathbb{R}})$), one has the subset $ \Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,\zeta)$ of $\Gamma(M,\zeta)$ consisting of elements that are strongly $G$-regular, elliptic support in $M({\mathbb{R}})$. Put $\Gamma(M_V,\zeta_V)= \Gamma(M,\zeta) \times \Gamma(M,\zeta)$ (corresponding to the two places $\infty_1,\infty_2$ in $V$). Then $\Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)$ is identified with the diagonal image of $\Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,\zeta)$ in $\Gamma(M_V,\zeta_V)$: $$\Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta) = \{(\gamma,\gamma):\gamma \in \Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,\zeta)\}.$$
The other side is the spectral side given by: $$I^G_{{\operatorname{disc}}}(f)=\int_{T_{{\operatorname{disc}}}(G,\zeta)}i^{G}(\tau) f_{G}(\tau) |R_{\pi}|^{-1} d\tau.$$ Here: $$T_{{\operatorname{disc}}}(G,\zeta)=\{\tau=(M,\pi,r)\in T(G,\zeta): W_{\pi}(r)_{{\operatorname{reg}}}\neq\emptyset, \pi\in\Pi_{2}(M({\mathbb{R}}),\zeta)\},$$
$W_{\pi}(r)_{{\operatorname{reg}}}=W_{\pi}(r)\cap W_{\pi,{\operatorname{reg}}}, W_{\pi}(r)=W_{\pi}^{\circ} \cdot r$ , $W_{\pi,{\operatorname{reg}}}=\{w\in W_{\pi}:\mathfrak{a}^{w}_{M}=\mathfrak{a}_{G}\},$ and $$i^{G}(\tau)=|W^{\circ}_{\pi}|^{-1}\sum_{w\in W_{\pi}(r)_{{\operatorname{reg}}}}\varepsilon_{\pi}(w)|\det(w-1)_{\mathfrak{a}^{G}_{M}}|^{-1}$$ which encode combinatorial data from Weyl groups. The sign $\varepsilon_{\pi}(w)$ stands for the sign of projection of $w$ onto the Weyl group $W^{\circ}_{\pi}$ taken relative to the decomposition $W_{\pi}=W_{\pi}^{\circ}\rtimes R_{\pi}$, and $\mathfrak{a}_M^G$ is the quotient of $\mathfrak{a}_M$ by $\mathfrak{a}_G$.
The other terms are defined as:
$$f_{G}(\tau)=\Theta(\tau,f_1) \Theta(\tau^{\vee},\bar{f}_2) = \Theta(\tau,f_1) \overline{\Theta(\tau,f_2) }$$
$$\Theta(\tau,f_i)={\operatorname{tr}}(R(r,\pi)I_{P}(\pi,f_{i})), \,\ i=1,2.$$
Finally the measure $d\tau$ on $T_{{\operatorname{disc}}}(G,\zeta)$ is defined by the formula ([*c.f.*]{} equation (3.5) of [@A2]):
$$\begin{aligned}
\int_{T_{{\operatorname{disc}}}(G,\zeta)}f(\tau)d\tau &=&\sum_{\tau\in T_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}} \int_{i\mathfrak{a}^{\ast}_{G,Z}}f(\tau_{\lambda})d\lambda.\end{aligned}$$
for $f \in C_c(T_{{\operatorname{disc}}}(G,\zeta))$.
Here when compared to equation (3.5) of [@A2], we first note that since we are in the archimedean case, the groups $R_{\pi}$ are abelian, and hence the groups $R_{\pi,r}$ in [*loc. cit.*]{}, namely the centralizer of $r$ in $R_{\pi}$, reduces to $R_{\pi}$.
Secondly, for our purpose, it would be more convenient to [*not*]{} to absorb the factor $|R_{\pi,r}|^{-1}=|R_{\pi}|^{-1}$ into the definition of the measure $d \tau$ on $T_{{\operatorname{disc}}}(G,\zeta)$, as was done in [@A2].
Endoscopic and stable objects on the spectral side
==================================================
From now on $G$ will always be assumed to be a quasi-split $K$-group. We first recall the Langlands parameters. These are admissible continuous homomorphisms: $$\phi: W_{{\mathbb{R}}} \longrightarrow {}^{L}G$$ where as usual ${}^{L}G = \widehat{G} \rtimes W_{{\mathbb{R}}}$ is the $L$-group of $G$, defined with respect to a splitting of $G$ (that we fix for the rest of the paper); $W_{{\mathbb{R}}}$ is the Weil group of ${\mathbb{R}}$: it is a non-split extension $1 \rightarrow {\mathbb{C}}^{\times} \rightarrow W_{{\mathbb{R}}} \rightarrow \Gamma_{{\mathbb{R}}} \rightarrow 1$, with $\Gamma_{{\mathbb{R}}}={\operatorname{Gal}}({\mathbb{C}}/{\mathbb{R}})$. The parameter $\phi$ is called bounded, if the image of $W_{{\mathbb{R}}}$ in $\widehat{G}$ is bounded. We denote by $\Phi(G)$ for the set of $\widehat{G}$-equivalence classes of bounded parameters (with respect to the conjugation action by $\widehat{G}$). For $\phi \in \Phi(G)$, we denote by $\Pi_{\phi}$ the $L$-packet of tempered representations of $G({\mathbb{R}})$ associated to $\phi$. The stable character $f \mapsto f(\phi):=\sum_{\Pi \in \Pi_{\phi}} {\operatorname{tr}}\Pi(f) $ is then a stable distribution on $G({\mathbb{R}})$ [@S1].
We denote by $\Phi_{2}(G)$, the set of (equivalence classes of) square-integrable parameters, for the subset of $\phi\in\Phi(G)$ that does not factor through ${}^{L}M$, for any proper Levi subgroup $M$ of $G$. For $\phi \in \Phi_2(G)$, the $L$-packet $\Pi_{\phi}$ consists of square-integrable representations of $G({\mathbb{R}})$.
For any $\phi\in \Phi(G)$, we set $$S_{\phi}={\operatorname{Cent}}(Im\phi,\widehat{G}),$$ $$\overline{S}_{\phi}=S_{\phi}/ Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}},$$ and $$\mathcal{S}_{\phi}=\pi_{0}(\overline{S}_{\phi}).$$ Since we are in the archimedean case, the component group $\mathcal{S}_{\phi}$ is a finite abelian elementary $2$-group [@S1]. In addition, since we are working in the context of a quasi-split $K$-group, one has that $\Pi_{\phi}$ is in bijection with the set of characters of $\mathcal{S}_{\phi}$; hence the cardinality of $\Pi_{\phi}$ is equal to the order of $\mathcal{S}_{\phi}$ [@S1; @S3].
One has $$\Phi_2(G) = \{\phi \in \Phi(G), |\overline{S}_{\phi}| < \infty \}.$$
We also define the following subsets of parameters $ \Phi_{{\operatorname{disc}}}(G) , \Phi_{s-{\operatorname{disc}}}(G), \Phi_{{\operatorname{ell}}}(G),$ of $\Phi(G)$: $$\Phi_{{\operatorname{disc}}}(G)=\{\phi \in \Phi(G),|Z(\overline{S}_{\phi})|<\infty\},$$ with $Z(\overline{S}_{\phi}) := {\operatorname{Cent}}(\overline{S}_{\phi},\overline{S}_{\phi}^{\circ})$, and:
$$\Phi_{s-{\operatorname{disc}}}(G)=\{\phi \in \Phi(G),|Z(\overline{S}^{\circ}_{\phi})|<\infty\},$$ with $Z(\overline{S}_{\phi}^{\circ})$ being the usual center of $\overline{S}_{\phi}^{\circ}$.
The set of elliptic parameters $\Phi_{{\operatorname{ell}}}(G)$ is defined as the subset of $\phi \in \Phi(G)$, such that $|\overline{S}_{\phi,s}|<\infty$ for some semi-simple element $s \in \overline{S}_{\phi}$; here $\overline{S}_{\phi,s}$ is being defined as: $$\overline{S}_{\phi,s} := {\operatorname{Cent}}(s,\overline{S}_{\phi}^{\circ} ).$$
One has: $$\Phi_2(G) \subset \Phi_{s-{\operatorname{disc}}}(G) \subset \Phi_{{\operatorname{disc}}}(G),$$
$$\Phi_2(G) \subset \Phi_{{\operatorname{ell}}}(G) \subset \Phi_{{\operatorname{disc}}}(G).$$
Also, for a central data $(Z,\zeta)$ of $G$ as before, we denote by $\Phi(G,\zeta)$ the set of parameters $\phi \in \Phi(G)$ that have character $\zeta$ with respect to $Z$, in the sense that the composition: $$W_{{\mathbb{R}}} \stackrel{\phi}{\rightarrow} {}^{L} G \rightarrow {}^{L} Z$$ corresponds to the character $\zeta$ of $Z$. Similar definition for the sets $\Phi_2(G,\zeta),\Phi_{{\operatorname{ell}}}(G,\zeta)$ etc.
We now define the endoscopic $R$-group, $$R_{\phi} :=W_{\phi}/ W^{\circ}_{\phi},$$ where the Weyl groups $W_{\phi},W_{\phi}^{\circ}$ are defined as: $$W_{\phi}={\operatorname{Norm}}(\overline{\mathcal{T}}_{\phi},\overline{S}_{\phi})/\overline{\mathcal{T}}_{\phi}.$$ Here $\overline{\mathcal{T}}_{\phi}$ is defined as $A_{\widehat{M}}/(A_{\widehat{M}}\cap Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}})$, with $A_{\widehat{M}}=(Z(\widehat{M})^{\Gamma_{{\mathbb{R}}}})^{\circ}$, and $M$ being the Levi subgroup of $G$ (which is unique up to conjugation) such that $\phi$ factors through ${}^{L}M$ as a square integrable parameter $\phi_M \in \Phi_2(M,\zeta)$ of $M$. Similarly $$W_{\phi}^{\circ}={\operatorname{Norm}}(\overline{\mathcal{T}}_{\phi},\overline{S}^{\circ}_{\phi})/\overline{\mathcal{T}}_{\phi}.$$
One also has, by the results in section 5 of [@S1], the split short exact sequence: $$\begin{aligned}
0 \rightarrow \mathcal{S}_{\phi_M} \rightarrow \mathcal{S}_{\phi} \rightarrow R_{\phi} \rightarrow 0.\end{aligned}$$
For $\phi \in \Phi(G,\zeta)$, we denote by $$T_{\phi}=\{\tau=(M,\pi,r) \in T(G,\zeta), \mbox{ such that } \pi\in\Pi_{\phi_{M}} \} .$$
If $\phi\in\Phi(G,\zeta)$, then we have canonical identification $R_{\pi}=R_{\phi}, W_{\pi}=W_{\phi}, W^{\circ}_{\pi}=W^{\circ}_{\phi}$, for $\tau = (M,\pi,r) \in T_{\phi}$.
Thus we have a natural surjective map of sets $T_{\phi} \rightarrow R_{\phi}$ by sending $\tau = (M,\pi,r)$ to $r \in R_{\pi} = R_{\phi}$.
We can also define a non-canonical bijection $\iota:T_{\phi} \rightarrow \mathcal{S}_{\phi}$ (which we fix once and for all) that respects the natural projection map to $R_{\phi}$.
This follow from the results of Knapp-Zuckerman [@KZ] and Shelstad [@S1]. If $\phi \in \Phi_{{\operatorname{ell}}}(G,\zeta)$ is elliptic, see the discussion before Proposition 5.2 of [@P]. In general, if $\phi\in\Phi(G,\zeta)$, then there is a Levi subgroup $\widetilde{M}$ of $G$ (unique up to conjugacy), that is maximal with respect to the property that the parameter $\phi$ factors through ${}^{L}\widetilde{M}$ as an elliptic parameter $\phi^{\widetilde{M}}$ of $\widetilde{M}$ (see section 5 of [@S1]). Under the Levi embedding ${}^{L} \widetilde{M} \hookrightarrow {}^{L}G$ on the dual side, have a canonical isomorphism $$\mathcal{S}_{\phi^{\widetilde{M}}}\longrightarrow\mathcal{S}_{\phi}.$$ However $\phi^{\widetilde{M}}$ is elliptic, and so the lemma is true for $\phi^{\widetilde{M}}$; thus we have canonical isomorphisms (noting that the $R$-group of $\pi$ with respect to $G$ is canonically identified with the $R$-group of $\pi$ with respect to $\widetilde{M}$, [*c.f. loc. cit.*]{}): $$R_{\pi}=R_{\phi^{\widetilde{M}}}=R_{\phi}.$$
To show that one has canonical identifications between $W_{\pi},W_{\phi}$ and $W_{\pi}^{\circ},W_{\phi}^{\circ}$, first note that $W_{\phi}$ is the stabilizer of $\phi_{M}$ in $W(\mathfrak{a}_M)$. It is then a consequence of the disjointness of tempered $L$-packets that $W_{\phi}$ contains $W_{\pi}$. On the other hand, we know that the intertwining operators on $I_{P}(\pi)$ coming from the subgroup $W^{\circ}_{\phi}$ of $W_{\phi}$ are scalars ([*c.f.*]{} section 5 of [@S1]). It follows that $W^{\circ}_{\phi}$ is contained in the subgroup $W^{\circ}_{\pi}$ of $W_{\pi}$. Now $R_{\pi}=W_{\pi}/W^{\circ}_{\pi}, R_{\pi}=W_{\phi}/W^{\circ}_{\phi}$, and we already know that $|R_{\pi}|=|R_{\phi}|$, so it follows that we must have $W_{\pi}=W_{\phi}$ and $W^{\circ}_{\pi}=W^{\circ}_{\phi}.$
Finally, one can thus constructs a bijection $\iota:T_{\phi} \rightarrow \mathcal{S}_{\phi}$ that respects the projection to $R_{\phi}$, from a bijection between $T_{\phi^{\widetilde{M}}}$ and $\mathcal{S}_{\phi^{\widetilde{M}}}$ that respects the projection to $R_{\phi^{\widetilde{M}}}$.
Next we recall some generalities from [@A1]. We shall consider $S$ to be any connected component of a complex reductive group. Given such $S$, we denote by $S^+$ the complex reductive group generated by $S$, and by $S^{\circ}$ the identity connected component of $S^+$. Put $$Z(S)={\operatorname{Cent}}(S,S^{\circ})$$ for the centralizer of $S$ in $S^{\circ}$. Then for a choice of maximal torus $\mathcal{T}$ of $S^{\circ}$, denote by $W(S^{\circ}) = {\operatorname{Norm}}(\mathcal{T},S^{\circ})/\mathcal{T}$ the usual Weyl group of $S^{\circ}$. We can also form the Weyl set $$W(S)={\operatorname{Norm}}(\mathcal{T},S)/ \mathcal{T}.$$ Denote by $W(S)_{{\operatorname{reg}}}$ the subset of elements $w\in W(S)$ that are regular, which means that the fixed point set of $w$ in $\mathcal{T}$ is finite. We can also regard $(w-1)$ as a linear transformation on the real vector space $$\mathfrak{a}_{\mathcal{T}}={\operatorname{Hom}}(X(\mathcal{T}),{\mathbb{R}}).$$ Then the condition for $w \in W(S)_{{\operatorname{reg}}}$ is equivalent to that $\det(w-1)_{\mathfrak{a}_{\mathcal{T}}} \neq 0$.
We denote by: $$s^{\circ}(w)=\pm 1$$ the sign of a element $w\in W$, to be $-1$ raised to the power of the number of the positive roots of $(S^{\circ},\mathcal{T})$ (with respect to some order) being mapped by $w$ to the negative roots. We then define the rational number $$i(S)=|W(S^{\circ})|^{-1}\sum_{w\in W(S)_{{\operatorname{reg}}}}s^{\circ}(w)|\det(w-1)_{\mathfrak{a}_{\mathcal{T}}}|^{-1}$$ associated to $S$.
Next we write $S_{ss}$ for the set of semisimple elements in $S$. For any $s\in S_{ss}$, we set $$S_{s}={\operatorname{Cent}}(s,S^{\circ})$$ the centralizer of $s$ in $S^{\circ}$. Then $S_{s}$ is also a complex reductive group, whose identity component is noted as: $$S^{\circ}_{s}=(S_{s})^{\circ}={\operatorname{Cent}}(s,S^{\circ})^{\circ}.$$ If $\Gamma$ is any subset of $S$ which is invariant under conjugation by $S^{\circ}$, then we shall denote by $\mathcal{E}(\Gamma)$ for the set of equivalence classes in $\Gamma_{ss}=\Gamma\cap S_{ss}$, with the equivalence relation defined by setting $s^{\prime}\sim s$ if $$s^{\prime}=s^{\circ}z s (s^{\circ})^{-1}, s^{\circ}\in S^{\circ},z\in Z(S^{\circ}_{s})^{\circ}.$$ The main interest is the subset $$S_{{\operatorname{ell}}}=\{s\in S_{ss}:|Z(S^{\circ}_{s})|<\infty\}$$ of elliptic elements of $S$. The equivalence relation on $S_{{\operatorname{ell}}}$ is then simply $S^{\circ}$-conjugation. Put: $$\mathcal{E}_{{\operatorname{ell}}}(S):=\mathcal{E}(S_{{\operatorname{ell}}}).$$
We have the following theorem which is a restatement of theorem $8.1$ of [@A1].
\[coefficient endos theorem\] There are unique constants $\sigma(S_{1})$, defined whenever $S_{1}$ is a connected complex reductive group, such that for any $S$ as above, the number $$\label{coefid}
e(S)=\sum_{s\in\mathcal{E}_{{\operatorname{ell}}}(S)}|\pi_{0}(S_{s})|^{-1}\sigma(S^{\circ}_{s})$$ equals $i(S)$, and such that $$\label{centv}
\sigma(S_{1})=\sigma(S_{1} / Z_{1})|Z_{1}|^{-1}$$ for any central subgroup $Z_{1}$ of $S_{1}$ (in particular $\sigma(S_1)=0$ if $Z_1$ is infinite).
Now back to the situation of Lemma 4.1. For $\tau \in T_{\phi}$, put $x = \iota(\tau)$. Then we can write the coefficient $i^G(\tau)$ as:
$$\begin{aligned}
i^{G}(\tau) &=& i_{\phi}(x)=|W^{\circ}_{\phi}|^{-1}\sum_{w\in W_{\phi}(x)_{{\operatorname{reg}}}}s^{\circ}_{\phi}(w)|\det(w-1)_{\mathfrak{a}^{G}_{M}}|^{-1} \\
&= & e_{\phi}(x) = \sum_{s\in \mathcal{E}_{\phi,{\operatorname{ell}}}(x)}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ}) \nonumber\end{aligned}$$
i.e. $i_{\phi}(x)$ is equal to the number $i(S_x)$ above, with $S_x$ being equal to the component of $\overline{S}_{\phi}$ that corresponds to $x \in \mathcal{S}_{\phi} = \pi_0(\overline{S}_{\phi})$, and $W_{\phi}(x)_{{\operatorname{reg}}}$ is the set $W(S_x)_{{\operatorname{reg}}}$ as defined above; similarly $e_{\phi}(x)$ is the number $e(S_x)$, with $\mathcal{E}_{\phi,{\operatorname{ell}}}(x)$ (resp. $\mathcal{E}_{\phi,{\operatorname{ell}}}$) being the set $\mathcal{E}_{{\operatorname{ell}}}(S_x)$ (resp. $\mathcal{E}_{{\operatorname{ell}}}(\overline{S}_{\phi})$) defined as above, etc.
Finally we note the following. Given $\phi \in \Phi(G,\zeta)$ as before, that factors through ${}^L M$ as a square-integrable parameter $\phi_M$ of Levi subgroup $M$ of $G$. Recall that one has a split short exact sequence: $$\begin{aligned}
0 \rightarrow \mathcal{S}_{\phi_M} \rightarrow \mathcal{S}_{\phi} \rightarrow R_{\phi} \rightarrow 0.\end{aligned}$$ Then for $x,y \in \mathcal{S}_{\phi}$, one has $i_{\phi}(x) = i_{\phi}(y)$ if $x = y \mod{\mathcal{S}_{\phi_M}}$. This follows easily from the definition of the number $i_{\phi}(x)$.
The constants $\sigma(\overline{S}_{\phi,s}^{\circ})$, for $s \in \mathcal{E}_{\phi,{\operatorname{ell}}} = \mathcal{E}_{{\operatorname{ell}}}(\overline{S}_{\phi})$, appear as the coefficients of the endoscopic and stable local trace formula, as we are going to see in the following sections.
Spectral transfer
=================
Suppose $({G^{\prime},s^{\prime},\mathcal{G}^{\prime},\xi^{\prime}})$ is an endoscopic datum for $G$. In the theory of endoscopy one chooses a $Z$-pair $(G^{\prime}_{1},\xi_{1}^{\prime})$, where $G^{\prime}_1$ is a $Z$-extension of $G^{\prime}$ and $\xi_1^{\prime}$ is an embedding of extensions $\mathcal{G}^{\prime} \hookrightarrow{}^{L}G_{1}^{\prime}$ that extends the embedding $\widehat{G}^{\prime} \hookrightarrow \widehat{G}^{\prime}_1$ dual to the surjection $G_1^{\prime} \rightarrow G^{\prime}$.
Given the endoscopic datum $({G^{\prime},s^{\prime},\mathcal{G}^{\prime},\xi^{\prime}})$ for $G$ as above, suppose that $\mathcal{G}^{\prime} = {}^{L}G^{\prime}$ (and thus $\xi^{\prime}: {}^{L}G^{\prime} \hookrightarrow {}^{L}G$ is an embedding of $L$-groups, and it is not needed to choose a $Z$-extension for $G^{\prime}$), we define a mapping $$\Phi(G^{\prime},\zeta)\longrightarrow \Phi(G,\zeta)$$ by $\phi^{\prime}\mapsto \phi=\xi^{\prime} \circ\phi^{\prime}$.
For the general case, we refer to Section 2 of [@A7] and also Section 2 of [@S2]. This construction gives a correspondence $(G^{\prime},\phi^{\prime}) \longleftrightarrow (\phi,s)$ between pairs $(\phi,s)$, where $\phi \in \Phi(G,\zeta), s \in \overline{S}_{\phi}$ semi-simple, and pairs $(G^{\prime},\phi^{\prime})$, where $G^{\prime}=({G^{\prime},s^{\prime},\mathcal{G}^{\prime},\xi^{\prime}})$ is an endoscopic datum of $G$, and $\phi^{\prime} \in \Phi(G^{\prime},\zeta)$, [*c.f. loc. cit.*]{} For simplicity we always assume that $\mathcal{G}^{\prime} = {}^{L}G^{\prime}$ in what follows.
For $\phi^{\prime} \in \Phi(G^{\prime},\zeta)$ and $\Pi \in \Pi_{{\operatorname{temp}}}(G,\zeta)$, we say that $(\phi^{\prime},\Pi)$ is a related pair if $\phi(\Pi)$ is the image of $\phi^{\prime}$ under the map $\Phi(G^{\prime},\zeta)\longrightarrow \Phi(G,\zeta)$ associated to $\xi^{\prime}$; here $\phi(\Pi) \in \Phi(G,\zeta)$ is the Langlands parameter of $\Pi$.
Given any $\phi^{\prime} \in \Phi(G^{\prime},\zeta)$, there is always a $\Pi \in \Pi_{{\operatorname{temp}}}(G,\zeta)$ that is related to $\phi^{\prime}$.
A related pair $(\phi^{\prime},\Pi)$ is $G$-regular if the parameter $\phi=\phi(\Pi)$ is $G$-regular, in the sense that we have that $${\operatorname{Cent}}(\phi({\mathbb{C}}^{\times}),\widehat{G})$$ is abelian.
Shelstad had built the spectral transfer factors $\Delta(\phi^{\prime},\Pi)$ [@S2] directly when the related pair $(\phi^{\prime},\Pi)$ is $G$-regular, and obtained spectral transfer identities (if the pair $(\phi^{\prime},\Pi)$ is not related, then one simply defines $\Delta(\phi^{\prime},\Pi)=0$). The general case is handled by using character identities of Hecht and Schmid, and coherent continuation of the identities from the $G$-regular case [@S2].
We have the following spectral transfer relations. For each $f\in \mathcal{H}(G({\mathbb{R}}),\zeta)$, and for any endoscopic datum $G^{\prime}$ of $G$, there exists $f^{\prime}\in \mathcal{H}(G^{\prime}({\mathbb{R}}),\zeta)$ such that the stable orbital integral of $f^{\prime}$ is equal to $f^{G^{\prime}}$, the Langlands-Shelstad transfer of $f$ to $G^{\prime}$ [@S1] (with respect to Whittaker normalization). Note that in particular when we take $G^{\prime}=G$, then $f^G$ is the stable orbital integral of $f$. In addition the following holds: for any endoscopic datum $G^{\prime}$ of $G$ and any tempered Langlands parameter $\phi^{\prime}$ of $G^{\prime}({\mathbb{R}})$, the stable character $f^{\prime}(\phi^{\prime})$ of $\phi^{\prime}$ (evaluated at $f^{\prime}$) satisfies: $$f^{\prime}(\phi^{\prime})=\sum_{\Pi\in\Pi_{{\operatorname{temp}}}(G,\zeta)}\Delta(\phi^{\prime},\Pi){\operatorname{tr}}\Pi(f).$$
Remark that, as the character $f^{\prime} \mapsto f^{\prime}(\phi^{\prime})$ is stable, it depends only on the stable orbital integral of $f^{\prime}$. Hence in the above, namely when the stable orbital integral of $f^{\prime}$ is equal to the Langlands-Sehlstad transfer $f^{G^{\prime}}$, the value $f^{\prime}(\phi^{\prime})$ depends only on $f^{G^{\prime}}$, and we will write $f^{\prime}(\phi^{\prime})$ as $f^{G^{\prime}}(\phi^{\prime})$.
Shelstad ([*c.f.*]{} [@S3], section 11) had also checked that the spectral transfer factors can be normalized (namely Whittaker normalization) to satisfy the Arthur’s conjecture, and we will always work with such normalization in what follows; more precisely given endoscopic data $(G^{\prime},s^{\prime},\mathcal{G}^{\prime},\xi^{\prime})$ as above, one has, whenever $(\phi^{\prime},\Pi)$ is a related pair, that the number $\Delta(\phi^{\prime},\Pi)$ depends only on $\Pi$ and on the image $x_s$ of $s$ in $\mathcal{S}_{\phi}$ (here $\phi$ is the Langlands parameter of $\Pi$).
Thus given $\phi \in \Phi(G,\zeta)$, and $s \in \overline{S}_{\phi}$ semi-simple, for $\Pi \in \Pi_{\phi}$, we will denote $$\Delta(\phi^s,\Pi) :=\Delta(\phi^{\prime},\Pi)$$ if the pair $(\phi,s)$ corresponds to $(G^{\prime},\phi^{\prime})$. For fixed $\Pi \in \Pi_{\phi}$, the function $s \mapsto \Delta(\phi^s,\Pi)$ factors through $\mathcal{S}_{\phi}$. Thus we have the function $x \mapsto \Delta(\phi^x,\Pi)$ for $x \in \mathcal{S}_{\phi}$. In addition, this function is a $\{ \pm 1 \}$-valued character of $\mathcal{S}_{\phi}$ [@S3].
Thus in particular, given $\phi \in \Phi(G,\zeta)$ and $s \in \overline{S}_{\phi}$ semi-simple, we can consider the linear form $f \mapsto f^{\prime}(\phi^{\prime})$, with the pair $(\phi,s)$ being corresponding to $(G^{\prime},\phi^{\prime})$. This linear form, which by construction depends only on $\phi$ and $s \in \overline{S}_{\phi}$, in fact depends only on $\phi$ and $x_s$ (where $x_s$ is as above the image of $s$ in $\mathcal{S}_{\phi}$). We will denote this linear form as $f \mapsto f^{\prime}(\phi,s)$ or $f^{\prime}(\phi^s)$, and also as $f \mapsto f^{\prime}(\phi,x)$ or $f \mapsto f^{\prime}(\phi^{x})$, i.e. for $x \in \mathcal{S}_{\phi}$: $$\begin{aligned}
f^{\prime}(\phi,x) &=& f^{\prime}(\phi,s) \\
f^{\prime}(\phi^{x}) &=& f^{\prime}(\phi^{s})\end{aligned}$$ for any $s \in \overline{S}_{\phi}$ such that $x_s =x$.
Shelstad had also constructed the adjoint spectral transfer factor $\Delta(\Pi,\phi^{s})$ and established the inversion formula [@S3], for $\Pi \in \Pi_{\phi}$: $$\begin{aligned}
{\operatorname{tr}}\Pi(f) = \sum_{x \in \mathcal{S}_{\phi}}\Delta(\Pi,\phi^{x})f^{\prime}(\phi,x).\end{aligned}$$ Here on the right hand side, for the adjoint spectral transfer factor, the dependence of $\Delta(\Pi,\phi^s)$ on $s$ again factors through the image $x_s$ of $s$ in $\mathcal{S}_{\phi}$, and so for $x \in \mathcal{S}_{\phi}$, we denote by $\Delta(\Pi,\phi^x)$ the value $\Delta(\Pi,\phi^s)$, for any $s \in \overline{S}_{\phi}$ such that $x_s=x$.
For our purpose, we need to construct the transfer factors $\Delta(\tau,\phi^{s})$ and $\Delta(\phi^{s},\tau)$, in the spirit of section 5 of [@A3]; again these depends only on the image $x_s$ of $s$ in $\mathcal{S}_{\phi}$, and so we similarly employ the notations $\Delta(\tau,\phi^x)$ and $\Delta(\phi^x,\tau)$. If $\phi \in \Phi_{{\operatorname{ell}}}(G,\zeta)$ is elliptic, then the transfer factors $\Delta(\tau,\phi^s)$ are defined in [@P]:
$$\Delta(\tau,\phi^{s}) =\sum_{\chi \in \widehat{R}_{\pi}} \chi(r) \Delta(\Pi^{\chi},\phi^s)$$ where $\tau=(M,\pi,r) \in T_{\phi}, s \in \overline{S}_{\phi}$, and $\Pi^{\chi}$ is the irreducible component of induced representation of $\pi$ whose character corresponds to that of $\chi$ (here recall that $R_{\pi}$ is an abelian elementary $2$-group, so $\chi \in \widehat{R}_{\pi}$ takes values in $\{\pm 1 \}$). We have the spectral transfer relation ([*c.f.*]{} equation (5.8) in [@P]): $$\label{inverse transfer}
\Theta(\tau,f)=\sum_{x \in \mathcal{S}_{\phi}}\Delta(\tau,\phi^{x})f^{\prime}(\phi,x).$$
We also define the adjoint transfer factor (still assuming $\phi \in \Phi_{{\operatorname{ell}}}(G,\zeta)$ being elliptic, and $\tau=(M,\pi,r) \in T_{\phi}$ as before): $$\Delta(\phi^{s},\tau)=\sum_{\chi\in\widehat{R}_{\pi}}\frac{1}{|R_{\pi}|}\chi(r) \Delta(\phi^s,\Pi^{\chi}).$$ We then have the spectral transfer formula, for $s \in \overline{S}_{\phi}$ ([*c.f.*]{} equation (5.9) in [@P]): $$\label{inverse 2}
f^{\prime}(\phi,s)=\sum_{\tau\in T_{\phi}}\Delta(\phi^{s},\tau)\Theta(\tau,f).$$ i.e. $$\begin{aligned}
f^{\prime}(\phi,x)=\sum_{\tau\in T_{\phi}}\Delta(\phi^{x},\tau)\Theta(\tau,f).\end{aligned}$$ for $x \in \mathcal{S}_{\phi}$.
We have the following lemma.
\[deeper lemma\] Suppose that $\phi \in \Phi(G,\zeta)$ is elliptic. Let $M$ be Levi subgroup of $G$ such that $\phi$ factors through ${}^{L}M$ as a square-integrable parameter $\phi_M \in \Phi_2(M,\zeta)$ for $M$, under which we have the split short exact sequence as in (4.1): $$\xymatrix@C=0.5cm{
0 \longrightarrow \mathcal{S}_{\phi_{M}} \longrightarrow \mathcal{S}_{\phi} \longrightarrow R_{\phi}\longrightarrow 0 }.$$
Given $x \in \mathcal{S}_{\phi}$, write $x$ with respect to the above split exact sequence as $x =x_{M} \cdot r^{\prime}$, where $x_M \in \mathcal{S}_{\phi_M}$, and $r^{\prime} \in R_{\phi}$. Then for $\tau =(M,\pi,r) \in T_{\phi}$, we have $\Delta(\phi^{x},\tau) =0$ unless $r=r^{\prime}$, in which case we have: $$\Delta(\phi^{x},\tau) = \Delta(\phi_M^{x_{M}} ,\pi)$$
This is a direct computation, using the properties of spectral transfer factors [@S3]: $$\begin{aligned}
& & \Delta(\phi^{x},\tau) = \sum_{\chi\in\widehat{R}_{\pi}}\frac{1}{|R_{\pi}|}\chi(r) \Delta(\phi^x,\Pi^{\chi}) \\
&=& \sum_{\chi\in\widehat{R}_{\pi}}\frac{1}{|R_{\pi}|}\chi(r) \chi(r^{\prime}) \Delta( \phi^{x_M} , \Pi^{\chi}) \nonumber\\
&=& \delta(r,r^{\prime}) \cdot \Delta(\phi_M^{x_M},\pi) \nonumber
\end{aligned}$$ where $\delta(\cdot,\cdot)$ is the Kronecker delta function. The lemma follows.
This property of spectral transfer factor is also studied in Chapters 2 and 6 of [@A9], in the form of local intertwining relations.
We now define the spectral transfer factors for general $\phi \in \Phi(G,\zeta)$, by reducing to the elliptic case. If $\phi\in\Phi(G,\zeta)$, then there exists a Levi subgroup $\widetilde{M}$ (unique up to conjugacy), that is maximal with respect to the property that $\phi$ factors through ${}^{L} \widetilde{M}$ as an elliptic parameter $\phi^{\widetilde{M}}$ for $\widetilde{M}$. With respect to the embedding ${}^{L}\widetilde{M}\hookrightarrow {}^{L}G$, the mapping $$\mathcal{S}^{}_{\phi^{\widetilde{M}}}\longrightarrow \mathcal{S}^{}_{\phi}$$ is an isomorphism ([*c.f.*]{} proof of Lemma 4.1).
If $\tau=(M,\pi,r)\in T_{\phi}$, then we denote by $\tau^{\widetilde{M}}=(M,\pi,r^{\widetilde{M}})\in T_{\phi^{\widetilde{M}}}$ the corresponding element of $T_{\phi^{\widetilde{M}}}$. Here $r^{\widetilde{M}}$ corresponds to $r$ under the canonical isomorphism $R_{\phi^{\widetilde{M}}}=R_{\phi}$ ([*c.f.*]{} proof of Lemma 4.1).
We define the spectral transfer factor: $$\Delta(\tau,\phi^{x})=\Delta(\tau^{\widetilde{M}},(\phi^{\widetilde{M}})^{x})$$ where on the right hand side, we still denote by $x$ the element in $\mathcal{S}_{\phi^{\widetilde{M}}}$ that corresponds to $x \in \mathcal{S}_{\phi}$ under the above isomorphism between $\mathcal{S}_{\phi^{\widetilde{M}}}$ and $\mathcal{S}_{\phi}$.
We have, with $f_{\widetilde{M}}$ being the descent of $f$ to $\widetilde{M}$, the identity: $$\Theta(\tau,f)=\Theta(\tau^{\widetilde{M}},f_{\widetilde{M}}).$$ Similarly we have, for $x \in \mathcal{S}_{\phi}$: $$f^{\prime}(\phi,x)=(f_{\widetilde{M}})^{\prime}(\phi^{\widetilde{M}} ,x).$$
It follows that we again have the transfer relation (generalization of (5.1)): $$\begin{aligned}
\Theta(\tau,f)=\sum_{x \in\mathcal{S}_{\phi}}\Delta(\tau,\phi^{x})f^{\prime}(\phi,x).
\end{aligned}$$
We can similarly define the adjoint transfer factor $\Delta(\phi^{x},\tau)$, and we have the inverse transfer relation, for $x \in \mathcal{S}_{\phi}$ (generalization of (5.2))): $$\begin{aligned}
f^{\prime}(\phi,x)=\sum_{\tau \in T_{\phi}} \Delta(\phi^{x},\tau)\Theta(\tau,f).
\end{aligned}$$
We also have the following properties about the transfer factors:
Let $\phi \in \Phi(G,\zeta)$ is a bounded Langlands parameter. Then we have:
- $\Delta(\tau,\phi^{x})=\frac{|R_{\phi}|}{|\mathcal{S}_{\phi}|} \overline{ \Delta(\phi^{x},\tau)} = \frac{|R_{\phi}|}{|\mathcal{S}_{\phi}|} \Delta(\phi^{x},\tau)$ for $\tau \in T_{\phi}$ and $x \in \mathcal{S}_{\phi}$.
- we have the adjoint relations $$\sum_{\tau\in T_{\phi}}\Delta(\phi^{x_{1}},\tau)\Delta(\tau,\phi^{x_{2}})=\delta(x_{1},x_{2}), \mbox{ for } x_1,x_2 \in \mathcal{S}_{\phi},$$ $$\sum_{x \in\mathcal{S}_{\phi}}\Delta(\tau_{1},\phi^{x})\Delta(\phi^{x},\tau_{2})=\delta(\tau_{1},\tau_{2}), \mbox{ for } \tau_1,\tau_2 \in T_{\phi}.$$
Where $\delta(\cdot,\cdot)$ is the Kronecker delta function.
This follows from the case where $\phi$ is an elliptic parameter, which is established in Proposition $5.2$ of [@P].
Finally to complete the discussion of this section, for $\tau \in T(G,\zeta), \phi \in \Phi(G,\zeta)$ and $x \in \mathcal{S}_{\phi}$, we define $\Delta(\tau,\phi^x)$ and $\Delta(\phi^x,\tau)$ to be zero if $\tau \notin T_{\phi}$.
Stabilization of the local trace formula
========================================
To summarize the discussion of the previous section, we have firstly, from equation (5.4) and (5.5), the following:
If $f\in \mathcal{H}(G({\mathbb{R}}),\zeta)$, then for $\tau \in T(G,\zeta)$, we have: $$\begin{aligned}
\label{general transfer}
\Theta(\tau,f)=\sum_{(\phi,x)}\Delta(\tau,\phi^{x})f^{\prime}(\phi,x)
\end{aligned}$$ where the summation is over $\phi \in\Phi(G,\zeta)$ and $x \in \mathcal{S}_{\phi}$.
Conversely, for $\phi \in \Phi(G,\zeta), x \in \mathcal{S}_{\phi}$, we have: $$\begin{aligned}
\label{general inverse}
f^{\prime}(\phi,x)=\sum_{\tau\in T(G,\zeta)}\Delta(\phi^{x},\tau)\Theta(\tau,f). \end{aligned}$$
The following also follows easily:
\[general deeper lemma\] The statement of Lemma 5.3 holds for general $\phi \in \Phi(G,\zeta)$.
Now for $\phi \in \Phi(G,\zeta)$, define $T_{\phi,{\operatorname{disc}}} := T_{\phi} \cap T_{{\operatorname{disc}}}(G,\zeta)$. We have $T_{\phi,{\operatorname{disc}}}$ is non-empty if and only if $\phi \in \Phi_{{\operatorname{disc}}}(G,\zeta)$ ([*c.f.*]{} Section 5 of [@S1]). Now recall from Lemma 4.1, one has the bijection $\iota: T_{\phi} \rightarrow \mathcal{S}_{\phi}$ that respects the projection to $R_{\phi}$. Thus for $\phi \in \Phi_{{\operatorname{disc}}}(G,\zeta)$, put $\mathcal{S}_{\phi,{\operatorname{disc}}}$ to be the image of $T_{\phi,{\operatorname{disc}}}$ under the bijection $\iota: T_{\phi} \rightarrow \mathcal{S}_{\phi}$. We define: $$\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)=\{(\phi,x):\phi\in\Phi_{{\operatorname{disc}}}(G,\zeta),x \in\mathcal{S}_{\phi,{\operatorname{disc}}}\}.$$
Thus for $\tau\in T_{{\operatorname{disc}}}(G,\zeta)$, on applying equation (6.1) and the above discussion, we have the transfer relation: $$\Theta(\tau,f)=\sum_{(\phi,x)\in\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\Delta(\tau,\phi^{x})f^{\prime}(\phi,x).$$ Similarly for $\phi \in \Phi_{{\operatorname{disc}}}(G,\zeta)$ and $x \in \mathcal{S}_{\phi,{\operatorname{disc}}}$, we have, on applying equation (6.2), the inversion formula: $$\begin{aligned}
f^{\prime}(\phi,x)=\sum_{\tau\in T_{{\operatorname{disc}}}(G,\zeta)}\Delta(\phi^{x},\tau)\Theta(\tau,f).
\end{aligned}$$
In the following, we also denote a pair $(\phi,x) \in \Phi^S_{{\operatorname{disc}}}(G,\zeta)$ as $\phi^x$.
The linear space $i\mathfrak{a}^{\ast}_{G,Z}$ acts on $\Phi_{{\operatorname{disc}}}(G,\zeta)$ through twisting: for $\phi \in \Phi_{{\operatorname{disc}}}(G,\zeta)$ and $\lambda \in i\mathfrak{a}^{\ast}_{G,Z}$ $$\phi_{\lambda}(\omega):=\phi(\omega)|\omega|^{\lambda}, \quad \omega\in W_{{\mathbb{R}}}$$ (here the element $|\omega|^{\lambda}\in Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}}$ is defined via the usual Nakayama-Tate duality). This induces the action of $i\mathfrak{a}^{\ast}_{G,Z}$ on $\Phi^S_{{\operatorname{disc}}}(G,\zeta)$, through twisting on $\Phi_{{\operatorname{disc}}}(G,\zeta)$: $(\phi^x)_{\lambda} :=(\phi_{\lambda})^x$; here we are identifying $\overline{S}_{\phi}$ and $\overline{S}_{\phi_{\lambda}}$, hence the identification for $\mathcal{S}_{\phi}$ and $\mathcal{S}_{\phi_{\lambda}}$, similarly identification for $\mathcal{S}_{\phi,{\operatorname{disc}}}$ and $\mathcal{S}_{\phi_{\lambda},{\operatorname{disc}}}$.
We define a measure on $\Phi_{{\operatorname{disc}}}(G,\zeta)$ and $\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)$ by setting: $$\int_{\Phi_{{\operatorname{disc}}}(G,\zeta)}\beta_1(\phi)d\phi=\sum_{\phi \in \Phi_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}}\int_{i\mathfrak{a}^{\ast}_{G,Z}}\beta_1(\phi_{\lambda})d\lambda,$$ $$\int_{\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\beta_2(\phi^{x})d\phi^{x}=\sum_{(\phi,x)\in \Phi^{S}_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}}\int_{i\mathfrak{a}^{\ast}_{G,Z}}\beta_2(\phi^{x}_{\lambda})d\lambda,$$ for any $\beta_1\in C_c(\Phi_{{\operatorname{disc}}}(G,\zeta)), \beta_2\in C_c(\Phi^{S}_{{\operatorname{disc}}}(G,\zeta))$. We have the following lemma, similar to Lemma 5.3 of [@A3]:
\[change integrable\] Suppose that $\alpha\in C_c(T_{{\operatorname{disc}}}(G,\zeta))$, and $\beta\in C_c(\Phi^{S}_{{\operatorname{disc}}}(G,\zeta))$. Then we have $$\label{chang measure}
\int_{T_{{\operatorname{disc}}}(G,\zeta)}\sum_{(\phi,x)\in\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\beta(\phi^{x})\Delta(\phi^{x},\tau)\alpha(\tau)d\tau,$$
$$\label{change measure 2}
=\int_{\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\sum_{\tau\in T_{{\operatorname{disc}}}(G,\zeta)}\beta(\phi^{x})\Delta(\phi^{x},\tau)\alpha(\tau)d\phi^{x}.$$
First note that for given a $\tau$, the first inner summation is a finite sum, by the definition of the transfer factors; similarly for a given $\phi^x$, the second summation is also finite. So the identity makes sense. From the definition of measure for $T_{{\operatorname{disc}}}(G,\zeta)$, we can write the left hand side of (6.5) as:
$$\label{change proof 1}
\sum_{\tau \in T_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}}\sum_{\phi^{x}\in\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}} \sum_{\mu \in i \mathfrak{a}^{\ast}_{G,Z}} \int_{i\mathfrak{a}^{\ast}_{G,Z}} \beta(\phi^{x}_{\mu})\Delta(\phi^{x}_{\mu},\tau_{\lambda})\alpha(\tau_{\lambda}) d\lambda$$
For a given representative $\tau \in T_{{\operatorname{disc}}}(G,\zeta)/ i \mathfrak{a}^{\ast}_{G,Z}$ in the outer sum, we may choose the representative $\phi^x \in \Phi^S(G,\zeta)/ i \mathfrak{a}^{\ast}_{G,Z}$ in the inner sum, that has the same central character on $A_G$.
Thus in the sum over $\mu$, we see that if $\lambda\neq\mu$, then $\phi^{x}_{\mu}$ and $\tau_{\lambda}$ are not related, and so by the definition of transfer factors, we have $\Delta(\phi^{x}_{\mu},\tau_{\lambda})=0$. So we can write (6.6) as:
$$\label{change proof 2}
\sum_{(\tau,\phi^x)} \int_{i\mathfrak{a}^{\ast}_{G,Z}}\beta(\phi^{x}_{\lambda})\Delta(\phi^{x}_{\lambda},\tau_{\lambda})\alpha(\tau_{\lambda})d\lambda.$$
where double sum of (6.7) is a sum over the subset $$(\tau,\phi^{x})\in (T_{{\operatorname{disc}}}(G,\zeta)\times\Phi^{S}_{{\operatorname{disc}}}(G,\zeta))/ i\mathfrak{a}^{\ast}_{G,Z}$$ consisting of those pairs that have the same central character on $A_G$.
In a parallel way, we can treat the right hand side of (6.5) along similar lines, using the definition of measure for $\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)$. On noting the symmetry of the sum-integral in (6.7), we thus conclude that the left hand side and the right hand side of (6.5) are equal.
We can now begin the stabilization of the spectral side of the local trace formula:
If $f=f_{1}\times\bar{f_{2}}$, $f_{i}\in \mathcal{H}(G({\mathbb{R}}),\zeta)$ for $i=1,2$, then we have:
$$I^G_{{\operatorname{disc}}}(f)=\int_{T_{{\operatorname{disc}}}(G,\zeta)}i^{G}(\tau) \Theta(\tau,f_1) \overline{\Theta(\tau,f_2)} |R_{\pi}|^{-1} d\tau$$
$$=\int_{\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\frac{1}{|\mathcal{S}_{\phi}|}i_{\phi}(x)f^{\prime}_{1}(\phi,x)\overline{f^{\prime}_{2}(\phi,x)}d\phi^{x}$$
Firstly, applying the spectral transfer, as given in equation (6.3), to the term $\Theta(\tau,f_1)$ in (6.8), we see that $I^G_{{\operatorname{disc}}}(f)$ equals $$\begin{aligned}
\int_{T_{{\operatorname{disc}}}(G,\zeta)}i^{G}(\tau)|R_{\pi}|^{-1}\sum_{(\phi,x)\in \Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\Delta(\tau,\phi^{x})f^{\prime}_{1}(\phi,x)\overline{\Theta(\tau,f_2)} d\tau \end{aligned}$$ $$= \int_{T_{{\operatorname{disc}}}(G,\zeta)}i^{G}(\tau)|R_{\pi}|^{-1}\sum_{(\phi,x)\in\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\frac{|R_{\phi}|}{|\mathcal{S}_{\phi}|} f^{\prime}_{1}(\phi,x)\overline{\Delta(\phi^{x},\tau) \Theta(\tau,f_2)}d\tau.$$ with the last equality follows from the first part of Lemma 5.5. Next, in order that $\Delta(\phi^x,\tau) \neq 0$, we must have $\tau \in T_{\phi}$ and hence $|R_{\pi}|=|R_{\phi}|$ in the integrand. Thus we see that the right hand side of (6.9) can be written as: $$\int_{T_{{\operatorname{disc}}}(G,\zeta)}\sum_{(\phi,x)\in\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\frac{i^G(\tau)}{|\mathcal{S}_{\phi}|} f^{\prime}_{1}(\phi,x)\overline{\Delta(\phi^{x},\tau) \Theta(\tau,f_2)}d\tau.$$ By Lemma 6.3, this is equal to:
$$\label{stable2}
\int_{\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\frac{1}{|\mathcal{S}_{\phi}|}f^{\prime}_{1}(\phi,x)\sum_{\tau\in T_{{\operatorname{disc}}}(G,\zeta)}i^{G}(\tau)\overline{\Delta(\phi^{x},\tau) \Theta(\tau,f_2)} d\phi^x.$$
Now, denote by $P: \mathcal{S}_{\phi}\rightarrow R_{\phi}$ for the surjective map from $\mathcal{S}_{\phi}$ to $R_{\phi}$ in the split short exact sequence (4.1). Given $\tau = (M,\pi,r) \in T_{\phi}$, and $\phi \in \Phi_{{\operatorname{disc}}}(G,\zeta)$, by Lemma 6.2, for $x \in \mathcal{S}_{\phi,{\operatorname{disc}}}$, if $P(x)\neq r,$ then $\Delta(\phi^{x},\tau)=0$. On the other hand, if $\Delta(\phi^{x},\tau)\neq 0$, then we must have $P(x)=r$, i.e. $\iota(\tau)$ and $x$ have the same image in $R_{\phi}$ under the map $P$. Thus we have: $$i^{G}(\tau)=i_{\phi}(\iota(\tau)) =i_{\phi}(x).$$ ([*c.f.*]{} the discussion after the statement of Theorem 4.2, near the end of Section 4).
Thus we can write (6.10) as $$\int_{\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\frac{1}{|\mathcal{S}_{\phi}|}i_{\phi}(x)f^{\prime}_{1}(\phi,x)\sum_{\tau\in T_{{\operatorname{disc}}}(G,\zeta)}\overline{\Delta(\phi^{x},\tau) \Theta(\tau,f_2)} d\phi^x.$$ Applying equation (6.4), we obtain the required formula: $$\label{stable lemma}
I^G_{{\operatorname{disc}}}(f)= \int_{\Phi^{S}_{{\operatorname{disc}}}(G,\zeta)}\frac{1}{|\mathcal{S}_{\phi}|}i_{\phi}(x)f^{\prime}_{1}(\phi,x)\overline{f^{\prime}_{2}(\phi,x)}d\phi^{x}.$$
We set $i_{\phi}(x)=0$, if $x \notin \mathcal{S}_{\phi,{\operatorname{disc}}}$. Then we have: $$\begin{aligned}
I^G_{{\operatorname{disc}}}(f)=\int_{\Phi_{{\operatorname{disc}}}(G,\zeta)}\sum_{x\in\mathcal{S}_{\phi}}|\mathcal{S}_{\phi}|^{-1}i_{\phi}(x)f^{\prime}_{1}(\phi,x)\overline{f^{\prime}_2(\phi,x)}d\phi.
\end{aligned}$$
Recall equation (4.4) from section 4, we write: $$i_{\phi}(x)=\sum_{s\in\mathcal{E}_{\phi,{\operatorname{ell}}}(x)}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ})$$
Thus we obtain $$I^G_{{\operatorname{disc}}}(f)=\int_{\Phi_{{\operatorname{disc}}}(G,\zeta)}\sum_{x\in \mathcal{S}_{\phi}}|\mathcal{S}_{\phi}|^{-1}\sum_{s\in\mathcal{E}_{\phi,{\operatorname{ell}}}(x)}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ})f^{\prime}_{1}(\phi,s)\overline{f^{\prime}_{2}(\phi,s)}d\phi$$ in other words, $$\begin{aligned}
I^G_{{\operatorname{disc}}}(f)=\int_{\Phi_{{\operatorname{disc}}}(G,\zeta)}\sum_{s \in \mathcal{E}_{\phi,{\operatorname{ell}}}}|\mathcal{S}_{\phi}|^{-1}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ})f^{\prime}_{1}(\phi,s)\overline{f^{\prime}_{2}(\phi,s)}d\phi.\end{aligned}$$
This is Theorem 1.1 as stated in the Introduction. In the next section, we are going to express the right hand side of (6.12) in the form of an endoscopic local trace formula.
Spectral side of endoscopic local trace formula
===============================================
To obtain Theorem 1.2 as stated in the Introduction, we need to rewrite the right hand side of (6.12), in terms of the endoscopic data of $G$. To do this we need to make precise the correspondence $(G^{\prime},\phi^{\prime}) \longleftrightarrow (\phi,s)$.
Denote by $E_{{\operatorname{ell}}}(G)$ for the set of elliptic endoscopic data of $G$. The set $\widehat{G} \bbslash E_{{\operatorname{ell}}}(G)$ of $\widehat{G}$-orbits in $E_{{\operatorname{ell}}}(G)$ is then equal to the set $\mathcal{E}_{{\operatorname{ell}}}(G)$ of equivalence classes of elliptic endoscopic data of $G$.
Write $F_{{\operatorname{disc}}}(G,\zeta)$ for the set of bounded Langlands parameters $\phi$ of $G$ that have character $\zeta$ with respect to $Z$ ([*not*]{} being regarded as up to $\widehat{G}$-equivalence), and such that $Z(\overline{S}_{\phi})$ is finite. Similarly for $G^{\prime} \in E_{{\operatorname{ell}}}(G)$, write $F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$ for the set of bounded Langlands parameters $\phi^{\prime}$ of $G^{\prime}$ that have character $\zeta$ with respect to $Z$ ([*not*]{} being regarded as up to $\widehat{G^{\prime}}$-equivalence), and such that $Z(\overline{S}_{\phi}^{\circ})$ is finite.
The set $\widehat{G} \bbslash F_{{\operatorname{disc}}}(G,\zeta)$ of $\widehat{G}$-orbits in $F_{{\operatorname{disc}}}(G,\zeta)$ is equal to $\Phi_{{\operatorname{disc}}}(G,\zeta)$. Similarly, for $G^{\prime} \in E_{{\operatorname{ell}}}(G)$, the set $\widehat{G^{\prime}} \bbslash F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$ of $\widehat{G^{\prime}}$-orbits in $F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$ is equal to $\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$.
We then define: $$X_{{\operatorname{disc}}}(G,\zeta):=\{(\phi,s):\phi\in F_{{\operatorname{disc}}}(G,\zeta),s\in\overline{S}_{\phi,{\operatorname{ell}}}\},$$ and $$Y_{{\operatorname{disc}}}(G,\zeta):=\{(G^{\prime},\phi^{\prime}):G^{\prime}\in E_{{\operatorname{ell}}}(G),\phi^{\prime}\in F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)\}.$$
The group $\widehat{G}$ acts on $X_{{\operatorname{disc}}}(G,\zeta)$ and $Y_{{\operatorname{disc}}}(G,\zeta)$ by conjugation. Denote by $\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)$ and $\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)$ for the set of $\widehat{G}$-orbits.
As in the previous section, the linear space $i \mathfrak{a}^{\ast}_{G,Z}$ acts on $\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)$ and $\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)$ by twisting, and we define measures on $\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)$ and $\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)$ by the same type of formula that define the measure on $\Phi_{{\operatorname{disc}}}(G,\zeta) = \widehat{G} \bbslash F_{{\operatorname{disc}}}(G,\zeta)$, i.e. $$\int_{ \widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)} =\sum_{ \widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}}\int_{i\mathfrak{a}^{\ast}_{G,Z}}$$ $$\int_{\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)}=\sum_{ \widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)/ i\mathfrak{a}^{\ast}_{G,Z}}\int_{i\mathfrak{a}^{\ast}_{G,Z}}$$
The correspondence $(\phi,s) \longleftrightarrow (G^{\prime},\phi^{\prime})$ induces the bijections: $$\begin{aligned}
\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta) & \longleftrightarrow &\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta) \\
\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta) / i \mathfrak{a}^*_{G,Z} & \longleftrightarrow & \widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta) / i \mathfrak{a}^*_{G,Z} \nonumber \end{aligned}$$ which is immediately seen to be measure preserving, and which is the focal point for the transformation of the right hand side of the expression (6.12). The argument we give below is parallel to that of Section 4.4 of [@A9].
The first step is to change the double sum-integral on the right hand side of (6.12) to an integral over $\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)$, using that $\Phi_{{\operatorname{disc}}}(G,\zeta) = \widehat{G} \bbslash F_{{\operatorname{disc}}}(G,\zeta)$, and also that the integrand is $\widehat{G}$-invariant. Given $\phi \in F_{{\operatorname{disc}}}(G,\zeta)$, the stabilizer of $\phi$ in $\widehat{G}$ is the centralizer $S_{\phi}$. Now the sum occurring in the integrand on the right hand side of (6.12), is over $\mathcal{E}_{\phi,{\operatorname{ell}}}=\overline{S}_{\phi}^{\circ} \bbslash \overline{S}_{\phi,{\operatorname{ell}}}$, the set of orbits in $\overline{S}_{\phi,{\operatorname{ell}}}$ under the conjugation action by the identity component $\overline{S}_{\phi}^{\circ}$ of $\overline{S}_{\phi}$. On the other hand, given $s \in \overline{S}_{\phi,{\operatorname{ell}}}$, the set of orbits under the conjugation action of $S_{\phi}$, or equivalently by $\overline{S}_{\phi}$, is bijective with the quotient of $\overline{S}_{\phi}$ by the subgroup $$\overline{S}^{+}_{\phi,s}={\operatorname{Cent}}(s,\overline{S}_{\phi}).$$
However, the $\overline{S}^{\circ}_{\phi}$-orbit of $s$ is bijective with the quotient of $\overline{S}^{\circ}_{\phi}$ by the subgroup
$$\overline{S}_{\phi,s}={\operatorname{Cent}}(s,\overline{S}_{\phi}^{\circ}).$$
We can therefore rewrite the sum-integral on the right hand side of (6.12), as an integral over $\widehat{G} \bbslash X_{{\operatorname{disc}}}(G,\zeta)$, if we multiply the summand on the right hand side of (6.12) by the number: $$|\overline{S}_{\phi}^{\circ}/\overline{S}_{\phi,s}|^{-1} |\overline{S}_{\phi}/\overline{S}^+_{\phi,s}| = |\overline{S}_{\phi,s}^{+}/\overline{S}_{\phi,s}|^{-1}|\overline{S}_{\phi}/\overline{S}_{\phi}^{\circ}|,$$ which is to say the number:
$$\begin{aligned}
\label{coefficint change 1}
|\overline{S}_{\phi,s}^{+}/\overline{S}_{\phi,s}|^{-1}|\mathcal{S}_{\phi}|.
\end{aligned}$$
The second step is to use the bijection (7.1) and write the right hand side of (6.12) as an integral over $\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)$. Recall firstly that: $$\mathcal{E}_{{\operatorname{ell}}}(G)= \widehat{G} \bbslash E_{{\operatorname{ell}}}(G).$$
Now, the stabilizer of a given $G^{\prime} \in E_{{\operatorname{ell}}}(G)$ in $\widehat{G}$ is the group ${\operatorname{Aut}}_{G}(G^{\prime})$. This means that the integral over $\widehat{G} \bbslash Y_{{\operatorname{disc}}}(G,\zeta)$ could be written as a double sum-integral: $$\sum_{G^{\prime} \in \mathcal{E}_{{\operatorname{ell}}}(G)} \int_{{\operatorname{Aut}}_G(G^{\prime}) \bbslash F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)}$$
Now the integral over $\phi^{\prime} \in {\operatorname{Aut}}_G(G^{\prime}) \bbslash F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$, could be replaced by the integral over $G^{\prime} \bbslash F_{s-{\operatorname{disc}}}(G^{\prime},\zeta)= \Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)$, so long as we multiply the integrand by the number:
$$\begin{aligned}
\label{coefficint change 2}
|{\operatorname{Out}}_{G}(G^{\prime})|^{-1}|{\operatorname{Out}}_{G}(G^{\prime},\phi^{\prime})|
\end{aligned}$$
where ${\operatorname{Out}}_{G}(G^{\prime},\phi^{\prime})$ is the stabilizer of $\phi^{\prime}$ in ${\operatorname{Out}}_{G}(G^{\prime})$.
We have now established that the double sum-integral on the right hand side of (6.12), can be replaced by the double sum-integral: $$\sum_{G^{\prime} \in \mathcal{E}_{{\operatorname{ell}}}(G)} \int_{\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)}$$ provided that the summand is multiplied by the product of the two numbers (\[coefficint change 1\]) and (\[coefficint change 2\]). Finally, the coefficient occurring in the summand on the right hand side of (6.12) is: $$\begin{aligned}
\label{coefficint change 3}
|\mathcal{S}_{\phi}|^{-1}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}^{\circ}_{\phi,s}).
\end{aligned}$$
Thus we have: $$I^G_{{\operatorname{disc}}}(f) = \sum_{G^{\prime} \in \mathcal{E}_{{\operatorname{ell}}}(G)} \int_{\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)} (7.2) \cdot (7.3) \cdot (7.4) \cdot f_1^{\prime}(\phi,s) \overline{f_2^{\prime}(\phi,s)} d \phi^{\prime} \cdot$$
We thus need to express the product of (7.2), (7.3), and (7.4), in terms of the pair $(G^{\prime},\phi^{\prime})$.
The product of (\[coefficint change 1\]), (\[coefficint change 2\]) and (\[coefficint change 3\]) is equal to the product of: $$|{\operatorname{Out}}_{G}(G^{\prime})|^{-1}$$ and
$$\begin{aligned}
\label{coefficint change 4}
|{\operatorname{Out}}_{G}(G^{\prime},\phi^{\prime})||\overline{S}_{\phi,s}^{+}/\overline{S}_{\phi,s}|^{-1}|\pi_{0}(\overline{S}_{\phi,s})|^{-1}\sigma(\overline{S}_{\phi,s}^{\circ}).
\end{aligned}$$
Now under the correspondence $(G^{\prime},\phi^{\prime}) \longleftrightarrow (\phi,s)$, we have: $$\begin{aligned}
|{\operatorname{Out}}_{G}(G^{\prime},\phi^{\prime})|&=|S_{\phi,s}^{+}/S_{\phi,s}^{+}\cap\widehat{G^{\prime}}Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}}| \\
&=|\overline{S}_{\phi,s}^{+}/\overline{S}_{\phi,s}^{+}\cap\overline{\widehat{G^{\prime}}}|\end{aligned}$$ where $\overline{\widehat{G^{\prime}}}$ denote the quotient $$\widehat{G^{\prime}}Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}}/Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}}\cong \widehat{G^{\prime}}/\widehat{G^{\prime}} \cap Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}}.$$
Also $$\sigma(\overline{S}_{\phi^{\prime}}^{\circ})=\sigma((S_{\phi^{\prime}}/Z(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}})^{\circ})=\sigma(\overline{S}_{\phi,s}^{\circ}/\overline{S}^{\circ}_{\phi,s}\cap\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}})$$ where $$\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}} =Z(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}} / Z(\widehat{G})^{\Gamma_{{\mathbb{R}}}}.$$
We thus have the identity $$\begin{aligned}
\label{coefficint change 5}
\sigma(\overline{S}^{\circ}_{\phi^{\prime}})=\sigma(\overline{S}^{\circ}_{\phi,s})|\overline{S}_{\phi,s}^{\circ}\cap\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}|\end{aligned}$$ by Theorem \[coefficient endos theorem\].
The term $$|\pi_{0}(\overline{S}_{\phi,s})|^{-1}|\overline{S}_{\phi,s}^{\circ}\cap\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}|^{-1}$$ equals $$\begin{aligned}
\label{coefficint change 6}
|\overline{S}_{\phi,s}/\overline{S}_{\phi,s}^{\circ}\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}|^{-1}|\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}|^{-1}\end{aligned}$$ as can be seen by using $$\overline{S}^{\circ}_{\phi,s}/\overline{S}_{\phi,s}^{\circ}\cap\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}\cong \overline{S}_{\phi,s}^{\circ}\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}/\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}.$$
Finally we can write: $$\begin{aligned}
\label{coefficint change 7}
|\mathcal{S}_{\phi^{\prime}}|&=|\pi_{0}(\overline{S}_{\phi^{\prime}})|\\
&=|\overline{S}_{\phi,s}^{+}\cap\overline{\widehat{G^{\prime}}}/(\overline{S}_{\phi,s}^{+})^{\circ}\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}| \nonumber \\
&=|\overline{S}_{\phi,s}^{+}\cap\overline{\widehat{G^{\prime}}}/ \overline{S}^{\circ}_{\phi,s} \overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}| \nonumber\end{aligned}$$ where the last equality follows on noting that $$\begin{aligned}
\label{coefficint change 8}
(\overline{S}^{+}_{\phi,s})^{\circ}=\overline{S}_{\phi,s}^{\circ}.\end{aligned}$$
Thus (7.5) is equal to:
$$|\overline{Z}(\widehat{G^{\prime}})^{\Gamma}|^{-1}|\mathcal{S}_{\phi^{\prime}}|^{-1}\sigma(\overline{S}_{\phi^{\prime}}^{\circ}).$$
Finally, under the correspondence $(G^{\prime},\phi^{\prime}) \longleftrightarrow (\phi,s)$, we have $$f_1^{\prime}(\phi,s) \overline{f_2^{\prime}(\phi,s)} = f_1^{G^{\prime}}(\phi^{\prime}) \overline{ f_2^{G^{\prime}}(\phi^{\prime})}.$$
We thus obtain the following main theorem:
If $f=f_{1}\times \bar{f}_{2}$, $f_{i}\in \mathcal{H}(G({\mathbb{R}}),\zeta),i=1,2$. Then we have $$I^G_{{\operatorname{disc}}}(f)=\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}(G)}\iota(G,G^{\prime})\cdot \int_{\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)}|\mathcal{S}_{\phi^{\prime}}|^{-1}\sigma(\overline{S}_{\phi^{\prime}}^{\circ})f_{1}^{G^{\prime}}(\phi^{\prime})\overline{f_{2}^{G^{\prime}}(\phi^{\prime})}d\phi^{\prime}$$ where $$\iota(G,G^{\prime})=|{\operatorname{Out}}_{G}(G^{\prime})|^{-1}|\overline{Z}(\widehat{G^{\prime}})^{\Gamma_{{\mathbb{R}}}}|^{-1}.$$
Now we define, for any quasi-split $K$-group $G$ over ${\mathbb{R}}$, the following stable distribution for $G$: $$S^G_{{\operatorname{disc}}}(f) := \int_{\Phi_{s-{\operatorname{disc}}}(G,\zeta)}|\mathcal{S}_{\phi}|^{-1}\sigma(\overline{S}_{\phi}^{\circ})f_{1}^{G}(\phi)\overline{f_{2}^{G}(\phi)}d\phi .$$ where as before $f=f_1 \times \bar{f}_2$, $f_1,f_2 \in \mathcal{H}(G({\mathbb{R}}),\zeta)$. In particular one has the stable distribution $S_{{\operatorname{disc}}}^{G^{\prime}}$ for $G^{\prime}$, with $G^{\prime}$ being an endoscopic datum of $G$.
Still with $f= f_1 \times \bar{f}_2$, $f_1, f_2 \in \mathcal{H}(G({\mathbb{R}}),\zeta)$, and $G^{\prime}$ an endoscopic datum of $G$, one has functions $f_1^{\prime},f_2^{\prime} \in \mathcal{H}(G^{\prime},\zeta)$ such that the stable orbital integrals of $f_i^{\prime}$ is equal to the Langlands-Shelstad transfer $f_i^{G^{\prime}}$ ($i=1,2$) [@S1]. Put $f^G := f_1^{G^{\prime}} \times \bar{f}_2^{G^{\prime}}$ and $f^{\prime} := f_1^{\prime} \times \bar{f}^{\prime}_2$. Then since $S_{{\operatorname{disc}}}^{G^{\prime}}$ is stable for $G^{\prime}$, one has that the value $S_{{\operatorname{disc}}}^{G^{\prime}}(f^{\prime})$ depends only on the stable orbital integral of $f^{\prime}$, i.e. only on the Langlands-Shelstad transfer $f^{G^{\prime}}$. We thus denote the value $S_{{\operatorname{disc}}}^{G^{\prime}}(f^{\prime})$ as $\widehat{S}_{{\operatorname{disc}}}^{G^{\prime}}(f^{G^{\prime}})$.
With these notations we have: $$\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}}) = \int_{\Phi_{s-{\operatorname{disc}}}(G^{\prime},\zeta)}|\mathcal{S}_{\phi^{\prime}}|^{-1}\sigma(\overline{S}_{\phi^{\prime}}^{\circ})f_{1}^{G^{\prime}}(\phi^{\prime})\overline{f_{2}^{G^{\prime}}(\phi^{\prime})}d\phi^{\prime}$$ hence in the context of Theorem 7.1, we can write equation (7.10) in the following form: $$I^G_{{\operatorname{disc}}}(f)=\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}(G)}\iota(G,G^{\prime})\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}}).$$
Spectral side of stable local trace formula
===========================================
With Theorem 7.1 in hand, we can now obtain the explicit formula for the spectral side of the stable local trace formula. As before $G$ is any $K$-group over ${\mathbb{R}}$ that is quasi-split, with central data $(Z,\zeta)$.
In [@A6], one has following stable distribution $S^G$ for $G$, which is the stable version of the geometric side of the local trace formula $I^G$, and is defined as follows. For $f=f_{1}\times\bar{f}_{2}$, with $f_1,f_2 \in \mathcal{H}(G({\mathbb{R}}),\zeta)$: $$S^{G}(f)=\sum_{M\in\mathcal{L}}|W^{M}_{0}||W^{G}_{0}|^{-1}(-1)^{{\operatorname{dim}}(A_{M}/ A_{G})}\int_{\Delta_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)}n(\delta)^{-1}S^G_{M}(\delta,f)d\delta$$ [*c.f.*]{} equation (10.11) of [@A6]. Here $n(\delta)$ is the order of the group $\mathcal{K}_{\delta}$ as defined on p. 509 of [@A3], and $\Delta_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)$ is the stable version of $\Gamma_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)$, similarly $S^G_M(\delta,f)$ is the stable version of $I^G_M(\gamma,f)$.
In particular one has the stable distribution $S^{G^{\prime}}$ for $G^{\prime}$, with $G^{\prime}$ being an endoscopic datum of $G$. It is then shown in [@A6] that, the geometric side of the local trace formula $I^G(f)$ for $G$ (where $f = f_1 \times \bar{f}_2$, $f_1,f_2 \in \mathcal{H}(G({\mathbb{R}}),\zeta)$ as above) satisfies the following endoscopic decomposition: $$I^G(f)=\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}(G)}\iota(G,G^{\prime})\widehat{S}^{G^{\prime}}(f^{G^{\prime}})$$ where as before $f^{G^{\prime}}$ is the Langlands-Shelstad transfer of $f$ to $G^{\prime}$, [*c.f.*]{} equation (10.16) of [@A6]. Here the meaning of $\widehat{S}^{G^{\prime}}(f^{G^{\prime}})$ is similar to that of concerning $\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}})$, [*c.f.*]{} the discussion near the end of Section 7.
We remark that, for the archimedean case of the local trace formula, the geometric transfer identities that are needed in [@A6] to establish the endoscopic decomposition (8.1)-(8.2), were established directly in [@A8], and so it is independent of global arguments (in the non-archimedean case, the endoscopic decomposition of the local trace formula was established in [@A6] using global arguments, and so the fundamental lemma is needed in the non-archimedean case). We also remark that, when one of the components of $f$ is cuspidal, then the arguments for the endoscopic decomposition (8.1)-(8.2) were already carried out in [@A4], Section 9 - 10.
We have the stable local trace formula: $$S^{G}_{{\operatorname{disc}}}(f) = S^{G}(f)$$ where as before $$S^{G}(f)=\sum_{M\in\mathcal{L}}|W^{M}_{0}||W^{G}_{0}|^{-1}(-1)^{{\operatorname{dim}}(A_{M}/ A_{G})}\int_{\Delta_{G-{\operatorname{reg}},{\operatorname{ell}}}(M,V,\zeta)}n(\delta)^{-1} S^G_{M}(\delta,f)d\delta,$$ and $$S^{G}_{{\operatorname{disc}}}(f)=\int_{\Phi_{s-{\operatorname{disc}}}(G,\zeta)}|\mathcal{S}_{\phi}|^{-1}\sigma(\overline{S}_{\phi}^{\circ})f_{1}^{G}(\phi)\overline{f_{2}^{G}(\phi)}d\phi.$$
By induction on ${\operatorname{dim}}(G_{{\operatorname{der}}})$. Put $\mathcal{E}_{{\operatorname{ell}}}^{\circ}(G) = \mathcal{E}_{{\operatorname{ell}}}(G) \backslash \{G\}$. Applying equation (7.11) and equation (8.2), we have $$S^{G}(f)= \widehat{S}^G(f^{G})= I^G(f)-\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}^{\circ}(G)}\iota(G,G^{\prime})\widehat{S}^{G^{\prime}}(f^{G^{\prime}})$$ and $$S_{{\operatorname{disc}}}^{G}(f)= \widehat{S}^G_{{\operatorname{disc}}}(f^G)= I^G_{{\operatorname{disc}}}(f)-\sum_{G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}^{\circ}(G)}\iota(G,G^{\prime})\widehat{S}_{{\operatorname{disc}}}^{G^{\prime}}(f^{G^{\prime}}).$$
Now for $G^{\prime}\in\mathcal{E}_{{\operatorname{ell}}}^{\circ}(G)$, one has ${\operatorname{dim}}(G^{\prime}_{{\operatorname{der}}})<{\operatorname{dim}}(G_{{\operatorname{der}}})$, so by the induction hypothesis, we have $$\widehat{S}^{G^{\prime}}_{{\operatorname{disc}}}(f^{G^{\prime}})=\widehat{S}^{G^{\prime}}(f^{G^{\prime}}).$$ Combining with the equality $I^G_{{\operatorname{disc}}}(f)=I^G(f)$, we have obtained the theorem.
\(1) The method and results of this paper could be extended to the case where $G$ is a connected reductive group over a $p$-adic field, as long as the local Langlands correspondence is known for $G$, and such that the $R$-groups $R_{\phi}$ and the component groups $\mathcal{S}_{\phi}$ attached to the Langlands parameters of $G$ are all abelian. For instance the case of classical groups by the works [@A9; @M] (with a slight complication in the even orthogonal case). We leave it to the reader to formulate the corresponding results.
\(2) For the geometric side of the stable local trace formula, the distributions $S^G_{M}(\delta,f)$ are defined inductively in terms of the invariant distributions $I^G_{M}(\gamma,f)$, which in turn is defined inductively by weighted orbital integrals, [*c.f.*]{} [@A4]. It is thus a highly non-trivial matter to obtain explicit formulas for the distributions $S^G_{M}(\delta,f)$, when $M \neq G$. In this regard, we refer the reader to [@P], where the stable local trace formula is used to obtain explicit formulas for $S^G_{M}(\delta,f)$, in the case where $f= f_1 \times \bar{f}_2$, where $f_2$ is the stable pseudo-coefficient of a square integrable parameter of $G$.
J. Arthur, *Unipotent automorphic representations: Global motivation, in Automorphic Forms, Shimura Varieties, and L-functions.* Academic Press, 1990, Vol. I, 1-75.
J. Arthur, *On elliptic tempered characters.* Acta Math. no. 171 (1993): 73-138.
J. Arthur, *On local character relations.* Selecta Mathematica (1996): 501-579.
J. Arthur, *On the transfer of distributions: weighted orbital integrals.* Duke Math. J. no. 99 (1999), 209-283.
J. Arthur, *A stable Trace Formula I. General expansions.* J. Inst. of Math. Jussieu 1(2) (2002), 175-277.
J. Arthur, *A stable Trace Formula III. Proof of the Main Theorems.* Ann. of Math. no. 158 (2003): 769-873.
J. Arthur, *A Note on $L$-packets.* Pure and Applied Mathematics Quarterly Vol. 2, No. 1 (2006), 199-217.
J. Arthur, *Parabolic transfer for real groups.* Journal of the American Mathematical Society, vol. 21, No. 1 (2008), pp. 171-234.
J. Arthur, *The Endoscopic Classification of Representations: Orthogonal and Symplectic Groups.* Colloquium Publications Volume: 61, 2013.
A,W. Knapp; G.Zuckerman, *Classification of irreducible tempered representations of semi-simple Lie groups.* Proc. Nat. Acad. Sci. U.S.A. 73 (1976), no. 7, 2178-2180.
C. P. Mok, *Endoscopic classification of representation of quasi-split unitary groups.* Memoirs of the American Mathematical Society, Volume 235 (2015), No. 1108.
Zhifeng Peng, *Multiplicity formula and stable trace formula.* Available at arXiv:1608.00055. To appear in the American Journal of Mathematics.
D.Shelstad, *$L$-indistinguishability for real groups.* Math. Ann. no. 259 (1982): 385-430.
D.Shelstad, *Tempered endoscopy for real groups II: spectral transfer factors.* International Press, 2008, 243-282.
D.Shelstad, *Tempered endoscopy for real groups III: Inversion of transfer and $L$-packet structure.* Represention Theory, no. 12 (2008): 369-402.
[^1]:
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Lausanne, Switzerland'
author:
- |
T.A.-Kh. Aushev [^1]$^{\,1,2}$\
for the Belle Collaboration\
$^1$\
$^2$\
title: '**MEASUREMENTS OF $\phi_1$ AND $\phi_2$ BY BELLE AND BABAR**'
---
Introduction
============
In the Standard Model (SM), $CP$ violation in $B^0$ meson decays originates from an irreducible complex phase in the $3\times3$ Cabibbo-Kobayashi-Maskawa (CKM) mixing matrix [@ckm]. The angles $\phi_1$ and $\phi_2$ of the CKM unitarity triangle have been measured in several $B$ decay modes [@jpsiks_belle; @jpsiks_babar; @phi2_belle; @phi2_babar]. Extra studies in different decay modes are important to check the self-consistence between measurements to probe the existence of New Physics.
The results reported in this paper were obtained by two experiments, Belle and BaBar, working at $e^+e^-$ asymmetric-energy colliders, KEKB [@KEKB] and PEP-II, correspondingly, with the center-of-mass (CM) energy at $\Upsilon(4S)$ resonance ($\sqrt s=10.58\,{\rm GeV}$). The Belle detector [@belledet] is a large-solid-angle magnetic spectrometer that consists of a silicon vertex detector (SVD), a 50-layer central drift chamber (CDC), a mosaic of aerogel threshold Cherenkov counters (ACC), time-of-flight scintillation counters (TOF), and an array of CsI(Tl) crystals (ECL) located inside a superconducting solenoid coil that provides a $1.5$ T magnetic field. An iron flux-return located outside of the coil is instrumented to detect $K_L$ mesons and to identify muons (KLM). For the results from Belle experiment the data sample of 657 million $B\bar B$ pairs is used.
The BaBar detector is described in detail elsewhere [@babardet]. Charged particle momenta are measured with a tracking system consisting of a five-layer silicon vertex tracker (SVT) and a 40-layer drift chamber (DCH) surrounded by a 1.5 T solenoidal magnet. An electromagnetic calorimeter (EMC) comprising 6580 CsI(Tl) crystals is used to measure photon energies and positions. Charged hadrons are identified with a detector of internally reflected Cherenkov light (DIRC) and ionization measurements in the tracking detectors. The results from BaBar experiment are based on 383 million $B\bar B$ pairs data sample.
Study of $B^+\to D^+\bar D^0$ and search for $B^0\to D^0\bar D^0$
=================================================================
Recently, evidence of direct $CP$ violation in the decay $B^0\to
D^+D^-$ was observed by Belle [@dpdm_belle], while BaBar measured an asymmetry consistent with zero [@dpdm_babar]. A similar effect might occur in the charged mode $B^+\to D^+\bar D^0$ [@cc]. This decay has already been observed by Belle [@dpd0_belle] and confirmed by BaBar [@dpd0_babar].
Now, Belle updated their result with larger data sample [@dd0_belle]. $366\pm32$ events were found from the fit to the $\Delta E-M_{\rm bc}$ distribution (Fig. \[dd\](a,b)), where $\Delta E=E_B-E_{\rm beam}$, $M_{\rm bc}=\sqrt{E_{\rm
beam}^2-p_B^{*2}}$, $E_B (p_B^*)$ is the energy (momentum) of $B$ candidate in the CM system, $E_{\rm beam}$ is the CM beam energy. The branching fraction of $B^+\to D^+\bar D^0$ is measured to be ${\cal
B}(B^+\to D^+\bar D^0)=(3.85\pm0.31\pm0.38)\times10^{-4}$, where the first error is statistical and the second one is systematic. The charge asymmetry for this decay is measured to be consistent with zero: $A_{CP}(B^+\to D^+\bar D^0)=0.00\pm0.08\pm0.02$. Belle also searched for the decay $B^0\to D^0\bar D^0$. An upper limit is established for the branching fraction: ${\cal B}(B^0\to D^0\bar
D^0)<0.43\times10^{-4}$ (Fig. \[dd\](c,d)).
![$\Delta E$ (a,c) and $M_{\rm bc}$ (b,d) distributions for the $B^+\to D^+\bar D^0$ (a,b) and $B^0\to D^0\bar D^0$ (c,d) candidates. Each distribution is the projection of the signal region of the other parameter. Points with errors represent the experimental data, open curves show projections from the $2D$ fits and cross-hatched curves show the $B\bar B$ background component only.[]{data-label="dd"}](dpd0_de "fig:"){width="35.00000%"} ![$\Delta E$ (a,c) and $M_{\rm bc}$ (b,d) distributions for the $B^+\to D^+\bar D^0$ (a,b) and $B^0\to D^0\bar D^0$ (c,d) candidates. Each distribution is the projection of the signal region of the other parameter. Points with errors represent the experimental data, open curves show projections from the $2D$ fits and cross-hatched curves show the $B\bar B$ background component only.[]{data-label="dd"}](dpd0_mbc "fig:"){width="35.00000%"}\
![$\Delta E$ (a,c) and $M_{\rm bc}$ (b,d) distributions for the $B^+\to D^+\bar D^0$ (a,b) and $B^0\to D^0\bar D^0$ (c,d) candidates. Each distribution is the projection of the signal region of the other parameter. Points with errors represent the experimental data, open curves show projections from the $2D$ fits and cross-hatched curves show the $B\bar B$ background component only.[]{data-label="dd"}](d0d0_de "fig:"){width="35.00000%"} ![$\Delta E$ (a,c) and $M_{\rm bc}$ (b,d) distributions for the $B^+\to D^+\bar D^0$ (a,b) and $B^0\to D^0\bar D^0$ (c,d) candidates. Each distribution is the projection of the signal region of the other parameter. Points with errors represent the experimental data, open curves show projections from the $2D$ fits and cross-hatched curves show the $B\bar B$ background component only.[]{data-label="dd"}](d0d0_mbc "fig:"){width="35.00000%"}
Study of $B^0\to D^{*+}D^{*-}$
==============================
Another interesting decay mode to study the $CP$ asymmetry is $B^0\to
D^{*+}D^{*-}$. Both experiments have updated their results for this decay mode and obtained high statistics signals shown in Fig. \[2dst\_signal\](a,c) [@2dst_babar]. The time-dependent decay rates of $B^0$ and $\bar B^0$ to a $CP$ eigenstate, like $D^{*+}D^{*-}$, is given by formula: $${\cal P}(\Delta t)=
\frac{e^{-\Delta t/\tau_{B^0}}}{4\tau_{B^0}}
\Big\{1+q\Big[{\cal S}_{f_{CP}}\sin(\Delta m_d\Delta t_{B^0})+
{\cal A}_{f_{CP}}\cos(\Delta m_d\Delta t_{B^0})\Big]\Big\},$$ where $q$ is the $b$-flavor charge: $q=+1(-1)$ when the tagging $B$ meson is a $B^0$ ($\bar B^0$), $\tau_{B^0}$ is the neutral $B$ lifetime, $\Delta m_d$ is the mass difference between two $B^0$ mass eigenstates, $\Delta t_{B^0}=t_{CP}-t_{\rm tag}$. The tree diagram dominates in this decay mode, which according to the SM gives ${\cal
S}_{f_{CP}}=\xi_{D^{*+}D^{*-}}\sin2\phi_1$ and ${\cal
A}_{f_{CP}}=0$. The parameter $\xi_{D^{*+}D^{*-}}$ is the $CP$ eigenvalue of the $D^{*+}D^{*-}$, which is $+1$ when the decay proceeds via $S$ and $D$ waves, or $-1$ for a $P$ wave. Therefore, the $CP$ measurement requires helicity study to obtain the $CP$-odd fraction $R_{\rm odd}$ of the decay. It is done in both analyses from Belle and BaBar in the so-called transversity basis. The fit results are presented in Fig. \[2dst\_signal\](b,d). The parameter $R_{\rm
odd}$ is found to be equal to $0.143\pm0.034({\rm
stat})\pm0.008({\rm syst})$ by BaBar and $0.116\pm0.042({\rm
stat})\pm0.004({\rm syst})$ by Belle.
![Measured distributions of $M_{\rm bc}$ $(a,c)$ and $\cos\theta_{\rm tr}$ in the region $M_{\rm bc}>5.27$ GeV/$c^2$ $(b,d)$ for BaBar $(a,b)$ and Belle $(c,d)$ of $B^0\to D^{*+}D^{*-}$ events. The solid lines are the projections of the fit results. The dotted lines represent the background components.[]{data-label="2dst_signal"}](2dst_mbc_polar_babar "fig:"){width="60.00000%"}\
![Measured distributions of $M_{\rm bc}$ $(a,c)$ and $\cos\theta_{\rm tr}$ in the region $M_{\rm bc}>5.27$ GeV/$c^2$ $(b,d)$ for BaBar $(a,b)$ and Belle $(c,d)$ of $B^0\to D^{*+}D^{*-}$ events. The solid lines are the projections of the fit results. The dotted lines represent the background components.[]{data-label="2dst_signal"}](2dst_mbc_belle "fig:"){width="30.00000%"} ![Measured distributions of $M_{\rm bc}$ $(a,c)$ and $\cos\theta_{\rm tr}$ in the region $M_{\rm bc}>5.27$ GeV/$c^2$ $(b,d)$ for BaBar $(a,b)$ and Belle $(c,d)$ of $B^0\to D^{*+}D^{*-}$ events. The solid lines are the projections of the fit results. The dotted lines represent the background components.[]{data-label="2dst_signal"}](2dst_polar_belle "fig:"){width="30.00000%" height="0.2\textheight"} (-240,120)[**(c)**]{} (-105,120)[**(d)**]{}
Finally, the unbinned maximum likelihood fit was performed to obtain the $CP$-violating parameters. The results of the fits are summarized in Table \[2dst\_table\] and presented in Fig. \[2dst\_cp\]. Both experiments obtained the results well consistent with each other in both the $CP$-odd fraction and the $CP$-violating parameters. Note that in the BaBar parametrization ${\cal A}=-{\cal C}$. The Belle results are preliminary.
Yield $R_{\rm odd}$ ${\cal A}=-{\cal C}$ ${\cal S}$
------- ------------ ------------------------- ----------------------- -----------------------
Belle $545\pm29$ $0.116\pm0.042\pm0.004$ $+0.16\pm0.13\pm0.02$ $-0.93\pm0.24\pm0.15$
BaBar $638\pm38$ $0.143\pm0.034\pm0.008$ $+0.02\pm0.11\pm0.02$ $-0.66\pm0.19\pm0.04$
: Results for $B^0\to D^{*+}D^{*-}$ decay mode.
\[2dst\_table\]
![The $\Delta t$ distributions of $B^0\to D^{*+}D^{*-}$ events in the region $M_{\rm bc}>5.27$ GeV/$c^2$ for $B^0(\bar B^0)$ tagged candidates $(a,c)$ and the raw asymmetry $(N_{B^0}-N_{\bar
B^0})/(N_{B^0}+N_{\bar B^0})$, as a function of $\Delta t$ $(b,d)$ for BaBar $(a,b)$ and Belle $(c,d)$. The lines represent the fit results.[]{data-label="2dst_cp"}](2dst_cp_babar "fig:"){width="35.00000%"} ![The $\Delta t$ distributions of $B^0\to D^{*+}D^{*-}$ events in the region $M_{\rm bc}>5.27$ GeV/$c^2$ for $B^0(\bar B^0)$ tagged candidates $(a,c)$ and the raw asymmetry $(N_{B^0}-N_{\bar
B^0})/(N_{B^0}+N_{\bar B^0})$, as a function of $\Delta t$ $(b,d)$ for BaBar $(a,b)$ and Belle $(c,d)$. The lines represent the fit results.[]{data-label="2dst_cp"}](2dst_cp_belle "fig:"){width="35.00000%" height=".22\textheight"} (-30,125)[**(c)**]{} (-30, 60)[**(d)**]{}
$CP$-violation in $B^0\to K_S\pi^0\pi^0$ and $B^0\to K_S\pi^0$
==============================================================
In the SM, the $CP$ violation parameters in $b\to s$ “penguin” and $b\to c$ “tree” transitions are predicted to be the same, ${\cal
S}_f\approx-\xi_f\sin2\phi_1$ and ${\cal A}_f\approx0$, with small theoretical uncertainties. Recent measurements however, indicate that the effective $\sin2\phi_1$ value, $\sin2\phi_1^{\rm eff}$, measured with penguin processes is different from $\sin2\phi_1=0.687\pm0.025$ measured in tree decays by $2.6\sigma$ [@b2s_diff]. New particles in loop diagrams may shift the weak phase.
Recently, Belle and BaBar measured the $CP$ asymmetry in $B^0\to
K_S^0\pi^0\pi^0$ and $B^0\to K_S\pi^0$ decays that proceed through $b\to s\bar qq (q=u,d)$ transitions [@kspi0pi0_belle; @kspi0pi0_babar; @kspi0_belle; @kspi0_babar]. The results of $CP$-violating parameters measurements are presented in Table \[kspi\_table\]. Both experiments are perfectly consistent with each other. In the case of $B^0\to K_S^0\pi^0\pi^0$ the central value of ${\cal S}$ has a sign opposite to what we expect from the SM, but the errors are still too large to claim the contradiction. The estimated deviation of the average value from the SM is more than $2\sigma$. The fit to the data for Belle for $B^0\to K_S^0\pi^0\pi^0$ is presented in Fig. \[kspi\](a-c) and the BaBar result for $B^0\to
K_S^0\pi^0$ is shown in Fig. \[kspi\](d-f).
${\cal A}=-{\cal C}$ ${\cal S}=-\sin2\phi_1$
-------------------------- ----------------------- -------------------------
$B^0\to K_S^0\pi^0\pi^0$
Belle $-0.17\pm0.24\pm0.06$ $+0.43\pm0.49\pm0.09$
BaBar $-0.23\pm0.52\pm0.13$ $+0.72\pm0.71\pm0.08$
Average $-0.18\pm0.22$ $+0.52\pm0.41$
$B^0\to K_S^0\pi^0$
Belle $-0.05\pm0.14\pm0.05$ $+0.33\pm0.35\pm0.08$
BaBar $-0.24\pm0.15\pm0.03$ $+0.40\pm0.23\pm0.03$
: Results for $B^0\to K_S^0\pi^0\pi^0$ and $B^0\to K_S^0\pi^0$ decay modes.
\[kspi\_table\]
![Distributions for the $M_{\rm bc}$ (a), the $\Delta t$ (b) and the raw asymmetry (c) for $B^0\to K_S^0\pi^0\pi^0$ decay mode from Belle. $\Delta t$ distributions for the $B^0$ (d) and $\bar
B^0$ (e) tagged events and the raw asymmetry (f) for the decay $B^0\to K_S^0\pi^0$ from BaBar. The lines represent the fit result.[]{data-label="kspi"}](kspi0pi0_mbc_belle "fig:"){width="25.00000%" height=".24\textheight"} (-40,150)[**(a)**]{} ![Distributions for the $M_{\rm bc}$ (a), the $\Delta t$ (b) and the raw asymmetry (c) for $B^0\to K_S^0\pi^0\pi^0$ decay mode from Belle. $\Delta t$ distributions for the $B^0$ (d) and $\bar
B^0$ (e) tagged events and the raw asymmetry (f) for the decay $B^0\to K_S^0\pi^0$ from BaBar. The lines represent the fit result.[]{data-label="kspi"}](kspi0pi0_dt_belle "fig:"){width="25.00000%" height=".27\textheight"} (-65,150)[**(b)**]{} ![Distributions for the $M_{\rm bc}$ (a), the $\Delta t$ (b) and the raw asymmetry (c) for $B^0\to K_S^0\pi^0\pi^0$ decay mode from Belle. $\Delta t$ distributions for the $B^0$ (d) and $\bar
B^0$ (e) tagged events and the raw asymmetry (f) for the decay $B^0\to K_S^0\pi^0$ from BaBar. The lines represent the fit result.[]{data-label="kspi"}](kspi0pi0_asym_belle "fig:"){width="25.00000%" height=".27\textheight"} (-40,150)[**(c)**]{} ![Distributions for the $M_{\rm bc}$ (a), the $\Delta t$ (b) and the raw asymmetry (c) for $B^0\to K_S^0\pi^0\pi^0$ decay mode from Belle. $\Delta t$ distributions for the $B^0$ (d) and $\bar
B^0$ (e) tagged events and the raw asymmetry (f) for the decay $B^0\to K_S^0\pi^0$ from BaBar. The lines represent the fit result.[]{data-label="kspi"}](kspi0_cp_babar "fig:"){width="25.00000%" height=".25\textheight"}
$\phi_2$ measurements
=====================
The CKM angle $\phi_2$ have been measured in decay modes like $B^0\to
\pi\pi, \rho\rho,\rho\pi$ [@pipirhorho]. Addition of new decay modes allows to improve an accuracy of $\phi_2$ measurement and to check a consistency of measurements in different final states. The decay $B^0\to a_1^\pm(1260)\pi^\mp$ proceeds through $b\to u$ transitions, hence its time-dependent $CP$ violation is also sensitive to $\phi_2$. Belle measured the branching fraction for this decay mode to be ${\cal B}(B^0\to a_1^\pm(1260)\pi^\mp){\cal
B}(a_1^\pm(1260)\to
\pi^\pm\pi^\pm\pi^\mp)=(14.9\pm1.6\pm2.3)\times10^{-6}$ [@a1ppim_belle], while BaBar has updated their previous measurements now with $CP$ violation study: ${\cal A}_{CP}=-0.07\pm0.07\pm0.02$ and ${\cal
S}=+0.37\pm0.21\pm0.07$ [@a1ppim_babar]. The angle $\phi_2$ was measured to be $\phi_2^{\rm eff}=78.6^o\pm7.3^o$. The result is presented in Fig. \[hh\](a-c).
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![$\Delta t$ distributions of the decay $B^0\to a_1^\pm\pi^\mp$ for $B^0$ (a) and $\bar B^0$ (b) tagged events, and the raw asymmetry (c). The solid lines show the fit results, while the dotted lines show the background component. Projection of the signal region onto (d) $\Delta E$ and (e) $M_{\rm bc}$ for $B^0\to\rho^0\rho^0$ candidates. The fit result is shown as the thick solid curve; the hatched region represents the signal component. The dotted, dot-dashed and dashed curves represent, respectively, the cumulative background contributions from continuum processes, $b\to c$ decays, and charmless $B$ decays.[]{data-label="hh"}](a1pi_cp_babar "fig:"){width="33.00000%"} ![$\Delta t$ distributions of the decay $B^0\to a_1^\pm\pi^\mp$ for $B^0$ (a) and $\bar B^0$ (b) tagged events, and the raw asymmetry (c). The solid lines show the fit results, while the dotted lines show the background component. Projection of the signal region onto (d) $\Delta E$ and (e) $M_{\rm bc}$ for $B^0\to\rho^0\rho^0$ candidates. The fit result is shown as the thick solid curve; the hatched region represents the signal component. The dotted, dot-dashed and dashed curves represent, respectively, the cumulative background contributions from continuum processes, $b\to c$ decays, and charmless $B$ decays.[]{data-label="hh"}](rho0rho0_de_belle "fig:"){width="33.00000%"} ![$\Delta t$ distributions of the decay $B^0\to a_1^\pm\pi^\mp$ for $B^0$ (a) and $\bar B^0$ (b) tagged events, and the raw asymmetry (c). The solid lines show the fit results, while the dotted lines show the background component. Projection of the signal region onto (d) $\Delta E$ and (e) $M_{\rm bc}$ for $B^0\to\rho^0\rho^0$ candidates. The fit result is shown as the thick solid curve; the hatched region represents the signal component. The dotted, dot-dashed and dashed curves represent, respectively, the cumulative background contributions from continuum processes, $b\to c$ decays, and charmless $B$ decays.[]{data-label="hh"}](rho0rho0_mbc_belle "fig:"){width="33.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
Belle also performed the search for the decay $B^0\to\rho^0\rho^0$ and other decay modes with four pions in the final state. In the absence of the signals, the upper limits on the branching fraction were established. The signal distributions for the $B^0\to\rho^0\rho^0$ are shown in Fig. \[hh\](a,b). All results are preliminary.
Also a number of the decay modes potentially usable for the $\phi_2$ measurements have been studied by BaBar [@a1pi_babar; @a1k_babar; @b1pik_babar]. All the results of these studies are summarized in Table \[hh\_table\].
Mode Yield $\epsilon$ (%) ${\cal S} (\sigma)$ ${\cal B}(\times10^{-6})$ UL$(\times10^{-6}) @ 90\%$ C.L.
-------------------- ----------------------------------- ---------------- --------------------- ------------------------------ ---------------------------------
Belle results
$\rho^0\rho^0$ $24.5^{+23.6+9.7}_{-22.1-9.9}$ 9.16 1.0 $0.4\pm0.4^{+0.2}_{-0.2}$ $<1.0$
$\rho^0\pi^+\pi^-$ $112.5^{+67.4+51.5}_{-65.6-53.7}$ 2.90 1.3 $5.9^{+3.5+2.7}_{-3.4-2.8}$ $<11.9$
$4\pi^\pm$ $161.2^{+61.2+26.0}_{-59.4-28.5}$ 1.98 2.5 $12.4^{+4.7+2.0}_{-4.6-2.2}$ $<19.0$
$\rho^0f^0$ $-11.8^{+14.5+4.9}_{-12.9-3.6}$ 5.10 $-$ $-$ $<0.6$
$f^0f^0$ $-7.7^{+4.7+3.0}_{-3.5-2.9}$ 2.75 $-$ $-$ $<0.4$
$f^0\pi^+\pi^-$ $6.3^{+37.0+18.0}_{-34.7-18.1}$ 1.55 $-$ $0.6^{+3.6}_{-3.4}\pm1.8$ $<7.3$
BaBar results
$b_1^0\pi^+$ $178^{+39}_{-37}$ 6.78 4.0 $6.7\pm1.7\pm1.0$
$b_1^0 K^+$ $219^{+38}_{-36}$ 6.73 5.3 $9.1\pm1.7\pm1.0$
$b_1^\mp\pi^\pm$ $387^{+41}_{-39}$ 9.54 8.9 $10.9\pm1.2\pm0.9$
$b_1^-K^+$ $267^{+33}_{-32}$ 9.43 6.1 $7.4\pm1.0\pm1.0$
$a_1^0\pi^+$ $382\pm79$ 7.2 3.8 $20.4\pm4.7\pm3.4$
$a_1^- K^0$ $241\pm32$ 9.6 6.2 $34.9\pm5.0\pm4.4$
$a_1^+\pi^0$ $459\pm78$ 12.5 4.2 $26.4\pm5.4\pm4.2$
$a_1^- K^+$ $272\pm44$ 7.9 5.1 $16.3\pm2.9\pm2.3$
: Fit results for decays relevant to $\phi_2$ measurements.
\[hh\_table\]
$CP$-violation in $\Upsilon(4S)$ decays
=======================================
In the decay $\Upsilon(4S)\to B^0\bar B^0\to f_1f_2$, where $f_1$ and $f_2$ are $CP$ eigenstates, the $CP$ eigenvalue of the final state $f_1f_2$ is $\xi=-\xi_1\xi_2$. Here the minus sign corresponds to odd parity from the angular momentum between $f_1$ and $f_2$. If $f_1$ and $f_2$ have the same $CP$ eigenvalue, i.e. $(\xi_1,\xi_2)=(+1,+1)$ or $(-1,-1)$, $\xi$ is equal to $-1$. Such decays, for example $(f_1,f_2)=(J/\psi K_S^0,J/\psi K_S^0)$, violate $CP$ conservation since the $\Upsilon(4S)$ meson has $J^{PC}=1^{--}$ and thus has $\xi_{\Upsilon(4S)}=+1$. The branching fraction within the SM is suppressed by the factor $$F\approx\frac{x^2}{1+x^2}(2\sin2\phi_1)^2=0.68\pm0.05,$$ where $x=\Delta m_d/\Gamma=0.776\pm0.008$ [@PDG].
This decay was studied by Belle. Due to a small branching fractions to the final state and low reconstruction efficiencies the expected yield is very small, $0.04$ events. In order to increase the signal yield, a partial reconstruction technique was used [@ups4s_belle]. One $B^0$ was fully reconstructed, while only $K_S^0$ was reconstructed from another one. The signal was searched in the recoil mass distribution to the reconstructed particles where, in principle, signals from $\eta_c,J/\psi,\chi_{c1}$, or $\psi(2S)$ can be seen. The method was checked using charged $B$ decay control samples, $\Upsilon(4S)\to B^+B^-\to(f_{B^+},J/\psi^{\rm tag}K^-$ and $\eta_c^{\rm tag}K^-)$, where $f_{B^+}$ stands for $J/\psi K^+$ and $\bar D^0\pi^+$. Also neutral $B$ decays were examined in the decay $\Upsilon(4S)\to B^0\bar B^0\to(f_{B^0},J/\psi^{\rm tag}K_S^0$ and $\eta_c^{\rm tag}K_S^0)$ with $f_{B^0}=B^0\to D^{(*)-}h^+$. The fit yields $206\pm57$ for charged $B$ and $35\pm16$ for neutral $B$ signal events, which is in good agreement with the MC expectation (Fig. \[ups4s\_belle\](a,b)). The results of the final fit are shown in Fig. \[ups4s\_belle\](c). The extracted signal yield, $-1.5^{+3.6}_{-2.8}$ events, is consistent with zero as well as with the SM prediction (1.7 events). An upper limit for the branching fraction was obtained ${\cal B}(\Upsilon(4S)\to B^0\bar B^0\to J/\psi
K_S^0, (J/\psi,\eta_c)K_S^0)<4\times10^{-7}$ at the $90\%$ confidence level, where the SM prediction is $1.4\times10^{-7}$. This corresponds to $F<2$ at the $90\%$ confidence level.
![Recoil mass distributions for samples reconstructed as $\Upsilon(4S)\to (B^+, (J/\psi,\eta_c)^{\rm tag}K^-)$ (a), $(B^0\to
D^{(*)-}h^+, (J/\psi,\eta_c)^{\rm tag}K_S^0)$ (b) and $(J/\psi
K_S^0, (J/\psi,\eta_c)^{\rm tag}K_S^0)$ (c). The solid lines show the fits to signal plus background distributions while the dashed lines show the background distributions.[]{data-label="ups4s_belle"}](ups4s_bp_cs "fig:"){width=".33\textwidth"} (-120,120)[**(a)**]{} ![Recoil mass distributions for samples reconstructed as $\Upsilon(4S)\to (B^+, (J/\psi,\eta_c)^{\rm tag}K^-)$ (a), $(B^0\to
D^{(*)-}h^+, (J/\psi,\eta_c)^{\rm tag}K_S^0)$ (b) and $(J/\psi
K_S^0, (J/\psi,\eta_c)^{\rm tag}K_S^0)$ (c). The solid lines show the fits to signal plus background distributions while the dashed lines show the background distributions.[]{data-label="ups4s_belle"}](ups4s_b0_cs "fig:"){width=".33\textwidth"} (-120,120)[**(b)**]{} ![Recoil mass distributions for samples reconstructed as $\Upsilon(4S)\to (B^+, (J/\psi,\eta_c)^{\rm tag}K^-)$ (a), $(B^0\to
D^{(*)-}h^+, (J/\psi,\eta_c)^{\rm tag}K_S^0)$ (b) and $(J/\psi
K_S^0, (J/\psi,\eta_c)^{\rm tag}K_S^0)$ (c). The solid lines show the fits to signal plus background distributions while the dashed lines show the background distributions.[]{data-label="ups4s_belle"}](ups4s_sig "fig:"){width=".33\textwidth"} (-120,120)[**(c)**]{}
Summary
=======
The $CP$ violating parameters have been measured in various decay modes. Most of the measurements are in a good agreement with the SM expectations. Although a room for New Physics becomes smaller and smaller, there is still some sign that it can be found in $b\to s$ transitions. More statistics is necessary to test these possibilities.
References {#references .unnumbered}
==========
[99]{}
N. Cabibbo, Phys. Rev. Lett. 10, 531 (1963); M. Kobayashi and T. Maskawa, Prog. Theor. Phys. [**49**]{}, 652 (1973).
K.-F. Chen [*et al.*]{} (The Belle Collaboration), Phys. Rev. Lett. [**98**]{}, 031802 (2007)
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**99**]{} 171803 (2007).
A. Kusaka, C.-C. Wang [*et al*]{}. (The Belle collaboration), Phys. Rev. D [**77**]{}, 072001 (2008).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. D [**76**]{}, 052007 (2007)
S. Kurokawa and E. Kikutani, Nucl. Instr. and Meth. A [**499**]{}, 1 (2003).
A. Abashian [*et al.*]{} (The Belle Collaboration), Nucl. Instr. and Meth. A[**479**]{}, 117-232 (2002).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Nucl. Instrum. Methods Phys. Res., Sect. A [**479**]{}, 1 (2002).
S. Fratina [*et al.*]{} (The Belle Collaboration), Phys. Rev. Lett. [**98**]{}, 221802 (2007).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**99**]{}, 071801 (2007).
The inclusion of charge conjugate modes is implied throughout this Letter.
G. Majumder [*et al.*]{} (The Belle Collaboration), Phys. Rev. Lett. [**95**]{}, 041803 (2005).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. D [**73**]{}, 112004 (2006).
I. Adachi [*et al.*]{} (The Belle Collaboration), Phys. Rev. D [**77**]{}, 091101(R) (2008).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. D [**76**]{}, 111102(R) (2007).
Heavy Flavor Averaging Group, “Winter 2007”, [`h`ttp://www.slac.stanford.edu/xorg/hfag]{}.
K. Abe [*et al.*]{} (The Belle Collaboration), arXiv:hep-ex/0708.1845.
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. D [**76**]{}, 071101(R) (2007).
K. Abe [*et al.*]{} (The Belle Collaboration), arXiv:hep-ex/0609006.
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. D [**77**]{}, 012003 (2008).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. D [**76**]{}, 052007 (2007). B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**99**]{}, 021603 (2007). A. Kusaka, C.C. Wang [*et al.*]{} (The Belle Collaboration), Phys. Rev. D [**77**]{}, 072001 (2008). A. Somov, A. Schwartz [*et al.*]{} (The Belle Collaboration), Phys. Rev. D [**76**]{}, 011104 (2007). H. Ishino [*et al.*]{} (The Belle Collaboration), Phys. Rev. Lett. [**98**]{}, 211801 (2007).
K. Abe [*et al.*]{} (The Belle Collaboration), arXiv:hep-ex/0706.3279.
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**98**]{}, 181803 (2007).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**99**]{}, 261801 (2007).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**100**]{}, 051803 (2008).
B. Aubert [*et al.*]{} (The BaBar Collaboration), Phys. Rev. Lett. [**99**]{}, 241803 (2007).
W.-M. Yao [*et al.*]{} (Particle Data Group), Journal of Phys. [**G 33**]{}, 1 (2006).
O. Tajima, M. Hazumi [*et al.*]{} (The Belle collaboration), Phys. Rev. Lett. [**99**]{}, 211601 (2007).
[^1]: [**e-mail**]{}: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyze the limit behavior as $s\to 1^-$ of the solution to the fractional Poisson equation ${(-\Delta)^su_s}=f_s$, $x\in\Omega$ with homogeneous Dirichlet boundary conditions $u_s\equiv 0$, $x\in\Omega^c$. We show that $\lim_{s\to 1^-} u_s =u$, with $-\Delta u =f$, $x\in\Omega$ and $u=0$, $x\in\partial\Omega$. Our results are complemented by a discussion on the rate of convergence and on extensions to the parabolic setting.'
address:
- '^1,2^DeustoTech, University of Deusto, 48007 Bilbao, Basque Country, Spain.'
- '^1,2^Facultad de Ingeniería, Universidad de Deusto, Avenida de las Universidades 24, 48007 Bilbao, Basque Country, Spain, +34 944139003 - 3282.'
author:
- Umberto Biccari^1^
- 'Víctor Hernández-Santamaría^2^'
bibliography:
- 'biblio.bib'
title: 'The Poisson equation from non-local to local'
---
Introduction and main result {#intro}
============================
Let $0<s<1$ and let $\Omega\subset\RR^N$ be a bounded and regular domain. Let us consider the following elliptic problem $$\label{PE}
\begin{cases}
{(-\Delta)^su} = f, &x\in\Omega\tag{$\mathcal P_s$}
\\
u\equiv 0, & x\in\Omega^c.
\end{cases}$$
In , with ${(-\Delta)^s}$ we indicate the fractional Laplace operator, defined for any function $u$ regular enough as the following singular integral $$\begin{aligned}
\label{frac_lapl}
{(-\Delta)^su}(x):={C_{N,s}}\,P.V.\int_{\RR^N} \frac{u(x)-u(y)}{|x-y|^{N+2s}}\,dy,\end{aligned}$$ where ${C_{N,s}}$ is a normalization constant given by $$\begin{aligned}
\label{cns}
{C_{N,s}}:= \frac{s2^{2s}\Gamma\left(\frac{N+2s}{2}\right)}{\pi^{N/2}\Gamma(1-s)},\end{aligned}$$ $\Gamma$ being the usual Gamma function. Moreover, we have to mention that, for having a completely rigorous definition of the fractional Laplace operator, it is necessary to introduce also the class of functions $u$ for which computing ${(-\Delta)^su}$ makes sense. We postpone this discussion to the next section.
Models involving the fractional Laplacian or other types of non-local operators have been widely used in the description of several complex phenomena for which the classical local approach turns up to be inappropriate or limited. Among others, we mention applications in turbulence ([@bakunin2008turbulence]), elasticity ([@dipierro2015dislocation]), image processing ([@gilboa2008nonlocal]), laser beams design ([@longhi2015fractional]), anomalous transport and diffusion ([@meerschaert2012fractional]), porous media flow ([@vazquez2012nonlinear]). Also, it is well known that the fractional Laplacian is the generator of s-stable processes, and it is often used in stochastic models with applications, for instance, in mathematical finance ([@levendorskii2004pricing]).
One of the main differences between these non-local models and classical Partial Differential Equations is that the fulfillment of a non-local equation at a point involves the values of the function far away from that point.
The Poisson problem is one of the most classical models involving the Fractional Laplacian, and it has been extensively studied in the past. Nowadays, there are many contributions concerning, but not limited to, existence and regularity of solutions, both local and global ([@biccari2017local; @cozzi2017interior; @grubb2015fractional; @leonori2015basic; @ros2016boundary; @ros2014dirichlet; @servadei2014weak]), unique continuation properties ([@fall2014unique]), Pohozaev identities ([@ros2014pohozaev]), spectral analysis ([@frank2016refined]) and numerics ([@acosta2017fractional]).
In the present paper, we are interested in analyzing the behavior of the solutions to under the limit $s\to 1^-$. Indeed, it is well-known (see, e.g., [@dihitchhiker; @stinga2010extension]) that, at least for regular enough functions, it holds
- $\lim_{s\to 0^+}{(-\Delta)^su} = u$.
- $\lim_{s\to 1^-}{(-\Delta)^su} = -\Delta u$.
In view of that, it is interesting to investigate whether, when $s\to 1^-$, a solution $u_s$ to converges to a solution to the classical Poisson equation $$\begin{aligned}
\label{poisson}
\begin{cases}
-\Delta u = f, &x\in\Omega\tag{$\mathcal P$}
\\
u = 0, & x\in\partial\Omega.
\end{cases}\end{aligned}$$
In our opinion, this is a very natural issue which, to the best of our knowledge, has never been fully addressed in the literature in the setting of weak solutions with minimal assumptions. As we will see, the answer to this question is positive.
Before introducing our main result, let us recall that we have the following definition of weak solutions.
\[weak\_sol\_def\] Let $f\in H^{-s}(\Omega)$. A function $u\in H_0^s(\Omega)$ is said to be a weak solution of the Dirichlet problem if $$\begin{aligned}
\label{weak-sol}
\frac{C_{N,s}}{2}\int_{\RR^N}\int_{\RR^N}\frac{(u(x)-u(y))(v(x)-v(y))}{|x-y|^{N+2s}}\;dxdy = \int_\Omega fv\,dx
\end{aligned}$$ holds for every $v\in\mathcal{D}(\Omega)$.
The main result of our work will be the following.
\[limit\_thm\] Let $\mathcal{F}_s=\{f_s\}_{0<s<1}\subset H^{-s}(\Omega)$ be a sequence satisfying the following assumptions:
- ${{\left\|f_s\right\|}_{H^{-s}(\Omega)}}\leq C$, for all $0<s<1$ and uniformly with respect to $s$;
- $f_s\rightharpoonup f$ weakly in $H^{-1}(\Omega)$ as $s\to 1^-$.
For all $f_s\in\mathcal{F}_s$, let $u_s\in H^s_0(\Omega)$ be the unique weak solution to the Dirichlet problem , in the sense of Definition \[weak\_sol\_def\]. Then, as $s\to 1^-$, $u_s\to u$ strongly in $H^{1-\delta}_0(\Omega)$ for all $0<\delta\leq 1$. Moreover, $u\in H^1_0(\Omega)$ and verifies $$\begin{aligned}
\int_{\Omega} \nabla u\cdot\nabla v\,dx = \int_{\Omega}fv\,dx, \;\;\; \forall v\in\mathcal{D}(\Omega),\end{aligned}$$ i.e. it is the unique weak solution to .
The proof of Theorem \[limit\_thm\] will be based on classical PDEs techniques. Moreover, the result will follow from the limit behavior as $s\to 1^-$ of the operator ${(-\Delta)^s}$ ([@dihitchhiker; @stinga2010extension]) and of the norm ${{\left\|\cdot\right\|}_{H^s(\Omega)}}$ ([@bourgain2001another]).
Furthermore, notice that Theorem \[limit\_thm\] requires the existence of a sequence $\mathcal{F}_s$ satisfying the assumptions $\textbf{H1}$ and $\textbf{H2}$. We point out that such sequence indeed exists, and that it is possible to construct it systematically. We will give a proof of this fact in Section \[prel\].
This paper will be organized as follows: Section \[prel\] will be devoted to introduce some preliminary definitions and results that will be needed in our analysis. In Section \[weak\_sec\], instead, we will present the proof of Theorem \[limit\_thm\], concerning the limit behavior of the solutions to . Finally, in Section \[rem\_sec\], we will present an additional result of convergence under weaker assumptions, a discussion on the rate of approximation and an extension to the the parabolic setting.
Preliminaries {#prel}
=============
In this section, we introduce some preliminary results that will be useful for the proof of our main theorem.
We start by giving a more rigorous definition of the fractional Laplace operator, as we have anticipated in Section \[intro\]. Define $$\begin{aligned}
\mathcal L_s^{1}(\RR^N):=\left\{u:\RR^N\to\RR\;\mbox{ measurable},\; \int_{\RR^N}\frac{|u(x)|}{(1+|x|)^{N+2s}}\;dx<\infty\right\}.\end{aligned}$$ For $u\in \mathcal L_s^{1}(\RR^N)$ and $\varepsilon>0$ we set $$\begin{aligned}
(-\Delta)_\varepsilon^s u(x):= C_{N,s}\int_{\{y\in\RR^N:\;|x-y|>\varepsilon\}}\frac{u(x)-u(y)}{|x-y|^{N+2s}}\;dy,\;\;x\in\RR^N.\end{aligned}$$
The fractional Laplace operator ${(-\Delta)^s}$ is then defined by the following singular integral: $$\begin{aligned}
\label{fl_def}
{(-\Delta)^su}(x)={C_{N,s}}\,\mbox{P.V.}\int_{\RR^N}\frac{u(x)-u(y)}{|x-y|^{N+2s}}\;dy=\lim_{\varepsilon\downarrow 0}(-\Delta)_\varepsilon^s u(x),\;\;x\in\RR^N,\end{aligned}$$ provided that the limit exists.
We notice that if $0<s<1/2$ and $u$ is smooth, for example bounded and Lipschitz continuous on $\RR^N$, then the integral in is in fact not really singular near $x$ (see e.g. [@dihitchhiker Remark 3.1]). Moreover, $\mathcal L_s^{1}(\RR^N)$ is the right space for which $ v:=(-\Delta)_\varepsilon^s u$ exists for every $\varepsilon>0$, $v$ being also continuous at the continuity points of $u$.
It is by now well-known (see, e.g., [@dihitchhiker]) that the natural functional setting for problems involving the Fractional Laplacian is the one of the fractional Sobolev spaces. Since these spaces are not so familiar as the classical integral order ones, for the sake of completeness, we recall here their definition.
Given $\Omega\subset\RR^N$ regular enough and $s\in(0,1)$, the fractional Sobolev space $H^s({\Omega})$ is defined as $$\begin{aligned}
H^s(\Omega):= \left\{u\in L^2(\Omega)\,:\, \frac{|u(x)-u(y)|}{|x-y|^{\frac N2+s}}\in L^2(\Omega\times\Omega)\right\}.\end{aligned}$$
It is classical that this is a Hilbert space, endowed with the norm (derived from the scalar product) $$\begin{aligned}
{{\left\|u\right\|}_{H^s(\Omega)}} := \left({{\left\|u\right\|}_{L^2(\Omega)}}^2 + |u|_{H^s(\Omega)}^2\right)^{\frac 12},\end{aligned}$$ where the term $$\begin{aligned}
|u|_{H^s(\Omega)}:= \left(\int_\Omega\int_\Omega \frac{|u(x)-u(y)|^2}{|x-y|^{N+2s}}\,dxdy\right)^{\frac 12}\end{aligned}$$ is the so-called Gagliardo seminorm of $u$. We set $$\begin{aligned}
H_0^s(\Omega):= \overline{C_0^\infty(\Omega)}^{\,H^s(\Omega)}\end{aligned}$$ the closure of the continuous infinitely differentiable functions compactly supported in $\Omega$ with respect to the $H^s(\Omega)$-norm. The following facts are well-known.
- For $0<s\leq\frac 12$, the identity $H_0^s(\Omega) = H^s(\Omega)$ holds. This is because, in this case, the $C_0^\infty(\Omega)$ functions are dense in $H^s(\Omega)$ (see, e.g., [@jllions1972non Theorem 11.1]).
- For $\frac 12<s<1$, we have $H_0^s(\Omega)=\left\{ u\in H^s(\RR^N)\,:\,u=0\textrm{ in } \Omega^c\right\}$ ([@fiscella2015density]).
Finally, in what follows we will indicate with $H^{-s}(\Omega)=\left(H^s(\Omega)\right)'$ the dual space of $H^s(\Omega)$ with respect tot the pivot space $L^2(\Omega)$.
A more exhaustive description of fractional Sobolev spaces and of their properties can be found in several classical references (see, e.g., [@adams2003sobolev; @dihitchhiker; @jllions1972non]).
Coming back to our problem, let us recall that the existence and uniqueness of weak solutions to is guaranteed by the following result (see, e.g., [@peradottolaplaciano Proposition 1.2.23]).
\[prop-ex\] Let $\Omega\subset\RR^N$ be an arbitrary bounded open set and $0<s<1$. Then for every $f\in H^{-s}(\Omega)$, the Dirichlet problem has a unique weak solution $u\in H_0^s(\Omega)$. Moreover, there exists a constant $C>0$ such that $$\begin{aligned}
\label{est-sol}
{{\left\|u\right\|}_{H_0^s(\Omega)}}\le C{{\left\|f\right\|}_{H^{-s}(\Omega)}}.\end{aligned}$$ In addition, we can take $C=\sqrt{2/{C_{N,s}}}$.
We remind that our main interest in the present work is the analysis of the behavior of the solutions of when $s\to 1^-$. The proof of Theorem \[limit\_thm\] is obtained employing classical techniques in functional analysis, as well as the following results.
\[limit\_prop\] For any $u\in C_0^\infty(\RR^N)$ the following statements hold:
- $\lim_{s\to 0^+}{(-\Delta)^su} = u$.
- $\lim_{s\to 1^-} {(-\Delta)^su} = -\Delta u$.
\[brezis\_prop\] For any $\varepsilon>0$, let $g_\varepsilon \in H^{1-\varepsilon}(\Omega)$. Assume that $$\begin{aligned}
\varepsilon{{\left\|g_\varepsilon\right\|}_{H^{1-\varepsilon}(\Omega)}}^2\leq C_0,\end{aligned}$$ where $C_0$ is a positive constant not depending on $\varepsilon$. Then, up to a subsequence, $\{g_\varepsilon\}_{\varepsilon>0}$ converges in $L^2(\Omega)$ (and, in fact, in $H^{1-\delta}(\Omega)$, for all $\delta > 0$) to some $g\in H^1(\Omega)$.
Finally, as we pointed out in Section \[intro\], our main result requires a sequence $\mathcal{F}_s$ satisfying the assumptions $\textbf{H1}$ and $\textbf{H2}$. The existence of such a sequence is guaranteed by the following.
\[limit\_f\] For any $f\in H^{-1}(\Omega)$ there exists a sequence $\mathcal{F}_s=\{f_s\}_{0<s<1}\subset H^{-s}(\Omega)$ verifying the assumptions
- ${{\left\|f_s\right\|}_{H^{-s}(\Omega)}}\leq C$, for all $0<s<1$ and uniformly with respect to $s$.
- $f_s\to f$ strongly in $H^{-1}(\Omega)$ as $s\to 1^-$.
Recall that any $f\in H^{-1}(\Omega)$ can be written as $f=\textrm{div}(g)$ with $g\in L^2(\Omega)$. Furthermore, let us introduce a standard mollifier $\rho_\varepsilon$ defined as $$\begin{aligned}
\rho_\varepsilon(x):= \begin{cases}
C\varepsilon^{-N}\exp\left(\frac{\varepsilon^2}{|x|^2-\varepsilon^2}\right), & \textrm{if } |x|<\varepsilon
\\
0, & \textrm{if } |x|\geq\varepsilon
\end{cases}\end{aligned}$$ and set $g_\varepsilon:=g\star\rho_\varepsilon$. It is classical that:
- $g_\varepsilon$ is well defined, since $g\in L^2(\Omega)$, hence it is locally integrable.
- $g_\varepsilon\in C_0^\infty(\Omega_\varepsilon)$, with $\Omega_\varepsilon:=\left\{x\in\Omega\,:\, \textrm{dist}(x,\partial\Omega)>\varepsilon\right\}.$
- $\partial_{x_i}g_\varepsilon$ is bounded uniformly with respect to $\varepsilon$ for all $i=1,\ldots,N$.
- $\lim_{\varepsilon\to 0^+} g_\varepsilon = g$, strongly in $L^2(\Omega)$.
Thus we can take $f_\varepsilon:=\textrm{div}(g_\varepsilon)$ and, from Property (iii) above, we immediately have that ${{\left\|f_\varepsilon\right\|}_{H^{-1+\varepsilon}(\Omega)}}$ is bounded uniformly with respect to $\varepsilon$. In addition, using Properties (ii) and (iv), it is straightforward that, for all $i=1,\ldots,N$, $\partial_{x_i} g_\varepsilon = \rho_\varepsilon\star g_{x_i} \to g_{x_i}$ as $\varepsilon\to 0^+$. Hence, $$\begin{aligned}
\lim_{\varepsilon\to 0^+}f_\varepsilon = \lim_{\varepsilon\to 0^+}\textrm{div}(g_\varepsilon) = \textrm{div}(g) = f,\end{aligned}$$ where the convergence is strong in $H^{-1}(\Omega)$. Therefore, by choosing $\varepsilon=1-s$, following the above argument we can construct a sequence $\{f_s\}_{0<s<1}\subset H^{-s}(\Omega)$ verifying $\textbf{H1}$ and $\widehat{\textbf{H2}}$.
Notice that $\widehat{\textbf{H2}}$ is a property of strong convergence in $H^{-1}(\Omega)$ which, clearly, implies the weak convergence in the same functional setting (property $\textbf{H2}$). Therefore, Proposition \[limit\_f\] provides a sequence $\mathcal{F}_s$ which is within the hypotheses of Theorem \[limit\_thm\].
The elliptic case: proof of Theorem \[limit\_thm\] {#weak_sec}
==================================================
In this Section, we give the proof of Theorem \[limit\_thm\] employing the definition of weak solution that we gave in Section \[prel\].
First of all, since we are interested in the behavior for $s\to 1^-$, until the end of the proof we will assume $s> 1/2$. Moreover, from $\textbf{H2}$ and the definition of weak convergence we get $$\begin{aligned}
\label{limit-rhs}
\lim_{s\to 1^-}\int_{\Omega} f_sv\,dx = \int_{\Omega} fv\,dx, \;\;\; \forall v\in\mathcal{D}(\Omega).\end{aligned}$$
For all $0<s<1$, let $u_s\in H_0^s(\Omega)$ be the solution to corresponding to the right-hand side $f_s$. According to Proposition \[prop-ex\], for $s$ sufficiently close to one we have the estimate $$\begin{aligned}
\label{norm_est}
\sqrt{1-s}{{\left\|u_s\right\|}_{H^s(\Omega)}}\leq\mathcal{C}(s,N){{\left\|f_s\right\|}_{H^{-s}(\Omega)}},\end{aligned}$$ with $$\begin{aligned}
\mathcal{C}(s,N) := \sqrt{\frac{2-2s}{{C_{N,s}}}}\end{aligned}$$
Moreover, for all $N$ fixed, the constant $\mathcal{C}(s,N)$ is decreasing as a function of $s$ (see Figure \[figure\]). This of course implies $$\begin{aligned}
\mathcal{C}(s,N) < \mathcal{C}\left(\frac 12,N\right) = \sqrt{\frac{\pi}{\Gamma\left(\frac{N+1}{2}\right)}}.\end{aligned}$$
table\[x=0,y=1\]; table\[x=0,y=3\]; table\[x=0,y=5\]; table\[x=0,y=7\]; table\[x=0,y=9\];
Therefore, from and the uniform boundedness of ${{\left\|f_s\right\|}_{H^{-s}(\Omega)}}$ we deduce that $$\begin{aligned}
\sqrt{1-s}{{\left\|u_s\right\|}_{H^s(\Omega)}}\leq C\end{aligned}$$ with $C$ depending only on $N$ and $\Omega$. This, thanks to Proposition \[brezis\_prop\], allows us to conclude that $u_s\to u$ strongly in $H^{1-\delta}_0(\Omega)$ for any $0<\delta\leq 1$, and that $u\in H_0^1(\Omega)$.
Notice that, according to [@warma2015fractional Section 6], for all $\phi\in H_0^s(\Omega)$ and $\psi\in\mathcal{D}(\Omega)$ we have the following identity $$\begin{aligned}
\big\langle {(-\Delta)^s\phi},\psi\big\rangle_{L^2(\Omega)} &=\frac{C_{N,s}}{2}\int_{\RR^N}\int_{\RR^N}\frac{(\phi(x)-\phi(y))(\psi(x)-\psi(y))}{|x-y|^{N+2s}}\;dxdy
\\
&= \big\langle \phi,{(-\Delta)^s\psi}\big\rangle_{L^2(\Omega)}.\end{aligned}$$
This can be applied to the variational formulation , which can thus be rewritten as $$\begin{aligned}
\label{weak-def-s}
\big\langle u_s,{(-\Delta)^sv}\big\rangle_{L^2(\Omega)} = \int_\Omega f_sv\,dx.\end{aligned}$$ Moreover, since $v\in\mathcal{D}(\Omega)$ we have $$\begin{aligned}
\big|u_s{(-\Delta)^sv}\big| \leq C|u_s|,\end{aligned}$$ where, clearly, $u_s\in L^2(\Omega)\hookrightarrow L^1(\Omega)$, being $\Omega$ a bounded domain. Hence we can use the Dominated Convergence Theorem and Proposition \[limit\_prop\] to conclude that $$\begin{aligned}
\lim_{s\to 1^-} \big\langle u_s,{(-\Delta)^sv}\big\rangle_{L^2(\Omega)} = \lim_{s\to 1^-} \int_\Omega u_s{(-\Delta)^sv}\,dx = -\int_\Omega u\Delta v\,dx = \int_\Omega \nabla u\cdot\nabla v\,dx.\end{aligned}$$ This, together with and implies that $u$ verifies $$\begin{aligned}
\int_\Omega \nabla u\cdot\nabla v\,dx = \int_{\Omega} fv\,dx, \;\;\; \forall v\in\mathcal{D}(\Omega),\end{aligned}$$ i.e. it is a weak solution to .
The result that we just proved is to some extent not surprising, due to the limit behavior of the fractional Laplacian as $s\to 1^-$. In fact, a hint that Theorem \[limit\_thm\] had to be true comes from the very classical example $$\begin{cases}
{(-\Delta)^su_s} =1, &x\in B(0,1)
\\
u_s\equiv 0, & x\in B(0,1)^c,
\end{cases}$$ whose solution is given explicitly by $$\begin{aligned}
u_s(x) = \frac{2^{-2s}\Gamma\left(\frac N2\right)}{\Gamma\left(\frac{N+2s}{2}\right)\Gamma(1+s)}\left(1-|x|^2\right)^s\chi_{B(0,1)}.\end{aligned}$$ Indeed, it can be readily checked that, for $x\in B(0,1)$, $$\begin{aligned}
\lim_{s\to 1^-} u_s(x) = \frac{1}{2N}\left(1-|x|^2\right):=u(x),\end{aligned}$$ which is the unique solution to the limit problem $$\begin{aligned}
\begin{cases}
-\Delta u =1 , &x\in B(0,1)
\\
u= 0, & x\in \partial B(0,1).
\end{cases}\end{aligned}$$
Of course, the above fact does not tell anything about the general case of problem . To the best of our knowledge, this is an issue that, although natural and probably expected, has not yet been fully addressed in the literature (at least, not in the setting of weak solutions with minimal assumptions) and our contribution helps to fill in this gap.
Additional results an further comments {#rem_sec}
======================================
Weakening the assumptions of Theorem \[limit\_thm\]
---------------------------------------------------
Scope of this section is to show that a convergence result in the spirit of Theorem \[limit\_thm\] can be obtained under weaker assumption on the sequence $\mathcal F_s$ of the right-hand sides of . In particular, we are going to prove the following.
\[limit\_thm\_weak\] Let $\mathcal F_s=\{f_s\}_{0<s<1}\subset H^{-1}(\Omega)$ be a sequence such that $f_s\rightharpoonup f$ weakly in $H^{-1}(\Omega)$. For all $f_s\in\mathcal F_s$, let $u_s$ be the corresponding solution to . Then, as $s\to 1^-$, $u_s\rightharpoonup u$ weakly in $L^2(\Omega)$, with $u$ solution to in the transposition sense.
First of all, since we are interesting in analyzing the behavior of $u_s$ as $s\to 1^-$, until the end of this proof we will always assume $s>1/2$. Moreover, observe that, the right-hand side $f_s$ belongs to $H^{-1}(\Omega)$, which is strictly greater than $H^{-s}(\Omega)$. Therefore, we cannot apply Lax-Milgram Theorem. Instead, we shall define the solution to in a different way.
For all $\phi\in L^2(\Omega)$, let $y$ be solution of the elliptic problem $$\begin{aligned}
\label{PE_transp}
\begin{cases}
{(-\Delta)^sy}=\phi, & x\in\Omega
\\
y \equiv 0, & x\in\Omega^c.
\end{cases}\end{aligned}$$
Recall that, due to the regularity of $\phi$ and to the results contained in [@biccari2017local; @cozzi2017interior], for all $\varepsilon>0$ we have $y\in H^{2s-\varepsilon}_0(\Omega)\hookrightarrow H^1_0(\Omega)$, with continuous and compact embedding.
Moreover, the map $\Lambda: \phi\mapsto y$ is linear and continuous from $L^2(\Omega)$ into $H^{2s-\varepsilon}_0(\Omega)$. Thus, $\Lambda$ is compact from $L^2(\Omega)$ into $H^1_0(\Omega)$ and its adjoint $\Lambda^*$ is a compact operator from $H^{-1}(\Omega)$ into $L^2(\Omega)$. In addition, $$\begin{aligned}
\langle f_s,y\rangle_{H^{-1}(\Omega),H^1_0(\Omega)} = \langle f_s,\Lambda \phi\rangle_{H^{-1}(\Omega),H^1_0(\Omega)} = (\Lambda^* f_s,\phi)_{L^2(\Omega)}.\end{aligned}$$
Therefore, $u_s:=\Lambda^*f_s\in L^2(\Omega)$ is a solution defined by transposition to , i.e. it satisfies $$\begin{aligned}
\label{transp_def_s}
\int_\Omega u_s\phi\,dx = \langle f_s,y\rangle_{H^{-1}(\Omega),H^1_0(\Omega)}. \end{aligned}$$ Moreover, we have $$\begin{aligned}
\label{us_norm}
{{\left\|u_s\right\|}_{L^2(\Omega)}} \leq C{{\left\|f_s\right\|}_{H^{-1}(\Omega)}}.\end{aligned}$$
In particular, $\{u_s\}_{0<s<1}$ is a bounded sequence in $L^2(\Omega)$, which implies that $u_s\rightharpoonup u$ weakly in $L^2(\Omega)$.
Notice that is obtained multiplying for $y$ and integrating over $\Omega$. Observe also that in this expression the functional spaces involved (namely $L^2(\Omega)$, $H_0^1(\Omega)$ and $H^{-1}(\Omega)$) do not depend on $s$. This, joint with and with the fact that $\phi\in L^2(\Omega)$, $f_s\in H^{-1}(\Omega)$ and $y\in H_0^1(\Omega)$, allows us to take the limit as $s\to 1^-$ in . Thanks to the Dominated Convergence Theorem, we then recover the expression $$\begin{aligned}
\label{transp_def}
\int_\Omega u\phi\,dx = \langle f,y\rangle_{H^{-1}(\Omega),H^1_0(\Omega)},\end{aligned}$$ i.e. $u$ is a solution by transposition to . Moreover, since the $L^2(\Omega)$-regularity of $u_s$ cannot be improved, its convergence to a solution to can be expected only in the weak sense.
Remarks on the convergence rate
-------------------------------
Our interest in the subject of this paper is motivated by previous results concerning the numerical approximation of the fractional Laplacian. In more detail, the issue that we addressed came from the observation that for the stiffness matrix $\mathcal A_h^s$ derived in [@biccari2017controllability] from the FE discretization of in dimension $N=1$ the following holds:
- $\lim_{s\to 0^+}\mathcal A_h^s = h\textrm{Tridiag}(1/6,2/3,1/6):=\mathcal I_h$, an approximation of the identity;
- $\lim_{s\to 1^-}\mathcal A_h^s = h^{-1}\textrm{Tridiag}(-1,2,-1):=\mathcal A_h$, the classical tridiagonal matrix for the FE approximation of the one-dimensional Laplacian.
The second property in particular implies that also the numerical solution $u_h^s$ associated to $\mathcal A_h^s$ converges to the one corresponding to $\mathcal A_h$. Therefore, investigating whether this still holds in the continuous case was a question that arose naturally.
While we answered to this question in Theorem \[limit\_thm\], there we did not specify under which rate this convergence occurs. In what follow, we present an informal discussion on this particular point.
During the proof of Theorem , we showed that the sequence $\{u_s\}_{0<s<1}$ of solutions to is bounded in $H_0^s(\Omega)$, with the following estimate $$\begin{aligned}
\label{us_norm_est}
\sqrt{1-s}{{\left\|u_s\right\|}_{H^s(\Omega)}}\leq C,\end{aligned}$$ with $C$ a constant uniform with respect to $s$. This last inequality, in turn, was obtained as a consequence of Proposition \[prop-ex\] and of the assumption $\mathbf{H1}$ on the sequence $\{f_s\}_{0<s<1}$ of the right-hand sides.
Moreover, the factor $\sqrt{1-s}$ in already appears in [@bourgain2001another] to correct the well-known defect of the seminorm $|\cdot|_{H^s(\Omega)}$ which, as $s\to 1^-$, does not converge to $|\cdot|_{H^1(\Omega)}$.
In fact, if $\zeta$ is any smooth non-constant function, then for all $1< p<\infty$ we have $|\zeta|_{W^{s,p}(\Omega)}\to +\infty$ as $s\to 1^-$. This situation may be rectified by multiplying by $(1 -s)^{1/p}$ in front of $|\zeta|_{W^{s,p}(\Omega)}\to +\infty$. IN particular, we have $$\begin{aligned}
\lim_{s\to 1^-} (1-s)^{\frac 1p}|\zeta|_{W^{s,p}(\Omega)} = \left(\int_\Omega |\nabla\zeta|^p\,dx\right)^{\frac 1p}. \end{aligned}$$
In view of these observations, we claim that the convergence that we obtained in Theorem \[limit\_thm\] satisfies the rate $$\begin{aligned}
\lim_{s\to 1^-}{{\left\|u_s-u\right\|}_{H^s(\Omega)}} \sim \mathcal O(\sqrt{1-s}).\end{aligned}$$
Indeed, if this convergence were slower, then we would still have blow-up phenomena in the $H^s(\Omega)$-seminorm. On the other hand, if the convergence were faster, then for some $\alpha>1/2$ $$\begin{aligned}
\lim_{s\to 1^-}(1-s)^{\alpha}|\cdot|_{H^s(\Omega)} = \lim_{s\to 1^-}\underbrace{(1-s)^{\alpha-\frac 12}}_{\to 0}\underbrace{\sqrt{1-s}\,|\cdot|_{H^s(\Omega)}}_{\to |\cdot|_{H^1(\Omega)}} = 0.\end{aligned}$$
Clearly, the discussion that we just presented is not a rigorous proof of our claim. Nevertheless, we believe that our statement is true, and a further confirmation is given by the following numerical simulations, where we compared the solution to and for different values of $s$ and we computed the approximation error in the $H^s(\Omega)$-norm. As expected, we observe a convergence of $u_s$ to $u$, with a rate of $\sqrt{1-s}$.
The parabolic case
------------------
As it most often happens, the properties of the solutions to elliptic problems can be naturally transferred into the parabolic setting. In our case, this translates in the fact that the solution $\phi_s$ to the fractional heat equation
$$\begin{aligned}
\label{FE}
\begin{cases}
\partial_t\phi_s + {(-\Delta)^s\phi_s} = g_s, &(x,t)\in\Omega\times(0,T)\tag{$\mathcal H_s$}
\\
\phi_s\equiv 0, & (x,t)\in\Omega^c\times(0,T)
\\
\phi_s(x,0) = 0, & x\in\Omega,
\end{cases}\end{aligned}$$
converges as $s\to 1^-$ to the one to the local problem $$\begin{aligned}
\label{HE}
\begin{cases}
\partial_t\phi -\Delta\phi = g, &(x,t)\in\Omega\times(0,T)\tag{$\mathcal H$}
\\
\phi= 0, & (x,t)\in\partial\Omega\times(0,T)
\\
\phi(x,0) = 0, & x\in\Omega.
\end{cases}\end{aligned}$$
First of all, let us recall that we have the following definition of weak solution for the parabolic problem (see, e.g., [@leonori2015basic]).
\[weak\_sol\_def\_parabolic\] Let $g_s\in L^2(0,T;H^{-s}(\Omega))$. A function $\phi_s\in L^2(0,T;H_0^s(\Omega))\cap C([0,T];L^2(\Omega))$ with $\partial_t\phi_s\in L^2(0,T;H^{-s}(\Omega))$ is said to be a weak solution to the parabolic problem if for every $\psi\in\mathcal{D}(\Omega\times(0,T))$, it holds the equality $$\begin{aligned}
\label{weak-sol-par}
\int_0^T \int_\Omega\partial_t\phi_s\psi\,dxdt &+ \frac{C_{N,s}}{2}\int_0^T\int_{\RR^N}\int_{\RR^N}\frac{(\phi_s(x)-\phi_s(y))(\psi(x)-\psi(y))}{|x-y|^{N+2s}}\;dxdydt \notag
\\
&= \int_0^T\int_\Omega g_s\psi\,dxdt.\end{aligned}$$
Moreover, thanks to [@leonori2015basic Theorem 26], existence and uniqueness of solutions is guaranteed. Namely, we have
Assume that $f_s\in L^2(0,T;H^{-s}(\Omega))$. Then problem has a unique finite energy solution, defined according to .
Then, adapting the methodology for the proof of Theorem , the following result is immediate.
\[limit\_thm\_parabolic\] Let $\mathcal{G}_s:=\{g_s\}_{0<s<1}\subset L^2(0,T;H^{-s}(\Omega))$ be a sequence satisfying the following assumptions for all $0<t<T$:
- ${{\left\|g_s(t)\right\|}_{H^{-s}(\Omega)}}\leq C$, for all $0<s<1$ and uniformly with respect to $s$.
- $g_s(t)\rightharpoonup g(t)$ weakly in $H^{-1}(\Omega)$ as $s\to 1^-$.
For any $f_s\in\mathcal{G}_s$, let $\phi_s\in L^2(0,T;H^s_0(\Omega))$ be the unique weak solution to the corresponding parabolic problem in the sense of Definition \[weak\_sol\_def\_parabolic\]. Then, as $s\to 1^-$, $(\phi_s,\partial_t\phi_s)\to(\phi,\partial_t\phi)$ strongly in $L^2(0,T;H^{1-\delta}_0(\Omega))\times L^2(0,T;H^{-1}(\Omega))$ for any $0<\delta\leq 1$. Moreover, $\phi\in L^2(0,T;H^1_0(\Omega))\times L^2(0,T;H^{-1}(\Omega))$ and verifies $$\begin{aligned}
\int_0^T \int_\Omega\partial_t\phi\psi\,dxdt + \int_0^T\int_\Omega \nabla\phi\cdot\nabla\psi\;dxdt = \int_0^T\int_\Omega g\psi\,dxdt, \;\;\; \forall\psi\in\mathcal{D}(\Omega\times(0,T)),\end{aligned}$$ i.e. it is the unique weak solution to .
First of all, notice that a sequence $\mathcal{G}_s$ verifying $\textbf{K1}$ and $\textbf{K2}$ exists. In fact, it can be constructed following the methodology of Proposition \[limit\_f\], since both properties are independent of the time variable. Moreover, it is evident that we shall only analyze the first term on the left-hand side of . This is due to the following two facts:
- The functional space in which the integration in time is carried out is fixed and does not depend on $s$. Therefore, the limit process does not affect the regularity in the time variable.
- For the remaining two terms in , the limit as $s\to 1^-$ can be addressed in an analogous way as in the proof of Theorem \[limit\_thm\].
On the other hand, since $\partial_t\phi_s\in L^2(0,T;H^{-s}(\Omega))$, the same argument previously developed for dealing with the term $$\int_0^T\int_\Omega g_s\psi\,dxdt$$ applies also to $$\int_0^T\int_{\Omega} \partial_t\phi_s\psi\,dxdt.$$
In this way, we immediately conclude that, as $s\to 1^-$, $(\phi_s,\partial_t\phi_s)\to (\phi,\partial_t\phi)$ strongly in $L^2(0,T;H^{1-\delta}_0(\Omega))\times L^2(0,T;H^{-1}(\Omega))$ for all $0<\delta\leq 1$ and, in particular, that $$\begin{aligned}
\lim_{s\to 1^-}\int_0^T\int_{\Omega} \partial_t\phi_s\psi\,dxdt = \int_0^T\int_{\Omega} \partial_t\phi\psi\,dxdt.\end{aligned}$$ This, together with the above remarks, implies that the function $\phi$ satisfies $$\begin{aligned}
\int_0^T \int_\Omega\partial_t\phi\psi\,dxdt + \int_0^T\int_\Omega \nabla\phi\cdot\nabla\psi\;dxdt = \int_0^T\int_\Omega g\psi\,dxdt, \;\;\; \forall\psi\in\mathcal{D}(\Omega\times(0,T)),\end{aligned}$$ i.e. it is the unique weak solution to .
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors wish to acknowledge Enrique Zuazua (Universidad Autoónoma de Madrid, DeustoTech and Laboratoire Jacques-Louis Lions) for having suggested the topic of this work. Moreover, a special thank goes to Xavier Ros-Oton (Universität Zürich) and Enrico Valdinoci (University of Melbourne) for interesting and clarifying discussions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate that if the universe is dominated by the massive cold dark matter, then besides the generally believed thermal distribution of the dark matter relics, there may exist some very energetic non-thermal relics of the dark matter particles in the universe from some unknown sources, such as from decay of supermassive X particle released from topological defect collapse or annihilation. Very interesting, we point out that these high energy dark matter particles may be observable in the current and future cosmic ray experiments.'
---
=0.1in =-0.3in =0.7truecm =0.2truecm tcilatex
=-0.4truecm =-0.4truecm
[[ **Cosmological Relics of the High Energy Cold Dark Matter Particle in the Universe**]{}]{}
[ **Xue-Qian Li$^{1,3}$ and Zhijian Tao$^{1,2}$**]{}
1\. China Center of Advanced Science and Technology, (World Laboratory)
P.O.Box 8730 Beijing 100080, China.
2\. Theory Division, Institute of High Energy Physics, Academia Sinica
Beijing 100039, China
3\. Department of Physics, Nankai University, Tianjin, 300071, China.
PACS: 95.35+d, 14.80-j, 13.85Tp, 98.80Cq
There is a great deal of evidence to indicate that the dominant component of the universe matter density is dark matter. The most direct evidence for existence of dark matter is from the astronomical observation of velocities of the spiral galaxies [@kol]. In our own galaxy, for example, in order to explain the observation, a local dark matter density $\rho ^{DM}=0.3{\rm GeV}/{\rm cm}^3$ is determined [@data]. In addition, the inflation theory predicts a flat universe, i.e. $\Omega =1$. The standard big-bang nucleosynthesis implies that the ordinary matter can contribute at most 15 percent of the critical density. It means that 90 percent of the matter in our universe may be dark [@tur]. The most attractive candidate of the dark matter is from the particle physics. One of them is the weakly-interacting massive particle (WIMP). The WIMP serves as a natural cold dark matter candidate to the universe, since if the WIMP is at a weak-scale mass it naturally provides a near closure density to the present universe as required by the inflation theory. In the present universe the WIMP dark matter particle is statically distributed. Its density is controlled by the Boltzman equation and the evolution of the universe, so it is called thermal relic abundance. So far experimental effort to search for the cold dark matter is made to find the signals of this distribution. The most promising experiments involve the direct detection at low background detectors, which look for the signal of dark matter particle as it collides with detector matter, and the indirect detection through observation of energetic neutrinos emerging from annihilation of cold dark matter particles which are accumulated in the Sun and the Earth [@kol]. Having not seen any signal in these experiments the parameter space of the WIMP models is seriously restricted. However to discover the WIMP dark matter particles or rule out the WIMP models much more efforts are needed, hopefully the next generation detectors will cover much larger parameter space of the WIMP models.
Instead, in this letter we discuss another possibility to search for cold dark matter. We start from a question, whether there exists a significant flux of very energetic cold dark matter particle in the universe. Since the thermal relics of cold dark matter are statically distributed in the universe, we are actually studying some non-thermal relics. It is believed that if heavy particles are produced at a temperature of the universe which is not much lower than the mass scale of the particle, the number of the high energy relics would reduce very fast due to thermalization processes and the expansion of the universe. As a result at present time there would be no heavy particles with high energies. However, there indeed is a possibility of existence of non-thermal relics at present universe. Its reason is not well understood, but they may wander around at present time. For example, at a later stage of the universe evolution when its temperature is sufficiently low, some topological defects, such as monopoles, cosmic strings are decoupled from the thermal bath of the universe. Later the superheavy fields, which form the defects, decay into some heavy particles. Since the parents are out of thermal equilibrium, the products are also non-thermally distributed over the space. That could serve as a new source for such non-thermal relics in our universe. Quite similar to the fact that there exists large flux of high energy cosmic ray components of nucleon and other stable particles of the standard model, it is natural to expect an existence of energetic cold dark matter particle in the cosmic rays as the dark matter particles dominate the universe density. The question is how much this kind of energetic dark matter particles remain in our universe. If the amount of the particles are very small, it is not interesting at all to study the potential of search for them. By a general and reasonable analysis we are suggesting there may exist a large amount of this kind of particles in the universe. Since the WIMP is electrically neutral and heavy, so the ordinary accelerating mechanisms for nucleon and neutrino in the cosmic rays cannot work for WIMP’s. Two mechanisms for generating energetic cold dark matter particles in the universe are considered below. One is through the collision of cosmic ray particles and statically distributed dark matter particles. Since the energy of the cosmic ray ranges from low to very high, the collision will result in some high energy cold dark matter particle fluxes. Another one is that high energy cold dark matter particles can also originate from the very early universe. Some examples are the decays of superheavy X particle released from destruction of topological defects formed during phase transition in the very high energy scale, like the grand unification scale. Recently some decays of the superheavy X particles with mass close to the grand unification scale released from topological defect to standard model light particles have been interpreted as a new mechanism to explain the observed extremely high energy cosmic rays, though it seems that the simple version of the topological defect model is not able to fit all the observed cosmic ray spectrum [@bla; @bha]. Similarly, the very massive X particle may also decay to cold dark matter particles resulting in the energetic particles in the universe. By considering these two sources we derive an evolution equation for the flux of the energetic cold dark matter in the universe. We assume an equilibrium solution to the equation, then we find that the flux of the high energy WIMP’s is sizable, therefore it should be detectable in some future large detectors. We stress that our scenario can also provide a possible explanation for the exotic event observed in Yunnan Cosmic Ray Station (YCRS) [@yun; @che], if it is confirmed.
It is a very interesting question to derive the effective density or flux of the high energy neutral WIMP particle in the present universe. We denote $E$ as the kinetic energy of the WIMP. For the static distribution of the thermal relic WIMP’s, $E=0$. Our task is to derive the density of WIMP’s with $E\not =0$, say from GeV to much higher energy. We assume that there already exists a non thermal relic density of WIMP’s with nonzero kinetic energy $n(E)$. As we stated, this density may result from some early universe sources, like decays of X particle released from topological defect. In principle , the scattering between the WIMP’s (including both static thermal and high energy non-thermal relics) and the energetic cosmic protons can make a new equilibrium with the existing high energy density $n(E)$.
We denote the differential WIMP density in the energy region from $E_2$ to $E_2+\Delta E_2$ as $\Delta n(E_2)$. The scattering cross section $\sigma(E_1,E_2,E_1^{\prime},E_2^{\prime})$ for a cosmic proton and WIMP into a WIMP plus other final state particles can be estimated, provided the model of WIMP interaction is given, here we just keep in mind this scattering is a weak interaction process. The equation for the distribution evolution of $n(E_2)$ is derived as the following
$$\begin{aligned}
{d\Delta n(E_2)\over dt} &=& -\int {d\sigma (E_1,E_2,E_1',E_2')\over dE_2'}
dE_2'
{dF(E_1)\over dE_1}dE_1\Delta n(E_2)(1-\delta_{E_2E_2'}) \nonumber\\
&+& \int
{d\Delta\sigma (E_1,E_2'',E_1',E_2)\over dE_2''}dE_2''
{dF(E_1)\over dE_1}dE_1n(E_2'')
(1-\delta_{E_2''E_2})\end{aligned}$$
where $E_1^{\prime },~E_2^{\prime }$ are the kinetic energies of the outgoing hadron and WIMP, $dF(E_1)/dE_1$ is the energy density of the charged cosmic ray particle flux and $\Delta \sigma (E_1,E_2^{\prime \prime
},E_1^{\prime },E_2)={\frac{d\sigma (E_1,E_2^{\prime \prime },E_1^{\prime
},E_2)}{dE_2}}\Delta E_2$. The first term on the left-hand side of the equation reduces $\Delta n(E_2)$ because a certain number of dark matter particles with energy from $E_2$ to $E_2+\Delta E_2$ are stricken out of the energy range by cosmic ray collision. On the other hand some dark matter particles with energy outside the range are stricken into the range by the cosmic ray collision. So the second term increases $\Delta n(E_2)$. We notice that the incident cosmic ray flux should include all possible sources in cosmic rays, charged particles, cosmic photons, neutrino background and etc. However the cross section of the WIMP and the very low temperature photon or neutrino scattering is much suppressed, so the effect can be neglected here. For the same reason we also neglect the contribution of cosmic rays with energy much lower than 1 GeV. The flux of energetic charged cosmic ray particles is measured on the earth and it has the spectrum in the outer space as $F(\Lambda )(E_1/\Lambda )^{-2.7}$ while $0.4{\rm GeV}\leq
E_1\leq 10^6$ GeV, the parameter $\Lambda $, $F(\Lambda )$ and the spectrum form for higher energy can be found in reference [@gai]. For our purpose the most relevant energy range should not be much higher than the weak scale, since for very high energy scale both the cosmic ray flux and the initial density $n(E_2)$ are too much suppressed (see below).
Obviously, after infinitely long evolution time, $d\Delta n(E_2)/dt=0$ to reach an equilibrium. However for a finite life time of the universe as long as $10^{10}$ years, the equilibrium is not necessarily reached. Our investigation is divided into two steps. Here we first assume the equilibrium condition is satisfied. We will come back to this point later. Moreover since the dominant density of dark matter is the static thermal relics, as a very good approximation one may use $n(E_2^{\prime\prime})=n(0)\delta_{E_2^{\prime\prime}0}$ in the above equation so one has $$\frac{\Delta n(E_2)}{\Delta E_2}= \frac{n(0)\int {\frac{d\sigma(E_1,0,E_1^{\prime},E_2) }{dE_2}} {\frac{dF(E_1)}{dE_1}}dE_1} {\int \sigma(E_1,E_2){\
\frac{dF(E_1)}{dE_1}}dE_1},$$ where $n(0)$ is the local Galactic halo density contributed by the WIMP dark matter, and $\sigma(E_1,E_2)$ means that all possible final state energies have been integrated over.
So long as the equilibrium condition is satisfied, the equilibrium solution indicates that the differential spectrum of WIMP’s in our galaxy is proportional to the local halo density. And it is independent of the details of the mechanism of the early universe evolution. No matter what the mechanism it is, it only needs to provide certain large initial $n(E_2)$ for the evolution equation, so that the equilibrium condition can be satisfied.
Without using detail information of any WIMP models we can very roughly estimate the equilibrium differential density spectrum of WIMP’s with large kinetic energy. If $E_2$ is much smaller than $10^6$ GeV while larger than a certain value, say 1 GeV, simply by dimension analysis one estimates $$\frac{dn(E_2)}{dE_2}\sim \frac{n(0)}{E_2}(\frac{1GeV}{E_2})^{2.7}$$ this corresponds to a flux $j(E_2\ge 10^2GeV)\sim 1{\rm cm}^{-2}{\rm s}^{-1}$. Though this flux is much smaller than the static dark matter flux across the Earth, which is about $10^5{\rm cm}^{-2}{\rm s}^{-1}$. This high energy WIMP flux may still cause some detectable signals in future large detectors, since it is a high energy involved process. To calculate the capture rate in the high energy process, one needs to specify a concrete WIMP model. However since the capture process is typically a weak interaction process, by assuming a weak interaction cross section as $10^{-35}-10^{-38}{\rm cm}^2$, we can reasonably estimate a capture rate as $10^{-5}-10^{-8}{\rm s}^{-1}$ per ton of target. The expected events in one year are 100 and 0.1 respectively.
There could be many new signatures to distinguish the desired processes from other background processes. We mention one possible signature which is to produce a heavy charged particle when the neutral dark matter hitting on the matter surrounding or inside the detector. For instance in SUSY model, if neutralino is the dark matter particle, there exist some charged supersymmetric particles. The mass difference of the charged particle and the neutralino should be comparable with the mass of neutralino itself, so if the incident kinetic energy of the neutralino is larger than the mass difference, the charged supersymmetric particle could be produced. Because the charged particle is heavy and unstable, it decays to some light products afterwards. The kinematics and decay properties of the charged particles may provide some clues for the incident dark matter particles.
In 1972, in the cloudy chamber of the YCRS, an exotic event with a heavy charged particle being tracked was observed [@yun]. In fact, three charged prongs were recorded and they correspondingly possessed three$-$momenta as $p_a=6.6_{-0.8}^{+10}$GeV/c, $p_b=62$GeV/c and $p_c=110$GeV/c with the spanned angles as $\theta _{ab}=3^{\circ }25\prime $, $\theta
_{bc}=1^{\circ }25\prime $and $\theta _{ac}=4^{\circ }55\prime $, namely they were coplanar trajectories. Later the tracks $a$ and $b$ were identified as $\pi ^{-}$ and proton while $c$ remained unknown.. A careful re-analysis which was done recently by Chen [*et al.*]{}[@che] indicates that the unidentified positively charged particle has a heavy mass as $$M^{+}>43\;{\rm GeV}$$ The authors of reference [@che] have suggested that the unknown charged heavy particle may be identified as a heavy elementary particle which was produced by a bombard of a heavy neutral particle coming from the universe on a proton. Obviously, in this case, the kinetic energy of the unknown neutral particles must be greater than the threshold $M^{+}-M^0+M_\pi$ where $M^0$ is the mass of the neutral particle. For so large momenta measured in the case, the neutral particle must be very energetic. If their postulation is valid, it would be a direct evidence for the existence of large flux of massive and high energy neutral particle in the cosmic ray, even though with a remarkable uncertainty. They also claimed to have found similar evidence at the data obtained by other experiments [@che]. Numerically the flux should be around $10 {\rm cm}^{-2}{\rm s}^{-1}$ in order to explain the observed capture rate. The question was where the high energy neutral particle flux is from. We see that the high energy WIMP flux derived above can give the necessary amount of flux.
Now we come back to discuss the equilibrium condition which is assumed before. We notice that an initial condition must be satisfied, otherwise the equilibrium cannot be reached for the present age of our universe. The condition is that the initial density $n(E_2)$ with $E_2$ being greater than a certain value is large enough. If one assumes that the initial $n(E_2)$ is negligibly small, from the evolution equation we can roughly estimate the order of the magnitude of the density, which is many order of the magnitudes smaller than the equilibrium density we obtained above. It means that by the collision of the high energy cosmic ray and the static WIMP dark matter particle alone, the produced high energy WIMP density in the universe is negligible. Therefore we need a large initial condition for the evolution equation. This implies that some new sources for generation of such initial condition should exist. We have named some possibilities such as decay of X particle released from topological defect[@bla]. Here the question is if these new mechanisms are indeed able to produce large enough $n(E_2)$ in the early universe. Certainly we are not trying to investigate all the possibilities, but considering only the case of cosmic string collapse [@bha1] and annihilation of monopole-antimonopole pairs [@bha2]. Following reference [@bha], the rate of release of massive particle X forming the topological defect from destruction of the defect can be expressed as $$\frac{dn_X(t)}{dt}=km_Xt^{-3}$$ where $m_X$ is the mass of the X particle, $t$ is the Hubble time, and $k$ is a dimensionless constant which is restricted to be less than about 0.2 by the cosmic ray spectrum observation [@bha]. It corresponds to $p=1$ in the equation (1) of reference [@bha]. Moreover, by assuming the massive particle X decay with a sizable branching ratio to stable WIMP particles, we can use the same formula of equation (5) in this reference for the differential particle flux $j(E_0)$. For the weak interaction feature of the WIMP we omit the energy loss effects when the particle propagates in the universe from some early time $t_i$ to the present time $t_0$, namely $dE_i/dE_0$ is dominantly determined by the expansion effect of the universe in the equation, here $E_i$ and $E_0$ are the energies of the particle at time $t_i$ and $t_0$ respectively. Then the flux is estimated as $$E_0\frac{dj(E_0)}{dE_0}\simeq ckm_X(\frac{m_X}{E_0})t_0^{-2}$$ where $c$ is the speed of light and $t_0$ is the age of the universe. Taking $m_X=10^{16}$ GeV, and $k=0.1$ we obtain the differential flux due to the X particle decay as $$E_0\frac{dj(E_0)}{dE_0}\sim \frac{100{\rm GeV}}{E_0}{\rm cm}^{-2}{\rm s}^{-1}$$ This flux is comparable with the equilibrium flux obtained above, so we conclude that it is possible to satisfy the equilibrium condition with the X particle decay as the source of high energy WIMP density in the early universe.
In this work we suggest to directly detect WIMP dark matter particles in the cosmic ray. The point is that we show that there may exist some sizable flux of high energy WIMP particles in the universe. Even though the flux is much smaller than that occurring in the direct dark matter search experiments, the detector for high energy experiments can be much bigger than that built for the low energy direct search experiments, therefore the numbers of events in high energy experiments are not necessarily less than that in the low energy experiments. And the YCRS event may be an indication of this kind of signal, even if the measurement on the event is not very reliable and needs new experiments to check. In fact some cosmic ray experiments in L3 are undergoing, one of the experimental purposes is to search for this kind of events [@ho]. Once the characteristics of those events are confirmed, the observation can be viewed as a strong evidence supporting the existence of the stable WIMP’s of high energy in the universe.
In order to understand where the high energy WIMP dark matter particles come from, we have derived an evolution equation for the differential density. With assumption of the equilibrium condition we estimate the differential density numerically and find that this density is interestingly large for future and even some current cosmic ray experiments. The interesting property of the equilibrium is that it does not depend on the details of any sources in the early universe, so long as the equilibrium condition is satisfied. We use the grand unification scale supermassive X particle released from $p=1$ type topological defect decay as an example to show that the equilibrium is possibly reached. One should have noticed that all the numerical estimation in this work is somehow crude, they may be one or two orders of magnitudes deviating from the stated values. And the decay mode of X particle is not clear, while we simply assume that it decays dominantly to the WIMP dark matter particle. Though there are those uncertainties, however, we believe that the qualitative features we discuss in this work are correct and may be very interesting to experimentalists. Finally we emphasize that the possibility of existence of high energy cold dark matter particle in the universes deserves much attention of both theoreticians and experimentalists. If the equilibrium state is not reached in the real world, so the flux is not as large as the equilibrium result we derived , the desired flux or density of the cold dark matter particle in the universe from some new sources, like X particle decay from topological defect destruction, can still be large enough for the future experimental detection.
It is a pleasure to express our gratitude to C. Qing and T. Ho for useful discussion. This work is partly supported by the National Natural Science Foundation of China (NSFC).
[99]{} E. Kolb and M. Turner, [ The Early Universe]{}, Addison-Wesley Pub.Co. (1990) New York; P.F. Smith and J.D. Lewin, Phys. Rep. 187, 203 (1990); G. Jungman, M. Kamionkowski and K. Greist, Phys. Rep. 267, 196 (1995); J. Parimack et al., Phys. Rev. Lett. 74, 2160 (1995).
Data Group, Phys.Rev. D54 (1996) 1.
M. Turner, in Proceeding of 17th International Symposium on lepton-photon Interaction, Beijing, 1995, edited by Zhi-Peng Zheng and He-Sheng Chen, (World Scientific, Singapore 1995), p. 693.
J. J. Blanco-Pillado, R. A. Vázquez and E. Zas; R. J. Protheroe and T. Stanev, Phys. Rev.Lett. 77, 3708 (1996); G. Gigl, S. Lee and David N. Schramm, astro-ph/9610221; G. Sigl and S. Lee, astro-ph/9604093. Recently some alternative new mechanisms are proposed to explain the observed extremely high energy cosmic ray. They involve some new heavy stable colored particles or qusistable massive particles as the sources of the cosmic rays. In principle these new sources could be the sources of high energy cold dark matter particles we discussed in this context. For the reference to see V. Berezinsky, M. Kachelrie$\beta $ and A. Vilenkin, astro-ph/9708217; R. N. Mohapatra and S. Nussinov, hep-ph/9708497; P.J.H. Chung, G.R. Farrar and E.W. Kolb, astro-ph/9707036.
P. Bhattacharjee, C. T. Hill and D. N. Schramm, Phys. Rev. Lett. 69, 567 (1992)
Yunnan Cosmic Ray Station, Institute of Atomic Energy, Academia Sinica, Phys. 1 (1972) 57. (in Chinese).
H. Chen [*et al*]{}, Phys. Rep. 282, 1(1997).
T. Gaisser, [*Cosmic Rays and Particle Physics*]{}, Cambridge University Press 1990.
P. Bhattarcharjee and N. C. Rana, Phys. Lett. B 246, 365 (1990).
P. Bhattarcharjee and N. C. Rana, Phys. Rev. D51, 4079 (1995).
T. Ho, private communication. A proposal is made to L3 experiments by a group of Chinese physicists.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The flow behavior of a semiflexible polymer in microchannels is studied using Multiparticle Collision Dynamics (MPC), a particle-based hydrodynamic simulation technique. Conformations, distributions, and radial cross-streamline migration are investigated for various bending rigidities, with persistence lengths $L_p$ in the range $0.5 \le L_p/L_r \le 30$. The flow behavior is governed by the competition between a hydrodynamic lift force and steric wall-repulsion, which lead to migration away from the wall, and a locally varying flow-induced orientation, which drives polymer away from the channel center and towards the wall. The different dependencies of these effects on the polymer bending rigidity and the flow velocity results in a complex dynamical behavior. However, a generic effect is the appearance of a maximum in the monomer and the center-of-mass distributions, which occurs in the channel center for small flow velocities, but moves off-center at higher velocities.'
address: '$^\dag$Institut für Festkörperforschung and $^{\ddag}$Institute for Advanced Simulation, Forschungszentrum Jülich, 52425 Jülich, Germany\'
author:
- 'Raghunath Chelakkot$^{\dag}$, Roland G. Winkler$^\ddag$, Gerhard Gompper$^{\dag, \ddag}$'
bibliography:
- 'polymer\_microfluidics.bib'
title: 'Semiflexible polymer conformation, distribution and migration in microcapillary flows'
---
Introduction
============
Polymers, Vesicles and Cells in Micro- and Nanochannels
-------------------------------------------------------
Complex fluids — such as polymer solutions, colloidal dispersions, and suspensions of vesicles or cells — exhibit an intriguing flow behavior, particularly in confined geometries. The interplay of internal degrees of freedom, e.g, conformational changes of polymers or shape changes of vesicles and cells, combined with fluid mediated interactions, inhomogeneous flow profiles, and wall interactions lead to novel and often unexpected effects [@wall:78; @broc:77; @vlie:90; @jend:03; @bald:07; @jend:03_1; @tege:04; @agar:94; @jend:04; @gg:gomp05g; @usta:06; @khar:06; @stei:06; @usta:07; @cann:08; @mark:09; @gg:gomp09d; @gg:gomp10d]. Insight into theses aspects provides not only an understanding of phenomena for conventional applications such as lubrication, adhesion, polymer processing, and blood flow, but also of emergent microfluidic devices [@squi05; @whit:06] and the lab-on-a-chip technologies [@crai06] with their micro- or nano-size length scale. In this context, the investigation and analysis of single molecules or cells is of particular interest, since on the one hand such knowledge is essential to develop and optimize micron- and nano-scale devices, while on the other hand the study of individual particles often provides a more detailed microscopic knowledge than can be extracted from an ensemble.
The size of red blood cells (RBCs), about $8\mu$m in diameter, is comparable to the size of microvessels and microcapillaries in a mammalian body. Thus, the deformability and dynamics of an individual cell determines the flow behavior of blood in the capillary network. Due to the physiological importance of this process and the relatively large size of the objects, detailed investigations of this system started already about 50 years ago [@skal69; @seco86; @prie96; @pozr05a]. Modern microfluidic techniques [@blaz:06; @whit:06; @utad:07] and new mesoscale simulation approaches [@dzwi03; @gg:gomp04h; @dupi07; @gg:gomp08d] provide new insight into the delicate interplay between flow forces, cell deformation, and hydrodynamic interactions [@gg:gomp05g; @gg:gomp09d; @abka08b; @toma09]. In particular, in microchannels slightly larger than the RBC diameter, it has been found that single red blood cells transform from their equilibrium shape of biconcave disks into a parachute shape at a critical flow velocity, which depends linearly on the bending rigidity and the shear modulus of the cell membrane [@gg:gomp05g]. At small volume fractions of red blood cells, they show a strong clustering tendency due to hydrodynamic interactions [@gg:gomp09d]. At higher volume fractions, several distinct phases are found, which range from disordered arrangement of discocytes over single-file motion of aligned parachutes to a staggered, zig-zag arrangement of slipper shapes [@gg:gomp09d]. In channels which are smaller than the RBC diameter, they assume bullet shapes at sufficiently high flow velocities [@seco86; @abka08b; @toma09].
Fluid vesicles are distinct from red blood cells since their membrane is fluid, and therefore has a vanishing shear modulus. The flow behavior of vesicles has been studied intensively in recent years, in particular in shear flows. Much less is known about fluid vesicles in capillary flows. In narrow, homogeneous channels, vesicles assume bullet shapes [@vitk04]. In structured, saw-tooth shaped channels, which are somewhat larger than the vesicle size, a complex conformational and dynamical behavior has been found [@gg:gomp10d]. For nearly spherical vesicles, a transition from symmetric shape oscillations to orientational oscillations has been predicted with decreasing flow velocity. For shapes, which deviate more strongly from a sphere, experiments and simulations show new shapes with two symmetric or a single asymmetric tail [@gg:gomp10d].
In narrower channels, with a diameter on the order of 100 nm, it is possible to investigate the static and dynamic behavior of single DNA molecules. This system is interesting for several reasons. First, microfluidic devices allow for the manipulation, sizing, and sorting of DNA fragments [@chou99; @levy10], and thereby a direct visualization of genomic information [@tege:04; @pers10]. Second, DNA has turned out to be an ideal model to study the properties of single polymers, because, when fluorescently labelled, its conformations, dynamics, and flow properties can be observed directly under a microscope [@wall:78; @perk95; @smit99; @reis:05; @schr:05; @gera:06] or by correlation spectroscopy [@petr:06]. Thus, DNA is also very well suited to study the behavior of semiflexible polymers in micro- and nanochannels.
With a persistence length of $L_p \simeq 17 \mu m$, filamentous actin is a biopolymer for which semiflexible behavior becomes important already in wider channels than for DNA (with its persistence length $L_p \simeq 50 nm$). Thus, experiments with actin are essential to elucidate the behavior of semiflexible polymers in microchannels without [@koes05; @koes08; @koes09] and with flow [@stei:08]. In equilibrium, the parallel extension of a semiflexible polymer shows scaling behavior with power-law dependencies on channel diameter and persistence length [@odij83; @reis:05; @gg:gomp07e; @levi07; @odij08].
Structured channels are particular interesting, because they allow to study time-dependent flows [@gg:gomp10d], threshold forces or flow rates for penetration into narrow pores [@gg:gomp95h; @saka05; @link06; @mark:09], and flow injection into a pore [@mark:09]. For fluid vesicles, it has been shown that the threshold strength of a driving field (e.g., an electric field) for narrow pores increases nearly linearly with membrane bending rigidity and vesicle area, and decreases rapidly with the pore radius [@gg:gomp95h]. For polymers, it has been predicted by scaling arguments [@saka05] and confirmed by simulations [@mark:09] that the threshold velocity flux for entry into a narrow pore is [*independent*]{} of both the polymer length and the pore radius.
Polymer Migration in Micro- and Nanochannels
--------------------------------------------
An interesting aspects of polymer transport in the presence of a zero-slip wall is cross-streamline migration. This effect has been observation in planar shear flow in the presence of walls as well as Poiseuille flows [@jend:03; @jend:03_1; @jend:04; @usta:06; @khar:06; @usta:07; @cann:08; @send:08] and in experiments [@tege:04; @stei:06; @bald:07], as a formation of depletion layer near the wall. The thickness of the depletion layer is found to increase with flow, which indicates that migration away from the wall increases with the flow rate. This migration phenomenon is explained as a results of polymer-wall hydrodynamic interactions. Analytical as well as simulation studies, which properly account for such hydrodynamic interactions, qualitatively reproduced the experimentally observed wall-induced migration.
In microchannels, confinement effects and a spatially varying shear rate due to the parabolic flow velocity profile imply additional features compared to simple shear flows. The steric interaction with the channel wall restricts the conformations of a flexible polymer and induce alignment to a semiflexible polymer. Moreover, the spatially varying local shear rate changes the conformations and alignment of a polymer as a function of its lateral position. In addition to the wall hydrodynamic interactions, these two effects also influence the qualitative behavior of cross-streamline migration. Simulations of flexible polymers have shown that under strong confinement there is a net migration away from the channel center, contrary to the predictions of wall induced inward migration [@usta:06]. The outward migration is also found in relatively weak flow [@cann:08]. In both cases, the outward migration has been attributed to the suppression of steric interactions with the wall due to an enhanced alignment by the flow. At large flow strengths, wall hydrodynamic interactions dominate, resulting a net inward migration of the polymer center-of-mass. Interestingly, the center-of-mass density at the channel center is also found to decrease with flow and at large flow strengths a density maximum is found at a distance away from the center. The reason behind the off-center peak of the center-of-mass distribution for flexible polymers is the flow-induced conformational change, which leads to stretched polymers close to the wall and coiled chains in the central part [@jend:03; @jend:03_1; @jend:04; @usta:06; @khar:06; @usta:07; @cann:08]. This give rise to an enhanced outward diffusion and ultimately leads to a concentration dip at the center.
Although the cross-streamline migration behavior of flexible polymers has been extensively studied, there are only a few studies, which focus on the migration behavior of semiflexible and rodlike polymers [@schi:96; @sain:06]. Understanding the migration properties of semiflexible and stiff polymers is important for the study of many biopolymers, such as actin filaments, microtubules, and intermediate filaments. Since polymer conformations are determined by its rigidity, the migration behavior of semiflexible and rodlike polymers can differ significantly from that of flexible polymers.
In this paper, we use computer simulations to study and discuss the migration properties of semiflexible polymers with persistence length ranging from half up to many times of its contour length. We employ a hybrid simulation scheme including Molecular Dynamics (MD) for the polymers and Multiparticle Collision Dynamics (MPC), a particle-based hydrodynamics simulation technique, for the solvent, so that wall hydrodynamic interactions and confinement effects are taken into account explicitly. The details of the method are given in Sec. \[sec:methods\]. Our results, presented in Secs. \[sec:concentration\] and \[sec:alignment\], show that polymer migration is indeed qualitatively influenced by bending rigidity. Furthermore, in Sec. \[sec:whi\], we investigate the importance of hydrodynamic interactions by comparing the results of simulations with MPC and Brownian solvents. Finally, the dynamics of polymer migration across the channel is studied in Sec. \[sec:migration\]
Model and Simulation Method {#sec:methods}
===========================
Semiflexible Polymer
--------------------
The linear polymer is comprised of $N_m$ point-like monomers of mass $M$ each, which are connected by the harmonic potential $$U_s = \frac{\kappa_s}{2} \sum_{i=1}^{N_m-1} \left(|{{\mbox{\bf {\em r}}}}_{i+1}
-{\mbox{\bf {\em r}}}_i | - b \right)^2,$$ where ${\mbox{\bf {\em r}}}_i$ denotes the position of particle $i$, $b$ the bond length, and $\kappa_s$ the force constant. Excluded-volume interactions are taken into account by the Lennard-Jones potential [@muss:05] $$U_{LJ} = \left\{ \begin{array}{ll}
4 \epsilon \left[ \left( \frac{ \displaystyle \sigma}{\displaystyle r}\right)^{12} -
\left( \frac{\displaystyle \sigma}{\displaystyle r}\right)^{6}\right] + \epsilon \ , & r < 2^{1/6} \sigma \\
0 \ , & \mbox{else}
\end{array} \right. .$$ To account for polymer stiffness, the bending potential is applied [@wink:04] $$U_b = \frac{\kappa_b}{2} \sum_{i=2}^{N_m-1} \left({{\mbox{\bf {\em R}}}}_{i+1}
-{\mbox{\bf {\em R}}}_{i} \right)^2 ,$$ where ${\mbox{\bf {\em R}}}_i ={\mbox{\bf {\em r}}}_i-{\mbox{\bf {\em r}}}_{i-1}$ is the bond vector and $\kappa_b$ is the bending rigidity. In the semiflexible limit, the bending rigidity is related to the persistence length by $L_p=\kappa_b/k_BT$.
The dynamics of the polymer is described by Newtons’ equations of motion, which are integrated using the velocity-Verlet algorithm.
![Snapshots of polymer conformations close to the channel center (top) and the wall (bottom) for the Peclet number $Pe =360$ and the persistence lengths $L_p/L_r \approx 0.5$ (a), (d), $2.1$ (b), (e), and $30.8$ (c), (f). Movies are available as Supporting Material [@movie]. \[fig:snapshot\]](fig1.eps){width="12cm"}
Hydrodynamic Solvent
--------------------
In the MPC algorithm, the fluid is described by a set of $N$ point-like particles of mass $m$ each, which move in continuous space with velocities determined by a stochastic process. Their dynamics evolves in two steps. In the streaming step, the solvent particles move ballistically for a time $h$, which we denote as collision time. In the collision step, particles are sorted into the cells of a cubic lattice of lattice constant $a$ and their relative velocities, with respect to the center-of-mass velocity of each cell, are rotated around a random axis by an angle $\alpha$. The orientation of the axis is chosen independently for every cell and collision step. For every cell, mass, momentum, and energy are conserved in this process. The algorithm is described in detail in refs. [@male:99; @male:00; @kapr:08; @gomp:09; @ripo:04; @ripo:05]. The fluid is confined in a cylindrical channel with periodic boundary conditions along the channel axis. No-slip boundary conditions are imposed on the channel walls by the bounce-back rule and virtual wall particles, as described in ref. [@lamu:01]. The flow is induced by a gravitational force ($mg$) acting on every fluid particle.
The interaction of a polymer with the solvent is realized by inclusion of its monomers in the MPC collision step [@male:00_1]. Between two MPC steps, several MD steps are performed to update the positions and velocities of the monomers. Extensive studies of polymer dynamics confirm the validity of this procedure [@ripo:05; @webs:05; @muss:05; @male:00_1].
To maintain a constant temperature, the velocity scaling algorithm is applied as described in Ref. [@huan:10]. Here, an kinetic energy $E_k$ is chosen from the gamma distribution independently for every cell, and the individual particle velocities of that cell are multiplied by the factor $(2 E_k/(m \sum
{\mbox{\bf {\em v}}}_i))^{1/2}$, where the sum runs over all particles in the considered cell. This assures that the velocity distribution of the particles is Maxwellian. Similarly, a Maxwellian velocity distribution is obtained by the Monte Carlo procedure described in Ref. [@hech:05].
Simulations of the pure solvent system yield velocity profiles which agree with the solution of Stokes’ equation for the considered geometry.
Brownian Solvent
----------------
An advantage of the MPC approach is that hydrodynamic interactions can easily be switched off, without altering the monomer diffusion significantly [@kiku:03; @ripo:07]. In this case, denoted as Brownian MPC, each monomer independently performs a stochastic collision with a phantom particle with a momentum taken from the Maxwell-Boltzmann distribution with variance $m \left\langle N_c
\right\rangle k_BT$, where $\left\langle N_c \right\rangle$ is the average number of solvent particles per collision cell [@ripo:07].
Parameters
----------
For the solvent, we employ the parameters $\alpha =130 ^\circ$, $h=0.1 \tau$, with $\tau=\sqrt{ma^2/k_{B}T}$ ($k_B$ is Boltzmann’s constant and $T$ is temperature), $\left\langle N_c \right\rangle
=10$, $M=m \left\langle N_c \right\rangle, b=\sigma =a$, the fluid mass density $\varrho = \left\langle N_c \right\rangle m/a^3$, and $k_BT/\epsilon =1$. The time step in the MD simulation is set to $h_{MD} = 5\times10^{-3} \tau$. A polymer with $N =14$ monomers is placed in a cylindrical channel of radius $R = 8 a$. With the length $L_r=(N-1)a=13a$, the polymer does not interact with the wall when its center of mass is near the channel center. In order to maintain a constant contour length of the polymer, we set $\kappa_s = 3 \times 10^3 /(k_BT/b^2)$. The persistence length $L_p$ is varied, by selecting the values $\kappa_b /(k_BT/b^2) =
7, \ 28, \ 50, \ 100, \ 200, \ 400$, which correspond to the persistence lengths $L_p/L_r \approx 0.5, \ 2.1, \ 3.8, \ 7.7, \
15.4, \ 30.8$. The channel length is $28 a$.
The strength of the applied pressure field is characterized by the Peclet number $Pe = {\dot \gamma} \tau_R$, where ${\dot \gamma} =
g \varrho R /(2 \eta)$ is the shear rate at the cylinder wall. The Reynolds number $Re = \varrho R v_m/ \eta = \varrho R^2 \dot
\gamma/(2 \eta)$, where $v_m$ is the maximum fluid velocity, depends linearly on the shear rate. For the above MPC parameters, the viscosity [@ripo:04] is such that $Re < 1$ for all considered $\dot \gamma$. Equilibrium simulations for a system with periodic boundary conditions yield the end-to-end vector relaxation time $\tau_R \approx 3200 \tau$. This value agrees within approximately $20\%$ with the relaxation time obtained theoretically for a semiflexible polymer with the same ratio $L_r/L_p$ [@wink:06_1]. The relaxation time of the Brownian MPC simulation is $\tau_R \approx 8200 \tau$.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial monomer distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:conc1\]](fig2a.eps "fig:"){width="60mm"} ![Radial monomer distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:conc1\]](fig2b.eps "fig:"){width="60mm"}
(c) (d)
![Radial monomer distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:conc1\]](fig2c.eps "fig:"){width="60mm"} ![Radial monomer distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:conc1\]](fig2d.eps "fig:"){width="60mm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The averages and probability distributions presented in the following sections are calculated in the stationary state for various independent initial conditions.
Radial Distribution Functions {#sec:concentration}
=============================
The imposed flow, with a parabolic flow profile, determines the polymer conformations, with a few examples shown in fig. \[fig:snapshot\], and the polymer distribution functions, such as the monomer distribution $P_{mo}$ and the center-of-mass distribution $P_{cm}$. Their dependence on flow strength will be characterized in the following sections.
Radial Monomer Distribution
---------------------------
Radial monomer distributions are presented in fig. \[fig:conc1\] for various flow rates and stiffnesses. They are normalized such that $$\int_0^{R/a} r P_{mo}(r/a) \ dr/a^2=1 .
\label{eqn2}$$ A feature in common with all polymer systems with hydrodynamic interactions is the decrease of the monomer concentration adjacent to the wall with increasing flow rate. Even without flow, there is a depletion zone for all stiffnesses, which extends approximately one radius of gyration into the channel. However, the distributions exhibit distinct differences to those obtained previously for flexible and rodlike polymers, respectively. For persistence lengths $L_p/L_r \lesssim 8$, the density at the center of the channel decreases with increasing Peclet number as shown in figs. \[fig:conc1\]a,b. In contrast, for stiffer polymers with $L_p/L_r > 8$ the density at the channel center first increases with increasing $Pe$ and then decreases at large Peclet numbers. Simultaneously, the wall induced migration towards the channel center causes an increase in concentration at a finite distance from the center for all stiffnesses. Two effects contribute to the formation of the maximum. On the one hand, there is cross-streamline migration due to hydrodynamic interactions, and, on the other hand, the flow field causes an alignment of a molecule (see section \[sec:alignment\]), which increases the local density at the radial positions of strong flow-induced alignment.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial monomer distributions for the Peclet numbers $Pe =73$ (a) and $360$ (b) and the persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$). \[fig:conc2\] ](fig3a.eps "fig:"){width="60mm"} ![Radial monomer distributions for the Peclet numbers $Pe =73$ (a) and $360$ (b) and the persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$). \[fig:conc2\] ](fig3b.eps "fig:"){width="60mm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The distribution functions are rather similar for the various stiffnesses at $Pe=0 $ within the accuracy of the simulation. Interestingly, stiffer polymers exhibit a more pronounced migration away from the wall for $Pe \lesssim 100$ as is shown in fig. \[fig:conc2\]a. Simultaneously, $P_{mo}$ decreases in the channel center for small stiffnesses, leading to an off-center maximum, and increases for larger ones, where the maximum is at the center for some of the larger stiffnesses. With increasing flow rate, off-center density maxima are obtained for all stiffnesses. For the flow rate $Pe=360$, we observe a decrease in the depletion zone adjacent to the wall with increasing stiffness for $L_p/L_r < 8$ and a reversion of the trend at larger stiffnesses as depicted in fig. \[fig:conc2\]b.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial center-of-mass distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:cm1\] ](fig4a.eps "fig:"){width="60mm"} ![Radial center-of-mass distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:cm1\] ](fig4b.eps "fig:"){width="60mm"}
(c) (d)
![Radial center-of-mass distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:cm1\] ](fig4c.eps "fig:"){width="60mm"} ![Radial center-of-mass distributions for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:cm1\] ](fig4d.eps "fig:"){width="60mm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The width of the distribution, $$\langle r^2 \rangle = \int_0^{R/a} r^3 P_{mo}(r/a) \ dr/a^2 ,$$ emphasizes the differences in migration behavior. As displayed in fig. \[fig:width\]a, stiffer polymers exhibit a smaller width at moderate Peclet numbers, indicating an enhanced net inward migration with increasing bending rigidity, but $\langle r^2
\rangle$ seems to saturate at large flow rates. In contrast, the widths of more flexible polymers show only minor changes at small $Pe$ and decrease more rapidly at larger Peclet numbers. Their widths can be smaller than those of stiffer polymers.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![ Radial center-of-mass distributions for the Peclet numbers $Pe =73$ (a) and $360$ (b) and persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$). \[fig:cm2\] ](fig5a.eps "fig:"){width="60mm"} ![ Radial center-of-mass distributions for the Peclet numbers $Pe =73$ (a) and $360$ (b) and persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$). \[fig:cm2\] ](fig5b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Widths $\langle r^2 \rangle$ of the radial monomer distributions (a) and center-of-mass distributions (b) for the persistence lengths $L_p/L_r \approx 0.5$ ($\bullet$), $2.1$ ($\square$), $3.8$ ($\diamond$), $7.7$ ($\vartriangle$), $15.4$ ($\star$), and $30.8$ ($+$). \[fig:width\] ](fig6a.eps "fig:"){width="60mm"} ![Widths $\langle r^2 \rangle$ of the radial monomer distributions (a) and center-of-mass distributions (b) for the persistence lengths $L_p/L_r \approx 0.5$ ($\bullet$), $2.1$ ($\square$), $3.8$ ($\diamond$), $7.7$ ($\vartriangle$), $15.4$ ($\star$), and $30.8$ ($+$). \[fig:width\] ](fig6b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Radial Center-of-Mass Distribution
----------------------------------
Radial polymer center-of-mass distributions are displayed in fig. \[fig:cm1\]. Qualitatively, $P_{cm}$ exhibits similar features as the monomer distribution $P_{mo}$, reflecting the same physical mechanisms. For $L_p/L_r \lesssim 8$, the concentration at the channel center decreases with increasing $Pe$. The density profiles for $50 \lesssim Pe \lesssim 150$ are very similar and indicate a weak dependence of the center-of-mass properties on the flow rate only. For high $Pe$ values, the densities at the channel center decrease significantly for all stiffnesses, and a clear maximum appears at $r_{cm}> 0$. The density decrease with flow near the wall is mainly due to wall induced cross-streamline migration of polymers and partly due to steric interactions of the very stiff polymers with the wall, while the density decrease at the channel center indicates a migration away from the center. The steric contribution to migration is evident from the Brownian MPC simulations presented in sec. \[sec:whi\]. The two competing migration mechanisms result in the formation of a maximum at a distance away from the center at large $Pe$. The maximum is less pronounced for the center-of-mass distribution than for the monomer distribution.
Figure \[fig:cm2\] shows center-of-mass distributions of polymers with different bending rigidities for $Pe \approx 73$ and $Pe \approx 360$. As for even smaller Peclet numbers, the distributions exhibit a maximum essentially in the channel center for $Pe \approx 73$. An off-center maximum only appears for larger flow rates, which is accompanied by a density decrease in the channel center. Moreover, a pronounced depletion layer is visible at the wall for large stiffnesses.
Flexible polymers exhibit a pronounced density increase adjacent to the wall with increasing flow rate and before migration sets in, whereas very stiff polymers exhibit an increased depletion layer at the wall with increasing flow rate, which we attribute to steric polymer-wall interactions. This is reflected in the flow rate dependence of the width of the center-of-mass distribution displayed in fig. \[fig:width\]b. For $L_p/L_r \lesssim 4$, the width increases with increasing Peclet numbers for small $Pe$. This increase is explained by the suppression of polymer-wall steric interactions due to flow alignment of polymers (cf. section \[sec:alignment\]). For flexible polymers such an alignment is enhanced by flow-induced polymer stretching along the channel axis. However, at larger $Pe$, the wall induced hydrodynamic lift force dominates the dynamics, which results in a decrease of width. This dependence is qualitatively different for polymers with $L_p/L_r \gtrsim 5$, where wall induced migration dominates even at low $Pe$, implying a decrease in width. However, at large Peclet numbers, we observe a saturation or even increase in width by outward migration due to hydrodynamic interactions, which counter balances the radially inward migration, as discussed in Sec. 6.
Alignment and Conformations {#sec:alignment}
===========================
Orientational Order Parameter
-----------------------------
In microchannel flows with no-slip boundary conditions the local shear rate changes linearly with radial position. Hence, the flow-induced force experienced by a polymer depends on its radial position. Its conformations and alignment, in response to the flow forces, is expected to vary with its bending rigidity and its radial center-of-mass position. Studies of flexible and semiflexible polymers [@cann:08; @chel:10] reveal large orientational changes by the imposed flow, which is important for cross-streamline migration [@send:08]. To study the influence of flow on the polymer orientation, we consider the orientational order parameter $$S(r_{cm})=\frac{1}{2}\langle 3 \cos^{2}\theta - 1 \rangle ,$$ where $\theta$ is the angle between the polymer end-to-end vector and the flow direction, as a function of its center-of-mass radial position. $S(r_{cm})$ is plotted for various Peclet numbers and stiffnesses in fig. \[fig:op1\]. In the absence of flow and for all $\kappa_b$, the polymer orientation is isotropic at the channel center. For distances close to the wall, confinement causes a preferred orientation along the channel axis with $S
\approx 1$. With increasing flow rate, the order parameter increases for all stiffness and reaches a plateau at large $Pe$ for certain radii. The plateau value and its extension depends on stiffness as shown in fig. \[fig:op2\].
The order parameter exhibits a non-monotonic dependence on flow strength close to the channel center and depends on polymer stiffness as shown in the inset of fig. \[fig:op2\]. An increase in stiffness leads to a decrease of the order parameter for $L_p/L_r < 15$. For larger stiffnesses, $S(0)$ is larger for larger $L_p/L_r$ at all radial distances. At $Pe=360$, $S$ decreases with increases stiffness and becomes even negative for $L_p/L_r \simeq 2$, compare the inset of fig. \[fig:op2\]. As the polymer stiffness increases, $S$ increases close to the channel center. The negative values of $S$ at the channel center are attributed to the formation of U-shaped conformations, compare fig. \[fig:snapshot\]. Such conformations are not possible for flexible polymers [@cann:08] and polymers with large bending rigidity, hence $S$ assumes higher values at the center as $\kappa_b$ increases.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Orientation order parameter for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:op1\] ](fig7a.eps "fig:"){width="60mm"} ![Orientation order parameter for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:op1\] ](fig7b.eps "fig:"){width="60mm"}
(c) (d)
![Orientation order parameter for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:op1\] ](fig7c.eps "fig:"){width="60mm"} ![Orientation order parameter for the Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with $L_p/L_r \approx 3.8$ (a), $7.7$ (b), $15.4$ (c), and $30.8$ (d). \[fig:op1\] ](fig7d.eps "fig:"){width="60mm"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Orientational order parameters $S$ for Peclet numbers $Pe =73$ (a) and $360$ (b) and persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$). Inset: $S$ close to the channel center for $Pe=360$. The line is a guide to the eye. \[fig:op2\]](fig8a.eps "fig:"){width="60mm"} ![Orientational order parameters $S$ for Peclet numbers $Pe =73$ (a) and $360$ (b) and persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$). Inset: $S$ close to the channel center for $Pe=360$. The line is a guide to the eye. \[fig:op2\]](fig8b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Bending energies $U_b$ for Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with persistence lengths $L_p/L_r \approx 3.8$ (a) and $30.8$ (b). \[fig:ub1\]](fig9a.eps "fig:"){width="60mm"} ![Bending energies $U_b$ for Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with persistence lengths $L_p/L_r \approx 3.8$ (a) and $30.8$ (b). \[fig:ub1\]](fig9b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Bending Energy
--------------
U-shaped conformations can be quantitatively analyzed by calculating the average bending energy $U_{B}$. Figure \[fig:ub1\] displays $U_b$ as function of the radial polymer center-of-mass position for various persistence lengths and for various Peclet numbers. For $Pe = 0$ the bending energy is nearly uniform across the channel cross section and decreases near the wall, due to wall induced alignment. The magnitude of $U_b$ is close to the thermal average $U_b = (L_p/a-1)k_B T$ of the nearly harmonic bending potential for $L_p/L_r > 1$, as expected. An increase in $Pe$ results in an increase in its absolute value, which is a result of the conformational changes of the polymer in response to the force exerted by the flow, compare fig. \[fig:snapshot\]. At a given $Pe$ and radial position, the flow exerts the same force on a polymer, but its conformations depend on its bending rigidity. Polymers with large bending rigidity hardly undergo conformational changes, while polymers with rather small $\kappa_b$ go through significant conformational changes, but their bending energies are similar. This is exemplified in fig. \[fig:ub2\], where $U_b/(k_BT)$ is shown for various bending rigidities and $Pe = 360$.
In figs. \[fig:ub1\] and \[fig:ub2\], the radial energy profiles are qualitatively similar for all $\kappa_b$ (with $L_p/L_r
>1$) for a given $Pe$. For $0.5 \lesssim L_p/L_r \lesssim 5$, the polymer adopts U-shaped conformations at the center and hence $U_b$ is large. Away from the center, $U_b$ decreases as U-shapes disappear. With increasing radial center-of-mass position, the local shear rate increases and a polymer with small persistence length assumes transient bent conformations. The average over individual configurations provides a high value for $U_b$ for such $r_{cm}$. The extent of bending decreases as $\kappa_b$ is increased, since for stiffer polymers $(1- \cos \vartheta) \sim
U_b/\kappa_b$, where $\vartheta$ is the angle between successive bond vectors. More aligned bonds and hence smaller angle $\vartheta$ imply that $U_b/\kappa_b$ decreases with increasing $\kappa_b$, in agreement with the results of fig. \[fig:ub1\].
![Bending energies for Peclet number $Pe=360$ and persistence lengths $L_p/L_r \approx 0.5$ ($\bullet$), $2.1$ ($\square$), $3.8$ ($\diamond$), $7.7$ ($\vartriangle$), $15.4$ ($\star$), and $30.8$ ($+$). \[fig:ub2\]](fig10.eps){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial monomer distributions for systems without hydrodynamic interactions. The persistence lengths are $L_p/L_r = 3.8$ (a) and $30.8$ (b), and the Peclet numbers $Pe =0$ ($\circ$), $147$ ($\square$), $360$ ($\vartriangle$), and $920$ ($\star$). \[fig:dis\_mon\_nohi\]](fig11a.eps "fig:"){width="60mm"} ![Radial monomer distributions for systems without hydrodynamic interactions. The persistence lengths are $L_p/L_r = 3.8$ (a) and $30.8$ (b), and the Peclet numbers $Pe =0$ ($\circ$), $147$ ($\square$), $360$ ($\vartriangle$), and $920$ ($\star$). \[fig:dis\_mon\_nohi\]](fig11b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial center-of mass distributions for systems without hydrodynamic interactions. The persistence lengths are $L_p/L_r = 3.8$ (a) and $30.8$ (b) and the Peclet numbers $Pe =0$ ($\circ$), $147$ ($\square$), $360$ ($\vartriangle$), and $920$ ($\star$). \[fig:dis\_cm\_nohi\] ](fig12a.eps "fig:"){width="60mm"} ![Radial center-of mass distributions for systems without hydrodynamic interactions. The persistence lengths are $L_p/L_r = 3.8$ (a) and $30.8$ (b) and the Peclet numbers $Pe =0$ ($\circ$), $147$ ($\square$), $360$ ($\vartriangle$), and $920$ ($\star$). \[fig:dis\_cm\_nohi\] ](fig12b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Structural Properties without Hydrodynamic Interactions {#sec:whi}
=======================================================
As is generally accepted by now, hydrodynamic interactions determine the behavior of polymers in microcapillary flows. In order to separate effects due to intramolecular and surface hydrodynamic interactions from those caused by the flow and steric interactions, we performed Brownian MPC simulations [@ripo:07]. Brownian dynamics simulation have also been performed in Ref. [@sain:06], with a polymer composed of rodlike segments. However, intramolecular hydrodynamic interactions are taken into account by employing anisotropic drag coefficients parallel and perpendicular to the rodlike segments. In contrast, our polymer model is composed of point-like monomers with isotropic friction. Hence, we consider a free draining chain, whereas in Ref. [@sain:06] intramolecular hydrodynamic interactions are taken into account implicitly.
Radial Concentration Distributions
----------------------------------
The radial monomer probability distributions for the persistence lengths $L_p/L_r \approx 2.1$ and $30.8$ are presented in fig. \[fig:dis\_mon\_nohi\] for various flow rates. For $Pe <
200$, the distribution functions are rather similar. Adjacent to the wall, the distributions for small stiffnesses are even similar for all flow rates. At large flow rates and persistence lengths $L_p/L_r < 3.5$, we find a monomer density increase at $r_{cm}/a \approx 4$, which we attribute to the induced alignment of the polymers by the flow. However, for larger stiffnesses, we find an inward migration adjacent to the wall due to increased steric polymer-surface interactions (cf. fig. \[fig:dis\_mon\_nohi\]b). Here, the polymer tumbling motion combined with the smaller number of conformational degrees of freedom at large stiffnesses leads to a strong repulsion from the wall. At the same time, the density in the channel center increases with a maximum at $r\approx 0$.
The radial center-of-mass distribution exhibits an outward migration even without hydrodynamic interactions for $L_p/L_r <
3.5$ due to polymer alignment (cf. fig. \[fig:dis\_cm\_nohi\]). In contrast, $P_{cm}$ becomes narrower with increasing flow rate for $L_p/L_r \approx 30.8$, similar to the monomer distribution. Hence, a qualitative different behavior is obtained depending on the ability of the polymer to adjust to the conformational restrictions by the walls.
The Brownian simulations of ref. [@sain:06] display in strong outward migration of the polymers, which leads to a density minimum at the channel center and a large density increase near the walls. Two effects contribute to this behavior. On the one hand alignment of the polymers and on the other hand anisotropic diffusion by the aligned segments. We do not observe such an increase in density adjacent to the wall, because no hydrodynamic effects are included in our simulations.
Alignment
---------
The orientational order parameters display qualitatively similar dependencies on flow rate as those for systems with hydrodynamic interactions (cf. figs. \[fig:op1\] and \[fig:op2\]). However, alignment is far less pronounced and much larger Peclet numbers are required to achieve significant flow alignment. For $Pe=360$, we find $S\approx 0.2$ at $L_p/L_r \approx 3.8$, compared to $S=0.6$ in the presence of hydrodynamic interactions. At the same time, we also observe the appearance of U-shaped structures in the channel center for larger Peclet numbers. The comparison shows that certain qualitative features of polymer alignment are determined by the flow profile, rather than hydrodynamic interactions. However, hydrodynamic interactions clearly influence the polymer orientation in a quantitative manner and lead to a more pronounced alignment.
Cross-Streamline Dynamics {#sec:migration}
=========================
The most striking effect of flow on a semiflexible polymer is the strong dependence of its orientation on $Pe$ and the radial distance. This aspect provides the key to understand the appearance of the off-center maximum in the center-of-mass distribution function.
As is well known, a rod in solution exhibits a larger diffusion coefficient parallel to its axis than perpendicular to it. In the absence of flow, a semiflexible polymer behaves very similar to a stiff rod, we therefore expect that the observed polymer orientational differences across the channel will lead to differences in the lateral diffusion behavior.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial center-of-mass displacements $\Delta r (r_{cm})$ and mean square displacements $\Delta r^2(r_{cm})$ (insets) for Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with persistence lengths $L_p/L_r \approx 3.8$ (a) and $30.8$ (b). \[fig:drift\_stiffness\]](fig13a.eps "fig:"){width="60mm"} ![Radial center-of-mass displacements $\Delta r (r_{cm})$ and mean square displacements $\Delta r^2(r_{cm})$ (insets) for Peclet numbers $Pe =0$ ($\circ$), $73$ ($\square$), $147$ ($\diamond$), and $360$ ($\vartriangle$) of polymers with persistence lengths $L_p/L_r \approx 3.8$ (a) and $30.8$ (b). \[fig:drift\_stiffness\]](fig13b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(a) (b)
![Radial center-of-mass displacements $\Delta r (r_{cm})$ and mean square displacements $\Delta r^2(r_{cm})$ (insets) for persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$), and $Pe=73$ (a) and $Pe=360$ (b). \[fig:drift\_peclet\]](fig14a.eps "fig:"){width="60mm"} ![Radial center-of-mass displacements $\Delta r (r_{cm})$ and mean square displacements $\Delta r^2(r_{cm})$ (insets) for persistence lengths $L_p/L_r \approx 0.5$ ($+$), $2.1$ ($\bullet$), $3.8$ ($\square$), $7.7$ ($\diamond$), $15.4$ ($\vartriangle$), and $30.8$ ($\star$), and $Pe=73$ (a) and $Pe=360$ (b). \[fig:drift\_peclet\]](fig14b.eps "fig:"){width="60mm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
In order to characterize this diffusive behavior, we calculate the radial mean displacement $\Delta r (r_{cm}) = \left\langle [{{\mbox{\bf {\em r}}}}_{cm}(t_0 + \Delta t) - {{\mbox{\bf {\em r}}}}_{cm}(t_0)] {{\mbox{\bf {\em e}}}}_{cm} (t_0)
\right\rangle$, where ${{\mbox{\bf {\em e}}}}_{cm}= {{\mbox{\bf {\em r}}}}_{cm}/ |{{\mbox{\bf {\em r}}}}_{cm}|$, and the center-of-mass mean square displacement $\Delta
r^2(r_{cm}) = \left\langle ({{\mbox{\bf {\em r}}}}_{cm}(t_0 + \Delta t) {{\mbox{\bf {\em e}}}}_{cm} (t_0) - \left\langle {{\mbox{\bf {\em r}}}}_{cm}(t_0) {{\mbox{\bf {\em e}}}}_{cm}
(t_0)\right\rangle )^2 \right\rangle$ within a certain time $\Delta t$ for a polymer with the center-of-mass position ${{\mbox{\bf {\em r}}}}_{cm}$ at $t_0$. We choose the time $\Delta t$ such that the center-of-mass displacement is smaller than $a$ and the mean square displacement is linear in time. For an homogeneous and isotropic system, $\Delta r$ and $\Delta r^2$ are independent of the initial position.
The mean values are the first and second moment of the conditional probability distribution function to find the polymer center-of-mass at position $r_{cm}(t_0+\Delta t)$ at the time $t_0
+ \Delta t$, when it has been at position $r_{cm}(t_0)$ at the time $t_0$. Initially, the distribution is very narrow, ideally a $\delta$ function, and broadens in time. In the limit $\Delta t
\to \infty$ the conditional probability distribution turns into the equilibrium distribution function as displayed in figs. \[fig:cm1\] and \[fig:cm2\].
Figures \[fig:drift\_stiffness\] and \[fig:drift\_peclet\] display $\Delta r$ and $\Delta r^2$ for various Peclet numbers and stiffnesses. The mean displacements evidently depend on flow rate and stiffness. At $Pe=0$, there is no drift in the channel center. Only for $r_{cm}/a \gtrsim 1.5$, we observe a drift, which we attribute to steric polymer wall interactions, and which increase adjacent to the wall. For larger flow rates, the drift for $r_{cm}
\lesssim 4$ is close to zero due to alignment of a polymer and lack of steric wall interactions (cf. fig. \[fig:drift\_peclet\] a). At large Peclet numbers, $\Delta r$ is positive for $r_{cm}/a
< 2.5$ (cf. fig. \[fig:drift\_peclet\]b) and all persistence lengths $L_p/L_r>2$, which indicates an outward migration. The wall interactions lead to a larger drift of polymers towards the channel center adjacent to the wall, which increases with increasing stiffness.
The mean square displacement decreases with increasing radial distance for $r_{cm} < 4 a$ at all stiffnesses and flow rates. Since $\Delta r$ is rather small for $r_{cm} < 4 a$, $\Delta r^2$ is determined by diffusion, i.e., the radial diffusion is larger in the channel center. This leads to a minimum in the center-of-mass distribution at large flow rates and stiffnesses. At larger radial distances and large stiffnesses, the mean square displacement is strongly affected by wall interactions. There is no simple explanation for the non-monotonic behavior of the mean square displacement. Various wall interactions contribute to the system behavior: the wall lift force, steric wall interactions, and an increase of the tumbling frequency with increasing flow rate [@chel:10; @wink:06_1].
[*Without hydrodynamic interactions*]{}, the drift at $Pe=0$ is comparable to that of the system with hydrodynamic interactions. It changes with increasing Peclet number due to flow alignment, but only slightly. The mean square displacement is independent of the radial position for distances $r_{cm} < 6 a$ for all Peclet numbers $Pe <1000$ and stiffnesses as shown in fig. \[msd\_nohi\]. Only close to the wall, there is a decrease due to steric wall interactions. This supports our conclusion that intramolecular hydrodynamic interactions are responsible for outward migration in systems with hydrodynamic interactions. At much larger Peclet numbers ($Pe \approx 2000$), the mean square displacement is constant for $r<4a$, increases then with the radial distance and decreases again beyond $r =6 a$, similar to the behavior shown in the inset of fig. \[fig:drift\_peclet\]a for $L_p/L_r =15.4$. This behavior is explained by the strong steric polymer wall interaction and fast tumbling motion at high flow rates. Note that the tumbling frequency increases with the shear rate [@chel:10; @wink:06_1]. Hence, systems with and without hydrodynamic interactions exhibit similar features, but larger Peclet numbers are required for systems without hydrodynamic interactions and the behavior of systems with hydrodynamic interactions is richer and additional features are displayed.
![Radial center-of-mass mean square displacements without hydrodynamic interactions for the persistence lengths $L_p/L_r \approx 30.8$ and the Peclet numbers $Pe=360$ ($\vartriangle$), $920$ ($\star$), and $1830$ ($\times$). \[msd\_nohi\]](fig15.eps){width="60mm"}
Summary and Conclusions
=======================
We have analyzed the flow behavior of semiflexible polymers confined in a cylindrical channel and systematically studied the influence of bending rigidity on structural properties and cross-streamline migration. The study includes polymers with moderate bending rigidity, comparable to actin filaments, to very stiff rodlike polymers.
Poiseuille flow induces a strong radial-dependent polymer alignment parallel to the channel associated with a radially outward migration of its center-of-masses at small Peclet numbers and, at large Peclet numbers, a wall induces inward migration.
Stiffness changes the polymer behavior qualitatively and quantitatively. Taking a system [*without*]{} hydrodynamic interactions as a reference, there is no outward migration and there are no wall hydrodynamic forces. Correspondingly, the center-of-mass density decreases in the channel center with increasing flow rate and up to a certain stiffness ($L_p/L_r <20$) due to polymer alignment. For large stiffnesses, however, steric polymer-wall interactions cause inward migration of the molecule and an increase in density at the channel center with increasing flow rate. We attribute this effect to the increase of the tumbling frequency at large flow rates and hence more polymer-wall contacts. Hence, even for such a system, a qualitative different behavior is found for large and small stiffnesses.
This picture is enriched by hydrodynamic interactions, which enhance outward migration. At the same time wall-lift forces lead to inward migration and an off-center density maximum is formed at large flow rates. Such a maximum has also been found in experiments using actin filaments [@stei:08]. The center-of-mass distribution strongly depends on stiffness and the presence of hydrodynamic interactions, as is reflected in its width. Flexible polymers exhibit a more complex migration behavior than stiffer polymers due to conformational changes, i.e, stretching and alignment. For large stiffnesses, wall steric interactions lead to a strong inward migration as for polymers without hydrodynamic interactions. The semiflexible polymers exhibit a non-universal behavior when changing stiffness and flow rate. At low flow rates, stiffer polymers display a stronger inward migration, whereas at high flow rates a reversed trend is observed initially and ultimately a strong inward migration at large stiffnesses. It is the interplay between confinement, hydrodynamic interactions, and stiffness which leads to the emergence of a complex structural and dynamical behavior of the polymer. This is reflected in the stiffness dependence of the mean radial and mean square radial displacements.
Although the understanding of the flow behavior of soft mesoscale particles in nano- and microchannels has made considerable progress in recent years, this field can be expected to attract far more attention in the future. For example, flow in structured channels has hardly been addressed so far. Also, interactions between polymers at finite concentration, and between polymers and other soft particles, needs to be investigated.
Acknowledgments
---------------
The financial support and the stimulating environment of the DFG priority program “Nano- and Microfluidics" is gratefully acknowledged. In particular, we thank R. Finken and U. Seifert (Stuttgart), L. Schmid and T. Franke (Augsburg), and D. Steinhauser and T. Pfohl (Göttingen & Basel) for many helpful and inspiring discussions.
References {#references .unnumbered}
==========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Some aspects of integrable field theories possessing purely transmitting defects are described. The main example is the sine-Gordon model and several striking features of a classical field theory containing one or more defects are pointed out. Similar features appearing in the associated quantum field theory are also reviewed briefly.'
title: Purely transmitting integrable defects
---
Classical picture
=================
The context of this topic is two-dimensional (i.e. one space – one time) integrable classical (or quantum) field theory, and the basic question concerns how to ‘sew’ together field theories defined in different segments of space.
The setup
---------
The simplest situation has two scalar fields, $u(x,t)$, $x<x_0$ and $v(x,t)$, $x>x_0$, with a Lagrangian density given formally by $$\label{lagrangian} {\cal L}=\theta(x_0-x){\cal L}_u +
\theta(x-x_0){\cal L}_v +\delta (x-x_0) {\cal B}(u,v)\,.$$ The first two terms are the bulk Lagrangian densities for the fields $u$ and $v$ respectively, while the third term provides the sewing conditions; it could in principle depend on $u$, $v$, $u_t$, $v_t$, $u_x$, $v_x$, … but the interesting question is how to choose ${\cal B}$ so that the resulting system remains integrable [@bczlandau].
With free fields, there are many ways to choose ${\cal B}$. For example, $$\label{deltaimpurity} {\cal B}(u,v)=-\sfrac12\,\sigma
uv+\sfrac{1}{2}(u_x+v_x)(u-v)\,,$$ with standard choices for the bulk Lagrangians, leads to the following set of field equations and sewing conditions, $$\begin{array}{rcll}
(\partial^2+m^2)u&=&0\,,\quad&x<x_0\,,\\
(\partial^2+m^2)v&=&0\,,\quad&x>x_0\,,\\
u&=&v\,,\quad&x=x_0\,,\\
v_x-u_x&=&\sigma u\,,\quad& x=x_0\,,
\end{array}$$ implying the fields are continuous with a discontinuity in the derivative. This is an example of a $\delta$-impurity. Typically, the sewing conditions at $x=x_0$ lead to reflection and transmission and (for $\sigma <0$) a bound state. However, if the fields on either side have nonlinear but integrable interactions (e.g. each is a sine-Gordon field), the $\delta$-impurity destroys the integrability (but still interesting [@Goodman02]).
If both $u$ and $v$ are sine-Gordon fields a suitable choice of Lagrangian would be to take $$\label{D}
\begin{array}{rcl}
{\cal B}(u,v)&=&\frac{1}{2}(vu_t-uv_t)+{\cal D}(u,v)\,,\\[9pt]
{\cal D}(u,v)&=&\disty
-2\biggl(\sigma\cos\frac{u+v}2+\frac{1}{\sigma} \cos \frac{u-v}{2}\biggr)\\
\end{array}$$ leading to the set of equations $$\label{sG}
\begin{array}{rcll}
\partial^2 u&=&-\sin u\,,\quad&x<x_0\,,\\
\partial^2 v&=&-\sin v\,,\quad&x>x_0\\[6pt]
u_x&=&\disty
v_t-\sigma \sin \frac{u+v}{2}-\frac{1}{\sigma}\sin\frac{u-v}{2}\,,\quad&x=x_0\\[9pt]
v_x&=&\disty
u_t+\sigma\sin\frac{u+v}{2}-\frac{1}{\sigma}\sin\frac{u-v}{2}\,,\quad&x=x_0\,.
\end{array}$$ (The bulk coupling and mass parameter have been scaled away for convenience.) This set up is not at all the same as the $\delta$-impurity: it is integrable, there is no bound state for any value of the parameter $\sigma$, and typically $u(x_0,t)-v(x_0,t)\ne 0$, implying a discontinuity in the fields. Clearly, the equations (\[sG\]) describe a ‘defect’ (occasionally called a ‘jump-defect’ to distinguish it from other types). Note also that the sewing conditions are strongly reminiscent of a Bäcklund transformation, and would be a Bäcklund transformation if they were not ‘frozen’ at $x=x_0$ (see, for example, [@Backlund]. That this setup is integrable can be verified by constructing Lax pairs using techniques similar to those described in [@bcdr] for boundary situations.
Since the setup (\[sG\]) is local, it is clear there may be many defects, with parameters $\sigma_i$, at different locations $x_i$ along the $x$-axis.
Energy and momentum
-------------------
Time translation invariance is not violated by the defect and therefore there is a conserved energy, which includes a contribution from the defect itself. On the other hand, space translation is violated by a defect and therefore momentum might not be expected to be conserved even allowing for a contribution from the defect. It is worth investigating this in more detail in terms of the quantity ${\cal D}(u,v)$ appearing in (\[D\]).
The momentum carried by the fields on either side of the defect at $x=x_0$ is given by $$\label{momentum}
P=\int_{-\infty}^{x_0}\D x\,u_xu_t+\int_{x_0}^\infty\D x\,v_xv_t$$ and thus, using the defect conditions coming from (\[D\]), $$\label{}
u_x=v_t-\frac{\partial{\cal D}}{\partial u}\,,\quad
v_x=u_t+\frac{\partial{\cal D}}{\partial v}\,,\quad\mathrm{at}\;\,
x=x_0\,,$$ one finds $$\label{Pdot}
\dot P=\left[-v_t \frac{\partial{\cal D}}{\partial u} -
u_t \frac{\partial{\cal D}}{\partial v}-V(u)+W(v)
+\frac{1}{2}\left(\frac{\partial{\cal D}}{\partial u}\right)^2 -
\frac{1}{2}\left(\frac{\partial{\cal D}}{\partial v}\right)^2 \right]_{x_0}.$$ In this expression the fields on either side of the defect have been allowed to have (possibly different) potentials. Clearly, (\[Pdot\]) is not generally a total time-derivative of a functional of the two fields. However, it will be provided, at $x=x_0$, $$\label{Dequations}
\frac{\partial^2{\cal D}}{\partial u^2}=\frac{\partial^2{\cal
D}}{\partial v^2}\,, \qquad \frac{1}{2}\left(\frac{\partial{\cal
D}}{\partial u}\right)^2 - \frac{1}{2}\left(\frac{\partial{\cal
D}}{\partial v}\right)^2 = V(u)-W(v)\,.$$ This set of conditions is satisfied by the sine-Gordon defect function (\[sG\]). However, there are other solutions too, for example Liouville–Liouville, Liouville–massless free, free–free. In fact, in many cases investigated so far, including cases with several scalar fields [@bczlandau; @bcztoda], it turns out that the requirements of integrability coincide with the requirement that there be a modified conserved momentum.
Classical scattering and solitons
---------------------------------
It is not difficult to check that the free-field limit of the sine-Gordon setup, given by $${\cal D}(u,v)\rightarrow \frac{\sigma}{4}(u+v)^2+\frac{1}{4\sigma}(u-v)^2,$$ leads to conditions describing a purely transmitting jump-defect (i.e. no reflection). Given that fact, it is natural to ask what might happen with solitons in the nonlinear sine-Gordon model (for details concerning solitons, see for example [@Scott73].
A soliton travelling in the positive $x$ direction (rapidity $\theta$) is given by expressions $$\E^{\I u/2}=\frac{1+\I E}{1-\I E}\,,\quad x<x_0\,;\qquad \E^{\I
v/2}=\frac{1+\I zE}{1-\I zE}\,,\quad x>x_0\,,$$ where $$E=\E^{ax+bt+c}\,,\quad a=\cosh\theta\,,\quad
b=-\sinh\theta\,,\quad \hbox{with $\E^c$ real}.$$ The defect conditions (\[sG\]) are satisfied provided ($\sigma
=\E^{-\eta}$) $$z=\frac{\E^{-\theta}+\sigma}{\E^{-\theta}-\sigma}\equiv \coth
\left(\frac{\eta-\theta}{2}\right),$$ and it is worth noting that $z^2$ would represent the delay experienced by a soliton of rapidity $\theta$ passing another of rapidity $\eta$. As it is, the quantity $z$ may change sign, meaning, in fact, that a soliton can convert to an anti-soliton, or vice-versa, besides being delayed, or even absorbed. In the latter case, the defect gains a unit of topological charge in addition to storing the energy and momentum of the soliton; in the former, the defect gains (or loses) two units of topological charge. Because the defect potential has period $4\pi$, all evenly charged defects have identical energy–momentum, as do all oddly charged defects. A fascinating possibility associated with this type of defect (if it can be realized in practice) would be the capacity to control solitons (see, for example [@cz]). Several defects affect progressing solitons independently; several solitons approaching a defect (inevitably possessing different rapidities) are affected independently, with at most one of the components being absorbed. Notice, too, that the situation is not time-reversal invariant owing to the presence of explicit time derivatives in eqs(\[sG\]). Starting with an odd charged defect, energy–momentum conservation would permit a single soliton to emerge. However, classically, there is nothing to determine the time at which the decay of the defect would occur. In that situation, quantum mechanics should supply a probability for the decay — and indeed it does.
Quantum picture
===============
The transmission matrix
-----------------------
Following the remarks made in the last section one expects two types of transmission matrix, one of them, $^{\rm even}T$, referring to even-labelled defects — and this is expected to be unitary, since these defects cannot decay — and the other, $^{\rm odd}T$, referring to odd-labelled defects. The latter is not expected to be unitary, yet would be expected to be related (via a bootstrap principle) to a complex bound state pole in the former. In fact this is precisely what happens and, remarkably enough, the relevant transmission matrices were described by Konik and LeClair some time ago [@Konik97]. Using roman labels to denote soliton states (taking the value $\pm 1$), and greek labels to label the charge on a defect, and assuming topological charge is conserved in every process, it is expected that both transmission matrices will satisfy ‘triangle’ compatibility relations with the bulk $S$-matrix, for example: $$\label{STT} S_{ab}^{cd}(\theta_1-\theta_2)\,
T_{d\alpha}^{f\beta}(\theta_1)\,T_{c\beta}^{e\gamma}(\theta_2)=
T_{b\alpha}^{d\beta}(\theta_2)\,T_{a\beta}^{c\gamma}(\theta_1)\,
S_{cd}^{ef}(\theta_1-\theta_2)\,.$$ Here, it is supposed the solitons are travelling along the positive $x$-axis ($\theta_1>\theta_2>0$). The bulk $S$-matrix depends on the bulk coupling $\beta$ via the quantity $\gamma=8\pi/\beta^2 -1$, and the conventions used are those adopted in [@bczsg]. The equations (\[STT\]) are well known in many contexts involving the notion of integrability (see [@Jimbo]), but were discussed first with reference to defects by Delfino, Mussardo and Simonetti [@Delf94a]; if the possibility of reflection were to be allowed an alternative framework (such as the one developed by Mintchev, Ragoucy and Sorba [@Mintchev02]), might be more appropriate. Here, the defect is expected to be purely transmitting.
The solution (for general $\beta$, and for even or odd labelled defects — note the labelling is never mixed by (\[STT\])), is given by $$\label{KL}
{\slacs{1.2ex}
T_{a\alpha}^{b\beta}(\theta)=f(q,x)\left(\begin{array}{cc}
\nu^{-1/2}Q^\alpha\delta_\alpha^\beta &
q^{-1/2}\E^{\gamma(\theta-\eta)}\delta_\alpha^{\beta-2}\\[5pt]
q^{-1/2}\E^{\gamma(\theta-\eta)}\delta_\alpha^{\beta+2}&
\nu^{1/2}Q^{-\alpha}\delta_\alpha^\beta\\
\end{array}\right).}$$ In (\[KL\]) a block form has been adopted with the labels $a$, $b$ labelling the four block elements on the right hand side, and where $\nu$ is a free parameter, as is $\eta$ (to be identified with the defect parameter introduced in the previous section), and $$q=\E^{\I\pi\gamma}\,,\quad x=\E^{\gamma\theta}\,,\quad
Q^2=-q=\E^{4\pi^2\I/\beta^2}\,.$$ In addition, $^{\rm even}T$ is a unitary matrix (for real $\theta$), and both types of transmission matrix must be compatible with soliton–anti-soliton annihilation as a virtual process. These two requirements place the following restrictions on the overall factor for the even transmission matrix, $^{\E}f(q,x)$: $$\left\{\begin{array}{l}
{}^{\E}\bar f(q,x)={^{\E}f}(q,qx)\,,\\[5pt]
{}^{\E}f(q,x)\;{^{\E}\!f}(q,qx)\left(1+\E^{2\gamma(\theta-\eta)}\right)=1\,.
\end{array}\right.$$ These do not determine $^{\E}f(q,x)$ uniquely but the ‘minimal’ solution determined by Konik–LeClair has $$\label{KLf} {}^{\E}f(q,x)=
\frac{\E^{\I\pi(1+\gamma)/4}}{1+\I\E^{\gamma(\theta-\eta)}}\,
\frac{r(x)}{\bar r(x)}\,,$$ with ($z=\I\gamma(\theta-\eta)/2\pi$), $$r(x)=\prod_{k=0}^\infty\,
\frac{\Gamma(k\gamma+\sfrac14-z)\,\Gamma((k+1)\gamma+\sfrac34-z)}
{\Gamma((k+\sfrac12)\gamma+\sfrac14-z)\,\Gamma((k+\sfrac12)\gamma+\sfrac34-z)}\,.$$ It is worth noting that the apparent pole in (\[KLf\]) at $1+\I\E^{\gamma(\theta-\eta)}=0$ is actually cancelled by a pole at the same location in $\bar r(x)$. However, there is another pole at $$\theta=\eta -\frac{\I\pi}{2\gamma} \rightarrow \eta\;\;{\rm
as}\;\;\beta\rightarrow 0\,,$$ uncancelled by a zero, and this does actually represent the expected unstable bound state alluded to in the first section.
Several brief remarks are in order. It is clear, on examining (\[KL\]), that the processes in which a classical soliton would inevitably convert to an anti-soliton are clearly dominant even in the quantum theory, yet suppressed if a classical soliton is merely delayed. This much is guaranteed by the factor $\E^{\gamma(\theta-\eta)}$ appearing in the off-diagonal terms. A curious feature is the different way solitons and anti-solitons are treated by the diagonal terms in (\[KL\]). They are treated identically by the bulk $S$-matrix yet one should not be surprised by this since the classical defect conditions (\[sG\]) do not respect all the usual discrete symmetries. Indeed, the dependence of the diagonal entries on the bulk coupling can be demonstrated to follow from the classical picture by using a functional integral type of argument, as explained more fully in [@bczsg]. The sine-Gordon spectrum contains bound states (breathers), and it is interesting to calculate their transmission factors. This much has been done [@bczsg]. However, it would also be interesting to attempt to match these breather transmission factors to perturbative calculations, and this has not yet been done. There are also open questions concerning how to treat defects in motion. From a classical perspective it seems quite natural that defects might move and scatter [@bczsg], however it is less clear how to describe this in the quantum field theory, or indeed to understand what these objects really are.
It is quite remarkable that the simple-looking question asked at the beginning has led to an interesting avenue of enquiry that does not appear to have been explored previously, that links with results, such as (\[KL\]), which had been obtained for seemingly quite different reasons, and that is not yet exhausted (for example, see [@Gomes], for an extension to supersymmetric sine-Gordon).
[**Acknowledgements.** I am very grateful to the organisers for giving me the opportunity to review this material, to Peter Bowcock and to Cristina Zambon for many discussions concerning this topic and for a longstanding collaboration, and to several other members of EUCLID, a Research Training Network funded by the European Commission (contract number HPRN-CT-2002-00325).]{}
[99]{}
P. Bowcock, E. Corrigan and C. Zambon: Int. J. Mod. Physics A **19** (Supplement) (2004) 82; [[hep-th/0305022]{}]{}. R.H. Goodman, P.J. Holmes and M.I. Weinstein: Physica D **161** (2002) 21. C. Rogers and W.K. Schief: *Bäcklund and Darboux Transformations: Geometry and Modern Applications in Soliton Theory*, Cambridge Text in Applied Mathematics, Cambridge University Press 2002. P. Bowcock, E. Corrigan, P.E. Dorey and R.H. Rietdijk: Nucl. Phys. B **445** (1995) 469; [[hep-th/9501098]{}]{}. P. Bowcock, E. Corrigan and C. Zambon: J. High Energy Phys. JHEP **01** (2004) 056; [[hep-th/0401020]{}]{}. A.C. Scott, F.Y.F. Chu and D.W. McLaughlin: IEEE Proc. **61** (1973) 1443. E. Corrigan and C. Zambon: J. Phys. A **37** (2004) L471; [[hep-th/0407199]{}]{}. R. Konik and A. LeClair: Nucl. Phys. B **538** (1999) 587; [[hep-th/9703085]{}]{}. P. Bowcock, E. Corrigan and C. Zambon: J. High Energy Phys. JHEP **0508** (2005) 023; [[hep-th/0506169]{}]{}. M. Jimbo: *Yang–Baxter Equation in Integrable Systems*, Advanced Series in Mathematical Physics 10, World Scientific 1989. G. Delfino, G. Mussardo and P. Simonetti: Nucl. Phys. B **432** (1994) 518;\
[[hep-th/9409076]{}]{}. M. Mintchev, E. Ragoucy and P. Sorba: Phys. Lett. B **547** (2002) 313;\
[[hep-th/0209052]{}]{}. J.F. Gomes, L.H. Ymai and A.H. Zimerman: J. Phys. A **39** (2006) 7471;\
[[hep-th/0601014]{}]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report anomalous features in the normal-state resistivity of single crystalline Bi$_2$Sr$_2$CaCu$_2$O$_8$ whiskers near $T^{*}\sim 250$ K. From these varied oxygen-doping samples, the resistance ($R$) was found to deviate [*upward*]{} from the linear-$T$ relation for $T < T^{*}$, while $R$ keeps decreasing with $T$. The deviation, $\Delta R \equiv R - (a + bT)$, reaches a maximum in a temperature range of $\sim 25$ K, just before $R-T$ follows the well-known descent associated with the pseudogap opening in underdoped samples. A second kink was also found in some samples immediately following the first one, resulting in an S-shaped feature in the $R-T$ curve. The resistance then resumes nearly linear-$T$ decrease below the anomalies in contrast to the first case. We interpret these features as being related to crystal structure transformation (or lattice distortion), as was evidenced in thermal and elastic property measurements. These structural instabilities seem to be connected to the subsequent dynamics as $T$ is further lowered.'
address: 'Department of Physics, University of Alberta, Edmonton T6G 2J1, Canada'
author:
- 'Weimin Chen, J. P. Franck, and J. Jung'
title: 'Evidence of structural instability near $T^{*}\sim 250$ K in the resistivity of Bi$_2$Sr$_2$CaCu$_2$O$_8$ whiskers'
---
The normal state of high-$T_{c}$ cuprates provides rich information for an understanding of superconducting mechanism. Two of the many interesting normal-state properties are the so-called pseudogap opening and crystal structural transformation. In resistivity ($\rho$) measurement, the former was believed to manifest itself as a downturn deviation from the linear $\rho-T$ relation below some characteristic temperature $T^{*}> T_{c}$ in underdoped samples. Evidence for crystal structure transformation (or lattice distortion) is more difficult to detect because of such factors as crystal quality and the weakness of its physical effects. Structural transformations are generally studied by elastic and acoustic methods [@1]. In most of the cuprates, anomalous features corresponding to structural instabilities in the normal state have been observed. In La$_{2-x}$Sr$_x$CuO$_4$, a tetragonal to orthorhombic structure transition upon cooling occurred at 220 K [@2]. Around the same temperature for YBa$_2$Cu$_3$O$_{7-\delta}$, the sound velocity shows a step-like change and was associated with the instability corresponding to a structural order-disorder transition [@3]. A step-like change in Young’s modulus (hysteresis) also appeared in elastic measurements of single crystals at 200 - 240 K [@4; @5]. In Bi cuprates, similar evidence was also accumulated. Dominec [*et al.*]{} [@6] reported sound velocity deviations around 240 K in Bi$_2$Sr$_2$CuO$_6$ (Bi-2201). Acoustic studies revealed sound velocity softening along the $a$- axis around 250 K in Bi$_2$Sr$_2$CaCu$_2$O$_8$ (Bi-2212) samples [@7; @8]. In elastic property measurements of similar Bi-2212 whiskers as studied in the present work, Tritt [*et al.*]{} [@9] found anomalous peaking of Young’s modulus and stress-strain hysteresis near 270 K, and interpreted it as evidence for a phase transition. Similar results were also obtained for Bi$_2$Sr$_2$Ca$_2$Cu$_3$O$_{10}$ (Bi-2223) [@10]. More directly, by using acoustic and x-ray diffraction methods, Titova [*et al.*]{} [@11] consistently found a small increase of lattice dimensions at $T \sim 190 - 270
$K for the cuprates: YBCO (123 and 124), Bi-2201, 2212, and 2223, and Ti-series compounds. The structural instabilities of the cuprates observed at comparable temperature ranges (about 200 to 300 K) demonstrate common features consistent with their similar crystalline structures. Obviously, structural transformation affects transport properties of the material and should be detectable by resistivity measurement. Indeed, resistivity jumps were observed above $T_{c}$ in La$_{2-x}$Sr$_x$CuO$_4$, which are caused by a structural phase transition [@12]. However, few results were reported for other cuprates in the temperature regions where lattice instabilities were found by mechanical methods described above. This situation can be understood in two ways. First, the influence of structural instabilities on sample resistivity is small, because at such structural transformation, the crystal symmetry usually remains the same, only the lattice parameters change (distortion), leaving weak traces on electron scattering. Secondly, the small change in resistivity can be easily covered up by effects such as percolation, and is undetectable unless the crystal is of very high quality. Therefore, in the present work on pure single crystalline Bi-2212, it is interesting that we observed unusual features in the normal-state resistivity in temperature regions where x-ray and mechanical studies had revealed anomalous structural properties. This is a clear indication for a correlation between these features probed by different methods. Moreover, drastically different temperature dependence of resistivity appeared below these anomalies, showing crucial connections between the phenomena in these temperature regimes.
In our extensive resistivity measurements of both Bi-2212 and Bi-2223 whiskers, we often noticed the data were fluctuating in a temperature region between 180 and 270 K. Outside this window, the signal is much more stable. Since this temperature range corresponds to the important regime where the resistivity shows dramatic changes and is central in the normal-state theory, we were motivated to check the region more closely. In this work we report five samples selected from many others which exhibit anomalous behavior in this temperature window.
Samples of Bi-2212 whiskers were grown by a sintering method [@13]. They are rectangular needle-like tapes with typical dimensions: 3 mm $\times$ 10 $\mu$m $\times$ 0.5 $\mu$m. The three crystalline axes ($a$, $b$ and $c$) are along these directions, respectively. Various characterization by scanning and transmission electron microscopy, and energy dispersive x-ray analysis indicate that the whiskers are good single crystals with perfect surface morphology throughout the whole length ($\sim$3 mm). The as-grown whiskers are usually oxygen overdoped. Before making the first measurements, some samples were annealed in flowing nitrogen, making them slightly underdoped. The resistivity was measured by the standard four-wire method. The electrical leads were made by magnetron sputtering of silver through a mask. A dc current (0.5 $\mu$ A) was driven along the $a$-axis and voltage was also measured in this direction.
Figure 1a shows the resistance ($R$)data of one whisker (sample A, as-grown), together with a segment of $R$ above 200 K. At high temperature, $R-T$ is well fitted to a linear line (fitting range: 260 - 308 K) with $R(\Omega) = 129.62 + 2.475T$. As the temperature is further lowered, $R$ shows a clear upward curving just before it follows the well-known downward deviation. To examine the curve in more detail, we plotted the temperature dependence of the resistance deviation $\Delta R \equiv R-(a +
bT)$, where $(a + bT)$ is the high temperature fit as described. The sharp turning point and the slope change are more evident in the plot of $\Delta R$ as shown in Fig. 1b. We can see that, while $R$ keeps decreasing with $T$, $\Delta R $ starts to deviate [*upward*]{} at $T_0 \sim 260$ K and reaches a maximum at $T^{*} =236$ K, it then falls and becomes negative, i.e. $R < a + bT$. The peak $\Delta R/R$(300 K) is about 0.5%. Although the choice of fitting range of temperature may slightly affect the kink, the two major features that $R$ curves upward from linear-$T$ dependence and reaches a sharp turning point are clearly identifiable. Such small bumps could be easily smeared out in samples with defects such as grain boundaries, where percolation is not negligible, to yield a fully linear-$T$ resistance. In our twined samples, which have a two-step superconducting transition, linear-$T$ dependence was observed without any feature in the normal state. This agrees with the studies by Gorlova and Timofeev [@14] who made combined electron diffraction and resistivity measurement on the same sample, and showed that a wider portion of linear-$T$ resistivity appears in twinned whiskers or samples covered with a polycrystalline film than in purer samples. This explains the often-observed discrepancy in bulk materials between transport and other measurements in which the sample’s polycrystalline nature is not important. The smallness in the change of resistivity puts high requirement on sample quality to observe these anomalies. The features were identified only after more significant kinks around this temperature region were found in our Bi-2223 single crystal samples [@15]. The kink in $\Delta R$ seems to follow a power-law relation $(1-T/T_0)^{2}$ (dotted line in Fig. 1b, where $T_0 = 260$ K). However, it could also fit to a weak exponential increase of the form: $R_{0} = \exp[c(T_{0} - T)]$, with $c = 0.1$ to 0.2.
In Fig.2 we plotted $\Delta R/R$(300) vs. $T$ for four more samples plus sample A from Fig. 1b. The annealing status of these samples are as follows: A and B: as-grown in ambient pressure; C: annealed in flowing N$_2$ ($450^\circ$C, 8 h); D: flowing N$_2$ annealing ($450^\circ$C, 4 h); and E: flowing O$_2$ annealing ($450^\circ$C, 10 h). The doping level may not solely depend upon the annealing conditions because the sample size varies. Nevertheless, data from long-time annealed samples are reliable. Two cases in the temperature dependence of $\Delta R$ can be classified, as shown in the separate panels of Fig. 2. In the lower panel, $\Delta R$ reveals one kink as discussed. $\Delta R$ falls steeply right after the sharp peak, which gives rise to the well-known downward turning $R-T$ in underdoped cuprates. Our identification of this transition temperature is much clearer than the data available in the literature [@16]. The curves in the top panel show another kink following the first one (samples B and C), which resulted in an S-shaped feature in the $R-T$ curve. At the end of this curving, $R$ nearly resumes the previous linear-$T$ descent. The origin of this anomaly is unclear. It does not show up in the underdoped cases. For the sample annealed in oxygen (E), the kink around 250 K is absent, but the kink at lower temperature ($\sim$ 220 K) is consistent with the second kinks in the other two samples. A very similar single kink was also found in the as-grown data of sample C (refer to the inset of Fig. 3), which is also oxygen overdoped. These features in the normal state thus seem to be intrinsic to the Bi-2212 samples. This result joins our previous results from Bi-2223 samples to give a consistent picture about the anomaly around this characteristic temperature region in Bi-series cuprates. The peaking temperature $T^{*}$ shows a trend to increase with the extent of underdoping. $T^{*}= 254$ K is the highest for sample C which experienced the longest N$_2$ annealing, whereas the oxygen-annealed sample E showed the lowest $T^{*}$ (220 K). The temperature $T_0$, at which $\Delta R$ starts to climb up, also shows the similar dependence. The detailed characteristic temperatures for the samples are listed in Table I. In light of the thermal and elastic measurement results in the literature, which showed evidence of structural phase transformation around the same temperature region as is reported here, we interpret such upward deviation as due to structural transformation or lattice distortion. Similar kinks in resistivity had been observed in La-214 due to structural transformation [@12]. The signal is often fluctuating immediately following the kink, in the region around 200 K as shown by the oscillating $\Delta R$. This is obviously correlated to the $\Delta R$ jump, and may well mean an influence on electron scattering due to structural instabilities.
The peak temperature $T^*$ clearly separates two regions in which different mechanisms are switched on. Especially for the noisy region below $T^*$, if it is caused by the opening of a pseudogap in the spin excitation, the initial fluctuating $R$ data may be a clue for this mechanism, when the antiferromagnetic correlation is gradually frozen out. In the scenario of stripe phase formation, on the other hand, the stripe phase may be triggered by the lattice distortion. Charge or spin excitation following a structure transformation (or lattice distortion) was established by neutron scattering studies [@17]. Bianconi [*et al.*]{} [@18] have shown for Bi-2212 that 1D modulation of the CuO$_2$ plane is formed by alternating stripes of low temperature orthorhombic and low temperature tetragonal lattices. However, our $T^{*} (\sim 3T_c)$ is higher than their data ($\sim 1.4T_c$). Whatever the model will be, the present result shows that the lattice structure undergoes some change in all the samples, which initiates new resistive mechanisms. However, the subsequent dynamics depends on the oxygen doping level, not on this structural change.
The upward deviation is not due to thermal stress in the sample, which is anchored onto the glass substrate by silver leads. We estimate the thermal stress to be a small value $\sim$ 0.018 GPa at 250 K, by taking the Young’s modulus $Y = 20$ GPa [@9], the thermal expansion coefficient along the $a$-axis $\alpha = 14.4
\times 10^{-6}$ K$^{-1}$ [@19], and a length between voltage leads of about 0.3 mm at room temperature. Chen [*et al.*]{} [@20] had measured the effect of uniaxial stress (strain) on the resistivity of similar whiskers, and found that the influence is very small ($\sim 1/12000$ at room temperature). Moreover, the pressure generally causes a decrease in resistivity.
To further study the influence of doping on $T^*$, we annealed one sample (C) successively in flowing nitrogen. The resistance data from this sample are as shown in Fig. 3. The result from the as-grown sample shows a slope change near 200 K, $R-T$ is linear above this temperature. After one annealing, a new feature appeared around $T^* \sim 220$ K, while the slope change around 200 K remains. After the second annealing, $T^*$ shifted to 245 K, as indicated by the arrows, showing the tendency that $T^*$ increases with further underdoping, which agrees with the widely observed behavior of resistivity anomaly and the pseudogap.
All the above discussion is based upon crystal structure instability in the normal state. In light of the magnetic structure of the CuO$_2$ planes, an alternative interpretation may come from the competition between quasiparticle scattering and the increasing correlation with antiferromagnetic (AF) ordering as the temperature is decreasing. It is interesting to compare the present results with the anomalous resistivity of magnetic materials [@21; @22; @23; @24]. In fact, resistivity kinks were commonly found near magnetic phase transitions in the latter. For example, in dysprosium single crystals [@21], a kink in the linearly decreasing $\rho-T$ appears near the Néel point when the sample transforms from the paramagnetic into the AF state. Similar anomalies were also reported in europium chalcogenides [@22]. It is surprising that the Néel temperature of perovskite KNiF$_3$ is about 253 K [@24], very close to the anomalous temperatures in the present work. This may point to a common origin of the resistivity kinks for the materials with similar crystal structures.
In the scenario of free electron scattering with spin excitations, the total resistivity consists of three additive terms: the $T$-independent residual resistivity $\rho_0$ due to elastic impurity scattering, the lattice contribution (phonon) $\rho_L \propto T$, and the spin-disorder resistivity $\rho_M$. The resistivity kink (exponentially increasing with decreasing $T$ ) is associated with the spin-disorder resistivity ($\rho_M$). In the critical region of magnetic transition, the behavior of $\rho_M$ reflects the strongly temperature dependent short-range spin-spin correlation. This is in the regime of spin-disorder due to the large long-range magnetic fluctuations near the transition temperature [@24]. The resistivity thus takes the form:
$$\rho(T)=\rho_0-b|t-1|\ln|t-1|^{-1},$$
where $t = T/T_c$. For $b>0$, $\rho(T)$ shows a concave cusp at the magnetic transition temperature $T_c$. If this idea is valid, then the observed anomaly near $T^* \sim 250$ K must come from the antiferromagnetic ordering in the CuO$_2$ planes. Especially, the sharp turning point in all the curves may well signify a similar magnetic phase transition. Currently only the spin fluctuation theory [@25; @26] is a close approach to this picture.
In conclusion, an upward deviation from the linear-$T$ dependence were observed around $T^* \sim 250$ K in the normal-state resistivity of Bi-2212 single crystals. This is usually followed by a fluctuating region near 200 K. We interpret such behavior as due to structural instabilities in accordance with many results from other measurement techniques (thermodynamic, elastic, acoustic, and x-ray). A peak in the resistance deviation ($\Delta
R$) is clearly identifiable. Below this sharp turning point, the resistivity drops more steeply, signifying a new mechanism in underdoped samples. This turning in resistivity had been widely identified as pseudogap opening. However, the scenario of stripe formation preceded with lattice instabilities was also proposed [@17]. The result shows that the lattice instability exists in all the samples regardless of their doping level. The same result was also found in our Bi-2223 samples. This gives a consistent picture intrinsic to similar crystal structures of Bi-cuprates. Although the structural transformation seems to trigger new mechanism, it is the oxygen doping that controls the subsequent behavior of $R-T$. It is noticeable that our $T^*$ data are consistently higher than the literature values (typically 120 -165 K, with the highest for Bi-2212 being 220 K [@16]). There is no identifiable feature (smooth $dR/dT$) in our $R-T$ curves below 180 K. An alternative interpretation of the resistance kinks is the correlation between electrons and AF spin fluctuations. Similar mechanisms in magnetic materials are known to give rise to the kinks in the resistivity curves at magnetic phase transition. Our work should stimulate theoretical interest in this direction.
For a review on ultrasonic measurements, see J. Dominec, Supercond. Sci. Technol. 6 (1993) 153. A. Migliori, W.M. Visscher, S. Wong, S. E. Brown, I. Tanaka, H. Kojima, P. B. Allen, Phys. Rev. Lett. 64 (1990) 2458. J. Thoulous, X. M.Wang, D. J. L. Hong, Phys. Rev. B 38 (1988) 7077. S. Hoen, L, C. Bourne, C. M. Kim, A. Zettle, Phys. Rev. B 38 (1988) 11949. X. D. Shi [*et al.*]{}, Phys. Rev. B 39 (1989) 827. J. Dominec, C. Laermans, V. Plechéček, Physica C 171 (1991) 373. Ye-Ning Wang [*et al.*]{}, Phys. Rev. B 41 (1990) 8981. O-M Nes [*et al.*]{}, Supercond. Sci. Technol. 4 (1991) S388. T. M. Tritt, M. Marone, A. C. Ehrlich, M. J. Skove, D. J. Gillespie, R. L. Jacobsen, G. X. Tessema, J. P. Franck, J. Jung, Phys. Rev. Lett. 68 (1992) 2531. Yusheng He, Jiong Xiang, Xin Wang, Aisheng He, Jinchang Zhang, Panggao Chang, Phys. Rev. B 40 (1989) 7384; T. Fukami, A. A. A. Youssef, Y. Horie, S Mase, Physica C 161 (1989) 34; V. Plechéček, J. Dominec, Solid State Commun. 74 (1990) 6339. S. Titova, I. Bryntse, J. Irvine, B. Mitchell, V. Balakirev, J supercond. 11 (1998) 471. Y. Nakamura, S. Uchida, Phys. Rev. B 46 (1992) 5841. J. Jung, J. P. Franck, D. F. Mitchell, H. Claus, Physica C 156 (1988) 494. I. G. Gorlova, V. N. Timofeev, Physica C 255 (1995) 131. Weimin Chen, J. P. Franck, J. Jung, Phys. Rev. B 60 (1999) 3527. T. Watanabe, T. Fujii, A. Mastuda, Phys. Rev. Lett. 79 (1997) 2113. J. M. Tranquada, J. D. Axe, N. Ichikawa, Y. Nakamura, S. Uchida, B. Nachuni, Phys. Rev. B 54 (1996) 7489. A. Bianconi [*et al.*]{}, Europhys. Lett. 31 (1995) 411; A. Bianconi, Physica C 235-240 (1994) 269. R. H. Ardent, M. F. Garbauskas, C. A. Meyer, F. J. Rotella, J. D. Jorgensen, R. L. Hitterman, Physica C 182 (1991) 73. Xin-Fen Chen, G. X. Tessema, M. J. Skove, Physica C 181 (1991) 340. P. M. Hall, S. Legvold, F. H. Spedding, Phys. Rev. 117 (1960) 971. S. von Molnar, IBM J. Res. Develop. 14 (1970) 269. S. Takada, Prog. Theor. Phys. 46 (1971) 15. P. G. de Gennes, J. Friedel, J. Phys. Chem. Solids 4 (1958) 71. D. Pines, in [*High Temperature Superconductors and the C$_{60}$ System*]{}, ed. H-C. Ren, (Gordon & Breach, 1995), p.1. Branko P. Stojkovic, D. Pines, Phys. Rev. B 55 (1997) 8576, and references therein.
---------------- ----------- ----------- -----------
Samples $T_c$ (K) $T^*$ (K) $T_0$ (K)
A (as-grown) 87.5 236 263
B (as-grown) 85 252 275
C (N$_2$/8h) 76 254 278
D (N$_2$/4h) 70 245 262
E (O$_2$/10 h) 76 220 246
---------------- ----------- ----------- -----------
: Characteristic temperatures of the samples: $T_c$ (from the maximum $dR/dT$), $T^*$ at the $\Delta R$ peaks, and $T_0$ where $\Delta R$ starts to deviate upward. Also shown are the thermal annealing conditions (always at $450^\circ$C).
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'bibliography.bib'
---
[**Software paper for submission to the Journal of Open Research Software**]{}\
------------------------------------------------------------------------
(1) Overview {#overview .unnumbered}
============
Title {#title .unnumbered}
=====
Probabilistic Inference on Noisy Time Series (PINTS).
Paper Authors {#paper-authors .unnumbered}
=============
1\. Clerx, Michael\* ORCiD:0000-0003-4062-3061\
2. Robinson, Martin\* ORCiD:0000-0002-1572-6782\
3. Lambert, Ben ORCiD:0000-0003-4274-4158\
4. Lei, Chon Lok ORCiD:0000-0003-0904-554X\
5. Ghosh, Sanmitra ORCiD:0000-0002-4879-7587\
6. Mirams, Gary R. ORCiD:0000-0002-4569-4312\
7. Gavaghan, David J. ORCiD:0000-0001-8311-3200
\*These authors contributed equally to this work.
Paper Author Roles and Affiliations {#paper-author-roles-and-affiliations .unnumbered}
===================================
1\. PDRA, Department of Computer Science, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, United Kingdom.\
2. PDRA, Department of Computer Science, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, United Kingdom.\
3. PDRA, MRC Centre for Global Infectious Disease Analysis, School of Public Health, Imperial College London, W2 1PG, United Kingdom.\
4. DPhil Student, Department of Computer Science, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, United Kingdom.\
5. PDRA, Department of Computer Science, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, United Kingdom.\
6. Research Fellow, Centre for Mathematical Medicine & Biology, School of Mathematical Sciences, University of Nottingham, Nottingham, NG7 2RD, United Kingdom.\
7. Professor, Department of Computer Science, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, United Kingdom.
Abstract {#abstract .unnumbered}
========
Time series models are ubiquitous in science, arising in any situation where researchers seek to understand how a system’s behaviour changes over time. A key problem in time series modelling is *inference*; determining properties of the underlying system based on observed time series. For both statistical and mechanistic models, inference involves finding parameter values, or distributions of parameters values, for which model outputs are consistent with observations. A wide variety of inference techniques are available and different approaches are suitable for different classes of problems. This variety presents a challenge for researchers, who may not have the resources or expertise to implement and experiment with these methods. PINTS (Probabilistic Inference on Noisy Time Series — <https://github.com/pints-team/pints>) is an open-source (BSD 3-clause license) Python library that provides researchers with a broad suite of non-linear optimisation and sampling methods. It allows users to wrap a model and data in a transparent and straightforward interface, which can then be used with custom or pre-defined error measures for optimisation, or with likelihood functions for Bayesian inference or maximum-likelihood estimation. Derivative-free optimisation algorithms — which work without harder-to-obtain gradient information — are included, as well as inference algorithms such as adaptive Markov chain Monte Carlo and nested sampling which estimate distributions over parameter values. By making these statistical techniques available in an open and easy-to-use framework, PINTS brings the power of modern statistical techniques to a wider scientific audience.
Keywords {#keywords .unnumbered}
========
Time series models; non-linear optimisation; MCMC sampling; nested sampling; Bayesian inference; Python
Introduction {#introduction .unnumbered}
============
Time series models are common in science, where they are used to describe the dynamics of system behaviours. In many cases, these models are non-linear and impossible to solve analytically, so that the *forward problem* (predicting the model output for a given set of parameters) is computationally hard. For such models, there is no single method which can reliably solve the *inverse problem* of estimating parameter values from a noisy time trace. Much like there is a variety of forward models, there is a diversity of approaches for parameter inference. Further, it is often unclear which approach to apply when, meaning that researchers are required to implement a range of methods before successfully fitting their model to data.
PINTS is a software framework that allows users to easily trial and apply different inference methods to their problem. The inference methods supplied by PINTS fall into two broad categories: *optimisers*, which attempt to find a single best parameter vector, and *samplers*, which aim to estimate a probability distribution over parameter values that are compatible with observed results. Users are expected to already have a forward model (for example, a simulation) at their disposal, which they make available to PINTS by writing a simple Python wrapper. They then define a [`Problem`]{} (a forward model plus a data set), on which either an [`ErrorMeasure`]{} (for optimisation) or a [`LogPDF`]{} (for optimisation and sampling) is defined. Currently available optimisers include CMA-ES [@hansen2003reducing], XNES [@glasmachers2010exponential], SNES [@schaul2011high], and Particle Swarm Optimisation (PSO) [@kennedy2011particle]. Sampling methods include, amongst other routines, Random Walk Metropolis Markov chain Monte Carlo (MCMC) [@metropolis1953equation; @lambert2018Student], adaptive covariance MCMC [@johnstone2016uncertainty] and Population MCMC [@jasra2007population]. In addition, ellipsoidal [@mukherjee2006nested] and rejection nested samplers [@skilling2006nested] are provided. Convenience plotting methods are provided to quickly visualise the results, as well as diagnostic tools to inspect the validity of the results.
PINTS was developed as a community effort by researchers in electrochemistry, cardiac electrophysiology, and statistics, to compare different methods for solving inverse problems in a common framework. It features a clean and transparent object-orientated API that is designed to accommodate easily new error measures, log-likelihoods, optimisers and samplers, allowing users to utilise pre-built components as much as possible, while adding their own code for problem-specific areas. The PINTS team aims for full test coverage, and includes unit testing and extended statistical tests to verify the correct operation of all methods.
Early research using PINTS in electrochemistry has included fitting a differential-algebraic equation (DAE) model of reduction-oxidation to voltammetry measurements of a Polyoxometalates molecule [@pom_paper], and the design and application of a custom hierarchical statistical model for repeat voltammetry experiments of a Ferricynide process [@robinson_bond_simonov_zhang_gavaghan_2018].
Optimisation algorithms are implemented in many different software packages, (see, for example, the Python [`scipy.optimize`]{} module), but are often biased towards gradient-based methods, which can perform poorly for many ordinary and partial differential equations used in time series modelling. PINTS therefore focuses on derivative-free optimisers, although we plan to add gradient-based methods for comparison. In contrast with more general-purpose optimisation software, PINTS contains a number of error measures specifically suited to time series models, and adds the ability to use any PINTS log-likelihood class as an error measure in order to perform maximum likelihood estimation (MLE).
Dakota [@adam2015dakota] is a widely regarded package for parameter fitting and uncertainty quantification and is most similar to PINTS in that it offers a generic interface to call an (assumed expensive) model, as well as a wide variety of optimisers and samplers. In contrast to the PINTS Python API, Dakota uses either a C++ or file input/output process for communication between user models and the library, and does not provide options for specifying either the error measures or the log-likelihoods of the inverse problem. However, Dakota has some features not yet available in PINTS, such as the option to train a surrogate model, useful for very expensive model evaluations.
Other software packages that enable parameter inference and sampling for ODE models include BioBayes [@vyshemirsky2008biobayes], ABC-SysBio [@liepe2010abc], SYSBIONS [@johnson2014sysbions] and Stan [@carpenter2017stan]. These packages use either a common model description format (for example, SBML) or their own language (for example, Stan’s probabilistic programming language) to specify the model, presenting additional learning hurdles for a user and often restricting the class of models which can be fit. By contrast, PINTS aims to be as general as possible to support a wider variety of models (for example, PDEs). PyMC3 [@salvatier2016probabilistic] does provide a similar generic model interface to PINTS but, as with the other packages, specialises in one sampling method, whereas PINTS aims to support a wide variety of methods with the assumption that no one sampling method is suitable for all models of interest. BCM [@thijssen2016bcm] offers both a generic interface (via C++) and a wide variety of samplers, but does not supply any likelihood functions and is unfortunately largely undocumented.
{width="\textwidth"}
Implementation and architecture {#implementation-and-architecture .unnumbered}
===============================
PINTS is designed around two core ideas: 1. PINTS should work with a wide range of time series models, and make no demands on how they are implemented other than a minimal input/output interface. 2. It is assumed that model evaluation (simulation) is the most costly step in any optimisation or sampling routine.
The decision to use Python fits both these criteria: Python interfaces well with C and C++, which are typically used for high-performance simulation, and any performance hit of using the high-level, easy to read and write language Python is overshadowed by simulation time.
#### Defining an optimisation or sampling problem
All optimisers operate on a callable [`ErrorMeasure`]{} object that describes a function to minimise, or on a callable [`LogPDF`]{} object that describes a probability density function (PDF) to maximise. Similarly, all samplers start from a callable [`LogPDF`]{}, so that the same probability function can be used with both optimisers and samplers. The natural logarithm of the PDF is used for computational efficiency and accuracy, and we assume that the probability density is unnormalised (i.e. its integral does not necessarily sum to 1).
Figure \[fig:uml\] shows how a user-defined model can be wrapped in a PINTS [`ForwardModel`]{} and combined with time points and measured values to create a [`Problem`]{} from which several standard [`ErrorMeasure`]{}s and [`LogPDF`]{}s can be created. For inference in a Bayesian context, a [`LogPosterior`]{} class and several [`LogPrior`]{} distributions are provided. If a given [`LogPDF`]{} or [`ErrorMeasure`]{} cannot be constructed from PINTS classes, users can also define their own classes.
{width="\textwidth"}
#### Implementation of optimisers and samplers
Most PINTS samplers and optimisers are implemented using a so-called *ask-and-tell interface*, inspired by the Python implementation of CMA-ES [@hansen2003reducing] (<https://github.com/CMA-ES/pycma>). In this framework, the details of solving the forward problem are partitioned away from the rest of the sampling or optimising algorithm. For each iteration the following steps are undertaken (Figure \[fig:ask\_tell\]): first, the user calls [`ask()`]{} to obtain one or more parameter values from their chosen method — these values are typically vectors generated stochastically conditional on an internal system state; second, the user solves the forward model and generates a score for each parameter vector, for example, an error measure or (unnormalised) posterior probability; third, the user calls [`tell()`]{} to pass the score back to the method, which can then update its internal state and finish the iteration. For example, in many MCMC methods each [`ask()`]{} call returns a single proposed sample to be evaluated by the user, and the following [`tell()`]{} then either accepts or rejects this point based on its probability. For optimisers such as CMA-ES, [`ask()`]{} returns a set of points in parameter space, and the scores passed in via [`tell()`]{} are then used to estimate the local gradient, which is used to move towards the estimated optimum. This framework has a number of advantages: since optimisation or sampling can take hours or even days, this allows programs using PINTS to provide regular user feedback and logging (which is not possible when the routine is implemented as a single monolithic function call); allows users with access to CPU clusters or GPU machines to implement their own parallelised evaluation of [`ErrorMeasure`]{}s and [`LogPDF`]{}s; lets users implement their own strategies (for example, by dynamically changing hyperparameters) and/or stopping criteria; and, finally, by delineating the sampling or optimisation algorithm’s steps from the methods used to solve the forward problem, encourages development of transparent and modular code.
For more casual users (whom we expect will be the majority), PINTS provides standard [`Optimisation`]{} and [`MCMCSampling`]{} methods that can be run for a fixed (user-specified) number of iterations or, alternatively, until user-specified stopping criteria are reached. These functions also allow logging and parallelised evaluations using Python multiprocessing.
{width="80.00000%"}
Quality control {#quality-control .unnumbered}
===============
PINTS has three levels of testing: unit testing, performance testing, and comparative testing. [Unit tests](https://github.com/pints-team/pints/tree/master/pints/tests) are used to test the functionality of simple (deterministic) methods, and to check that complex (pseudo-random) methods run without raising exceptions. All the unit tests are available to be run by a user to ensure the software is working correctly. Continuous integration is carried out using [Travis CI](https://travis-ci.org/pints-team/pints) (Ubuntu Trusty distribution with Python versions 2.7, 3.4, 3.5 and 3.6, and OS/X with Python version 2.7) and [AppVeyor](https://ci.appveyor.com/project/MichaelClerx/pints) (Windows Server 2011 R2 with Python versions 2.7, 3.4, 3.5 and 3.6). PINTS uses Flake8 linter tests to ensure that contributed code conforms to best practices, and code coverage tests to ensure that all code is sufficiently tested and documented. [Functional testing](https://github.com/pints-team/functional-testing) of the methods is performed separately, and is used to test the method’s behaviour from different (pseudo-random) initial conditions. Analysis of functional checking is done both visually and statistically (for example, if recent results deviate significantly from previous results this indicates the possible introduction of a bug). Finally, in [comparative testing](https://github.com/pints-team/performance-testing) a number of problems are set up and solved with different methods, in order to compare the solutions they return and evaluate their performance.
(2) Availability {#availability .unnumbered}
================
Operating system {#operating-system .unnumbered}
================
PINTS uses no functions specific to any operating system (OS), and so can run on any OS that provides Python. Optional parallelisation is provided that uses the Python [`multiprocessing`]{} module, which works best on UNIX-based systems (for example, Linux and OS/X), but runs on Windows with slightly reduced performance.
Programming language {#programming-language .unnumbered}
====================
PINTS requires Python 2.7 or higher, or Python 3.4 or higher.
Additional system requirements {#additional-system-requirements .unnumbered}
==============================
PINTS has a minimal disk space footprint (approximately 2MB) and can be run on single-processor devices or headless on multi-processor machines (for example, via [`ssh`]{}).
Dependencies {#dependencies .unnumbered}
============
PINTS uses the NumPy (version 1.8 or higher) and SciPy (version 0.14 or higher, [@jones2014scipy]) libraries extensively. The default optimisation method is CMA-ES, for which the [`cma`]{} package (version 2 or higher) is used. The remaining optimisation and sampling methods require no further dependencies. Finally, utility functions for plotting with Matplotlib (version 1.5 or higher [@hunter2007matplotlib]) are provided, but it is possible to use PINTS without Matplotlib or with a different plotting library.
List of contributors {#list-of-contributors .unnumbered}
====================
Michael Clerx, Sanmitra Ghosh, Ben Lambert, Chon Lok Lei, Martin Robinson.
Software location: {#software-location .unnumbered}
==================
Name:
: GitHub (release v0.1.1)
Persistent identifier:
: https://github.com/pints-team/pints/releases/tag/v0.1.1
Licence:
: BSD 3-clause
Publisher:
: Pints team
Version published:
: 0.1.1
Date published:
: 01/11/18
[**Code repository**]{}
Name:
: GitHub (develop)
Persistent identifier:
: https://github.com/pints-team/pints
Licence:
: BSD 3-clause
Date published:
:
Language {#language .unnumbered}
========
English.
(3) Reuse potential {#reuse-potential .unnumbered}
===================
Detailed documentation is provided on using PINTS with user-supplied models (see, for example, the [`writing-a-model`]{} example on the GitHub repository). While PINTS was designed primarily with biological and electrochemical problems in mind, there is nothing to prohibit its use on time series models from other fields. Similarly, while the implemented [`ErrorMeasure`]{} and [`LogPDF`]{} classes were chosen to work well with time series problems, PINTS can be used outside this setting (in fact, utility functions [`fmin`]{} and [`curve_fit`]{} are provided specifically for this purpose). Users are also free to create their own [`ErrorMeasure`]{} or [`LogPDF`]{} classes which do not rely on PINTS [`ForwardModel`]{} or [`Problem`]{} classes. A link to the full API documentation can be found on the GitHub repository, which also contains a list of examples for all PINTS’ main features (<https://github.com/pints-team/pints/examples>). We have found that these examples, rather than our API documentation, serve to kick-start any new project based on PINTS. We welcome questions, bug reports, and user contributions via the same repository, which acts as a central communication platform for PINTS.
Funding statement {#funding-statement .unnumbered}
=================
M.C., G.R.M. and D.J.G. acknowledge support from the UK Biotechnology and Biological Sciences Research Council \[BBSRC grant number BB/P010008/1\]; M.R., S.G. and D.J.G. gratefully acknowledge research support from the UK Engineering and Physical Sciences Research Council Cross-Disciplinary Interface Programme \[EPSRC grant number EP/I017909/1\]; C.L. acknowledges support from the Clarendon Scholarship Fund, the EPSRC and the UK Medical Research Council (MRC) \[EPSRC grant number EP/L016044/1\]; BL acknowledges support from the UK Engineering and Physical Sciences Research Council \[EPSRC grant number EP/F500394/1\]; and S.G. and G.R.M. acknowledge support from the Wellcome Trust & Royal Society \[Wellcome Trust grant number 101222/Z/13/Z\].
Competing interests {#competing-interests .unnumbered}
===================
The authors declare that they have no competing interests.
------------------------------------------------------------------------
[ **Copyright Notice**]{}\
Authors who publish with this journal agree to the following terms:\
Authors retain copyright and grant the journal right of first publication with the work simultaneously licensed under a [Creative Commons Attribution License](http://creativecommons.org/licenses/by/3.0/) that allows others to share the work with an acknowledgement of the work’s authorship and initial publication in this journal.\
Authors are able to enter into separate, additional contractual arrangements for the non-exclusive distribution of the journal’s published version of the work (e.g., post it to an institutional repository or publish it in a book), with an acknowledgement of its initial publication in this journal.\
By submitting this paper you agree to the terms of this Copyright Notice, which will apply to this submission if and when it is published by this journal.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For SU(3) lattice gauge theory we study properties of static quark sources represented by local Polyakov loops. We find that for temperatures both below and above $T_c$ coherent domains exist where the phases of the local loops have similar values in the vicinity of the center values $0, \pm 2 \pi/3$. The cluster properties of these domains are studied numerically. We demonstrate that the deconfinement transition of SU(3) may be characterized by the percolation of suitably defined clusters.'
---
[ Coherent center domains in SU(3) gluodynamics and their percolation at $T_c$]{} Christof Gattringer Institut für Physik, Unversität Graz,\
Universitätsplatz 5, 8010 Graz, Austria
[*To appear in Physics Letters B*]{}
[**Introductory remarks**]{} Confinement and the transition to a deconfining phase at high temperatures are important, but not yet sufficiently well understood properties of QCD. With the running and upcoming experiments at the RHIC, LHC and GSI facilities, it is important to also obtain a deeper theoretical understanding of the mechanisms that drive the various transitions in the QCD phase diagram.
An influential idea is the Svetitsky-Jaffe conjecture [@znbreaking] which states that for pure gluodynamics the critical behavior can be described by an effective spin model in 3 dimensions which is invariant under the center group $\mathds{Z}_3$ (for gauge group SU(3)). The spin degrees of freedom are related [@ploopeff] to static quark sources represented by Polyakov loops, which in a lattice regularization are given by $$L(\vec{x}) \, = \,
\mbox{tr}_c \prod_{t=1}^{N} U_4(\vec{x},t) \; .$$ The Polyakov loop $L(\vec{x})$ is defined as the ordered product of the SU(3) valued temporal gauge variables $U_4(\vec{x},t)$ at a fixed spatial position $\vec{x}$, where $N$ is the number of lattice points in time direction and $\mbox{tr}_c$ is the trace over color indices. The loop $L(\vec{x})$ thus is a gauge transporter that closes around compactified time. Often also the spatially averaged loop $P = 1/V
\sum_{\vec{x}} L(\vec{x})$ is considered, where $V$ is the spatial volume. Due to translational invariance $P$ and $L(\vec{x})$ have the same vacuum expectation value.
The Polyakov loop corresponds to a static quark source and its vacuum expectation value is (after a suitable renormalization) related to the free energy $F_q$ of a single quark, $\langle P \rangle
\propto \exp(-F_q/T)$, where $T$ is the temperature (the Boltzmann constant is set to 1 in our units). Below the critical temperature $T_c$ quarks are confined and $F_q$ is infinite, implying $\langle P \rangle = 0$. This is evident in the lhs. plot of Fig. \[ploopscatter\] where we show scatter plots of the values of the Polyakov loop $P$ in the complex plane for 100 configurations below (lhs. panel) and above $T_c$ (rhs.) [^1]. In the high temperature phase quarks become deconfined leading to a finite $F_q$ which gives rise to a non-vanishing Polyakov loop (rhs. in Fig. \[ploopscatter\]).
On a finite lattice, above $T_c$ the phase of the Polyakov loop assumes values near the center phases which for SU(3) are $0, \pm 2\pi/3$ (rhs. plot of Fig. \[ploopscatter\]). This is a reflection of the underlying center symmetry which is a symmetry of the action and the path integral measure of gluodynamics, that is broken spontaneously above the deconfinement temperature $T_c$. As long as the volume is finite all three sectors are populated, while in an infinite volume only one of the three phase values survives. This center symmetry and its spontaneous breaking are the basis for the above mentioned Svetitsky-Jaffe conjecture [@znbreaking].
![Scatter plots of the spatially averaged Polyakov loop $P$ in the complex plane for configurations below (lhs. panel) and above $T_c$ (rhs.). We show the results for our $40^3 \times 6$ ensembles. \[ploopscatter\]](Ploopscatter_40x6.eps){width="10.5cm"}
The relation of the deconfinement transition of SU($N$) gauge theory to $\mathds{Z}_N$-symmetric spin models has an interesting implication: For such spin models it is known that suitably defined clusters made from neighboring spins that point in the same direction show the onset of percolation at the same temperature where the $\mathds{Z}_N$-symmetry is broken spontaneously. For, e.g., the Ising system these percolating clusters were identified [@coniglio] as the Fortuin-Kasteleyn clusters [@fkclusters]. An interesting question is whether the cluster- and percolation properties can be directly observed in a lattice simulation of gluodynamics – without the intermediate step of the effective spin theory [@ploopeff] for the Polyakov loops.
For the case of gauge group SU(2) such cluster structures were analyzed in a series of papers [@satz; @fortunato], while for SU(3) the formation of center clusters has not yet been explored. In this paper we try to close this gap and study the behavior of the local loops $L(\vec{x})$ and the possible formation of center clusters. Furthermore, we study center clusters not only near $T_c$ (where they directly can be expected from the Svetitsky-Yaffe conjecture) but explore their emergence and properties in a window of temperatures ranging from 0.63 $T_c$ to 1.32 $T_c$.
[**Properties of local Polyakov loops**]{} For analyzing spatial structures of $L(\vec{x})$ on individual configurations we write the local loops in terms of a modulus $\rho(\vec{x})$ and a phase $\varphi(\vec{x})$, $$L(\vec{x}) \; = \; \rho(\vec{x}) \, e^{ \, i \, \varphi(\vec{x})} \; .$$ The first step of our investigation is to study the behavior of the modulus $\rho(\vec{x})$. In Fig. \[rhohistogram\] we show histograms for the distribution of $\rho(\vec{x})$ in the confined (lhs. plot) and the deconfined phase (rhs.). It is obvious, that the distributions of the modulus $\rho(\vec{x})$ below and above $T_c$ are almost indistinguishable. Furthermore we find that the distribution follows very closely the distribution according to Haar measure, which we show as a full curve. Only above $T_c$ we observe a very small deviation from the Haar measure distribution. The Haar measure distribution curves for the modulus and the phase are defined as $$P(\rho) \; = \; \int D[U] \, \delta(\rho - |\,\mbox{Tr}\,[U]|) \; , \;
P(\varphi) \; = \; \int D[U] \, \delta(\varphi - \mbox{arg Tr\, }[U]) \; ,$$ where $\delta$ is the Dirac delta-function and $D[U]$ is the Haar integration measure for group elements $U \in$ SU(3). These two distributions are obtained from a single group element and thus do not depend on any lattice parameters.
From the fact that the change of the modulus is very small we conclude that the jump of $\langle P \rangle$ at $T_c$, signaling the first order deconfinement transition, is not driven by a changing modulus of the local loops $L(\vec{x})$.
![Histograms for the distribution of the modulus $\rho(\vec{x})$ of the local loops $L(\vec{x})$ for temperatures below and above $T_c$. The full curve is the distribution according to Haar measure ($40 \times 6$ ensembles). \[rhohistogram\]](Pr_histo_40x6.eps){width="9cm"}
![Histograms for the distribution of the phase $\varphi(\vec{x})$ of the local loops $L(\vec{x})$. We compare the distribution below $T_c$ (lhs. plot) to the distribution in the deconfined phase (rhs.) for the sector of configurations with phases of the averaged loop $P$ near $-2\pi/3$. The full curve in the lhs. plot is the distribution according to Haar measure ($40 \times 6$ ensembles). \[phasehisto\]](histo_thetaP_rot_40x6.eps){width="14cm"}
Thus we focus on the behavior of the phase $\varphi(\vec{x})$, and again study histograms for its distribution. In Fig. \[phasehisto\] we compare the distribution below $T_c$ (lhs. plot) to the one in the deconfined phase (rhs.). For the latter we show the distribution for the sector of configurations characterized by phases of the averaged Polyakov loop $P$ in the vicinity of $-2\pi/3$ (compare the rhs. of Fig. \[ploopscatter\]).
The distribution of the phases $\varphi(\vec{x})$ is rather interesting: Also in the confined low temperature phase (lhs. plot in Fig. \[phasehisto\]) the distribution clearly is peaked at the center phases $-2\pi/3$, $0$ and $+2\pi/3$, and again perfectly follows the Haar measure distribution (full curve in the lhs. plot). The distribution is identical around these three phases and the vanishing result for $\langle P \rangle $ below $T_c$ comes from a phase average, $1 + e^{i2\pi/3} + e^{-i2\pi/3} = 0$.
Above $T_c$ (rhs. plot in Fig. \[phasehisto\]) the distribution singles out one of the phases. In our case, where configurations in the sector with phases of the averaged loop $P$ near $-2\pi/3$ are used for the plot, it is the value $-2\pi/3$ which is singled out. For configurations in one of the other two sectors (see rhs. plot of Fig. \[ploopscatter\]) the distribution is shifted periodically by $\pm 2\pi/3$. Obviously, above $T_c$ the distribution is not equal for the three center phases and the cancellation of phases does no longer work, resulting in a non-zero $\langle
P \rangle$.
The histograms for the phases $\varphi(\vec{x})$ suggest that at the critical temperature the local loops $L(\vec{x})$ start to favor phases near one spontaneously selected center value, while phases near the other two center values are depleted. This is illustrated in more detail in Fig. \[abundance\] where we show the abundance $A$ of lattice points with phases of $L(\vec{x})$ near the dominant and subdominant center values. To define the abundance $A$ we cut the interval $(-\pi, \pi)$ at the minima of the distribution of Fig. \[phasehisto\] into the three sub-intervals $(-\pi,-\pi/3)$, $(-\pi/3,\pi/3)$, $(\pi/3, \pi)$, which we refer to as “center sectors”. We count the number of lattice points with phases in each of the three center sectors and obtain their abundance $A$ by normalizing these counts with the volume. Fig. \[abundance\] shows that at low temperatures all three center sectors are populated with probability 1/3. Near $T_c$ one of the sectors starts to dominate while the other sectors are depleted.
[**Coherent center domains**]{} We have demonstrated for a wide range of temperatures that the center sectors play an important role for the phases $\varphi(\vec{x})$ of the local loops $L(\vec{x})$, which cluster near the center phases $0, \pm 2\pi/3$ at all temperatures. The deconfinement transition is manifest in the onset of a dominance of one spontaneously selected center sector. We now analyze whether the values of the phases $\varphi(\vec{x})$ are distributed homogeneously in space, or if instead there exist spatial domains with coherent phase values in the same sector.
![Abundance of lattice points $\vec{x}$ with phases of the local loop $L(\vec{x})$ in the dominant (triangles) and subdominant center sectors (circles, squares) as a function of temperature ($40 \times 6$ ensembles). \[abundance\]](abundance_vs_ttc_40x6.eps){width="8.9cm"}
In order to study such domains, we use sub-intervals that divide the interval $(-\pi,\pi)$ for the values of the $\varphi(\vec{x})$. For a more general analysis we introduce the cutting parameter $\delta \geq 0$ and define the three sub-intervals as $(-\pi + \delta, -\pi/3 -\delta)$, $(-\pi/3 + \delta, \pi/3 -\delta)$ and $(\pi/3 + \delta, \pi -\delta)$ (which we again refer to as “center sectors”). For $\delta = 0$ we obtain the old sub-intervals, while a value of $\delta > 0$ allows us to cut out lattice points where the phases are near the minima of the distributions shown in Fig. \[phasehisto\].
The definition of the clusters slightly differs from those that have been used for the analysis [@satz; @fortunato] of clusters in SU(2) gauge theory. Besides modifications of the Fortuin-Kasteleyn prescription studied in [@satz], in [@fortunato] the bonding probability between neighboring sites with same sign Polyakov loops[^2] was introduced as a free parameter. This parameter could then be tuned such that the onset of percolation agrees with the deconfinement temperature. In our definition the parameter $\delta$ allows one to reduce the lattice to a skeleton of points with phases close to the center elements (in intervals of width $2(\pi/3 - \delta)$ around the center values). For the plots shown in Figs. \[maxcluster\] and \[percplot\] we choose $\delta$ such that roughly those 40 % of lattice points are cut where the phases do not strongly lean towards one of the center values. We found that near $T_c$ the critical properties of the clusters (behavior of largest cluster and percolation) are stable when $\delta$ is varied in a small interval around that value [@inprep] (compare also the discussion in [@fortunato]). For example the curve for the weight of the largest cluster (see Fig. \[maxcluster\] below, where we show a comparison of different spatial volumes for a cut of 39 %) is form-invariant in a range of cuts from 30 % to 45 % and only is rescaled by a change of the amplitude of less than 15 %.
![Weight of the largest center cluster as a function of temperature. \[maxcluster\]](maxcluster_finsize.eps){width="9cm"}
In a next step we define clusters by assigning neighboring lattice sites with phases $\varphi(\vec{x})$ in the same center sector to the same cluster. Once these center clusters are defined we can study their properties and behavior with temperature using concepts developed for the percolation problem [@percolation]. In Fig. \[maxcluster\] we show the weight (i.e., the number of sites) of the largest cluster as a function of the temperature. For low temperatures all clusters are small, while as $T$ is increased towards the deconfinement temperature the largest cluster starts to grow quickly and above $T_c$ scales with the volume. This property indicates that in the deconfined phase the system has developed a percolating cluster. The onset of percolation at $T_c$ is confirmed in Fig. \[percplot\] where we show the percolation probability $p_\infty$ as a function of $T/T_c$. The percolation probability is computed by averaging an observable which is 1 if a spanning cluster exists and 0 otherwise. In our case, where we have periodic spatial boundary conditions, a spanning cluster is defined as a cluster who has at least one member site in every $y$-$z$ plane. In other words, we analyze percolation in $x$ direction, which is, however, no loss of generality as we have invariance under discrete spatial rotations. Varying $\delta$ in a range where the number of points we cut varies between 30 and 45 % (Fig. \[percplot\] is for 39 %), slightly roundens the transition curve, but leaves the onset of percolation unchanged at $T/T_c = 1$.
![Percolation probability $p_\infty$ of the dominant center clusters as a function of temperature.[]{data-label="percplot"}](percolation_finsize.eps){width="9cm"}
An interesting question is the size of the clusters in physical units in the confining phase, which could be related to some hadronic scale (see also the discussion in the next section). In order to study this cluster size below $T_c$ we computed 2-point correlation functions of points within the individual clusters. These correlators decay exponentially $\propto \exp(- r/\xi)$ with distance $r$, and the factor $\xi$ defines a linear size of the clusters in lattice units. We then analyze $d \equiv 2 \, \xi \, a$, which gives a definition of the cluster diameter in physical units ($a$ is the lattice spacing in fm). We find that up to $T = 0.85 T_c$ this diameter is essentially independent of the temperature, with a value of $d = 0.46(5)$ fm at a cut of 39 % and $d = 0.62(7)$ fm at a cut of 30 %. Compared to the expected sizes of rougly 0.5 fm for heavy quark mesons this is a quite reasonable result for the linear scale of the clusters which suggests that the physical role of the clusters below $T_c$ should be studied in more detail (see [@inprep]).
[**Summary and discussion**]{} We have explored the clustering of the phases $\varphi(\vec{x})$ of the local quark sources $L(\vec{x})$ near the center values, both below and above $T_c$. We find that in the range of temperatures we consider, $T = 0.63\, T_c$ to $T = 1.32\, T_c$, the local Polyakov loop phases prefer values near the center values and corresponding clusters may be identified for these temperatures. Using the parameter $\delta$ we can construct clusters such that the deconfinement transition is characterized by percolation of the clusters in the dominant sector.
From the cluster properties a simple qualitative picture for confinement and the deconfinement transition emerges. Below $T_c$ the clusters of lattice points which have the same center phase information are small. Only if a quark- and an anti-quark source are sufficiently close to each other they fit into the same cluster and can have a non-vanishing expectation value. Sources at distances larger than a typical cluster size receive the independent center fluctuations from different clusters and the correlator averages to zero. Above $T_c$ the clusters percolate and coherent center information is available also for larger distances allowing for non-vanishing correlation at large separation of the sources. In this picture deconfinement is a direct consequence of a percolating center cluster.
A possible role of local center structures for confinement has been addressed also in a different approach, using a projection of the link variables $U_\mu(\vec{x},t)$ at all points in space and time to a center element after fixing to a suitable gauge (see, e.g., [@centervortices] for a selection of recent results). This analysis is motivated by understanding the role of topological objects for the QCD phase transition. It would be highly interesting to study a possible connection of the percolation aspects of the transition to the dynamics of such topological objects. Of particular relavance would be an analysis of a possible relation to calorons which induce strong local variations of the Polyakov loop that might play an important role in the formation of the center clusters [@calorons].
We conclude with a few comments on the extension of the center domain picture to the case of full QCD: The fermion determinant describing the dynamical quarks can be expressed as a sum over closed loops, which may be viewed as generalized Polyakov loops and are sensitive to the center properties of the gauge fields [@candet]. The fermion determinant breaks the center symmetry explicitly and acts like an external magnetic field which favors the real sector (phase 0) for the Polyakov loop $P$. However, preliminary numerical results with dynamical fermions [@inprep] show that locally also the two complex sectors (phases $\pm 2\pi/3$) remain populated. The corresponding clusters will again lead to a coherent phase information for sufficiently close quark lines. As for the pure gauge theory studied in this letter, the preliminary results [@inprep] show that the transition to confinement is again accompanied by a pronounced increase of the abundance for the dominant (i.e., real) sector. However, the explicit symmetry breaking through the determinant leads to a crossover type of behavior in the dynamical case. An interesting related question, which has already been raised in the literature [@chiral], is whether also the chiral transition may be characterized as a percolation phenomenon.\
\
[**Acknowledgments:**]{} The author thanks Mike Creutz, Julia Danzer, Mitja Diakonov, Christian Lang, Ludovit Liptak, Axel Maas, Stefan Olejnik, Alexander Schmidt and Andreas Wipf for valuable comments. The numerical calculations were done at the ZID clusters of the University Graz.
[1234567]{}
L.G. Yaffe, B. Svetitsky, Phys. Rev. D [**26**]{} (1982) 963; Nucl. Phys. B [**210**]{} (1982) 423. J. Polonyi, K. Szlachanyi, Phys. Lett. B [**110**]{} (1982) 395; M. Ogilvie, Phys. Rev. Lett. [**52**]{} (1984) 1369; F. Green, F. Karsch, Nucl. Phys. B [**238**]{} (1984) 297; S. Fortunato, F. Karsch, P. Petreczky and H. Satz, Phys. Lett. B [**502**]{} (2001) 321; C. Wozar, T. Kaestner, A. Wipf, T. Heinzl, B. Pozsgay, Phys. Rev. D [**74**]{} (2006) 114501; A. Wipf, T. Kaestner, C. Wozar, T. Heinzl, SIGMA [**3**]{} (2007) 006; C. Wozar, T. Kaestner, A. Wipf, T. Heinzl, Phys. Rev. D [**76**]{} (2007) 085004. M. Lüscher, P. Weisz, Commun. Math. Phys. [**97**]{} (1985) 59 \[Err.: [**98**]{} (1985) 433\]; G. Curci, P. Menotti, G. Paffuti, Phys. Lett. B [**130**]{} (1983) 205 \[Err.: B [**135**]{} (1984) 516\]. C. Gattringer, R. Hoffmann, S. Schaefer, Phys. Rev. D [**65**]{} (2002) 094503. C. Gattringer, P.E.L. Rakow, A. Schäfer, W. Söldner, Phys. Rev. D [**66**]{} (2002) 054502. A. Coniglio and W. Klein, J. Phys. A [**13**]{} (1980) 2775.
C.M. Fortuin and P. W. Kasteleyn, Physica [**57**]{} (1972) 536. S. Fortunato, H. Satz, Phys. Lett. B [**475**]{} (2000) 311; S. Fortunato and H. Satz, Nucl. Phys. A [**681**]{} (2001) 466. S. Fortunato, J. Phys. A [**36**]{} (2003) 4269. G.R. Grimmett, [*Percolation*]{}, Springer, New York, 1999.
C. Gattringer et al, work in preparation.
R. Bertle, M. Faber, J. Greensite, S. Olejnik, JHEP [**9903**]{} (1999) 019; M. Engelhardt, K. Langfeld, H. Reinhardt, O. Tennert, Phys. Rev. D [**61**]{} (2000) 054504; J. Greensite, Prog. Part. Nucl. Phys. [**51**]{} (2003) 1; M. Engelhardt, Nucl. Phys. Proc. Suppl. [**140**]{} (2005) 92; M. Pepe, PoS [**LAT2005**]{} (2006) 017; J. Greensite, K. Langfeld, S. Olejnik, H. Reinhardt, T. Tok, Phys. Rev. D [**75**]{} (2007) 034501; V.G. Bornyakov, E.M. Ilgenfritz, B.V. Martemyanov, S.M. Morozov, M. Müller-Preussker and A. I. Veselov, Phys. Rev. D [**77**]{} (2008) 074507. E.M. Ilgenfritz, M. Müller-Preussker and D. Peschka, Phys. Rev. D [**71**]{} (2005) 116003; F. Bruckmann, E.M. Ilgenfritz, B. Martemyanov and B. Zhang, Phys. Rev. D [**81**]{} (2010) 074501. C. Gattringer, L. Liptak, arXiv:0906.1088 \[hep-lat\]. M. Beccaria and A. Moro, Phys. Rev. D [**66**]{} (2002) 037502 \[Erratum-ibid. D [**72**]{} (2005) 029901\].
[^1]: The numerical results we show are from a Monte Carlo simulation of SU(3) lattice gauge theory using the Lüscher-Weisz gauge action [@LuWe]. We work on various lattice sizes ranging from $20^3 \times 6$ to $40^3\times 12$. The scale was set [@scale] using the Sommer parameter. In our figures we always use the dimensionless ratio $T/T_c$ with the critical temperature $T_c = 296$ MeV calculated for this action in [@TCdet]. All errors we show are statistical errors determined with a single elimination jackknife analysis.
[^2]: For SU(2) the $L(\vec{x})$ are real and the center phases are either $+1$ or $-1$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove a surgery formula and an excision formula for the Furuta-Ohta invariant $\lambda_{FO}$ defined on homology $S^1 \times S^3$, which provides more evidence on its equivalence with the Casson-Seiberg-Witten invariant $\lambda_{SW}$. These formulae are applied to compute $\lambda_{FO}$ of certain families of manifolds obtained as mapping tori under diffeomorphisms of $3$-manifolds. In the course of the proof, we give a complete description of the degree-zero moduli space of ASD instantons on $4$-manifolds of homology $H_*(D^2 \times T^2; {\mathbb{Z}})$ with a cylindrical end modeled on $[0, \infty) \times T^3$.'
address: 'MS 050 Department of Mathematics, Brandeis University, 415 South St., Waltham MA 02453'
author:
- Langte Ma
bibliography:
- 'ReferencesLFO.bib'
title: 'Surgery and Excision for Furuta-Ohta invariants on Homology $S^1 \times S^3$ '
---
Introduction
============
The Furuta-Ohta invariant was introduced by Furuta and Ohta in [@FO93] to study exotic structures on punctured four manifolds. Originally the Furuta-Ohta invariant is defined on manifolds called ${\mathbb{Z}}[{\mathbb{Z}}]$-homology $S^1 \times S^3$, i.e. closed $4$-manifolds $X$ with the same homology as $S^1 \times S^3$ whose infinite cyclic cover $\tilde{X}$ has the same homology as $S^3$. The conjecture is that the Furuta-Ohta invariant $\lambda_{FO}(X) \mod 2$ reduces to the Rohlin invariant $\mu(X)$ associated to the $4$-manifold. Instead of approaching the conjecture directly, Mrowka-Ruberman-Saveliev [@MRS11] considered the Seiberg-Witten correspondence $\lambda_{SW}$ defined over homology $S^1 \times S^3$ where they manage to show that $\lambda_{SW}(X) \mod 2$ reduces to the Rohlin invariant. So the problem has been transformed to prove the equivalence between the Casson-Seiberg-Witten invariant $\lambda_{SW}$ and the Furuta-Ohta invariant $\lambda_{FO}$. The motivation of this article is to study how the Furuta-Ohta invariants change under certain topological operations, which in turn provides more evidence on this equivalence.
In this article we allow the Furuta-Ohta invariants to be defined on a slightly larger class of manifolds which we refer to as admissible homology $S^1 \times S^3$.
Let $X$ be a smooth oriented closed $4$-manifold with $H_*(X; {\mathbb{Z}}) \cong H_*(S^1 \times S^3; {\mathbb{Z}})$. We call $X$ an admissible homology $S^1 \times S^3$ if it further satisfies the following property: for all non-trivial $U(1)$-representations $\rho: \pi_1(X) \to U(1)$, one has $$H^1(X; {\mathbb{C}}_{\rho}) = 0.$$
Now let $X$ be an admissible homology $S^1 \times S^3$. After fixing a generator $1_X \in H^1(X; {\mathbb{Z}})$ as the homology orientation, the Furuta-Ohta invariant is defined to be a quarter of the counting of degree-zero irreducible anti-self-dual $SU(2)$-instantons on $X$, which is written as $$\lambda_{FO}(X):= {1 \over 4}\# {\mathcal{M}}^*_{\sigma}(X).$$ It’s proved in [@RS1] that for a generic small holonomy perturbation $\sigma$, the moduli space ${\mathcal{M}}^*_{\sigma}(X)$ is a compact oriented $0$-manifold and the counting is well-defined. We note that when $X=S^1 \times Y$ is given by the product of $S^1$ with an integral homology sphere, the Furuta-Ohta invariant coincides with the Casson invariant of $Y$ [@RS1], i.e. $\lambda_{FO}(S^1 \times Y) = \lambda(Y)$.
The first topological operation we consider is the torus surgery. We give a brief description here, and a detailed one in Sectoin \[sffo\]. Let $X$ be an admissible integral homology $S^1 \times S^3$ with a fixed generator $1_X \in H^1(X; {\mathbb{Z}})$, and $\mathcal{T} \hookrightarrow X$ an embedded $2$-torus satisfying that the induced map on first homology $$H_1(\mathcal{T}; {\mathbb{Z}}) \to H_1(X; {\mathbb{Z}})$$ is surjective. We will refer to such a torus as an essentially embedded torus. We write $\nu(\mathcal{T})$ for a tubular neighborhood of $\mathcal{T}$ in $X$, and fix an indentification $\nu(\mathcal{T}) \cong D^2 \times T^2$ as a framing. Let $M=\overline{X \backslash \nu(\mathcal{T})}$ be the closure of the complement of the neighborhood. It’s straightforward to compute that $H_*(M; {\mathbb{Z}}) \cong H_*(D^2 \times T^2; {\mathbb{Z}})$. The framing provides us with a basis $\{\mu, \lambda, \gamma\}$ for $H_1(\partial \nu(\mathcal{T}); {\mathbb{Z}})$. We require $[\gamma] \in H_1(X; {\mathbb{Z}})$ to be the dual of the generator $1_X \in H^1(X; {\mathbb{Z}})$, $\mu$ to be the meridian of $\mathcal{T}$, and $[\lambda]$ to be null-homologous in $M$. Given a relatively prime pair $(p, q)$, performing $(p,q)$-surgery along $\mathcal{T}$ results in the $4$-manifold $$X_{p, q} = M \cup_{\varphi_{p,q}} D^2 \times T^2,$$ where with respect to the basis $\{\mu, \lambda, \gamma\}$ the gluing map is given by $$\varphi_{p, q}=
\begin{pmatrix}
p & r & 0 \\
q & s & 0 \\
0 & 0 & 1
\end{pmatrix}
\in SL(3, {\mathbb{Z}}).$$ We will see later in Section \[sffo\] that the $(1, q)$-surgered manifold $X_{1,q}$ is still an admissible integral homology $S^1 \times S^3$ with a homology orientation induced from that of $X$. We note that $H_*(X_{0,1}; {\mathbb{Z}}) \cong H_*(S^2 \times T^2; {\mathbb{Z}})$. We denote by $w_{\mathcal{T}} \in H^2(X_{0,1}; {\mathbb{Z}}/2)$ the class that’s dual to the mod $2$ class of the core $T^2$ in the gluing $D^2 \times T^2$. We write $D^0_{w_{\mathcal{T}}}(X_{0,1})$ for the counting of gauge equivalence classes of the irreducible anti-self-dual $SO(3)$-connections on the bundle $SO(3)$-bundle $P$ over $X_{0,1}$ characterized by $$p_1(P)=0 \text{ and } w_2(P)=w_{\mathcal{T}}.$$ The torus surgery formula relates $\lambda_{FO}(X)$, $\lambda_{FO}(X_{1,q})$, and $D^0_{w_{\mathcal{T}}}(X_{0,1})$ as follows.
\[surq\] After fixing appropriate homology orientations, one has $$\lambda_{FO}(X_{1, q}) = \lambda_{FO}(X) + {q \over 2}D^0_{w_{\mathcal{T}}}(X_{0,1}), \; q \in {\mathbb{Z}}.$$
In the product case $X=S^1 \times Y$ with $Y$ an integral homology sphere, the Furuta-Ohta invariant of $X$ coincides with the Casson invariant $Y$. We recall that the surgery formula of the Casson invariant is $$\lambda(Y_{{1 \over q}}(\mathcal{K})) = \lambda(Y) + {q \over 2}\Delta''_{\mathcal{K}}(1),$$ where $\mathcal{K} \subset Y$ is a knot, and $\Delta_{\mathcal{K}}(t)$ is the symmetrized Alexander polynomial of $\mathcal{K}$. Comparing with the surgery formula of the Furuta-Ohta invariant we get the following result.
\[0d0s\] Let $\mathcal{K} \subset Y$ be a knot in an integral homology sphere. Then $$D^0_{w_{\mathcal{T}}}(S^1 \times Y_0(K)) = \Delta''_{\mathcal{K}}(1).$$
The surgery formula of the Furuta-Ohta invariant should be compared with that of the Casson-Seiberg-Witten invariant proved in [@M1]: $$\lambda_{SW}(X_{1, q}) = \lambda_{SW}(X) + q \mathcal{SW}(X_{0,1}),$$ where $\mathcal{SW}(X_{0,1})$ is the Seiberg-Witten invariant of $X_{0,1}$ computed in the chamber specified by small perturbations. The Casson-Seiberg-Witten invariant is defined using Seiberg-Witten theory [@MRS11] combining the counting of irreducible monopoles and an index-theoretical correction term. However, the Witten conjecture has not been proved in the case for non-simply connected $4$-manifolds with $b^+=1$. So one does not get the equivalence of $\lambda_{FO}$ and $\lambda_{SW}$ for families of admissible homology $S^1 \times S^3$ by appealing to the surgery formulae directly.
On the other hand, we apply the surgery formula to give an independent computation of the Furuta-Ohta invariant for manifolds given by the mapping torus of finite order diffeomorphism as in [@LRS20]. More precisely, we let $\mathcal{K} \subset Y$ be a knot in an integral homology $S^1 \times S^3$. Fix an integer $n>1$, we denote by $\Sigma_n(Y, \mathcal{K})$ the $n$-fold cyclic cover of $Y$ branched along $\mathcal{K}$, and $\tau_n: \Sigma_n(Y, \mathcal{K}) \to \Sigma_n(Y, \mathcal{K})$ the covering translation. It’s shown in [@LRS20 Proposition 6.1] that the mapping torus $X_n(Y, \mathcal{K})$ of $\Sigma_n(Y, \mathcal{K})$ under the map $\tau_n$ is an admissible homology $S^1 \times S^3$ whenever $\Sigma_n(Y, \mathcal{K})$ is a rational homology sphere. We can apply the surgery formula to compute $\lambda_{FO}(X_{n}(Y, \mathcal{K})$ as follows.
\[fod\] Assume that $\Sigma_n(Y, \mathcal{K})$ is a rational homology sphere. Then $$\lambda_{FO}(X_n(Y, \mathcal{K})) = n\lambda(Y) + {1 \over 8} \sum_{m=1}^{n-1}\operatorname{sign}^{m/ n}(Y, \mathcal{K}),$$ where $\operatorname{sign}^{m/n}(Y, \mathcal{K})$ is the Tristram-Levine signature of $\mathcal{K}$.
This computation is carried out in [@LRS20 Theorem 6.4] using a more diret method by relating to the equivariant Casson invariant of $\Sigma_n(Y, \mathcal{K})$. In [@M1], the author also computed the Casson-Seiberg-Witten invariant for $X_n(Y, \mathcal{K})$ without assuming $\Sigma_n(Y, \mathcal{K})$ is a rational homology sphere, which turns out to be the same formula. However, when $\Sigma_n(Y, \mathcal{K})$ fails to be a rational homology sphere, $X_n(Y, \mathcal{K})$ is not admissible. So the Furuta-Ohta invariant is not defined.
We move to the next topological operation which we call torus excision in this article. The idea is to replace the torus neighborhood $D^2 \times T^2$ with a homology $D^2 \times T^2$ and glue it to the complement by further applying a diffeomorphism. Let $(X_1, \mathcal{T}_1)$ and $(X_2, \mathcal{T}_2)$ be two pairs of essentially embedded torus in an admissible homology $S^1 \times S^3$ as above. We also fix framings for both $\nu(\mathcal{T}_1)$ and $\nu(\mathcal{T}_2)$ to get a basis $\{\mu_i, \lambda_i, \gamma_i\}$ of $H_1(\partial\nu(\mathcal{T}_i))$ as before. To emphasize the orientation, we write $X_1 = M_1 \cup \nu(\mathcal{T}_1)$ and $X_2 = \nu(\mathcal{T}_2) \cup M_2$, i.e. the left parts $\partial M_1 = -\nu(\mathcal{T}_1), \partial \nu(\mathcal{T}_2) =-M_2$ are identified with $T^3$ with a fixed orientation. Let $\varphi: \partial M_2 \to \partial M_1$ be a diffeomorphism so that the glued manifold $$X_1 \#_{\varphi} X_2:=M_1 \cup_{\varphi} M_2$$ is an admissible homology $S^1 \times S^3$. We let $X_{1, \varphi} = M_1 \cup_{\varphi} D^2 \times T^2$ and $X_{2, \varphi}=D^2 \times T^2 \cup_{\varphi} M_2$. We will give more explanations on what the gluing map means in this context later in Section \[fsfo\]. Roughly $D^2 \times T^2$ is glued to $M_1$ the same way as $M_2$ does, so is to $M_2$. Since the admissible assumption is purely homological, we see that both $X_{1, \varphi}$ and $X_{2, \varphi}$ are admissible. The excision formula of the Furuta-Ohta invariants states as follows.
\[exif\] After fixing appropriate homology orientations, one has $$\lambda_{FO}(X_1 \#_{\varphi} X_2) = \lambda_{FO} (X_{1, \varphi}) + \lambda_{FO}(X_{2, \varphi}).$$
Note that $X_{i, \varphi}$ is obtained from $X_i$ via a torus surgery. When the gluing map $\varphi$ has the form we considered in Theorem \[surq\], we further expand the formula as $$\lambda_{FO}(X_1 \#_{\varphi_{1, q}} X_2) = \lambda_{FO} (X_1) + {q \over 2}D^0_{w_{\mathcal{T}}}(X_{1, \varphi_{0,1}}) + \lambda_{FO}(X_2) + {q \over 2}D^0_{w_{\mathcal{T}}}(X_{2, \varphi_{0,1}}).$$ We will see later in examples there are certain interesting gluing maps that are not of the form we considered in the surgery formula. Thus we need a generalized surgery formula to compare $\lambda_{FO}(X_{i, \varphi})$ with $\lambda_{FO}(X_i)$. It turns out there is an extra term coming out in the formula as we go through the proof of Theorem \[surq\], which is caused by the contribution of some ‘bifurcation points’ (c.f. Definition \[bfp\]) in the moduli space of the torus complement. We hope to formulate this extra term in a future article.
We note that the fiber sum operation considered in [@M2] is a special case of the excision. The fiber sum of $(X_1, \mathcal{T}_1)$ and $(X_2, \mathcal{T}_2)$ is given by gluing the torus complements by a map interchanging the meridian $\mu$ and longitude $\lambda$: $$X_1 \#_{\mathcal{T}} X_2:=M_1 \cup_{\varphi_{\mathcal{T}}} M_2,$$ where with respect to the basis $\{\mu_i, \lambda_i, \gamma_i\}$, $\varphi_{\mathcal{T}}$ is given by $$\label{fsrm}
\varphi_{\mathcal{T}}=
\begin{pmatrix}
0 & 1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1
\end{pmatrix}.$$ Note that $M_1 \cup_{\varphi_{\mathcal{T}}} D^2 \times T^2= X_1$, $D^2 \times T^2 \cup_{\varphi_{\mathcal{T}}} M_2 = X_2$. We conclude that the Furuta-Ohta invariant is additive under taking torus fiber sum.
\[fibs\] After fixing appropriate homology orientations, one has $$\lambda_{FO}(X_1 \#_{\mathcal{T}} X_2)= \lambda_{FO}(X_1) + \lambda_{FO}(X_2).$$
In the product case $X_i=S^1 \times Y_i$, $\mathcal{T}_i = S^1 \times \mathcal{K}_i$, i=1, 2, the fiber sum $X_1 \#_{\mathcal{T}} X_2$ is the product of $S^1$ with the knot splicing $Y_1 \#_{\mathcal{K}} Y_2$ of the pairs $(Y_1, \mathcal{K}_1)$ and $(Y_2, \mathcal{K}_2)$. Then the fiber sum formula for the Furuta-Ohta invariant recovers the additivity of the Casson invariant under knot splicing. We also note that the same fiber sum formula holds for the Casson-Seiberg-Witten invariant proved in [@M2]: $$\lambda_{SW}(X_1 \#_{\mathcal{T}} X_2)= \lambda_{SW}(X_1) + \lambda_{SW}(X_2).$$
Both the proofs of Theorem \[surq\] and Theorem \[exif\] rely on understanding the anti-self-dual moduli space of the torus complement. The difficulty of analyzing the moduli space of the torus complement arises from two parts. One is $b^+=0$, which prevents us from using metric perturbations to get rid of the reducible locus. The other is that the gluing boundary is $T^3$ whose moduli space has certain non-degeneracy, especially when we consider the trajectories on the torus complement flows to the singuler points in the moduli space. To deal with the first issue, we consider holonomy perturbations and analyze the local structure of the moduli space near the reducible locus. To deal with the second issue, we adopt the ‘center manifold’ technique developed in [@MMR].
For the rest of this section, we state the results on the degree zero anti-self-dual moduli space. We let $Z$ be a $4$-manifold with cylindrical end and $E = Z \times {\mathbb{C}}^2$ the trivialized ${\mathbb{C}}^2$-bundle. The degree-zero perturbed moduli space ${\mathcal{M}}_{\sigma}(Z)$ of ASD instantons on $Z$ consists of gauge equivalence classes of $SU(2)$-connections $A$ on $E$ satisfying the following:
1. The self-dual part of the curvature equals the perturbation, i.e. $F^+_A=\sigma(A)$.
2. The curvature of $A$ is of finite energy, i.e. $\int_Z |F_A|^2 < \infty$.
3. The Chern-Weil integral vanishes, i.e. $\int_Z \operatorname{tr}(F_A \wedge F_A) =0$.
The reason for calling this moduli space degree zero is due to the third requirement on the vanishing of the Chern-Weil integral. The perturbation function $\sigma$ is gauge-equivariant and satisfy an exponential decay condition along the end: $$\| \sigma(A) \|_{L^{\infty}(\{t \} \times Y)} \leq Ce^{-\mu t},$$ where $C$ and $\mu$ are some positive constant independent of $A$. The space $\mathcal{P}_{\mu}$ of perturbations is parametrized by a Banach space denoted by $(W, \| \cdot \|_W)$. We will write $\sigma=\sigma_{\pmb{\omega}}$ for some $\pmb{\omega} \in W$, and $\|\sigma\|=\|\pmb{\omega}\|_W$. We write $Z=M \cup [0, \infty) \times Y$ where $M$ is a compact $4$-manifold with boundary, $Y$ is a $3$-manifold. $[0, \infty) \times Y$ is refered to as the cylindrical part of $Z$. Let’s write $$\chi(Y):=\operatorname{Hom}(\pi_1(Y), SU(2)) /Ad$$ for the $SU(2)$-character variety of the $3$-manifold $Y$ in the cylindrical end. Via the holonomy map, $\chi(Y)$ is identified with the gauge equivalence classes of the flat connections on $Y$. The first step to deduce a structure theorem is to establish the existence of the asymptotic map on ${\mathcal{M}}_{\sigma}(Z)$.
\[asyom\] Given $[A] \in {\mathcal{M}}_{\sigma}(Z)$, the limit $\lim_{t \to \infty} [A|_{\{t \} \times Y}]$ exists and lies in $\chi(Y)$. The assignment of $[A]$ to its limit defines a continuous map $$\partial_+: {\mathcal{M}}_{\sigma}(Z) \longrightarrow \chi(Y).$$
To our best knowledge, the existence of the limit of finite energy instantons, or in general the gradient flowlines in banach spaces, has been established in the following cases, none of which fit into our setting. Let’s write $A|_{\{ t\} \times Y} = B(t)$. The gauge-fixed perturbed ASD equation restricted on the end has the form $$\dot{B}(t)= -\operatorname{grad}\operatorname{cs}(B(t)) + \mathfrak{p}(B(t)),$$ where $\operatorname{cs}$ is the Chern-Simons functional. Consider the following cases
1. The critical points of $\operatorname{cs}$ are isolated and non-degenerate, and the integral $\int_{[0, \infty)} \|\mathfrak{p}(B(t))\|$ is finite.
2. There is a perturbed functional $\operatorname{cs}_{\mathfrak{p}}$ satisfying $$-\operatorname{grad}\operatorname{cs}_{\mathfrak{p}}(B(t)) = -\operatorname{grad}\operatorname{cs}(B(t)) + \mathfrak{p}(B(t)).$$ Moreover, either the critical points of $\operatorname{cs}_p$ are Morse-Bott or the perturbed functional $\operatorname{cs}_{\mathfrak{p}}$ is analytic.
3. $\|\mathfrak{p}(B(t))\| \leq \alpha \| \operatorname{grad}\operatorname{cs}(B(t))\|$ for some $\alpha <1$ after $t >>0$.
Perturbations of the form (i) and (ii) are usually considered in establishing the Floer homology. (iii) was considered in [@MMR] where they used metric perturbations. The existence of the limit follows from a modified argument of Simon’s [@S83]. Sometimes people also consider perturbations of compact support, which falls into case (iii). In the end, generic compact perturbations suffice for our purpose. But to show various transversality results, one has to put perturbations into a Banach space. So we will essentially show the ‘center manifold’ technique in [@MMR] works for perturbations in a completed space. We also note Theorem \[asyom\] is proved in a slightly more general context in Theorem \[eam\] where we shall consider the based moduli space.
We refer to $\partial_+$ as the asymptotic map. Now we focus on the type of manifolds of our primary interests. Let $Z$ be a manifold with cylindrical end satisfying the following:
1. The integral homology of $Z$ is the same as that of $D^2 \times T^2$, i.e. $H_*(Z; {\mathbb{Z}}) \cong H_*(D^2 \times T^2; {\mathbb{Z}})$.
2. The cylindrical end of $Z$ is modeled on $[0, \infty) \times T^3$. The torus $T^3$ is endowed with a flat metric $h$.
We shall refer to a pillowcase as the quotient of a torus $T^2$ under the hypoelliptic involution. Thus a pillowcase is an orbifold smoothable to $S^2$ with $4$ singular points having ${\mathbb{Z}}/2$ as the isotropy group. One will see later in Section \[tims\] that each central connection in $\chi(T^3)$ is a singular point, and there are eight of them up to gauge equivalence. We denote by $\mathfrak{C} \subset \chi(T^3)$ the set of these central classes. We now state the structure theorem for the moduli space over $Z$.
\[str\] Let $\sigma$ be a generic perturbation in $\mathcal{P}_{\mu}$ with $\|\sigma\| < \mathfrak{c}$ for some constant $\mathfrak{c} >0$. We fix an orientation of $H^1(Z; {\mathbb{Z}})$. The degree zero moduli space of perturbed anti-self-dual instantons ${\mathcal{M}}_{\sigma}(Z)$ is a compact smooth oriented stratified space with the following structures:
1. The reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ is a pillowcase whose singular points consist of the four gauge equivalence classes of central flat $SU(2)$-connections on $Z$.
2. The irreducible locus ${\mathcal{M}}^*_{\sigma}(Z)$ is a smooth oriented $1$-manifold of fintie components, each of which is diffeomorphic to either the circle $S^1$ or the open interval $(0,1)$.
3. The ends of the closure of the open arcs in ${\mathcal{M}}^*_{\sigma}(Z)$ lie in ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ away from the singular points. Near each end $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$, the moduli space ${\mathcal{M}}_{\sigma}(Z)$ is modeled on a neighborhood of $0$ in the zero set $\mathfrak{o}^{-1}(0)$, where $$\begin{split}
\mathfrak{o}: {\mathbb{R}}^2 \oplus {\mathbb{R}}_+ &\longrightarrow {\mathbb{C}}\\
(x_1, x_2, r) & \longmapsto (x_1 + ix_2) \cdot r.
\end{split}$$
4. The asymptotic map restricted on the irreducible locus $\partial_+: {\mathcal{M}}^*_{\sigma}(Z) \to \chi(T^3)$ is $C^2$ and transverse to a given submanifold (the choice of the perturbation $\sigma$ depends on the given submanifold).
5. The asymptotic values of irreducible instantons miss the central classes, i.e. $\partial_+\big ({\mathcal{M}}^*_{\sigma}(Z)\big) \cap \mathfrak{C} = \varnothing$.
The proof of Theorem \[str\] is divided into Proposition \[TSC\], Proposition \[TIS\], Corollary \[RPI\], and Proposition \[nbr\].
The techniques in the proof of Theorem \[str\] can be applied to deduce the structure of moduli spaces of any degree over any end-cylindrical $4$-manifold with $b^+=0$ consisting of instantons asymptotic away from the singular points along the end. The picture will be a space stratified by the type of stabilizers on both the $4$-manifold and its asymptotic $3$-manifold. The irreducible locus will be a smooth manifold of dimension given by the index of the deformation complex, and the reducible locus will admit a local cone bundle neighborhood. The case we considered in Theorem \[str\] is described completely due to the fact that the dimension of the irreducible locus is low, thus after perturbations, the asymptotic values avoid the singular points.
Outline {#outline .unnumbered}
-------
Here we give an outline of this article. Section \[pre\] introduces the set-up for the moduli space including the holonomy perturbations. Section \[EAM\] establishes the existence of the asymptotic value for the perturbed instantons, i.e. Theorem \[asyom\]. Section \[tims\] deduces the transversality of the irreducible moduli space and the asymptotic map. Section \[trl\] is devoted to describing the reducible locus together with its neighborhood under small generic perturbations. Section \[sffo\] and Section \[fsfo\] prove the surgery formula Theorem \[surq\] and the excision formula Theorem \[exif\] respectively.
Acknowledgement {#acknowledgement .unnumbered}
---------------
The author would like to express his gratitude to his advisor Daniel Ruberman for helpful discussions and encouragement. He also wants to thank Cliff Taubes for sharing his expertise and communicating a proof of Proposition \[lgf\].
Prelinminaries {#pre}
==============
The Unperturbed Moduli Space of Manifolds with Cylindrical End
--------------------------------------------------------------
A Riemannian manifold $(Z, g)$ with cylindrical end consists of the following data:
1. A compact manifold $M \subset Z$ with boundary $\partial M=Y$.
2. A cylindrical end $[0, \infty) \times Y$ is attached to the boundary $\partial M$ so that $$Z = M \cup [0, \infty) \times Y.$$
3. Over the end the metric $g$ has the form $$g|_{[0, \infty) \times Y} = dt^2 + h,$$ where $h$ is a metric on $Y$, $t$ is the coordinate function on $[0, \infty)$.
4. Over a collar neighborhood $(-1, 0] \times Y$ of $\partial M$ in $M$ the metric is identified as $$g|_{(-1, 0] \times Y}=dt^2 +h.$$
Let $(Z, g)$ be a smooth manifold with cylindrical end as above. Consider the trivial ${\mathbb{C}}^2$-bundle $E \to Z$ satisfying $$E|_{[0, \infty) \times Y} = \pi^* E',$$ where $E' \to Y$ is the trivial ${\mathbb{C}}^2$-bundle on $Y$, $\pi: [0, \infty) \times Y \to Y$ is the projection map. Fix $k \geq 3$, we write ${\mathcal{A}}_{k, loc}$ for the space of $L^2_{k, loc}$ $SU(2)$-connections on $E$. The gauge group ${\mathcal{G}}_{k+1, loc}$ consists of $L^2_{k+1, loc}$ automorphisms of the associated principle bundle $P$, which is identified with $L^2_{k+1, loc}(Z, SU(2))$. The gauge action is given by $$\label{GA}
\begin{split}
{\mathcal{G}}_{k+1, loc} \times {\mathcal{A}}_{k, loc} & \longrightarrow A_{k, loc} \\
(u, A) & \longmapsto u\cdot A := A- u^{-1}d_Au
\end{split}$$ We say $A$ is irreducible if $\operatorname{Stab}(A)={\mathbb{Z}}/2$, and reducible if $\operatorname{Stab}(A) \supset U(1)$. In particular when $\operatorname{Stab}(A)=SU(2)$, we say $A$ is central. The configuration space ${\mathcal{A}}_{k, loc}$ decomposes into the irreducible and reducible parts: $${\mathcal{A}}_{k, loc}={\mathcal{A}}^*_{k, loc} \cup {\mathcal{A}}^{\operatorname{red}}_{k, loc}.$$ Given a connection $A$ with $F_A \in L^2(Z, \Lambda^2T^*Z\otimes {\mathfrak{su}}(2))$, we denote its Chern-Weil integral by $$\kappa(A):={1 \over {8\pi^2}} \int_Z \operatorname{tr}(F_A \wedge F_A).$$ We denote the energy of a connection $A$ by $$\mathcal{E}(A):=\int_Z |F_A|^2.$$
The unperturbed moduli space of finite energy instantons is defined to be $${\mathcal{M}}_k(Z):= \{A \in {\mathcal{A}}_{k, loc}: F^+_A=0, \mathcal{E}(A) < \infty, \kappa(A)=0 \} / {\mathcal{G}}_{k+1, loc}.$$
Unlike the case of closed manifolds, any $SU(2)$-bundle over a manifold with cylindrical end is trivial. But this fact does not imply any ASD connection on $Z$ is flat. So we have imposed the vanishing of the Chern-Weil integral on the definition to ensure the setting is the same when the closed case as we run the neck-stretching argument later. In this way every connection $[A] \in {\mathcal{M}}_k(Z)$ is actually flat.
To simplify notation, we usually omit $k$ in the notation unless it becomes important. Note that the gauge action ${\mathcal{G}}$ on ${\mathcal{A}}$ is not free. To get rid of this issue one can consider the based moduli space defined as follows. Let’s fix a basepoint $z_0=(0, y_0) \in \{ 0 \} \times Y \subset Z$. Then the gauge group ${\mathcal{G}}$ acts on the fiber $E_{z_0}$, and acts freely on the product ${\mathcal{M}}(Z) \times E_{z_0}$. We define the based moduli space to be $$\tilde{{\mathcal{M}}}(Z):=\{ (A, v) \in {\mathcal{A}}\times E_{z_0} : F^+_A=0, \mathcal{E}(A) < \infty, \kappa(A)=0 \}/ {\mathcal{G}}.$$
Let $l=k-{1 \over 2} \geq {5 \over 2}$. Over the trivial ${\mathbb{C}}^2$-bundle $E' \to Y$, one can consider the space of $L^2_l$ $SU(2)$-connections ${\mathcal{A}}_l(Y)$ and the gauge group ${\mathcal{G}}_{l+1}(Y)$. The moduli space of $Y$ consists of equivalence classes of flat connections on $E'$: $${\mathcal{M}}(Y):= \{ B \in {\mathcal{A}}_{l}(Y): F_B=0 \}/ {\mathcal{G}}_{l+1}(Y).$$ Usually we use $B$ to represent a generic connection on $E'$, and $\Gamma$ a generic flat connection. Via the holonomy morphism the moduli space ${\mathcal{M}}(Y)$ is identified with the character variety $$\chi(Y):=\operatorname{Hom}(\pi_1(Y), SU(2)) /Ad.$$ Fixing a point $y_0 \in Y$, one can also define the based moduli space to be $$\tilde{{\mathcal{M}}}(Y):=\{ (B, v) \in {\mathcal{A}}(Y) \times E'_{y_0} : F_B=0\} / \mathcal{G}$$ Via the holonomy morphism the based moduli space is identified with the representation variety $$\mathcal{R}(Y):=\operatorname{Hom}(\pi_1(Y), SU(2)).$$ Over the cylindrical end $[0, \infty) \times Y$, the ASD equation is related to flat connections on $Y$ in the following way. Let’s fix a smooth flat connection $\Gamma_0$ on $E'$ as a reference connection. The Chern-Simons functional on ${\mathcal{A}}(Y)$ is $$\operatorname{cs}(B)=- \int_Y \operatorname{tr}({1 \over 2} b \wedge d_{\Gamma_0}b + { 1 \over 3} b \wedge b \wedge b),$$ where $B=\Gamma_0+b$ with $b \in L^2_l(T^*Y \otimes {\mathfrak{su}}(2))$. It’s straightforward to compute that the formal gradient and Hessian are given respectively by $$\operatorname{grad}\operatorname{cs}(B) = *F_B \text{ and } \operatorname{Hess}\operatorname{cs}|_B (b) =* d_B b.$$ Thus the critical points of the Chern-Simons functional consists of flat connections on $Y$. The restricted Hessian $*d_B|_{\ker d_B^*}$ has real, discrete, and unbounded spectrum (c.f. [@MMR Lemma 2.1.1]).
Given a gauge transformation $u \in {\mathcal{G}}(Y)$ and $B \in {\mathcal{A}}(Y)$, we have $$\operatorname{cs}(u\cdot B) - \operatorname{cs}(B) = -4\pi^2 \deg u.$$ Thus the Chern-Simons functional descends to an $S^1$-valued function on the quotient space ${\mathcal{B}}(Y)$.
Any connection $A$ on the cylindrical end $[0, \infty) \times Y$ has the form $$A= B(t) + \beta(t)dt,$$ where $\beta(t) \in L^2_l(Y, {\mathfrak{su}}(2))$ is a time-dependent ${\mathfrak{su}}(2)$-valued $0$-form on $Y$. Then $$F^+_A={1 \over 2}(*(*F_{B}+\dot{B}-d_B\beta) + dt \wedge (*F_B+\dot{B} - d_B \beta)).$$ Thus the ASD equation on the cylindrical end reads as $$\label{asdeq}
\dot{B}=-*F_B+d_B \beta.$$ Thus up to gauge transformations the ASD equation on the cylindrical end is the downward gradient flow equation of the Chern-Simons functional. There are two useful methods to choose representatives for a gauge equivalence class. One way is to put $A|_{[0, \infty) \times Y}$ in temporal gauge, i.e. trivialize the bundle $E|_{[0, \infty) \times Y}$ via parallel transport of $A$ so that $A=B(t)+dt$. The other way is to restrict the analysis to a local slice given by a flat connection $\Gamma$ on $Y$. More precisely let’s fix a smooth flat connection $\Gamma \in {\mathcal{A}}(Y)$. Let $$\mathcal{K}_{\Gamma}:= \ker d_{\Gamma}^*\subset L^2_l(T^*Y \otimes {\mathfrak{su}}(2)) \text{ and } \mathcal{S}_{\Gamma}:= \Gamma+ \mathcal{K}_{\Gamma} \subset {\mathcal{A}}(Y).$$
We say a connection $A=B(t)+ \beta(t)dt$ on $[0, \infty) \times Y$ is in standard form with respect to $\Gamma$ if for all $t \in [0, \infty)$ one has $$B(t) \in \mathcal{S}_{\Gamma} \text{ and } \beta(t) \in (\ker \Delta_{\Gamma})^{\perp},$$ where $\Delta_{\Gamma}:=d^*_{\Gamma}d_{\Gamma}: L^2_l(Y, {\mathfrak{su}}(2)) \to L^2_{l-2}(Y, {\mathfrak{su}}(2))$ is the Laplacian twisted by $\Gamma$.
Let’s write $\operatorname{cs}_{\Gamma}:=\operatorname{cs}|_{\mathcal{S}_{\Gamma}}$ for the restriction on the Chern-Simons functional on the slice. Then we have the following result:
[[@MMR Lemma 2.5.1]]{}\[CPG\] Let $\Gamma$ be a smooth flat connection on $E'$. Then there exists a $\operatorname{Stab}(\Gamma)$-invariant neighborhood $U_{\Gamma}$ of $\Gamma$ in $\mathcal{S}_{\Gamma}$ and a unique smooth $\operatorname{Stab}(\Gamma)$-equivariant map $\Theta: U_{\Gamma} \to L^2_l(Y, {\mathfrak{su}}(2))$ such that for all $B \in U_{\Gamma}$ one has $$*F_B - d_B (\Theta(B)) \in \mathcal{K}_{\Gamma} \text{ and } \Theta(B) \in (\ker \Delta_{\Gamma})^{\perp}.$$ Furthermore the map $\Theta$ has the following properties. Let $B=\Gamma+b \in U_{\Gamma}$ be a connection in the slice neighborhood of $\Gamma$.
1. The formal gradient of $\operatorname{cs}_{\Gamma}$ at $B$ is $$\operatorname{grad}\operatorname{cs}_{\Gamma} (B) = -*F_B + d_B (\Theta(B)).$$
2. One has the bounds: $$\|\Theta(B)\|_{L^2_l} \leq \mathfrak{c}(l) \|b\|_{L^2_l} \|\operatorname{grad}\operatorname{cs}_{\Gamma} (B)\|_{L^2_{l-2}},$$ where $\mathfrak{c}(l) > 0$ is a constant only depending on $l$.
Only (ii) is different from Lemma 2.5.1 in [@MMR]. However the argument there goes through without any change. We note that $$d^*_{\Gamma}d_B(\Theta(B)) = d^*_{\Gamma}(*F_B) = *[b, F_B].$$ Due to the Sobolev multiplication $L^2_l \times L^2_j \to L^2_j$ for any $j \leq l$, we have $$\|d^*_{\Gamma}d_B(\Theta(B))\|_{L^2_{l-2}} \leq const. \|b\|_{L^2_l}\|F_B\|_{L^2_{l-2}}.$$ Note that $d^*_{\Gamma}d_B: L^2_{l} \cap (\ker \Delta_{\Gamma})^{\perp} \to L^2_{l-2}$ is uniformly invertible for $B$ close to $\Gamma$ in $L^2_l$ norm, we conclude that $$\| \Theta(B)\|_{L^2_l} \leq const. \|b\|_{L^2_l}\|F_B\|_{L^2_{l-2}}.$$ From (i) it follows that $$\|F_B\|_{L^2_{l-2}} \leq \|\operatorname{grad}\operatorname{cs}_{\Gamma}(B)\|_{L^2_{l-2}} + \|d_B(\Theta(B))\|_{L^2_{l-2}}.$$ Since $d_B(\Theta(B)) = d_{\Gamma} (\Theta(B)) + [b, \Theta(B)]$, the Sobolev multiplication tells us that $$\|d_B(\Theta(B))\|_{L^2_{l-2}} \leq const. (1+\|b\|_{L^2_l}) \| \Theta(B)\|_{L^2_l}.$$ In summary we have $$\| \Theta(B)\|_{L^2_l} \leq const. \|b\|_{L^2_l}\big(\|\operatorname{grad}\operatorname{cs}_{\Gamma}(B)\|_{L^2_{l-2}} +(1+\|b\|_{L^2_l}) \| \Theta(B)\|_{L^2_l} \big)$$ Thus when $\|b\|_{L^2_l}$ is small, which can be achieved by shrinking $U_{\Gamma}$, we get the desired estimate.
Uhlenbeck’s gauge fixing tells us that if the $L^2$-norm of the curvature $F_B$ is small, one can find a smooth flat connectoin $\Gamma$ such that $B \in \mathcal{S}_{\Gamma}$. Combining [@MMR Lemma 2.4.3] and the regularity result on ASD connections, we conclude that if the curvature of a connection $A$ on a cylinder $[t_1, t_2] \times Y$ is small, one can find a gauge tranformation to tranform $A$ into a standard form $B(t)+\beta(t) dt$ with respect to a flat connection $\Gamma$ on $E'$. From the ASD equation (\[asdeq\]), we see that $$d^*_{\Gamma} *F_B = d^*_{\Gamma}d_B \beta.$$ Thus by Lemma \[CPG\] $\beta(t) = \Theta(B(t))$ and the ASD equation reads as $$\dot{B}(t) = -\operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)).$$
Holonomy Perturbations
----------------------
To get transversality for moduli spaces, we need to introduce perturbations. Since the manifolds we are interested in have $b_+=0$, we cannot use merely metric perturbations as in [@MMR]. We adopt holonomy perturbations instead, following the lines in [@D87] and [@K04].
We start with an embedded ball $N \subset Z$ and a smooth map $q: S^1 \times N \to Z$ satisfying
1. $q$ is a submersion;
2. $q(1, x)=x$ for any $x \in N$.
Given $x \in N$, we denote by $\operatorname{Hol}_{A_x} \in SU(2)$ the holonomy of a connection $A$ around the loop $q(-, x)$. Denote by $\operatorname{\mathfrak{h}ol}_{A_x} \in {\mathfrak{su}}(2)$ the traceless part of $\operatorname{Hol}_{A_x}$. We then get a section $\operatorname{\mathfrak{h}ol}_A \in C^{\infty}(N, {\mathfrak{su}}(2))$. Let $\omega \in \Omega^+(Z)$ be a self-dual $2$-form supported on $N$. We form a ${\mathfrak{su}}(2)$-valued $2$-form $$V_{q, \omega}:=\omega \otimes \operatorname{\mathfrak{h}ol}_A \in \Omega^+(Z; {\mathfrak{su}}(2))$$ supported on $N \subset Z$. The key estimates of $V_{q, \omega}$ are derived in [@K04].
([@K04 Proposition 3.1]) \[ehp\] Given a submersion $q$, there exist constants $K_n$ such that for any $A \in {\mathcal{A}}_{k, loc}$, $\omega \in \Omega^+(Z)$ supported on $N$, one has $$\|D^n V_{q, \omega}|_A (a_1, ..., a_n)\|_{L^2_k(N)} \leq K_n \|\omega\|_{C^k} \prod_{i=1}^n \|a_i\|_{L^2_k(N)},$$ where $D^n V_{q, \omega}$ is the $n$-th differential of $V_{q, \omega}$.
In particular we get a smooth ${\mathcal{G}}_{k+1, loc}$-equivariant map $$V_{q, \omega}: {\mathcal{A}}_{k, loc} \longrightarrow L^2_{k, loc}(Z, \Lambda^+ \otimes {\mathfrak{su}}(2)).$$ Now we take a countable family of embedded balls $\{N_{\alpha}\}_{\alpha \in {\mathbb{N}}}$ in $Z$ together with submersions $\{q_{\alpha}\}_{\alpha \in {\mathbb{N}}}$ satisfying the condition that for any $x \in Z$, the countable family of maps $$\{q_{\alpha}(-, x) : \alpha \in {\mathbb{N}}, x \in \operatorname{Int}(N_{\alpha}) \}$$ is $C^1$-dense in the space of smooth loops in $Z$ based at $x$.
\[hped\] Let $\{q_{\alpha}\}$ be a family of submersions as above. Given $\mu >0$, the space $\mathcal{P}_{\mu}$ of holonomy perturbations consists of perturbations of the form $$\sigma_{\pmb{\omega}}= \sum_{\alpha} V_{q_{\alpha}, \omega_{\alpha}},$$ where $\pmb{\omega}=\{\omega_{\alpha}\}$ is a family of self-dual $2$-form supported on $N_{\alpha}$ satisfying the following.
1. Denote by $C_{\alpha}:=\sup \{K_{n, \alpha} : 0 \leq n \leq \alpha \}$, where $K_{n, \alpha}$ is the constant for $q_{\alpha}$ in Proposition \[ehp\]. Then $$\sum_{\alpha} C_{\alpha} \|\omega_{\alpha} \|_{C^k}$$ converges.
2. For $A \in {\mathcal{A}}_{k, loc}$, one has $$\|\nabla^j \sigma(A)|_{\{t\} \times Y}\|_{C^0} \leq C'_je^{-\mu t} \sum_{\alpha} C_{\alpha} \|\omega_{\alpha} \|_{C^k}, \; \; j \leq k,$$ where $C'_j > 0$ only depends on and $\{q_{\alpha} \}$ and $j$.
When there is no confusion, we simply write $\sigma$ for $\sigma_{\pmb{\omega}}$. We denote by $W_{\mu}$ the space of sequences $\pmb{\omega}=\{\omega_{\alpha}\}$ satisfying $(1)$ and $(2)$ in Definition \[hped\]. Note that $W$ forms a Banach space with respect to the norm $$\| \pmb{\omega}\|_W := \sum_{\alpha} C_{\alpha} \|\omega_{\alpha}\|_{C^k}.$$ Since $\sigma$ depends linearly on $\pmb{\omega}$, $\mathcal{P}_{\mu}$ is also a Banach space. Moreover each $\sigma \in \mathcal{P}_{\mu}$ gives rise to a smooth ${\mathcal{G}}_{k+1, loc}$-equivariant map $$\sigma: {\mathcal{A}}_{k, loc} \longrightarrow L^2_{k, \mu}(Z, \Lambda^+ \otimes {\mathfrak{su}}(2)).$$ Given $\sigma \in \mathcal{P}_{\mu}$, we define the perturbed moduli space of finite energy instantons to be $${\mathcal{M}}_{\sigma}(Z):= \{ A \in {\mathcal{A}}_{k, loc}: F^+_A = \sigma(A), \mathcal{E}(A) < \infty, \kappa(A)=0 \} / {\mathcal{G}}_{k+1, loc}.$$
The Asymptotic Map {#EAM}
==================
In this section we deduce the asymptotic behaviors of the perturbed moduli space ${\mathcal{M}}_{\sigma}(Z)$ by modifying the corresponding arguments in [@MMR].
Let $\Gamma$ be a smooth flat connection on $E'$. Denote by $\mu_{\Gamma}$ the smallest nonzero absolute value of eigenvalues of the restricted Hessian $*d_{\Gamma}|_{\ker d^*_{\Gamma}}$.
Since $\chi(Y)$ is compact, and $\mu_{\Gamma}$ is gauge invariant, we can fix a finite number $\mu > 0$ satisfying $$\mu \geq \max \{ \mu_{\Gamma}: \Gamma \text{ is a smooth flat connection} \}.$$
Given $A \in {\mathcal{A}}_{k, loc}$, we write $$\sigma(A)|_{\{t \} \times Y} = *\rho_A(t) +dt \wedge \rho_A(t),$$ where $\rho_A(t)$ is a ${\mathfrak{su}}(2)$-valued $1$-form on $Y$. Suppose $A|_{[0, \infty) \times Y} = B(t) +\beta(t) dt $ is in standard form with respect to $\Gamma$. Then the perturbed ASD equation restricted on the end $[0, \infty) \times Y$ reads as $$\dot{B}(t)=-*F_{B(t)} +d_B \beta(t) + 2\rho_A(t).$$ The decay condition in Definition \[hped\] for $\sigma_{\pmb{\omega}} \in \mathcal{P}_{\mu}$ implies $p_A(t)$ decays expoentially on the end as well: $$\|\rho_A(t)\|_{L^2_j(Y)} \leq \mathfrak{c}_0(j)\|\pmb{\omega}\|_W \cdot e^{-\mu t}, \; j \leq l,$$ where the constant $\mathfrak{c}_0>0$ depends on neither $A$ nor $\pmb{\omega}$.
Note that in the non-perturbed case, once $A$ is in standard form we identify $\beta(t) = \Theta(B(t))$, thus the equation has the form of the downward gradient flow equation. In the perturbed case, this is no longer true. However we still have estimates on how far $\beta(t)$ is from $\Theta(B(t))$.
\[dgl\] Suppose $A|_{[0, \infty)} =B(t) +\beta(t)dt$ is the restriction of a perturbed ASD connection $A$ on $Z$ such that $A$ is in standard form with respect to $\Gamma$ and $B(t) \in U_{\Gamma}$ as in Lemma \[CPG\] . Then $$\|\beta(t) - \Theta(B(t))\|_{L^2_j} \leq 2
_0(j-2) \|\pmb{\omega}\|_W e^{-\mu t}, \; j \leq l.$$
We note that the perturbed ASD equation gives us $$\dot{B} = -\operatorname{grad}\operatorname{cs}_{\Gamma}(B) + d_B\big(\beta(t) - \Theta(B(t)) \big) + 2\rho_A(t).$$ By construction $\beta(t) - \Theta(B(t)) \in (\ker \Delta_{\Gamma})^{\perp}$ and $\dot{B} + \operatorname{grad}\operatorname{cs}_{\Gamma}(B) \in \ker d^*_{\Gamma}$. Since $d^*_{\Gamma}d_B: L^2_{l} \cap (\ker \Delta_{\Gamma})^{\perp} \to L^2_{l-2}$ is uniformly invertible for $B$ close to $\Gamma$ in $L^2_l$ norm, we have $$\begin{split}
\|\beta(t) - \Theta(B(t))\|_{L^2_j} & \leq const. \|d^*_{\Gamma} d_B \beta(t) - \Theta(B(t))\|_{L^2_{j-2}} \\
& \leq const. \|p_A(t)\|_{L^2_{j-2}} \\
&\leq 2\mathfrak{c}_0(j-2) \|\pmb{\omega}\|_W e^{-\mu t}.
\end{split}$$
To simplify notations, we write $$p_A(t) := d_B \big( \beta(t) - \Theta(B(t)) \big) + 2\rho_A(t).$$ Then the perturbed ASD equation over the end in standard form reads as $$\label{pas}
\dot{B}(t)= - \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) + p_A(t),$$ where the extra pertubration term satisfies $$\label{pwc}
\|p_A(t)\|_{L^2(Y)} \leq \mathfrak{c}_0 \| \pmb{\omega}\|_W e^{-\mu t}$$ for some constant $\mathfrak{c}_0 > 0$ independent of $A$ and $\pmb{\omega}$.
To derive the convergence of a perturbed ASD connection $[A] \in {\mathcal{M}}_{\sigma}(Z)$, we start with a sequence of lemmas establishing the estimate for the length of the flowline $B(t)$ corresponding to $A$.
[[@MMR Proposition 4.2.1]]{}\[gcf\] Let $\Gamma$ be a smooth flat connection on $E'$. Then there exists a neighborhood $U_{\Gamma} \subset \mathcal{S}_{\Gamma}$ of $\Gamma$, and constant $\theta \in (0, {1 \over 2}]$ so that for any connection $B \in U_{\Gamma}$ one has $$|\operatorname{cs}(B)-\operatorname{cs}(\Gamma)|^{1- \theta} \leq \| \operatorname{grad}\operatorname{cs}_{\Gamma}(B) \|_{L^2(Y)}.$$
This is a Łojasiewicz type inequality in the infinite dimensional setting originally proved by Simon [@S83].
\[lis\] Let $\Gamma$, $U_{\Gamma}$, and $\theta$ be as in Lemma \[gcf\]. Let $A$ be a perturbed instanton on $E$ of the form in (\[pas\]) such that $B(t) \in U_{\Gamma}$ for $t \in [t_1, t_2]$. Suppose $$\label{cpd}
\| p_A(t) \|_{L^2} \leq \alpha \| \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \|_{L^2},$$ for some constant $\alpha \in (0, 1)$. Then $$\int_{t_1}^{t_2} \| \dot{B}(t) \|_{L^2} dt \leq {1 \over {\theta^2 \sqrt{1- \alpha^2}}} |\operatorname{cs}(B(t_1)) - \operatorname{cs}(B(t_2)) |^{\theta}.$$
From (\[pas\]) we get $$\begin{split}
\| \dot{B}(t) \|^2_{L^2} + \| \operatorname{grad}\operatorname{cs}_{\Gamma} (B(t)) \|^2_{L^2} & = -2 \langle \dot{B}(t), \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \rangle + \| p_A(t) \|^2_{L^2} \\
& \leq -2 \langle \dot{B}(t), \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \rangle + \alpha^2 \| \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \|^2_{L^2}.
\end{split}$$ Thus $$\begin{split}
2\sqrt{1- \alpha^2} \|\dot{B}(t)\| \cdot \|\operatorname{grad}\operatorname{cs}_{\Gamma} (B(t)) \| & \leq \| \dot{B}(t) \|^2 + (1 -\alpha^2) \| \operatorname{grad}\operatorname{cs}_{\Gamma} (B(t)) \|^2 \\
& \leq -2 \langle \dot{B}(t), \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \rangle \\
& = -2 {d \over dt} \operatorname{cs}(B(t)).
\end{split}$$ In particular we conclude that Chern-Simons functional $\operatorname{cs}$ is non-increasing along a path solving (\[pas\]). Let’s assume that $\operatorname{cs}(B(t_1)) >\operatorname{cs}(B(t_2) > \operatorname{cs}(\Gamma)$. The other cases can be proved similarly. Then applying Lemma \[gcf\] we get $$\begin{split}
-{d \over dt} (\operatorname{cs}(B(t)) - \operatorname{cs}(\Gamma))^{\theta} & \geq \theta \sqrt{1-\alpha^2} |\operatorname{cs}(B(t)) - \operatorname{cs}(\Gamma)|^{\theta -1} \|\dot{B}(t)\| \cdot \|\operatorname{grad}\operatorname{cs}_{\Gamma} (B(t)) \| \\
& \geq \theta \sqrt{1-\alpha^2} \|\dot{B}(t)\|.
\end{split}$$ Thus integrating both sides we get $$\begin{split}
\int_{t_1}^{t_2} \|\dot{B}(t)\| dt & \leq {1 \over {\theta \sqrt{1-\alpha^2}}} \big( (\operatorname{cs}(B(t_1)) -\operatorname{cs}(\Gamma))^{\theta} - (\operatorname{cs}(B(t_2)) -\operatorname{cs}(\Gamma))^{\theta} \big) \\
& \leq {1 \over {\theta^2 \sqrt{1-\alpha^2}}} |\operatorname{cs}(B(t_1)) - \operatorname{cs}(B(t_2))|^{\theta}.
\end{split}$$
Although the Chern-Simons functional $\operatorname{cs}$ do not necessarily decay along the path $B(t)$ corresponding to a perturbed ASD connection $A$, one can show after adding a term of exponential decay it’s decreasing.
\[l2.8\] Let $\Gamma$, $U_{\Gamma}$, and $\theta$ be as in Lemma \[gcf\]. Let $A$ be a perturbed instanton on $E$ of the form in (\[pas\]) such that $B(t) \in U_{\Gamma}$ for $t \in [t_1, t_2]$. Then the function $$\operatorname{cs}_A(t):= \operatorname{cs}(B(t)) + {\mathfrak{c}^2_0 \|\pmb{\omega}\|^2_W \over {2\mu}} e^{-2\mu t}$$ is non-increasing.
We compute that $$\begin{split}
- {d \over {dt}} \operatorname{cs}_A(t) & = -\langle \dot{B}(t), \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \rangle + \mathfrak{c}^2_0\|\pmb{\omega}\|^2_W e^{-2\mu t} \\
& \geq - \langle \dot{B}(t), \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \rangle + \|p_A(t)\|^2_{L^2} \\
& = {1 \over 2}(\| \dot{B}(t) \|^2_{L^2} + \| \operatorname{grad}\operatorname{cs}_{\Gamma} (B(t)) \|^2_{L^2}) \\
& \geq \| \dot{B}(t) \|_{L^2} \cdot \| \operatorname{grad}\operatorname{cs}_{\Gamma} (B(t)) \|_{L^2} \\
& \geq 0.
\end{split}$$
Now we give an estimate of the length of a perturbed flowline $B(t)$, which is based on discussion with Cliff Taubes.
\[lgf\] Let $\Gamma$, $U_{\Gamma}$, and $\theta$ be as in Lemma \[gcf\]. Let $A$ be a perturbed instanton on $E$ of the form in (\[pas\]) such that $B(t) \in U_{\Gamma}$ for $t \in [0, \infty)$. Then one can find $T_1 > 0$ such that
1. either $$\int_{T_1+1}^{\infty} \|\dot{B}(t)\| dt \leq \mathfrak{c}'_1 (\|F_A\|^{\theta}_{L^2([T_1+1, \infty) \times Y)} + e^{-2\mu \theta T_1}),$$ where $\mathfrak{c}'_1$ only depends on $\theta$ and $\mu$;
2. or $$\int_{T_1+1}^{\infty} \| \dot{B}(t) \|_{L^2} \leq \mathfrak{c}''_1 e^{- \mu \theta T_1},$$ where $\mathfrak{c}''_1$ only depends on $\mathfrak{c}_0$, $\|\pmb{\omega}\|_W$, $\mu$, $\theta$ and $\|F_A\|_{L^2([T_1+1, \infty) \times Y)}$.
We first explain how the time $T_1$ comes into the picture. By the standard elliptic theory for ASD connections, see for our case explicitly [@MMR Lemma 3.5.1], there is a positive number $\epsilon_0 >0$ such that whenever a perturbed ASD connection $A$ on $E$ satisfies $\|F_A \|_{L^2([T_1, \infty) \times Y)} < \epsilon_0$, one has $$\|F_{B(s)} \|^2_{L^2_1} \leq const. (\|F_A \|^2_{L^2([T_1, \infty) \times Y)} + e^{-2\mu s}), \text{ when } s \geq T_1+1.$$ The finiteness of $\|F_A\|_{L^2(Z)}$ guarantees us the existence of such a $T_1$.
Now consider the following two cases:
1. $\| p_A(t) \|_{L^2} \leq {1 \over 2} \| \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \|_{L^2}$.
2. $\| p_A(t) \|_{L^2} \geq {1 \over 2} \| \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) \|_{L^2}$.
If ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {1};}}$ holds for all $t \in [0, \infty)$, then Lemma \[lis\] tells us that for any $T' > T_1+1$ we have $$\begin{split}
\int_{T_1+1}^{T'} \|\dot{B}(t)\| dt & \leq const. |\operatorname{cs}(B(T_1) - \operatorname{cs}(B(T'))|^{\theta} \\
& =const. \Big({1 \over 2} \int_{T_1+1}^{T'} \int_Y \operatorname{tr}(F_A \wedge F_A) \Big)^{\theta}\\
& =const. \Big({1 \over 2} \int_{T_1+1}^{T'} \int_Y |F_A|^2 +2\operatorname{tr}(F_A^+ \wedge F_A^+) \Big)^{\theta} \\
& \leq const. (\|F_A\|^{\theta}_{L^2([T_1+1, T'] \times Y)} + e^{-2\mu \theta T_1}).
\end{split}$$ We conclude that $$\int_{T_1+1}^{\infty} \|\dot{B}(t)\| dt \leq \mathfrak{c}_1 (\|F_A\|^{\theta}_{L^2([T_1+1, \infty) \times Y)} + e^{-2\mu \theta T_1}).$$
If ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}}$ holds for all $t \in [0, \infty)$, then the exponential decay on both $p_A(t)$ and $\operatorname{grad}\operatorname{cs}_{\Gamma}(B(t))$ implies the reseult.
Now we may assume ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}}$ holds at $t=T_1$. Let $[a_0, b_0], ..., [a_n, b_n], ...$ be a sequence of intervals with integer end points such that
1. $a_0 > T_1$, and $a_i > b_{i-1}$ for $i \geq 1$.
2. ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {1};}}$ holds for all $t \in [a_i, b_i]$, $i \geq 0$.
3. For any $k$ with $b_i \leq k < a_{i+1}$, there exists $t_k \in [k, k+1]$ such that ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}}$ holds at $t=t_k$.
We need to estimate the the length of $B(t)$ over $[a_i, b_i]$ and $[b_i, a_{i+1}]$ respectively.\
Step 1. Let’s first consider the case over $[a_i, b_i]$ where ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {1};}}$ always holds. We choose $$t_{a_i} = \max \{ t \in [a_i - 1 , a_i] : {\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}} \text{ holds at } t \}$$ and $$t_{b_i} = \min \{t \in [b_i, b_i+1]: {\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}} \text{ holds at } t\}.$$ From Lemma \[lis\] and Lemma \[gcf\] we know that $$\begin{split}
\int_{t_{a_i}}^{t_{b_i}} \|\dot{B}(t) \| dt & \leq {2 \over {\sqrt{3}\theta}}(| \operatorname{cs}(B(a_i) - \operatorname{cs}(\Gamma) |^{\theta} + |\operatorname{cs}(B(b_i)) - \operatorname{cs}(\Gamma)|^{\theta}) \\
& \leq {2 \over {\sqrt{3}\theta}} (\| \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t_{a_i}) \|^{\theta \over {1-\theta}} + \| \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t_{b_i}) \|^{\theta \over {1-\theta}} ) \\
& \leq const. (e^{- \mu t_{a_i} \cdot {\theta \over{1- \theta}}}+ e^{- \mu t_{b_i} \cdot {\theta \over{1- \theta}}}) \\
& \leq const. e^{- {\mu \theta \over {1-\theta}} \cdot (a_i -1)}
\end{split}$$
[Step 2.]{} Next we consider the case over $[b_i, a_{i+1}]$. Let $k$ be an integer in $[b_i, a_{i+1})$, and $t_k \in [k, k+1]$ such that ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}}$ holds at $t=t_k$. Applying the Cauchy-Schwarz inequality we get $$\label{2.19}
\begin{split}
\int_{t_k}^{k+2} \|\dot{B}(t)\| dt & \leq \sqrt{2} \int_{t_k}^{k+2} \langle \dot{B}(t), -\operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) +p_A(t) \rangle dt \\
& \leq \sqrt{2} \Big(|\operatorname{cs}(B(t_k)) - \operatorname{cs}(B(k+2))|^{1 \over 2} \\
&+ |\int_{t_k}^{k+2} \langle -\operatorname{grad}\operatorname{cs}_{\Gamma}(B(t)) + p_A(t), p_A(t) \rangle dt|^{1 \over 2} \Big).
\end{split}$$ Corollary 2.5.2 in [@MMR] tells us that $$\| \operatorname{grad}\operatorname{cs}_{\Gamma}(B) \|_{L^2} \leq {4 \over 3}\|F_B\|_L^2,$$ when $B \in U_{\Gamma}$ (one may shrink $U_{\Gamma}$ in the first place to apply the result). We continue with (\[2.19\]) to get $$\begin{split}
\int_{t_k}^{k+2} \|\dot{B}(t)\| dt & \leq const. \Big(|\operatorname{cs}(B(t_k)) - \operatorname{cs}(\Gamma)|^{1 \over 2} + |\operatorname{cs}(B(k+2)) - \operatorname{cs}(\Gamma)|^{1 \over 2} +e^{-\mu t_k} \Big) \\
&\leq const. \Big(\|\operatorname{grad}\operatorname{cs}_{\Gamma}(B(t_k))\|^{1 \over {2- 2 \theta}} + e^{- {\mu \over {1 - \theta}}t_k} + e^{- \mu t_k} \Big) \\
&\leq const. (e^{-{\mu \over {2- 2 \theta}} t_k } + e^{- {\mu \over {1 - \theta}}t_k} + e^{-\mu t_k}),
\end{split}$$ where the second inequality uses Lemma \[l2.8\] to bound $\operatorname{cs}(B(k+2)) - \operatorname{cs}(\Gamma)$ via $\operatorname{cs}(B(t_k)) - \operatorname{cs}(\Gamma)$.
Combining Step 1 and Step 2 we conclude that if neither of ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {1};}}$ nor ${\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=2pt] (char) {2};}}$ holds for all $t \in [0, \infty)$, one has $$\begin{split}
\int_{T+1}^{\infty} \|\dot{B}(t) \| & \leq const. \sum_{n=N}^{\infty} e^{-{\mu \theta \over {1- \theta}} n} + e^{-{\mu \over {2- 2 \theta}} n } + e^{- {\mu \over {1 - \theta}} n} + e^{-\mu n} \\
& \leq \mathfrak{c}''_1 e^{- \mu \theta T},
\end{split}$$ where the constant $\mathfrak{c}''_1$ only depends on $\mathfrak{c}_0, \|\pmb{\omega}\|_W$, $\mu$, and $\|F_A\|_{L^2([T_1+1, \infty) \times Y)}$.
With the estimate of the length of the perturbed gradient flowlines, we are able to obtain the existence of the asymptotic map from ${\mathcal{M}}_{\sigma}(Z)$ to the character variety $\chi(Y)$.
\[eam\] Let $A$ be a finite energy perturbed ASD connection on the bundle $E \to Z$ of a manifold with cylindrical end $[0, \infty) \times Y$. Given $v \in E_{z_0}$ we get a path $v(t) \in E'_{y_0}$ by parallel transporting $v$ via $A$ along the path $[0, \infty) \times \{y_0\}$. Let $B(t)=A(t)|_{{t} \times Y}$. Then the path of equivalence classes $[(B(t), v(t)]$ has a limit $[B_o, v_o] \in \mathcal{R}(Y)$ . This defines a continuous $SU(2)$-equivariant map $$\tilde{\partial}_+: \tilde{{\mathcal{M}}}_{\sigma}(Z) \longrightarrow \mathcal{R}(Y).$$
This is the main result of Chapter 4 in [@MMR] where they have given a complete proof in the case when one adopts metric perturbations. We sketch the proof here and point out the modification we need in our case. For each flat connection $\Gamma$ on $E'$, one choose a neighborhood $U_{\Gamma} \subset \mathcal{S}_{\Gamma}$ such that Proposition \[lgf\] holds. Since $\mathcal{R}(Y)$ is compact, it follows from Uhlenbeck’s gauge fixing that one can choose $\epsilon_0 >0$ such that whenever $\|F_B\|_{L^2(Y)} < \epsilon_0$ for a connection $B$ on $E'$, one can find a gauge transformation $u$ such that $u\cdot B \in U_{\Gamma}$.
Given a perturbed ASD connection $A$ on $E$, we write $B(t)=A|_{Y_t}$. Since $A$ has finite energy, one can find $T_1$ and $\epsilon_1$ such that $\|F_A\|_{L^2([T_1, \infty) \times Y)} < \epsilon_1$ and $\|F_{B(T_1)}\|_{L^2(Y)} < \epsilon_0$ due to the elliptic theory of ASD connections as mentioned in the proof of Proposition \[lgf\]. Thus $B(T_1) \in U_{\Gamma}$ for some flat connection $\Gamma$. Following the proof of Proposition 4.3.1 in [@MMR], the length estimate in Proposition \[lgf\] implies that after gauge transformations $A|_{[T_1, \infty) \times Y}$ is in standard form with respect to $\Gamma$ and $B(t) \in U_{\Gamma}$ for all $t >T_1$ by possibly choosing larger $T_1$ and smaller $\epsilon_1$. Due to the non-increasingness of the modified function $\operatorname{cs}_A(t)$, we know $\lim_{t \to \infty} \operatorname{cs}(B(t))$ exists. Again we can shrink $U_{\Gamma}$ if necessary so that $U_{\Gamma} \cap U_{\Gamma'} =\varnothing$ if $\operatorname{cs}(\Gamma)=\operatorname{cs}(\Gamma')$ and $[\Gamma]$, $[\Gamma']$ are in different component. Thus the limit set of $[B(t)]$ is contained in a single component of $\chi(Y)$. After gauge transformations, $B(t)$ now has finite length, which implies that the limit set has to be a point. In this way we get a map $$\partial_+: {\mathcal{M}}_{\sigma}(Z) \longrightarrow \chi(Y).$$ The continuity of this map follows from the argument in [@MMR Page 71] verbatimly. The proof of the based version $\tilde{\partial}_+$ is the same as that of the original version in [@MMR Theorem 4.6.1] once the existenc of $\partial_+$ is established.
After establishing the existence of the asymptotic map, we can apply the Uhlenbeck’s compactness argument to the perturbed ASD connections as in [@K04]. Since the bundle $E$ in our case is trivial, there is no bubbling nor energy escape. Thus we obtain the compactness of ${\mathcal{M}}_{\sigma}(Z)$.
The perturbed moduli space ${\mathcal{M}}_{\sigma}(Z)$ is compact.
To study the behavior of a perturbed flowline $[B(t)]$ given by an instanton $[A]$ asymtotic to a limit $[\Gamma_A] \in \chi(Y)$, we recall the notion of center manifolds in [@MMR Definition 5.1.2].
Let $H=H_0 \oplus H^{\perp}_0$ be an orthogonal decomposition of a Hilbert space $H$ with $H_0$ a finite dimensional subspace. Let $U \subset H$ be a neighborhood of $0$ in $H$, and $\nu: U \to H$ a vector field over $U$. A $C^k$-center manifold for the pair $(U, \nu)$ is a submanifold $\mathcal{H} \subset H$ given by the graph of a $C^k$-map $f: U_0 \to H^{\perp}_0$, where $U_0 \subset H_0$ is a neighborhood of $0$ in $H_0$ satisfying the following conditions:\
(1) $\mathcal{H} \subset U$, and $T_0 \mathcal{H} = H_0$,\
(2) $\nu_{(x, f(x))} \in T_{(x, f(x))} \mathcal{H}$ for any $x \in U_0$,\
(3) $\operatorname{Crit}(\nu) \cap U' \subset \mathcal{H}$, where $U' \subset U$ is a smaller open neighborhood of $0$ in $H$. Here $Crit(\nu)$ is a set of critical points of the vector field $\nu$.
Roughly speaking a center manifold is a finite-dimensional submanifold that is locally preserved by the flow of $\nu$ and contains all nearby critical points. Now we take $\Gamma$ to be a smooth flat connection on $Y$ with a neighborhood $U_{\Gamma}$ in $\mathcal{S}_{\Gamma}$ satisfying Lemma \[CPG\]. We take $U \subset \mathcal{K}_{\Gamma}$ so that $U_{\Gamma}=\{\Gamma\} + U$. The deformation complex at $\Gamma$ is given by $$\label{3cx}
L^2_{l+1}({\mathfrak{su}}(2)) \xrightarrow{d_{\Gamma}} L^2_l(T^*Y \otimes {\mathfrak{su}}(2)) \xrightarrow{d_{\Gamma}} L_{l-1}^2(\Lambda^2 T^*Y \otimes {\mathfrak{su}}(2)).$$ We identify $H^1(Y; \operatorname{ad}\Gamma)=\ker d_{\Gamma} \cap \ker d^*_{\Gamma}$ as a subspace in $\mathcal{K}_{\Gamma}$. Let $H^{\perp}_{\Gamma} \subset \mathcal{K}_{\Gamma}$ be the $L^2$-orthogonal complement of $H^1(Y; \operatorname{ad}\Gamma)$. Corollary 5.1.4 in [@MMR] ensures the existence of a $\operatorname{Stab}(\Gamma)$-invariant $C^2$-center manifold $\mathcal{H}_{\Gamma}$ for the pair $(U, -\operatorname{grad}\operatorname{cs}_{\Gamma})$. We write $W^s_{\Gamma} \subset \mathcal{H}_{\Gamma}$ for the stable set of $-\operatorname{grad}\operatorname{cs}_{\Gamma}$ on the center manifold, i.e. for any $B \in W^s_{\Gamma}$ the flowline of $-\operatorname{grad}\operatorname{cs}_{\Gamma}$ starting at $B$ converges to some point in $\mathcal{H}_{\Gamma}$. The most important property of the center manfiold is that any perturbed ASD connection on the end can be approximated exponentially closely by a gradient flowline on the center manifold after a sufficiently long time.
\[EPA\] Let $\mathcal{H}_{\Gamma}$ be a center manifold of $\Gamma$ with respect to $(U, \operatorname{grad}\operatorname{cs}_{\Gamma})$. Let $[A] \in {\mathcal{M}}_{\sigma_{\pmb{\omega}}}(Z)$ with $\|\pmb{\omega}\|_W \leq 1$ such that $A=B(t) + \beta(t)dt$ is in standard form with respect to $\Gamma$ over the end $[0, \infty) \times Y$. Then there exists a neighborhood $V_{\Gamma} \subset U_{\Gamma}$ of $\Gamma$, and positive constants $T_{\Gamma}, \epsilon_{\Gamma},\kappa_{\Gamma}>0$ so that the following are satisfied.
1. $B(t) \in V_{\Gamma}$ for all $t \geq T_{\Gamma} -1$.
2. $\|F_A|_{[T_{\Gamma}-1, \infty) \times Y}\|_{L^2} < \epsilon_{\Gamma}.$
3. There is a unique downward gradient flowline, $B_{\Gamma}: [T_{\Gamma}-1, \infty) \to \mathcal{H}_{\Gamma}$, of $\operatorname{cs}_{\Gamma}$ on the center manifold of the following property. The ASD connection $A_{\Gamma}=B_{\Gamma}(t) + \Theta(B_{\Gamma}(t)) dt$ induced from $B_{\Gamma}$ satisfies $$\|A - A_{\Gamma}\|_{L^2_k([t-{1 \over 2}, t+{1 \over 2}] \times Y)} < \kappa_{\Gamma} e^{-{\mu_{\Gamma} \over 2}(t-T_{\Gamma})}, \; \forall t \geq T_{\Gamma},$$ where $\mu_{\Gamma}$ is the smallest nonzero absolute value of eigenvalues of the Hessian $*d_{\Gamma}|_{\ker d^*_{\Gamma}}$.
This is Theorem 5.2.2 in [@MMR] when one adopts metric perturbations. Its proof carries through our case without any change. From the finite length of $B(t)$ and finite energy of $A$, (1) and (2) follow immediately. Over $[T_{\Gamma}-1, \infty) \times Y$, we decompose $A=\Gamma+b(t)+c(t)+\beta(t)dt$, where $b(t) \in \mathcal{H}_{\Gamma}$, $c(t) \in H_{\Gamma}^{\perp}$. [@MMR Lemma 5.4.1] tells us that $$\|c(t)\|_{L^2} \leq const. e^{-{\mu_{\Gamma} \over 2}(t-T_{\Gamma})}.$$ From (\[pas\]) and $\mu > {\mu_{\Gamma} \over 2}$, we conclude that $$\| \dot{b}(t) + \operatorname{grad}\operatorname{cs}_{\Gamma}(B(t))\|_{L^2} \leq const. e^{-{\mu_{\Gamma} \over 2}(t-T_{\Gamma})}$$ Then [@MMR Lemma 5.3.1] gives us the unique gradient flowline $B_{\Gamma}=\Gamma + b_{\Gamma}: [T_{\Gamma} -1, \infty) \to \mathcal{H}_{\Gamma}$ such that $$\| b_{\Gamma}(t) - b(t) \|_{L^2} \leq const. e^{-{\mu_{\Gamma} \over 2}(t-T_{\Gamma})}.$$ From Lemma \[CPG\] and Lemma \[dgl\], we get $$\begin{split}
\|\beta(t) - \Theta(B_{\Gamma}(t))\|_{L^2} & \leq \|\beta(t) - \Theta(B(t))\|_{L^2} + \|\Theta(B) - \Theta(B_{\Gamma})\|_{L^2} \\
& \leq const. e^{-{\mu_{\Gamma} \over 2}(t-T_{\Gamma})}.
\end{split}$$ Now the result follows from the standard bootstrapping argument, see [@MMR Lemma 3.3.2] for our particular case.
\[rt0\] Note that $T_{\Gamma}, \epsilon_{\Gamma}, \kappa_{\Gamma}$ depend continuously on $[\Gamma]$ and $[A]$. The compactness of $\chi(Y)$ and ${\mathcal{M}}_{\sigma}(Z)$ implies that those parameters can be chosen uniformly, which we denote by $T_0, \epsilon_0$, and $\kappa_0$ respectively. We also choose $V_{\Gamma}$ for each $\Gamma$ uniform to all $[A] \in {\mathcal{M}}_{\sigma_{\pmb{\omega}}}(Z)$ with $\|\pmb{\omega}\|_W \leq 1$. However there is certain dependence among $V_{\Gamma}, T_0$, and $\epsilon_0$. Choosing $\epsilon_0$ smaller forces $T_0$ larger, which in turn enables us to shrink $V_{\Gamma}$ to ensure the estimates hold. In this way, we can choose three neighborhoods $V_{\Gamma} \subset V'_{\Gamma} \subset U_{\Gamma}$ so that all perturbed ASD connections $[A]$ would enter $V_{\Gamma}$ at $T_0$ and stay within $V'_{\Gamma}$ ever since for some $\Gamma$. Meanwhile the center manifold $\mathcal{H}_{\Gamma}$ is defined inside $U_{\Gamma}$.
We write ${\mathcal{M}}_{\sigma}(Z, V_{\Gamma})$ for the set of equivalence classes $[A]$ of ASD connections on $Z$ for which one can pick up a gauge class representative $A$ whose restriction on the end $[T_0, \infty) \times Y$ satisfies Proposition \[EPA\]. Then the assignment $$\begin{split}
Q_{\Gamma}: {\mathcal{M}}_{\sigma}(Z, V_{\Gamma}) & \longrightarrow W^s_{\Gamma} \\
[A] & \longmapsto B_{\Gamma}(T_0)
\end{split}$$ defines a continuous map following the same argument as in [@MMR Proposition 5.2.2]. Note that any two gauge transformations transforming the restriction of $A$ on the end into a connection of standard form with respect to $\Gamma$ differ by a constant gauge transformation in $\operatorname{Stab}(\Gamma)$, thus the map $Q_{\Gamma}$ is well-defined. The map $Q_{\Gamma}$ refines the asymptotic map $\partial_+$ in the sense that $\partial_+$ is the composition of $Q_{\Gamma}$ with the map sending $B_{\Gamma}(T_0)$ to the limit point in $V_{\Gamma}$ following the downward gradient flowline of $\operatorname{cs}_{\Gamma}$. We also have the based version $$\tilde{Q}_{\Gamma}: \tilde{{\mathcal{M}}}_{\sigma}(Z, V_{\Gamma}) \to W^s_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}.$$
Transversality on the Irreducible Moduli Space {#tims}
==============================================
As considered by Morgan-Mrowka-Ruberman in [@MMR], to improve the regularity of the map $Q_{\Gamma}$ one can first embed the moduli space $\tilde{{\mathcal{M}}}_{\sigma}(Z, V_{\Gamma})$ into a larger one which they refer to as a thickened moduli space, then prove transversality results there.
The thickening moduli space is defined with the help of thickening data about smooth flat connections on $E'$, which we now recall from [@MMR]. The motivation for introducing the thickening data is to resolve the issue that the gradient vector field $\operatorname{grad}\operatorname{cs}_{\Gamma}$ is incomplete. So one artificially truncates this vector field via a cut-off function. In this way all the local properties near $\Gamma$ are preserved, and one can apply analysis tools without worrying about the incompleteness.
Let $\Gamma$ be a smooth flat connection on $E'$. We choose
1. a center manifold $\mathcal{H}_{\Gamma}$ of $\Gamma$,
2. neighborhoods $V_{\Gamma} \subset V_{\Gamma}' \subset U_{\Gamma}$ as in Remark \[rt0\],
3. a cut-off function $\varphi_{\Gamma}: \mathcal{H}_{\Gamma} \to [0,1]$ such that $\varphi_{\Gamma} \equiv 1$ on $V'_{\Gamma} \cap \mathcal{H}_{\Gamma}$ and $\operatorname{supp}\varphi_{\Gamma} \subset U'_{\Gamma} \cap \mathcal{H}_{\Gamma}$ for some $U'_{\Gamma} \subset U_{\Gamma}$.
We refer to the triple $\mathcal{T}_{\Gamma} = (\mathcal{H}_{\Gamma}, V_{\Gamma}, \varphi_{\Gamma})$ as a set of thickening data about $\Gamma$.
Given a thickening triple $\mathcal{T}_{\Gamma}$, we write $$\mathcal{H}_{\Gamma}^{out} := \varphi_{\Gamma}^{-1}(0,1] \text{ and } \mathcal{H}_{\Gamma}^{in}:= \varphi_{\Gamma}^{-1}(1).$$ We denote by $$\Xi^{tr}_{\Gamma} :=- \varphi_{\Gamma} \cdot \operatorname{grad}\operatorname{cs}_{\Gamma}|_{\mathcal{H}_{\Gamma}}$$ the truncated downward gradient vector field over the center manifold $\mathcal{H}_{\Gamma}$. Then $\Xi^{tr}_{\Gamma}$ is a complete vector field over $\mathcal{H}_{\Gamma}$ despite that $\mathcal{H}_{\Gamma}$ is only defined near $\Gamma$. For each $h \in \mathcal{H}^{out}_{\Gamma}$, we let $B_h:[T_0, \infty) \to \mathcal{H}_{\Gamma}$ be the unique flowline of $\Xi^{tr}_{\Gamma}$ such that $B_h(T_0)=h$. Now we extend the connection $B_h(t) + \Theta(B_h(t))dt$ smoothly to a connection $A_h$ on the entire manifold $Z$ so that over the compact part $A_h|_{Z \backslash [T_0, \infty) \times Y}$ depends on $h$ smoothly. To put the weighted Sobolev space into the package, we choose a weight $\delta_{\Gamma} \in (0, {\mu_{\Gamma} \over 2} )$ for each $[\Gamma] \in \chi(Y)$.
\[d2.17\] Given $h \in \mathcal{H}_{\Gamma}^{out}$, we write $${\mathcal{A}}_{k, \delta_{\Gamma} }(Z, \mathcal{T}_{\Gamma}, h):=\{ A \in {\mathcal{A}}_{k, loc}(Z) : A-A_h \in L^2_{k, \delta_{\Gamma}}(Z, T^*Z \otimes {\mathfrak{su}}(2))\}.$$ We denote the union by $${\mathcal{A}}_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}) := \bigcup_{h \in \mathcal{H}_{\Gamma}^{out}} {\mathcal{A}}_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}, h).$$ The gauge group ${\mathcal{G}}_{k+1, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma})$ that preserves ${\mathcal{A}}_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma})$ consists of all $L^2_{k+1, loc}$ gauge transformations $u$ such that there exists $\tau \in \operatorname{Stab}(\Gamma)$ satisfying $$u|_{[T_0, \infty) \times Y} \circ \tau - \operatorname{id}\in L^2_{k+1, \delta_{\Gamma}}([T_0, \infty) \times Y, SU(2)).$$ Finally we pick a cut-off function $\varphi: Z \to [0, 1]$ such that $\varphi|_{[T_0+1, \infty) \times Y} \equiv 1$ and $\varphi|_{[0, T_0] \times Y} \equiv 0$. The thickened moduli space with respect to the thickening data $\mathcal{T}_{\Gamma}$ perturbed by $\sigma \in \mathcal{P}_{\mu}$ is defined to be $${\mathcal{M}}_{\sigma}(Z, \mathcal{T}_{\Gamma}):= \{ A \in {\mathcal{A}}_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}) : F^+_A - \varphi F^+_{A_h} = \sigma(A), \kappa(A)=0 \} / {\mathcal{G}}_{k+1, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}).$$ The thickened based moduli space $\tilde{{\mathcal{M}}}_{\sigma}(Z, \mathcal{T}_{\Gamma})$ is defined similarly.
For the rest of this section, we are concerned with irreducible connections. The construction of the thickened moduli space gives us a map $$\begin{split}
P_{\Gamma}: {\mathcal{M}}^*_{\sigma}(Z, \mathcal{T}_{\Gamma}) & \longrightarrow \mathcal{H}_{\Gamma} \\
[A] & \longmapsto h,
\end{split}$$ where $h$ is the element specified in Definition \[d2.17\]. We also have the based version $$\tilde{P}_{\Gamma}: \tilde{{\mathcal{M}}}^*_{\sigma}(Z, \mathcal{T}_{\Gamma}) \to \mathcal{H}_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}.$$ Note that the perturbations $\sigma: {\mathcal{A}}_{k, loc} \to L^2_{k, \mu}$ are smooth maps. Following from [@MMR Section 7.3] the based thickened moduli space $\tilde{{\mathcal{M}}}^*_{\sigma}(Z, \mathcal{T}_{\Gamma})$ is a $C^2$-manifold and $\tilde{P}_{\Gamma}$ is a $C^2$-map. From Proposition \[EPA\], for $\|\pmb{\omega}\|_W \leq 1$ we have an identification $$\tilde{{\mathcal{M}}}^*_{\sigma_{\pmb{\omega}}}(Z, V_{\Gamma}) \simeq \tilde{P}_{\Gamma}^{-1} (W^s_{\Gamma} \cap V_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}).$$ Denote the embedding by $j: \tilde{{\mathcal{M}}}_{\sigma}(Z, V_{\Gamma}) \hookrightarrow \tilde{M}_{\sigma}(Z, \mathcal{T}_{\Gamma})$. We obtain the following commutative diagram: $$\label{ic}
\begin{tikzcd}
\tilde{{\mathcal{M}}}^*_{\sigma}(Z, V_{\Gamma}) \arrow[hookrightarrow]{r}{j} \ar[d, "\tilde{Q}_{\Gamma}"] & \tilde{{\mathcal{M}}}^*_{\sigma}(Z, \mathcal{T}_{\Gamma}) \ar[d, "\tilde{P}_{\Gamma}"] \\
W^s_{\Gamma}\times_{\operatorname{Stab}{\Gamma}} E'_{y_0} \ar[hookrightarrow]{r} & \mathcal{H}_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}
\end{tikzcd}$$
As mentioned above, we introduce the thickened moduli space mainly to establish the transversality result. The following result is a variance of [@MMR Theorem 9.0.1] in our case.
\[TSC\] The map $\tilde{P}_{\Gamma}$ is transverse to any finite set of smooth submanifold in $\mathcal{H}_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}$ with respect to a generic perturbation $\sigma$. Moreover the dimension of the based thickened irreducible moduli space $\tilde{{\mathcal{M}}}^*_{\sigma}(Z, \mathcal{T}_{\Gamma})$ is given by $$\label{dimm}
-{3 \over 2}(\chi(Z)+\sigma(Z))+{h^1_{\Gamma}- h^0_{\Gamma} \over 2}+{\rho(\Gamma) \over 2}+ 3,$$ where $h^i_{\Gamma}=\dim H^i(Y; \operatorname{ad}\Gamma)$, $i=0, 1$, $\rho(\Gamma)$ is the Atiyah-Potadi-Singer $\rho$-invariant of the odd signature operator twisted by $\Gamma$ in [@APS2], $\chi(Z)$ is the Euler characteristic, and $\sigma(Z)$ is the signature.
It only remains to show the transversality of $\tilde{P}_{\Gamma}$. The computation of the formal dimension is the same as that in [@MMR Chapter 8]. Consider the map $$\begin{split}
\mathcal{F}: \mathcal{P}_{\mu} \times {\mathcal{A}}^*_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}) &\longmapsto L^2_{k-1, \delta_{\Gamma}}(\Lambda^+T^*Z \otimes {\mathfrak{su}}(2)) \\
(\sigma_{\pmb{\omega}}, A) & \longmapsto F^+_A - \varphi F^+_{A_h} - \sigma_{\pmb{\omega}}(A).
\end{split}$$ Denote by $\mathcal{F}_h:=\mathcal{F}|_{\mathcal{P}_{\mu} \times {\mathcal{A}}_{k, \mu}(Z, \mathcal{T}_{\Gamma}, h)}$ the restricited map. Then the differential of $\mathcal{F}_h$ is $$D\mathcal{F}_h|_{(\sigma_{\pmb{\omega}}, A)}(\pmb{\nu}, a) = d^+_Aa - D\sigma_{\pmb{\omega}}|_A a - \sigma_{\pmb{\nu}}(A),$$ where $a \in L^2_{k,\mu}(T^*Z \otimes {\mathfrak{su}}(2))$, $\pmb{\nu} \in W_{\mu}$. Recall that the perturbation has the form $\sigma_{\pmb{\nu}}(A)=\sum_{\alpha} V_{q_{\alpha, \nu_{\alpha}}}(A)$, where $q_{\alpha}$ gives us a dense family of loops in the loop space of $Z$. As in [@K04 Lemma 13], the irreducibility of $A$ implies the image of $\sigma_{\pmb{\nu}}(A)$ is dense as we vary $\pmb{\nu}$. On the other hand, the operator $$d_A^+ \oplus e^{-\tau \delta_{\Gamma}}d^*_A e^{\tau \delta_{\Gamma}}: L^2_{k, \delta_{\Gamma}}(T^*Z \otimes {\mathfrak{su}}(2)) \to L^2_{k-1, \delta_{\Gamma}}(\Lambda^+T^*Z \otimes {\mathfrak{su}}(2)) \oplus L^2_{k-1, \delta_{\Gamma}}({\mathfrak{su}}(2)),$$ where $\tau: Z \to {\mathbb{R}}$ is a smooth function such that $\tau|_{\{t\} \times Y} =t$, is Fredholm except at a discrete set of ${\mathbb{R}}$ from the theory of Atiyah-Patodi-Singer [@APS1]. We may choose $\delta_{\Gamma}$ in the first place to make this operator Fredholm. In particular we see that the image of $d^+_A$ is closed and has finite dimensional cokernel. Thus we conclude that $\mathcal{F}_h$ is a submersion.
Consider the map $$\tilde{P}_{\Gamma}': \mathcal{P}_{\mu} \times \big( {\mathcal{A}}^*_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}) \times_{{\mathcal{G}}} E_{z_0} \big) \longrightarrow \mathcal{H}_{\Gamma} \times_{\operatorname{Stab}(\Gamma)} E'_{y_0}$$ sending $(\sigma, [A, v])$ to the pair $[h, v']$, where $v'$ is the limit of $v$ under the parallel transport by $A$. By the construction of ${\mathcal{A}}^*_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma})$, the map $\tilde{P}_{\Gamma}'$ is a submersion. Since $\mathcal{F}$ is gauge equivariant and $\mathcal{F}_h$ is a submersion, we conclude that the restricted map $$\tilde{P}'_{\Gamma}: \mathcal{F}^{-1}(0) \times_{{\mathcal{G}}} E_{z_0} \longmapsto \mathcal{H}_{\Gamma} \times_{\operatorname{Stab}(\Gamma)} E'_{y_0}$$ is also a submersion. Denote by $\Pi: \mathcal{P}_{\mu} \times \big( {\mathcal{A}}^*_{k, \delta_{\Gamma}}(Z, \mathcal{T}_{\Gamma}) \times_{{\mathcal{G}}} E_{z_0} \big) \to \mathcal{P}_{\mu}$ the projection onto the first factor. Then the Sard-Smale theorem tells us that $\tilde{P}_{\Gamma}=\tilde{P}'_{\Gamma}|_{\Pi^{-1}(\sigma)}$ is transeverse to a given smooth submanifold for a generic perturbation $\sigma$.
Before extracting more information of the asymptotic map from Proposition \[TSC\], let’s first recall the Kuranishi obstruction map at a flat connection $\Gamma$.
[@MMR Theorem 12.1.1]\[Kur\] Let $\Gamma$ be a $C^{\infty}$ flat connection on $Y$. Then there exists a $\operatorname{Stab}(\Gamma)$-invariant neighborhood $V$ of $0$ in $H^1(Y; \operatorname{ad}\Gamma)$, a $\operatorname{Stab}(\Gamma)$-invariant neighborhood $U$ of $\Gamma$ in $\mathcal{S}_{\Gamma}$, and $C^{\infty}$ $\operatorname{Stab}(\Gamma)$-equivariant maps $$\mathfrak{p}_{\Gamma}: V \to U \text{ and }
\mathfrak{o}_{\Gamma}: V \to H^2(Y; \operatorname{ad}\Gamma)$$ satisfying
1. $\mathfrak{p}_{\Gamma}$ is an embedding whose differential at $0$ is the inclusion $H^1(Y; \operatorname{ad}\Gamma) \hookrightarrow \mathcal{K}_{\Gamma}$.
2. The restriction of $\mathfrak{p}_{\Gamma}|_{\mathfrak{o}_{\Gamma}^{-1}(0)}$ is a homeomorpshim onto the space of flat connections in $U$.
\[RKur\] Roughly speaking, the proof of Theorem \[Kur\] makes use of the implicit function theorem to obtain a map $$\mathfrak{q}_{\Gamma}: V \to \operatorname{im}d^*_{\Gamma} \subset L^2_1(T^*Y \otimes {\mathfrak{su}}(2)),$$ which is characterized by the fact that $$\Pi'_{\Gamma} F_{\Gamma+b+\mathfrak{q}_{\Gamma}(b)}=0,$$ where $\Pi'_{\Gamma}: L^2_{l-1}(\Lambda^2T^*Y \otimes {\mathfrak{su}}(2)) \to \operatorname{im}d_{\Gamma}$ is the $L^2$-orthogonal projection. Then $$\mathfrak{p}_{\Gamma}(b)=\Gamma+b+\mathfrak{q}_{\Gamma}(b) \text{ and }
\mathfrak{o}_{\Gamma}(b)=\Pi_{\Gamma} F_{\mathfrak{p}_{\Gamma}(b)},$$ where $\Pi_{\Gamma}: L^2_{l-1}(\Lambda^2T^*Y \otimes {\mathfrak{su}}(2)) \to \ker d_{\Gamma}$ is the $L^2$-orthogonal projection. The zero set of map $\mathfrak{o}_{\Gamma}$ provides a local structure of the character variety $\chi(Y)$ near $[\Gamma]$.
\[smooth\] For a flat connection $\Gamma$ on $Y$, the map $\mathfrak{o}_{\Gamma}$ in Theorem \[Kur\] is called the Kuranishi obstruction map. $[\Gamma] \in \chi(Y)$ is said to be a smooth point if the Kuranishi map of $\Gamma$ vanishes on $V_{\Gamma}$, i.e. $\mathfrak{o}_{\Gamma}\equiv 0$, otherwise a singular point.
[@MMR Corollary 9.3.1]\[TIS\] Let $[\Gamma] \in \chi(Y)$ be a smooth point. Then there exists a center manifold $\mathcal{H}_{\Gamma}$ consisting of flat connections. Moreover for a generic perturbation $\sigma$ the asymptotic map $$\tilde{\partial}_+: \tilde{M}_{\sigma}(Z, V_{\Gamma}) \to \mathcal{R}(Y)$$ is $C^2$ and tranverse to a given submanifold in its range.
When $[\Gamma] \in \chi(Y)$ is a smooth point, one can take a center manifold $\mathcal{H}_{\Gamma}$ to be the graph of the map $\mathfrak{q}_{\Gamma}$. Indeed in this case $\mathcal{H}_{\Gamma}$ consists of flat connections which are the critical points of $\operatorname{cs}_{\Gamma}$ near $\Gamma$, and are preserved by the gradient flow of $\operatorname{cs}_{\Gamma}$. Since every point on the center manifold $\mathcal{H}_{\Gamma}$ is a critical point, all gradient flowlines on $\mathcal{H}_{\Gamma}$ is constant. Thus the map $\tilde{Q}_{\Gamma}$ coincides with the asymptotic map $\partial^0_+$. Then result now follows from Proposition \[TSC\].
Now we restrict our attention to the case when $Y=T^3$. For each singular point in $\chi(T^3)$, Gompf and Mrowka [@GM] have constructed a center manifold. We recall their construction below and use it to prove that there are no irreducible ASD connections asymptotic to those singular points.
The character variety $\chi(T^3)$ is identified as a copy of the quotient $T^3/ \sim$, where $\sim$ is given by the hypoellipticinvolution consisting of $8$ fixed points corresponding to central connections . When $[\Gamma] \in \chi(Y)$ is a noncentral connection, the first homology $H^1(T^3; \operatorname{ad}\Gamma)$ is computed as $$H^1(T^3; \operatorname{ad}\Gamma) \cong \mathcal{H}^1(T^3) \otimes H^0(T^3; \operatorname{ad}\Gamma),$$ where $\mathcal{H}^1(T^3)$ is the space of harmonic $1$-forms on $T^3$, $ H^0(T^3; \operatorname{ad}\Gamma) \cong i{\mathbb{R}}$ is the Lie algebra of the stabilizer of $\Gamma$. Thus each $b \in H^1(T^3; \operatorname{ad}\Gamma)$ gives a flat connection $\Gamma+b$ due to $b \wedge b =0$. This implies that $\mathfrak{q}_{\Gamma}(b)=0$, and then the Kuranishi map $\mathfrak{o}_{\Gamma}(b) =0$ with $b$ in a small neighborhood $V_{\Gamma}$ of $0$. We conclude that $\Gamma$ is a smooth point in the sense of Definition \[smooth\]. When $[\Gamma] \in \chi(Y)$ is a central connection, we have $$H^1(T^3; \operatorname{ad}\Gamma) \cong \mathcal{H}^1(T^3) \otimes {\mathfrak{su}}(2).$$ Now we fix an orthonormal frame $\{ e^1, e^2, e^3\}$ of $\mathcal{H}^1(T^3)$ with respect to the product metric. Then each $b \in H^1(T^3; \operatorname{ad}\Gamma)$ has the form $$b=\sum_i e^i \otimes X_i, \; X_i \in {\mathfrak{su}}(2).$$ Thus the curvature of $\Gamma+b$ has the form $$F_{\Gamma+b}={1 \over 2} \sum_{i, j} e^i \wedge e^j \otimes [X_i, X_j].$$ In particular $F_{\Gamma+b} \in \mathcal{H}^2(T^3)$. From Remark \[RKur\] we conclude that $\mathfrak{q}_{\Gamma}(b)=0$. Thus the Kuranishi map is $$\mathfrak{o}_{\Gamma}(b)={1 \over 2} \sum_{i, j} e^i \wedge e^j \otimes [X_i, X_j].$$
[@GM Proposition 15.2]\[STAB\] Let $\Gamma$ be a smooth central flat connection on $E' \to T^3$. Then $H^1(T^3; \operatorname{ad}\Gamma)$ is a center manifold of $\Gamma$. Moreover the stable manifold of the origin is given by $$\label{dims}
\begin{split}
W^0_{\Gamma}=\{& b= \sum_i e^i \otimes X_i \in \mathcal{H}^1(T^3) \otimes {\mathfrak{su}}(2): \\
&\|X_i\|=\|X_j\|, \langle X_i, X_j \rangle=0, \langle X_1, [X_2, X_3] \rangle \leq 0 \}.
\end{split}$$
\[asms\] Let $\Gamma$ be a central flat connection on the trivial $SU(2)$-bundle $E'$ over $T^3$. Then for a generic perturbation $\sigma$, one has $$\partial_+^{-1}([\Gamma]) \cap {\mathcal{M}}^*_{\sigma}(Z) =\varnothing,$$ where $Z=M \cup [0, \infty) \times T^3$ is a homology $D^2 \times T^2$.
From the commutative diagram (\[ic\]) and Proposition \[TSC\] we know that the map $$\tilde{P}_{\Gamma}: \tilde{{\mathcal{M}}}_{\sigma}^*(Z, \mathcal{T}_{\Gamma}) \to \mathcal{H}_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}$$ is transverse to the stable set $W^0_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0}$ for a generic perturbation $\sigma$. Moreover $\dim \tilde{{\mathcal{M}}}^*_{\sigma}(Z, \mathcal{T}_{\Gamma}) = 6$ from (\[dimm\]), and the stratified space $W^0_{\Gamma}$ has codimension $4$ in $\mathcal{H}_{\Gamma}$ from (\[dims\]). Thus $\tilde{\partial}_{+}^{-1}([\Gamma]) \cap \tilde{{\mathcal{M}}}^*_{\sigma}(Z)$ lies in a $2$-dimensional $C^2$-manifold $\tilde{P}^{-1}_{\Gamma}(W^0_{\Gamma} \times_{\operatorname{Stab}{\Gamma}} E'_{y_0})$. Since $\partial_+^{-1}([\Gamma]) \cap {\mathcal{M}}^*_{\sigma}(Z)$ is the quotient of the free smooth $SO(3)$-action on $\tilde{\partial}_{+}^{-1}([\Gamma]) \cap \tilde{{\mathcal{M}}}^*_{\sigma}(Z)$, we conclude that it has to be empty due to dimension counting.
So far we have a complete description of the irreducible moduli space ${\mathcal{M}}^*_{\sigma}(Z)$ for a generic perturbation $\sigma$:
1. ${\mathcal{M}}^*_{\sigma}(Z)$ is a smooth oriented $1$-manifold.
2. ${\mathcal{M}}^*_{\sigma}(Z)$ misses all central connections in $\chi(T^3)$ under the asymptotic map.
3. $\partial_+: {\mathcal{M}}^*_{\sigma}(Z) \to \chi(T^3)$ is transverse to a any prefixed subcomplex in $\chi(T^3)$.
The Reducible Locus {#trl}
===================
In this section, we study the structure of the reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ with respect to a generic perturbation. We continue to assume that $Y=T^3$ and $H_*(Z;{\mathbb{Z}}) \cong H_*(D^2 \times T^2; {\mathbb{Z}})$. The first part is to give a global description of ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$. The second part is to give a local description of the moduli space near points that are approached by a sequence of irreducible instantons.
Before diving into the details, we would like to comment on the choice of the weight $\delta_{\Gamma}$. There are two things we need to take care of. The first is to ensure the moduli space lies in the thickened moduli space. This is achieved by choosing $\delta_{\Gamma} \in (0, {\mu_{\Gamma} \over 2})$, where $\mu_{\Gamma}$ is the smallest nonzero absolute value of eigenvalues of the restricted Hessian $*d_{\Gamma}|_{\ker d^*_{\Gamma}}$. The second is to ensure the deformation complex at an instanton $[A] \in {\mathcal{M}}_{\sigma}(Z)$ is Fredholm. Later we will see it’s equivalent to the Fredholmness of the following complex: $$\tag{$F_{\delta_{\Gamma}, \sigma}$}\label{ssu2tc}
L^2_{k+1, \delta_{\Gamma}}(Z, {\mathfrak{su}}(2)) \xrightarrow{-d_A} L^2_{k, \delta_{\Gamma}}(T^*Z \otimes {\mathfrak{su}}(2)) \xrightarrow{d^+_{A, \sigma}} L^2_{k-1, \delta_{\Gamma}}(\Lambda^+T^*Z \otimes {\mathfrak{su}}(2)),$$ where $d^+_{A, \sigma}=d^+_A - D \sigma|_A$. We may ignore the perturbation part, since the Freholmness is preserved under small or compact perturbations. According to [@MMR Lemma 8.3.1] the complex ($F_{\delta_{\Gamma}}$) is Fredholm if and only if ${\delta_{\Gamma} \over 2}$ is not an eigenvalue of $-*d_{\Gamma}|_{\operatorname{im}d^*_{\Gamma}|_{\Omega^2}}$. Since $\operatorname{im}d^*_{\Gamma}|_{\Omega^2} \subset \ker d^*_{\Gamma}|_{\Omega^1}$, we may simply consider the smallest absolute value of eigenvalues of $*d_{\Gamma}|_{\ker d^*_{\Gamma}}$. Note that when $\Gamma$ is not a central connection, we have $H^1(T^3, \operatorname{ad}\Gamma) =1$. Thus only when approaching the central connections can the smallest absolute value of eigenvalues approach $0$. Let’s fix a neighborhood $\mathcal{O}_{c}$ of the central connections in $\chi(T^3)$. Then we may choose a uniform weight $\delta > 0$ for all instantons $[A]$ with $\partial_+([A]) \in \chi(T^3) \backslash \mathcal{O}_c$.
The Global Picture
------------------
Recall that an $SU(2)$-connection $A$ on $E$ is reducible if $A$ preserves a splitting $E=L \oplus L^*$ for some line bundle $L$. We write $A=A_L \oplus A^*_L$ corresponding to the splitting of $E$. The holonomy perturbation decomposes correspondingly as $$\sigma(A):=
\begin{pmatrix}
\sigma_L(A_L) & 0 \\
0 & \sigma_{L}(A_L^*)
\end{pmatrix}
\in \Omega^+(Z, {\mathfrak{su}}(2)),$$ where $\sigma_L$ is defined by $$\label{rlp}
\sigma_L(A_L) = {1 \over 2} \sum_{\alpha} (\operatorname{Hol}_{q_{\alpha}} A_L - \operatorname{Hol}_{q_{\alpha}} A^*_L) \otimes \omega_{\alpha}.$$ Note that $\sigma_L(A_L)=-\sigma_L(A^*_L)$. Thus the perturbed ASD equation $F^+_A=\sigma(A)$ is equivalent to $$F^+_{A_L}= \sigma_L(A_L),$$ where $\sigma_L(A_L) \in \Omega^+(Z, i{\mathbb{R}})$.
\[rrl\] Let $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma_{\pmb{\omega}}}(Z)$ be a reducible class of perturbed ASD connections. Then one can choose a representative $A$ which preserves the splitting $E=\underline{{\mathbb{C}}} \oplus \underline{{\mathbb{C}}}$.
As in the proof of Proposition \[eam\], one can choose a reprentative $A$ such that $A|_{[T_1, \infty) \times T^3}$ is in standard form with respect to a flat connection $\Gamma$ on $E'$. Let $E=L \oplus L^*$ be the decomposition that $A$ preserves. Let’s write $A=B+\beta dt$ and $A_L=B_L+ \beta_L dt$. Then we have $$F_{A_L} = F_{B_L} + dt \wedge (\dot{B}_L - d_{B_L} \beta_L).$$ We denote by $T_s \subset \{s\} \times T^3$ the $2$-torus representing a generator $1_Z \in H_2(Z; {\mathbb{Z}}) \cong {\mathbb{Z}}$. Then the Chern-Weil formula gives us that $$c_1(L) \cdot 1_Z= {1 \over {2\pi i}} \int_{T_s} F_{A_L} = {1 \over {2\pi i}} \int_{T_s} F_{B_L(s)} .$$ Due to the finite energy of $A$, elliptic theory tells us that for $T >>0$ one has $$\|F_{B(s)} \|^2_{L^2_l} \leq const. (\|F_A \|^2_{L^2([T, \infty) \times Y)} + e^{-2\mu s}), \text{ when } s \geq T+1.$$ We may take $l = 2$. Then the Sobolev embedding $L^2_2 \hookrightarrow C_0$ for $3$-manifolds implies that $\|F_{B(s)}\|_{C^0} \to 0$ as $s \to \infty$. Since $\|F_{B_L(s)}\|_{C^0} \leq \|F_{B(s)}\|_{C^0}$, we conclude that $$c_1(L) \cdot 1_Z = \lim_{s \to \infty} {1 \over {2\pi i}} \int_{T_s} F_{B_L(s)} =0.$$ Thus $c_1(L)=0 \in H^2(Z; {\mathbb{Z}})$ meaning $L$ is equivalent to the trivial bundle $\underline{{\mathbb{C}}}$.
Lemma \[rrl\] leads us to consider the moduli space of ASD $U(1)$-connections on the trivial bundle $\underline{{\mathbb{C}}}$. We write ${\mathcal{A}}_{k, loc}^{U(1)}(Z)$ for the space of $L^2_{k, loc}$ $U(1)$-connections on the trivial line bundle $\underline{{\mathbb{C}}}$ of $Z$. Fixing the product connection as the reference connection, ${\mathcal{A}}_{k, loc}^{U(1)}$ is identified with $L^2_{k, loc}(T^*Z \otimes i{\mathbb{R}})$. We write ${\mathcal{G}}^{U(1)}_{k+1,loc}$ for the $L^2_{k+1, loc}$ gauge transformations of the $U(1)$-bundle $\underline{{\mathbb{C}}}$. Then ${\mathcal{G}}^{U(1)}_{k+1,loc}$ is given by $L^2_{k+1, loc}(Z, U(1))$.
The moduli space of perturbed anti-self-dual $U(1)$-connections is defined to be $${\mathcal{M}}^{U(1)}_{\sigma}(Z):= \{ A_L \in {\mathcal{A}}_{k, loc}^{U(1)}: F^+_{A_L} = \sigma_L(A_L), \int_Z |F_{A_L}|^2 < \infty \} /{\mathcal{G}}^{U(1)}_{k+1,loc}.$$
If we write $A_L = d + a_L$ with $a_L \in L^2_{k, loc}(T*Z \otimes i{\mathbb{R}})$, the induced dual connection has the form $A^*_L= d-a_L$. Thus $F_{A_L} = -F_{A^*_L}$. Combining with the fact that $\sigma_L(A_L) = -\sigma_L(A^*_L)$, we get an involution $\tau$ on ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ given by $\tau([A]) = [A^*]$.
\[pci\] The quotient of ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ under $\tau$ is the reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$. Moreover the set of fixed points of $\tau$ consists of classes of flat connections whose holonomy groups lie in $\{-1, 1\} \subset U(1)$.
Note that the map $A_L \mapsto A_L \oplus A^*_L$ descends to a surjective map ${\mathcal{M}}^{U(1)}_{\sigma}(Z) \to {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ by Lemma \[rrl\]. We also note the Weyl group of $SU(2)$ is ${\mathbb{Z}}/2$ generated by the matrix representative $$\eta=
\begin{pmatrix}
0 & -1 \\
1 & 0
\end{pmatrix}.$$ Thus any $SU(2)$ gauge transformation of $E$ preserves the splitting $E= \underline{{\mathbb{C}}} \oplus \underline{{\mathbb{C}}}$ is either a $U(1)$ gauge transformation or a $U(1)$ gauge transformation multiplied by $\eta$. The effect of multiplying $\eta$ is applying the involution $\tau$. This identifies ${\mathcal{M}}^{U(1)}_{\sigma}(Z) /\tau = {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$.
Let $[A_L] \in {\mathcal{M}}_{\sigma}^{U(1)}$ be a fixed point of $\tau$. Then there exists $u \in {\mathcal{G}}^{U(1)}$ such that $A^*_L = u \cdot A_L$. Let’s write $A_L=d + a_L$. Then this is equivalent to $$-a_L = a_L - u^{-1}du.$$ Since $u^{-1}du$ is a closed $1$-form, we conclude that $d a_L=0$, which implies that $A_L$ is a flat connection. In particular we have $F^+_{A_L}=0$. From (\[rlp\]) we see that this requires $\operatorname{Hol}_{\gamma} A_L = \operatorname{Hol}_{\gamma} A^*_L$ for any loop $\gamma$ in $Z$. We conclude that the holonomy of $A_L$ has to be real in $U(1)$, which lies in $\{-1, 1\}$.
We say a $U(1)$-connection $A_L$ is a central connection if the holonomy group of $A$ lies in $\{\pm 1\} \subset U(1)$. Otherwise we say $A_L$ is non-central. We denote by ${\mathcal{M}}^{U(1),*}_{\sigma}(Z)$ the non-central part of the moduli space ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$.
The analysis of the structure of the $U(1)$-moduli space ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ is essentially simpler than that of the $SU(2)$-case due to the commutativity of the group $U(1)$. We first point out that one does not need to consider the thickened moduli space and the center manifold looks as simple as one would hope. To see this, let’s consider the deformation complex at a flat $U(1)$-connection $\Gamma_L$ over $T^3$: $$\label{3cx1}
L^2_l(T^3, i{\mathbb{R}}) \xrightarrow{-d} L^2_{l-1}(T^*T^3 \otimes i{\mathbb{R}}) \xrightarrow{*d} L^2_{l-2}(T^*T^3 \otimes i{\mathbb{R}}).$$ The $U(1)$-version Chern-Simons functional is $$\operatorname{cs}^{U(1)}(B_L) = -{1 \over 2} \int_{T^3} b_L \wedge db_L,$$ where $B_L= \nabla_L + b_L$, $\nabla_L$ is the product connection, and $b_L \in L^2_l(T^*T^3 \otimes i{\mathbb{R}})$. The gradient of $\operatorname{cs}^{U(1)}$ is given by $$\operatorname{grad}\operatorname{cs}^{U(1)}|_{B_L} = *d b_L.$$ We denote by $H^1_{\Gamma_L}:=\ker d \cap \ker d^* \in L^2_l(T^*T^3 \otimes i{\mathbb{R}})$. We claim that $\Gamma_L + H^1_{\Gamma_L}$ is the center manifold for the pair $(H^1_{\Gamma_L}, -\operatorname{grad}\operatorname{cs}^{U(1)})$, i.e. the center manifold is the graph of the zero map. Indeed, $\Gamma_L + H^1_{\Gamma_L}$ consists of all critical points of $\operatorname{grad}\operatorname{cs}^{U(1)}$ in $\ker d^*$ and is preseved by the gradient flowlines. Moreover, the gradient vector field $\operatorname{grad}\operatorname{cs}^{U(1)}$ is already complete over $H^1_{\Gamma_L}$. So there is no need to consider the thickened moduli space. The argument in Proposition \[EPA\] implies that all connections in ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ have exponential decay to their asymptotic value. We denote by $\mathcal{R}^{U(1)}(T^3):=\operatorname{Hom}(T^3, U(1))$ the space of $U(1)$-representations of $T^3$, and write the asymptotic map as $$\partial_+: {\mathcal{M}}^{U(1)}_{\sigma}(Z) \longrightarrow \mathcal{R}^{U(1)}(T^3).$$ Since the center manifold is $C^{\infty}$, the asymptotic map is $C^{\infty}$ as well. If we only consider the non-central stratum ${\mathcal{M}}^{U(1), *}_{\sigma}(Z)$ where the holonomy perturbations are nonzero, the argument of Proposition \[TSC\] implies that $$\partial_+: {\mathcal{M}}^{U(1), *}_{\sigma}(Z) \longrightarrow \mathcal{R}^{U(1)}(T^3)$$ is transverse to any given submanifold with respect to generic perturbations. The computation of the dimension is given by considering the deformation complex at $[A_L] \in {\mathcal{M}}^{U(1),*}_{\sigma}(Z)$ as in [@MMR Chapter 8]: $$d=-\big(\chi(Z) + \sigma(Z) \big) + {h^1 + h^0 \over 2} + {\rho(\Gamma_L) \over 2},$$ where $\Gamma_L$ is the asymptotic value of $A_L$, $h^i$ is the dimension of the $i$-the homology of the deformation complex (\[3cx1\]) at $\Gamma_L$ as above, and $\rho(\Gamma_L)$ is $\rho$-invariant of the odd signature operator twisted by $\Gamma_L$. Note that $T^3$ admits an orientation-reversing diffeomorphism, thus $\rho(\Gamma_L)=0$. since $h^1=3, h^0=1$, we conclude that $d=2$. Combining with the transversality result, we see that the image $\partial_+({\mathcal{M}}_{\sigma}^{U(1),*}(Z))$ misses the points in $\mathcal{R}^{U(1)}(T^3)$ whose holonomy groups lie in $\{ \pm 1\}$. We summarize the above discussion as follows.
Let $Z$ be a Riemannian smooth manifold with a cylindrical end modeled on $[0, \infty) \times T^3$ satisfying $H_*(Z; {\mathbb{Z}}) \cong H_*(D^2 \times T^2; {\mathbb{Z}})$. Then given a generic perturbation $\sigma \in \mathcal{P}_{\mu}$ we have the following description for the non-central moduli space ${\mathcal{M}}^{U(1),*}_{\sigma}(Z)$.
1. ${\mathcal{M}}^{U(1),*}_{\sigma}(Z)$ is an oriented smooth $2$-manifold.
2. The asymptotic map $\partial_+: {\mathcal{M}}^{U(1),*}_{\sigma}(Z) \to \mathcal{R}^{U(1)}(T^3)$ is smooth and transverse to any given submanifold of $\mathcal{R}^{U(1)}(T^3)$.
3. The image $\partial_+({\mathcal{M}}_{\sigma}^{U(1),*}(Z))$ misses all connections in $\mathcal{R}^{U(1)}(T^3)$ whose holonomy groups lie in $\{ \pm 1\}$.
Note that when there are no perturbations, the $U(1)$-moduli space ${\mathcal{M}}^{U(1)}(Z)$ is the space of gauge equivalence classes of flat $U(1)$-connections on $Z$. The holonomy map identifies ${\mathcal{M}}^{U(1)}(Z)$ with the space $\mathcal{R}^{U(1)}(Z)$ consisting of $U(1)$-representations of $\pi_1(Z)$, thus is identified as a $2$-torus. Our next step is to show that this feature is preserved under small generic perturbations.
Note that all connections in ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ has exponential decay after a fixed time $T_0$ given by Remark \[rt0\]. Moreover the decay rate is given by ${\mu \over 2}$, where $\mu$ is the smallest abosolute value of eigenvalues of $*d|_{\ker d^*}$. We choose $\delta < {\mu \over 2}$ as the weight. Now we can narrow down the ambient connection space to be $$\begin{split}
{\mathcal{A}}^{U(1)}_{k, \delta}(Z):=\{A_L \in {\mathcal{A}}^{U(1)}_{k, loc}(Z): & \exists b_L \in \mathcal{H}^1(T^*T^3; i{\mathbb{R}}) \text{ such that } \\
& A_L- \varphi b_L \in L^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}}) \},
\end{split}$$ where $\varphi: Z \to {\mathbb{R}}$ is the cut-off function in Definition \[d2.17\]. The gauge group that preserves ${\mathcal{A}}^{U(1)}_{k, \delta}(Z)$ is $$\begin{split}
{\mathcal{G}}_{k+1, \delta}:=\{ u \in {\mathcal{G}}_{k+1, loc}(Z):& u|_{[T_0 ,\infty) \times T^3} =u_0 \cdot e^{\xi}, \text{ where } u_0 \in S^1 \\
& \xi \in L^2_{k+1, \delta}([T_0, \infty) \times T^3, i{\mathbb{R}}) \}.
\end{split}$$ Let $[A_L] \in {\mathcal{M}}^{U(1)}_{\sigma}(Z)$. The Lie algebra of the gauge group ${\mathcal{G}}^{U(1)}_{k+1, \delta}(Z)$ is $$\hat{L}^2_{k+1, \delta}(Z, i{\mathbb{R}}):=\{ \xi \in L^2_{k+1, loc}(Z, i{\mathbb{R}}): d\xi \in L^2_{k, \delta}(Z, i{\mathbb{R}}) \}.$$ The tangent space $T_{A_L} {\mathcal{A}}^{U(1)}_{k, \delta}(Z)$ is $$\hat{L}^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}}):= \{ a_L + \varphi b_L : a_L \in L^2_{k, \delta}(T^*Z \otimes i {\mathbb{R}}), b_L \in \mathcal{H}^1(T^3; i{\mathbb{R}}) \}.$$ Finally the deformation complex at $A_L$ is $$\tag{$E^{U(1)}_{\delta, \sigma}$}\label{u1tc}
\hat{L}^2_{k+1, \delta}(Z, i{\mathbb{R}}) \xrightarrow{-d} \hat{L}^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}}) \xrightarrow{d^+_{\sigma}} L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes i{\mathbb{R}}),$$ where $d^+_{\sigma}:=d^+ - D \sigma_L|_{A_L}$ is the linearization of the perturbed ASD equation at $A_L$. Sitting inside of the complex (\[u1tc\]) is a subcomplex: $$\tag{$F^{U(1)}_{\delta, \sigma}$}\label{u1sc}
L^2_{k+1, \delta}(Z, i{\mathbb{R}}) \xrightarrow{-d} L^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}}) \xrightarrow{d^+_{\sigma}} L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes i{\mathbb{R}}).$$ When $\sigma=0$, the quotient of ($E^{U(1)}_{\delta}$) by ($F^{U(1)}_{\delta}$) is identified as $$i{\mathbb{R}}\xrightarrow{0} \mathcal{H}^1(T^3; i{\mathbb{R}}) \to 0.$$ As mentioned in the beginning of this section, the choice of $\delta$ also ensures that the complex (\[u1tc\]) is Fredholm with respect to small or compact perturbations. Now let $[A_L] \in {\mathcal{M}}^{U(1)}_{\sigma}(Z)$ be a central connection. The differential of $\sigma_L$ at $A_L$ is $$\big( D \sigma_L|_{A_L} (a) \big)_x = {1 \over 2} \sum_{\alpha} (\int_{q_{\alpha, x}} a) \cdot (\operatorname{Hol}_{q_{\alpha, x}} A^*_L - \operatorname{Hol}_{q_{\alpha, x}} A_L) \otimes \omega_{\alpha}|_x,$$ where $a \in \hat{L}^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}})$, $x\in Z$. Since $\operatorname{Hol}_{\gamma}A_L = \operatorname{Hol}_{\gamma} A^*_L$, we conclude that $D \sigma_L|_{A_L}=0$ whenever $A_L$ is central. [@M1 Proposition 3.12] identifies $H^2(E^{U(1)}_{\delta}) \cong \hat{H}^+_c(Z; i{\mathbb{R}})$, where $\hat{H}^+_c(Z; i{\mathbb{R}})$ is the image of $H^+_c(Z; i{\mathbb{R}})$ in $H^2(Z; i{\mathbb{R}})$ under the inclusion map. Due to the fact that $b^+(Z) =0$, we conclude that $H^2(E^{U(1)}_{\delta})=0$. From this fact we learn two things:
1. Each central class $[A_L]$ is a smooth point in ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$.
2. The non-perturbed $U(1)$-moduli space ${\mathcal{M}}^{U(1)}(Z)$ is regular, i.e. it’s smoothly cut-out by the defining equation.
\[TRS\] ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ is diffeomorphic to a $2$-torus with respect to a generic perturbation $\sigma_{\pmb{\omega}}$ with $\| \pmb{\omega} \|_W \leq \mathfrak{c}_2$ for some constant $\mathfrak{c}_2 >0$.
The proof is similar to the transversality result as in Proposition \[TSC\]. Let $\sigma_{\pmb{\omega}} \in \mathcal{P}_\mu$ be a generic perturbation so that ${\mathcal{M}}^{U(1)}_{\sigma_{\pmb{\omega}}}(Z)$ is regular. Pick a path $\sigma_t$ from $0$ to $\sigma_{\pmb{\omega}}$ in $\mathcal{P}_{\mu}$. Now we consider the map $$\begin{split}
\mathcal{F}^{U(1)}: \mathcal{P}_{\mu} \times \ker d^*_{\delta} &\longrightarrow L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes i{\mathbb{R}}) \\
(\sigma, a_L) & \longmapsto d^+_{\sigma} a_L,
\end{split}$$ where $d^*_{\delta} = e^{-\delta\tau} d^* e^{\delta \tau}: L^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}}) \to L^2_{k-1, \delta}(Z, i{\mathbb{R}})$ is the formal $L^2_{\delta}$-adjoint of $d$. Our discussion above implies that $\mathcal{F}^{U(1)}$ is a submersion. We denote by $\mathcal{Z}^{U(1)} = (\mathcal{F}^{U(1)})^{-1}(0)$. By construction, $\sigma_0$ and $\sigma_1$ are two regular values of the projection map $\pi: \mathcal{Z}^{U(1)} \to \mathcal{P}_{\mu}$. We approximate the path $\sigma_{t}$ relative to boundary by a generic path $\sigma'_{t}$ transverse to the map $\pi: \mathcal{Z}^{U(1)} \to \mathcal{P}_{\mu}$. Then the union $$\mathcal{Z}^{U(1)}_I:=\bigcup_{t \in [0, 1]} \pi^{-1} (\sigma'_t) \cap \mathcal{Z}^{U(1)}$$ is a cobordism from ${\mathcal{M}}^{U(1)}(Z)$ to ${\mathcal{M}}^{U(1)}_{\sigma_1}(Z)$. Since $\sigma_0=0$ is a regular value of $\pi|_{\mathcal{Z}^{U(1)}}$, we conclude that whennever $\|D \sigma_L\|$ is small, $\sigma$ is a regular value as well. Since $\|D\sigma_{\pmb{\omega},L}\| \leq const. \|\pmb{\omega}\|_W$, we can choose $\pmb{\omega}$ to have small norm, say less than $\mathfrak{c}_2$ so that each point $\sigma'_t$ in the path is a regular value of $\pi|_{\mathcal{Z}^{U(1)}}$. Thus the cobordism $\mathcal{Z}^{U(1)}_I$ is a product. This shows that ${\mathcal{M}}^{U(1)}_{\sigma_{\pmb{\omega}}}(Z)$ is diffeomorphic to ${\mathcal{M}}^{U(1)}(Z)$ which is a $2$-torus.
\[RPI\] Given a generic perturbation $\sigma_{\pmb{\omega}}$ satisfying $\|\pmb{\omega}\|_W < \mathfrak{c}_2$, the reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ is identified as a pillowcase, i.e. the quotient of $T^2$ by the hypoellipticinvolution.
Lemma \[pci\] tells us that ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ is the quotient of ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ under an involution whose fixed point set consists of flat connection on $Z$ with holonomy group inside $\{\pm 1\}$. Since $|b^1(Z; {\mathbb{Z}}/2)|=4$, there are four of them. Moreover each of them are smooth in ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$. Now we know ${\mathcal{M}}^{U(1)}_{\sigma}(Z)$ is a $2$-torus. The result follows.
The Kuranishi Picture
---------------------
Now we analyze the Kuranishi picture about a reducible instanton $[A]$ inside the entire moduli space ${\mathcal{M}}_{\sigma}(Z)$. Let $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ with the form $A=A_L \oplus A_L^*$ with respect to a reduction $E=\underline{{\mathbb{C}}} \oplus \underline{{\mathbb{C}}}$.
\[iic\] Any central instanton $[A] \in {\mathcal{M}}_{\sigma}(Z)$ is isolated from the irreducible locus ${\mathcal{M}}_{\sigma}^*(Z)$ for a small perturbation $\sigma \in \mathcal{P}_{\mu}$.
Since $\sigma(A)=0$ for any perturbation $\sigma \in \mathcal{P}_{\mu}$, we know that $A$ is actually flat. Then it suffices to prove the result in the non-perturbed case.
Note that the non-perturbed moduli space ${\mathcal{M}}(Z)$ is identified with the space of gauge equivalent classes of flat connections. Following the same argument as in the paragraph above Proposition \[STAB\], we see the Kuranishi obstruction map at $[A]$ in the ${\mathcal{M}}(Z)$ is given by $$\begin{split}
\mathfrak{o}_A: H^1(Z; {\mathfrak{su}}(2)) & \longrightarrow H^2(Z; {\mathfrak{su}}(2)) \\
e^1 \otimes X_1 + e^2 \otimes X_2 & \longmapsto e^1 \wedge e^2 \otimes [X_1, X_2],
\end{split}$$ where $\{e^1, e^2\}$ is an orthonormal frame of $\mathcal{H}^1(Z)$. Since the stabilizer of $A$ is $SU(2)$, a neighborhood of $[A]$ in ${\mathcal{M}}(Z)$ is identified with a neighborhood in the $SU(2)$-quotient $\mathfrak{o}^{-1}_A(0) / SU(2) \simeq {\mathbb{R}}^2 / ({\mathbb{Z}}/2)$. This proves that a neighborhood of $[A]$ in ${\mathcal{M}}(Z)$ is the same as a neigborhood of $[A]$ in the reducible locus ${\mathcal{M}}^{\operatorname{red}}(Z)$. Thus $[A]$ is isolated from the irreducible locus.
Given a generic small perturbation $\sigma$, any instanton $[A] \in {\mathcal{M}}_{\sigma}(Z)$ asymptotic to a central connection in $\chi(T^3)$ has to be central itself. Moreover Lemma \[iic\] tells us that the central instantons $[A] \in {\mathcal{M}}_{\sigma}(Z)$ is isolated from the irreducible locus. Thus we can choose the neighborhood $\mathcal{O}_c$ of central connections in $\chi(T^3)$ such that all irreducible instantons $[A] \in {\mathcal{M}}^*_{\sigma}(Z)$ has their asymptotic values outside $\mathcal{O}_c$. This in turn enables us to pick a weight $\delta > 0$ uniformly for all irreducible instantons in the Fredholm package.
Now let $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ be a non-central reducible instanton satisfying $\partial_+([A]) \notin \mathcal{O}_c$. We may write $A=d+a$ with $a \in \hat{L}^2_{k, \delta}(Z, {\mathfrak{su}}(2))$ and $A_L=d+a_L$ with $a_L \in \hat{L}^2_{k, \delta}(Z, i{\mathbb{R}})$. The perturbed deformation complex at $[A]$ is $$\tag{$E_{\delta, \sigma}$}\label{su2t}
\hat{L}^2_{k+1, \delta}(Z, {\mathfrak{su}}(2)) \xrightarrow{-d_A} \hat{L}^2_{k, \delta}(T^*Z \otimes {\mathfrak{su}}(2)) \xrightarrow{d^+_{A, \sigma}} L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes {\mathfrak{su}}(2)),$$ where $d^+_{A, \sigma}=d^+_A - D \sigma|_A$. With respect to the isomorphism $$\begin{split}
i{\mathbb{R}}\oplus {\mathbb{C}}& \longrightarrow {\mathfrak{su}}(2) \\
(v, z) & \longmapsto
\begin{pmatrix}
v & z \\
-\bar{z} & -v
\end{pmatrix}
\end{split}$$ the induced connection on ${\mathfrak{su}}(2)$-forms $\Omega^j(Z, {\mathfrak{su}}(2))$ splits as $d \oplus A_{{\mathbb{C}}}$ with $A_{{\mathbb{C}}} = A^{\otimes2}_L$, the holonomy perturbation splits as $\sigma_L \oplus \sigma_{{\mathbb{C}}}$, and the deformation complex (\[su2t\]) splits as the direct sum of the following two complexes: $$\tag{$E^{U(1)}_{\delta, \sigma}$}
\hat{L}^2_{k+1, \delta}(Z, i{\mathbb{R}}) \xrightarrow{-d} \hat{L}^2_{k, \delta}(T^*Z \otimes i{\mathbb{R}}) \xrightarrow{d^+_{\sigma}} L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes i{\mathbb{R}})$$ and $$\tag{$E^{{\mathbb{C}}}_{\delta, \sigma}$}\label{su2tc}
L^2_{k+1, \delta}(Z, {\mathbb{C}}) \xrightarrow{-d_{A_{{\mathbb{C}}}}} L^2_{k, \delta}(T^*Z \otimes {\mathbb{C}}) \xrightarrow{d^+_{A_{{\mathbb{C}}}, \sigma}} L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes {\mathbb{C}}),$$ where $d^+_{A_{{\mathbb{C}}}, \sigma} = d^+_{A_{{\mathbb{C}}}} - D \sigma_{{\mathbb{C}}}|_{A_{{\mathbb{C}}}}$. More precisely, let $a \in L^2_{k, \delta}(T^*Z \otimes {\mathbb{C}})$. The differential $D \sigma_{{\mathbb{C}}}|_{A_{{\mathbb{C}}}}$ evaluating at $a$ is given by $$\big(D \sigma_{{\mathbb{C}}}|_{A_{{\mathbb{C}}}} (a) \big)_x = - \sum_{\alpha} (\int_{q_{\alpha, x}} a) \cdot (\operatorname{Hol}_{q_{\alpha, x}} A_L)^2 \otimes \omega_{\alpha}|_x.$$ Since we are working with weighted Sobolev spaces, the homology of the complex is defined to be $$H^0_A(E_{\delta , \sigma}):= \ker d_A, \; H^1_A(E_{\delta, \sigma}):= \ker d^+_{A, \sigma} \cap \ker d^*_{A, \delta}, H^2_A(E_{\delta, \sigma}): = \ker d^{+, *}_{A, \sigma, \delta},$$ where $$d^*_{A, \delta} = e^{-\delta \tau} d^*_A e^{\delta \tau}, \; d^{+, *}_{A, \sigma, \delta} = e^{-\delta \tau} d^{+, *}_{A, \sigma} e^{\delta\tau}$$ are the $L^2_{\delta}$-adjoints. We further note that the complex ($E_{\delta, \sigma}$) is $U(1)$-equivariant, where the $U(1)$-action on ${\mathfrak{su}}(2)$-valued forms are induced by the action on the Lie algebra ${\mathfrak{su}}(2)$: $$e^{i\theta} \cdot (v, z) = (v, e^{i2\theta}z).$$ Earlier in Theorem \[Kur\] we have considered the Kuranishi obstruction map at a flat connection on a $3$-manifold to study the local structure. The same strategy can be applied to the four dimensional case as well.
Following [@MMR Theorem 12.1.1] there exists a $U(1)$-invariant neighborhood $V_A$ of $0$ in $H^1_A(E_{\delta, \sigma})$ together with a $U(1)$-equivariant map $$\mathfrak{o}_A : V_A \to H^2_A(E_{\delta, \sigma})$$ such that the $U(1)$-quotient of $\mathfrak{o}^{-1}_A(0)$ is isomorphic to a neighborhood of $[A] \in {\mathcal{M}}_{\sigma}(Z)$ as a stratified space. In particular when $H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma})=H^2(E^{{\mathbb{C}}}_{\delta, \sigma})=0$ at $A$, we have $H^2_A(E_{\delta, \sigma})=0$. We see that $$\mathfrak{o}_A^{-1}(0) = H^1_A(E^{U(1)}_{\delta, \sigma}) \simeq H^2(Z; i{\mathbb{R}}) = i{\mathbb{R}}\oplus i{\mathbb{R}}.$$ Thus $[A]$ is isolated from the irreducible part ${\mathcal{M}}^*_{\sigma}(Z)$. The next proposition shows this situation fits with all but finitely many reducibles $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$.
\[gph\] With respect to a small generic perturbation $\sigma \in \mathcal{P}_{\mu}$, for all but finitely many noncentral reducible instantons $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ satisfying $\partial_+([A]) \notin \mathcal{O}_c$ one has $$H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma}) = 0.$$ Moreover $H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma}) \cong {\mathbb{C}}$ for the finitely many exceptional reducibles.
Given $A_L \in {\mathcal{A}}^{U(1)}_{k, \delta}(Z)$, we write $A_{{\mathbb{C}}}:=A_L^{\otimes 2}$ for the connection on the trivial line bundle $\underline{{\mathbb{C}}} \to Z$. Let $\bar{A}=\bar{A}_{L} \oplus \bar{A}^*_{L}$ be a noncentral flat connection satisfying $\partial_+([\bar{A}]) \notin \mathcal{O}_c$ and $H^1_{\bar{A}}(E^{{\mathbb{C}}}_{\delta}) \neq 0$. We write $\mathcal{H}^1:=H^1_{\bar{A}}(E^{{\mathbb{C}}}_{\delta})$, $\mathcal{H}^2:=H^2_{\bar{A}}(E^{{\mathbb{C}}}_{\delta})$. Note that $i {\mathbb{R}}= H^0(Z; \operatorname{ad}\bar{A}) = H^0(Z; i{\mathbb{R}}) \oplus H^0_{\bar{A}}(E^{{\mathbb{C}}}_{\delta})$. Thus $H^0_{\bar{A}}(E^{{\mathbb{C}}}_{\delta})=0$. Denote by $\Pi: L^2_{k-1}(\Lambda^+T^*Z \otimes {\mathbb{C}}) \to \operatorname{im}d^+_{\bar{A}_{{\mathbb{C}}}}$ the orthogonal projection to the image of $d^+_{\bar{A}_{{\mathbb{C}}}}$. Note that $$\operatorname{ind}E^{{\mathbb{C}}}_{\delta} = \operatorname{ind}E_{\delta} - \operatorname{ind}E^{U(1)}_{\delta} = 1 -1 =0$$ for all $A \in {\mathcal{M}}^{\operatorname{red}}(Z)$ with $\partial_+([A]) \notin \mathcal{O}_c$ . We conclude $\dim_{{\mathbb{C}}} \mathcal{H}^1= \dim_{{\mathbb{C}}} \mathcal{H}^2$. Let’s consider the map $$\begin{split}
\eta: \mathcal{P}_{\mu} \times {\mathcal{A}}^{U(1)}_{k, \delta}(Z) \times \ker d^*_{\bar{A}, \delta} & \longrightarrow \operatorname{im}d^+_{\bar{A}_{{\mathbb{C}}}} \\
(\sigma, A_L, b) & \longmapsto \Pi (d^+_{A_{{\mathbb{C}}}, \sigma} b).
\end{split}$$ The differential of $\eta$ at $(0, \bar{A}_L, b)$ on the third component is given by $$D\eta|_{(0, \bar{A}_L, b)} (0, 0, \beta) = \Pi (d^+_{\bar{A}_{{\mathbb{C}}}} b),$$ which is surjective. By the implicit function theorem we can find a neighborhood $U \times V \subset \mathcal{P}_{\mu} \times {\mathcal{A}}^{U(1)}_{k, \delta}(Z)$ of $(0, \bar{A}_L)$ and a map $h: U \times V \times \mathcal{H}^1 \to \ker d^*_{\bar{A}, \delta}$ such that for all $(\sigma, A_L, b) \in U \times V \times \mathcal{H}^1$ one has $$\eta(\sigma, A_L, b + h(\sigma, A_L, b)) = 0.$$ In particular $d^+_{A_{{\mathbb{C}}}, \sigma}(b + h(\sigma, A_L, b)) \in \mathcal{H}^2$. This leads us to a map $$\begin{split}
\xi: U \times V & \longrightarrow L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes i{\mathbb{R}}) \times \operatorname{Hom}_{{\mathbb{C}}}(\mathcal{H}^1, \mathcal{H}^2) \\
(\sigma, A_L) & \longmapsto \big(F^+_{A_L} - \sigma_L(A_L), b \mapsto d^+_{A_{{\mathbb{C}}}, \sigma}(b + h(\sigma, A_L, b)) \big)
\end{split}$$ We write $\xi = \xi_1 \times \xi_2$ for its decomposition into the two factors in its range. The argument in Proposition \[TSC\] implies that $\xi_1$ is a submersion. Since the loops $q_{\alpha}$ constructed in the holonomy perturbation are dense at each point, Proposition 65 in [@H94] implies that we only need to vary finitely many components in $\pmb{\omega}=\{\omega_{\alpha}\}$ to ensure that $\xi_2$ is a submersion. Thus we conclude the map $\xi$ is a submersion. Let $S_i \subset \operatorname{Hom}_{{\mathbb{C}}}(\mathcal{H}^1, \mathcal{H}^2)$ be the stratum consisting of linear maps of complex codimension-$i$ kernel. Then the projection map $\pi: \xi^{-1}( \{0 \} \times S_i) \to U$ is Fredholm of real index $2-2i^2$. By the Sard-Smale theorem, for a generic perturbation $\sigma \in U$ only the top two strata $S_0$ and $S_1$ survive in the image of $\xi|_{\{\sigma\} \times V}$, which corresponds to $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ having $\dim_{{\mathbb{C}}} H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma})=0, 1$ respectively. Moreover set of connections $A$ attaining $\dim_{{\mathbb{C}}} H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma}) = 1$ is discrete in $V$.
Since ${\mathcal{M}}^{\operatorname{red}}(Z)$ is compact, we only need to run the argument above for finitely many $[A] \in {\mathcal{M}}^{\operatorname{red}}(Z)$. Due to the compactness of ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$, there are only finitely many reducible instantons $[A]$ satisfying $H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma}) = {\mathbb{C}}$.
\[gph1\] The only reason we impose the condition that $\partial_+{[A]} \notin \mathcal{O}_c$ is to ensure the complex (\[su2t\]) is Fredholm. Since we know all central instantons are isolated from the irreducible part ${\mathcal{M}}^*(Z)$, any reducible instanton $[A] \in {\mathcal{M}}^{\operatorname{red}}(Z)$ with $\partial_+{[A]} \in \mathcal{O}_c$ is also isolated from ${\mathcal{M}}^*(Z)$ once we choose $\mathcal{O}_c$ small enough. From the perspective of representation variety, this is equivalent to the vanishing of the twisted cohomology $H^1(Z; {\mathbb{C}}_{A_{{\mathbb{C}}}}) =0$. This property is also preserved under small perturbations. However in the perturbed case we need to vary the weight $\delta$ to define the cohomology.
With the help of Proposition \[gph\], we have the following description of a neighborhood of the reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ in the total moduli space ${\mathcal{M}}_{\sigma}(Z)$.
\[nbr\] Given a small generic perturbation $\sigma \in \mathcal{P}_{\mu}$, all but finitely many reducible instantons $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ are isolated from the irreducible moduli space ${\mathcal{M}}^*_{\sigma}(Z)$. Moreover any reducible instanton $[A]$ not isolated from ${\mathcal{M}}^*_{\sigma}(Z)$ is noncentral, and has a neighborhood $U_{[A]}$ in ${\mathcal{M}}_{\sigma}(Z)$ such that $U_{[A]} \cap {\mathcal{M}}^*_{\sigma}(Z) \simeq [0, \epsilon)$ for some $\epsilon >0$.
Let $[A] \in {\mathcal{M}}_{\sigma}(Z)$ have $H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma})={\mathbb{C}}$. From Proposition \[gph\] and Remark \[gph1\] there are only finitely many such instantons, all of which are noncentral. The irreducible part of a neighborhood $U_{[A]}$ is identified with the $U(1)$-quotient of $\mathfrak{o}_A^{-1}(0) \cap V_A$, where $V_A \subset H^1_A(E_{\delta, \sigma})$ is a neighborhood of the origin. We identify $H^1_A(E_{\delta, \sigma}) \cong i{\mathbb{R}}\oplus i{\mathbb{R}}\oplus {\mathbb{C}}$, and $H^2(E_{\delta, \sigma}) \cong {\mathbb{C}}$ so that the $U(1)$-action are given respectively by $$e^{i\theta} \cdot (x_1, x_2, z) = (a, b, e^{2i\theta}z), \; \; e^{i\theta} \cdot w = e^{2i\theta}w.$$ To get a better understanding of how $\mathfrak{o}_A$ looks like, we recall its construction as follows. One first considers the map $$\begin{split}
\ker d^*_{A, \delta} & \longrightarrow \operatorname{im}d^+_{A, \sigma} \\
a & \longmapsto \Pi(F^+_{A+a} - \sigma(A+a)),
\end{split}$$ where $\Pi: L^2_{k-1, \delta}(\Lambda^+T^*Z \otimes {\mathfrak{su}}(2)) \to \operatorname{im}d^+_{A, \sigma}$ is the $L^2_{\delta}$ orthogonal projection onto the image of $d^+_{A, \sigma}$. Since this map is a submersion, the implicit function theorem gives us a function $\mathfrak{q}_A: V_A \to \operatorname{im}d^{+, *}_{A, \sigma, \delta}$ so that $A+a+\mathfrak{q}_A(a) \in H^2(E_{\delta, \sigma})$. Then we let $$\mathfrak{o}_A(a):=F^+_{A+a+\mathfrak{q}_A(a)} - \sigma(A+a+\mathfrak{q}_A(a)).$$ Note that $\mathfrak{o}_A$ is analytic and vanishes at least up to second order by the virtue of its construction. Thus the $U(1)$-equivariance forces the Kuranishi map to take the following form $$\mathfrak{o}_A(x_1, x_2, z) =f(x_1, x_2, |z|) \cdot z.$$ where $f: i{\mathbb{R}}\oplus i{\mathbb{R}}\oplus {\mathbb{C}}\to {\mathbb{C}}$ vanishes at least up to first order. We further write $$f(x_1, x_2, |z|) := \sum_{i \geq 0} f_i(x_1, x_2)|z|^i.$$ We note that $\mathfrak{q}_A$ vanishes at least to second order. Thus the second order term of $\mathfrak{o}_A$ at $0$ is given by $Dd^+_{A, \sigma}|_0$, which is nonvanishing due to the transversality in Proposition \[gph\]. So up to an orientation-preserving change of coordinates, we may take $f_0(x_1, x_2)=x_1 \pm ix_2$. Since the complex $(E^{{\mathbb{C}}}_{\delta, \sigma})$ is complex linear, we know that $\mathfrak{o}_A(0, 0, z) = f(0, 0, |z|) \cdot z$ is complex linear. Thus $f_i(0, 0)=0$ for all $i \geq 1$. It now follows that the zero set of $\mathfrak{o}_A$ is given by $$\mathfrak{o}_A^{-1}(0) = \{x_1=x_2=0\} \cup \{z=0\}.$$ We then conclude that the normal part is identified with ${\mathbb{C}}/ U(1) \simeq [0, \infty)$.
\[bfp\] Any reducible instanton $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ in Proposition \[gph\] satisfying $H^1_A(E^{{\mathbb{C}}}_{\delta, \sigma}) = {\mathbb{C}}$ is called a bifurcation point of ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$.
The Orientation
---------------
At the end of this section, we discuss how we orient the perturbed moduli spaces. Formally an orientation of the moduli space ${\mathcal{M}}_{\sigma}(Z)$ is a trivialization of the determinant line of the index bundle associated with the deformation complex parametrized by connections in ${\mathcal{M}}_{\sigma}(Z)$. As we noted above, one cannot choose a uniform weight $\delta$ so that the deformation complex $(E_{\delta, \sigma, })$ is Fredholm for all instantons $[A] \in {\mathcal{M}}_{\sigma}(Z)$. So we only orient the portion of the moduli space that makes $(E_{\delta, \sigma})$ Fredholm.
Choose a weight $\delta$ and a neighborhood $\mathcal{O}_c \subset \chi(T^3)$ of the central connections as before. We write $${\mathcal{M}}_{\sigma}(Z, \mathcal{O}_c^c):=\{ [A] \in {\mathcal{M}}_{\sigma}(Z) : \partial_+[A] \notin \mathcal{O}_c\}$$ for the portion of the moduli space consisting of instantons $[A]$ that are not asymptotic to an element in $\mathcal{O}_c$. We first orient the unperturbed reducible locus ${\mathcal{M}}^{\operatorname{red}}(Z)$ as follows. Note the unperturbed deformation complex $(E^{U(1)}_{\delta})$ is independent of $[A] \in {\mathcal{M}}^{\operatorname{red}}(Z)$, thus the corresponding index bundle is trivialized automatically once we fix a trivialization at a single point. According to [@M1 Proposition 3.12], its determinant is identified with $$\det \operatorname{Ind}(E^{U(1)}_{\delta}) = \Lambda^{\max} H^0(Z; i{\mathbb{R}})^* \otimes \Lambda^{\max} H^1(Z; i{\mathbb{R}}).$$ Following the path $(E^{U(1)}_{\delta, t\sigma})$, $t \in [0, 1]$, we use the oriention on $\operatorname{Ind}(E^{U(1)}_{\delta})$ to orient the perturbed index bundle $\operatorname{Ind}(E^{U(1)}_{\delta, \sigma})$ for all small perturbations $\sigma$. The $SU(2)$ deformation complex $(E_{\delta, \sigma})$ splits into the direct sum of two complexes $(E^{U(1)}_{\delta, \sigma})$ and $(E^{{\mathbb{C}}}_{\delta, \sigma})$ at a reducible instanton $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z, \mathcal{O}_c^c)$. The complex structure on $(E^{{\mathbb{C}}}_{\delta, \sigma})$ provides us with a canonical orientation. Combining with an orientation on $(E^{U(1)}_{\delta, \sigma})$, we get a trivialization of $\det \operatorname{Ind}(E_{\delta, \sigma})$ on the reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z, \mathcal{O}_c^c)$.
Now we discuss how we orient the irreducible moduli space ${\mathcal{M}}^*_{\sigma}(Z)$. Let $[A] \in M^*_{\sigma}(Z)$. Recall that we only allow small perturbations $\sigma$ so that $\partial_+[A] \notin \mathcal{O}_c$. We choose a path $[A_t]$ in the space of connections asymptotic to flat connections not in $\mathcal{O}_c$ with exponential decay rate $-\delta$ so that $[A_1]=[A]$ and $[A_0] \in {\mathcal{M}}^{\operatorname{red}}(Z, \mathcal{O}_c^c)$. Then the orientation at $[A_0]$ will induce one at $[A]$. The orientation does not depend on the choice of $[A_0]$ since the index bundle $\operatorname{Ind}(E_{\delta})$ is trivialized over ${\mathcal{M}}^{\operatorname{red}}(Z, \mathcal{O}^c_c)$.
To conclude, an orientation of $H^0(Z; i{\mathbb{R}})^* \oplus H^1(Z; i{\mathbb{R}})$ induces an orientation on ${\mathcal{M}}_{\sigma}(Z, \mathcal{O}_c^c)$ for all small perturbations $\sigma$. The orientation of $Z$ induces an orientation on $H^0(Z; i{\mathbb{R}})$. We fix an orientation on $H^1(Z; i{\mathbb{R}}) \cong i{\mathbb{R}}\oplus i{\mathbb{R}}$ which is referred to as a homology orientation. Moreover to each bifurcation point $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ we assign a sign as follows.
Let $[A] \in {\mathcal{M}}^{\operatorname{red}}_{\sigma}(Z)$ be a bifurcation point as in Proposition \[gph\]. We assign $+1$ (resp. $-1$) to $[A]$ if $f_0: i{\mathbb{R}}\oplus i{\mathbb{R}}\to {\mathbb{C}}$ is orientation-preserving (resp. orientation-reversing), where $f_0$ is given by the Kuranishi obstruction map $\mathfrak{o}_A$ in the proof of Proposition \[gph\].
Since the local structure the moduli space near a bifurcation $[A]$ is modeled on $\mathfrak{o}_A^{-1}(0)$, the ‘$+1$’ assignment describes the case when the path of irreducible instantons is pointing away from $[A]$, and the ‘$-1$’ assignment corresponds to the case when the path is pointing into $[A]$.
The Surgery Formula {#sffo}
===================
The Set-Up
----------
We first give a more explicit description of the surgery operation. Let $X$ be an admissible integral homology $S^1 \times S^3$, and $\mathcal{T} \hookrightarrow X$ an embedded torus inducing a surjective map on first homology, i.e. the map $H_1(\mathcal{T}; {\mathbb{Z}}) \to H_1(X; {\mathbb{Z}})$ given by the inclusion is surjective. We fix a generator $1_X \in H^1(X; {\mathbb{Z}})$ serving as a homology orientation. We fix a framing of $\mathcal{T}$ by choosing an identification $\nu(\mathcal{T}) \cong D^2 \times T^2$. We write $$\mu=\partial D^2 \times \{pt.\} \times \{pt.\}, \lambda=\{pt.\} \times S^1 \times \{pt.\}, \gamma=\{pt.\} \times \{pt.\} \times S^1.$$ Let’s denote by $M:=\overline{X \backslash \nu(\mathcal{T})}$ the closure of the complement of the tubular neighborhood. Then we have $H_*(M; {\mathbb{Z}}) \cong H_*(D^2 \times T^2; {\mathbb{Z}})$. We require the framing is chosen so that $[\lambda]$ generates $\ker \big(H_1(\partial M; {\mathbb{Z}}) \to H_1(M; {\mathbb{Z}}) \big)$ and $1_X \cdot [\gamma]=1$. Under this choice, the isotopy class of $\mu$ and $\lambda$ are fixed, but there is still ambiguity in choosing $\gamma$ which we will allow. Since the diffeomorphism type of the surgered manifold $X_{p,q}=M \cup_{\varphi_{p,q}} D^2 \times T^2$ is determined by the isotopy class of $[\varphi_{p,q}(\mu)] = p[\mu] + q[\lambda]$, the surgery operation is well-defined despite the framing ambiguity.
Note that only when $p=1$ can the $(p,q)$-surgered manifold have the same homology as that of $S^1 \times S^3$. For simplicity, we write $$X_q=X_{1, q}, \; X_0=X_{0,1} \text{ and } \varphi_q=\varphi_{1, q}, \; \varphi_0=\varphi_{0, 1}.$$ We also write $N=D^2 \times T^2$ and identify $T^3=\partial M=-\partial N$. In this way, $X_q = M \cup_{\varphi_q} N$. Since the gluing map $\varphi_q$ preserves $[\gamma]$ for all $q$, we abuse the notation $[\gamma] \in H_1(X_q; {\mathbb{Z}})$ as a chosen generator. To define the Furuta-Ohta invariant one needs $X_q$ to be admissible for $q \neq 0$. This can be seen as follows. Let $q \neq 0$. Any representation $\rho: \pi_1(X_q) \to U(1)$ is determined by the image $\rho([\gamma]) \in U(1)$ of the generator $[\gamma]$. So every representation on $X_q$ comes from one on $X$. Consider the following portion of Mayer-Vietoris sequence: $$0 \to H^1(X_q; {\mathbb{C}}_{\rho}) \to H^1(M; {\mathbb{C}}_{\rho}) \oplus H^1(N; {\mathbb{C}}_{\rho}) \xrightarrow{j_q} H^1(T^3; {\mathbb{C}}_{\rho}),$$ where $j_q(\alpha, \beta) = \alpha|_{\partial M} - \varphi^*_q(\beta|_{\partial N})$. Since $H^1(X, {\mathbb{C}}_{\rho})=0$, we conclude that $\forall \alpha \in H^1(M, {\mathbb{C}}_{\rho})$, $\beta \in H^1(N, {\mathbb{C}}_{\rho})$ $$\alpha|_{\partial M} - \beta|_{\partial N} = 0 \Longleftrightarrow \alpha=0, \; \beta=0.$$ Denote by $r_M: H^1(M; {\mathbb{C}}_{\rho}) \to H^1(T^3; {\mathbb{C}}_{\rho})$ and $r_N: H^1(N; {\mathbb{C}}_{\rho}) \to H^1(T^3; {\mathbb{C}}_{\rho})$ the restriction map. Note that $\operatorname{im}r_M \cap \operatorname{im}r_N =\operatorname{im}r_M \cap \operatorname{im}\varphi^*_q \circ r_N.$ Thus $j_q(\alpha, \beta) = 0$ implies that $\exists \beta' \in H^1(N; {\mathbb{C}}_{\rho})$ such that $\beta'|_{\partial N} = \varphi^*_q(\beta|_{\partial N})$, which further implies that $\alpha=0, \beta'=0$, thus $\beta=0$. This shows that $H^1(X_q; {\mathbb{C}}_{\rho})=0$.
Let $E={\mathbb{C}}^2 \times X$ be a trivialized ${\mathbb{C}}^2$-bundle over an admissible integral homology $S^1 \times S^3$. We denote by ${\mathcal{M}}_{\sigma}(X)$ the moduli space of perturbed ASD $SU(2$)-connections on $E$. The vanishing of the twisted first homology ensures that the reducible locus ${\mathcal{M}}^{\operatorname{red}}_{\sigma}(X)$ is isolated from the irreducible locus ${\mathcal{M}}^*_{\sigma}(X)$ for all small perturbations. Moreover the irreducible locus ${\mathcal{M}}^*_{\sigma}(X)$ is an oriented compact $0$-manifold. The Furuta-Ohta invariant [@RS1] is defined to be the signed count of irreducible instantons under a generic small perturbation: $$\lambda_{FO}(X):= {1 \over 4} \#{\mathcal{M}}^*_{\sigma}(X).$$
The proof of the surgery formula is based on a neck-stretching argument which we set up as follows. Recall we have the decomposition $X=M \cup N$ with $N=\nu(\mathcal{T})$ the tubular neighborhood of the embedded torus, and $T^3=\partial M = -\partial N$. Identify a neighborhood of $T^3$ in $X$ by $(-1, 1) \times T^3$. Let $h$ be a flat metric on $T^3$. We pick a metric $g$ on $X$ so that $$g|_{(-1, 1) \times T^3} = dt^2 + h.$$ Given $L > 0$, we stretch the neck $(-1, 1) \times T^3$ of $X$ to obtain $(X_L, g_L)$: $$X_L = M \cup [-L, L] \times T^3 \cup N.$$ The geometric limit is denoted by $M_o:=M \cup [0, \infty) \times T^3$, $N_o:=(- \infty, 0] \times T^3 \cup N$, and $X_o = M_o \cup N_o$.
Instead of proving Theorem \[surq\] directly, we prove the following special case when $q=1$.
\[sur1\] After fixing appropriate homology orientations, one has $$\lambda_{FO}(X_1) = \lambda_{FO}(X) + {1 \over 2}D^0_{w_{\mathcal{T}}}(X_0).$$
We explain why the special case is sufficient. Let’s denote by $\mathcal{T}_q \subset X_q$ the image of the core $\{0\} \times T^2 \subset D^2 \times T^2$ in $X_q$ after the surgery performed. Then with respect to the framing of $\mathcal{T}_q$ given by the gluing copy $D^2 \times T^2$, $(1,1)$-surgery along $\mathcal{T}_q$ results in $X_{q+1}$, and $(0,1)$-surgery along $\mathcal{T}_q$ results in $X_0$. Thus Theorem \[surq\] is derived by applying Theorem \[sur1\] repetitively.
The Proof of Theorem \[sur1\]
-----------------------------
Recall $(X, g)$ is an admissible integral homology $S^1 \times S^3$ decomposed as $X=M \cup_{T^3} N$, where $N= \nu(\mathcal{T})$ is a tubular neighborhood of an embedded torus $\mathcal{T} \subset X$. We have framed $N \cong D^2 \times T^2$ with a basis $\{\mu, \lambda, \gamma\}$ on $\partial N = -T^3$. The surgered manifolds $X_1$ and $X_0$ are given respetively by $$X_1 = M \cup_{\varphi_1} N \text{ and } X_0= M \cup_{\varphi_0} N,$$ where under the basis $\{ \mu, \lambda, \gamma \}$ on $-T^3$ we have $$\varphi_1=
\begin{pmatrix}
1 & 0 & 0 \\
1& 1 & 0 \\
0 & 0 & 1
\end{pmatrix} \text{ and }
\varphi_0=
\begin{pmatrix}
0 & -1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1
\end{pmatrix}.$$ We put a metric $g_1$ and $g_0$ on $N$ such that $(\varphi^1)^*(g_1|_{\partial N}) =(\varphi^0)^*(g_0|_{\partial N}) = h$. Then we get the neck-stretched manifolds $X_{1, L}$ and $X_{0, L}$. Theorem \[str\] tells us that $\partial_+: {\mathcal{M}}^*_{\sigma}(M_o) \to \chi(T^3)$ is transverse to any given submanifold in $\chi(T^3)$ and misses the singular points in $\chi(T^3)$ for small perturbations. Note that $\pi_1(N)$ is abelian. Thus the unperturbed moduli space consists of only reducible instantons. Let’s write $${\mathcal{M}}_1:={\mathcal{M}}^*_{\sigma}(M_o) \text{ and } {\mathcal{M}}_2:={\mathcal{M}}^{\operatorname{red}}(N_o).$$ It follows from the standard gluing theorem (see for instance [@D02], [@MM93], [@M88]) that $$\begin{split}
\# {\mathcal{M}}^*_{\sigma}(X_L) & = \# \big( \partial_+({\mathcal{M}}_1) \cap \partial_-({\mathcal{M}}_2) \big) \\
\# {\mathcal{M}}^*_{\sigma}(X_{1, L}) & = \# \big( \partial_+({\mathcal{M}}_1) \cap \varphi^*_1 \circ \partial_-({\mathcal{M}}_2) \big),
\end{split}$$ where $\varphi^*_1: \chi(T^3) \to \chi(T^3)$ is the map induced by $\varphi_1$. To compare the difference we now put coordinates on the character variety $\chi(T^3)$.
Let’s first consider the $U(1)$-character variety $\mathcal{R}^{U(1)}(T^3)$ which is a double cover of $\chi(T^3)$. Recall that we have fixed a basis $\{ \mu, \lambda, \gamma\}$ for $\partial M = T^3$. To any $U(1)$-connection $A_L$ we assign a coordinate $(x(A_L), y(A_L), z(A_L))$ given by $$x(A_L) = {1 \over {2\pi i }} \int_{\mu} a_L, \; y(A_L) = {1 \over {2\pi i }} \int_{\lambda} a_L, \; z(A_L) = {1 \over {2\pi i }} \int_{\gamma} a_L,$$ where $a_L = A - d \in \Omega^1(T^3, i{\mathbb{R}})$ is the difference between $A_L$ and the product connection. Modulo $U(1)$-gauge transformations, the coordinates $x, y, z$ take values in ${\mathbb{R}}/{\mathbb{Z}}$. Then the holonomies of $A_L$ around $\mu, \lambda, \gamma$ are given respectively by $$\operatorname{Hol}_{\mu}A_L = e^{-2\pi i x}, \;\operatorname{Hol}_{\lambda}A_L = e^{-2\pi i y}, \; \operatorname{Hol}_{\gamma}A_L = e^{-2\pi i z}.$$ If we restricts to the fundamental cube $$\mathcal{C}_{T^3}:=\{(x, y, z): x, y, z\in [-{1 \over 2}, {1 \over 2}]\},$$ the $SU(2)$-character vareity $\chi(T^3)$ is identified with the quotient of $\mathcal{C}_{T^3}$ under the equivalence relations $(x, y, z) \sim (-x, -y, -z)$, $(-{1 \over 2}, y, z) \sim ({1 \over 2}, y, z)$, $(x, -{1 \over 2}, z) \sim (x, {1 \over 2}, z)$, and $(x, y, -{1 \over 2}) \sim (x, y, {1 \over 2})$. We shall restrict further to the following portion of the fundamental cube involved in the proof: $$\mathcal{C}^o_{T^3}:= \{(x, y, z): x \in [-{1 \over 2}, 0], y \in [0, {1 \over 2}], z\in [-{1 \over 2}, {1 \over 2}] \}.$$
(230, 230) (-100, -35)[{width="100.00000%"}]{} (-85, 40)[$y$]{} (-40, 30)[$z$]{} (0, 0)[$x$]{} (180, 215)[$P_0$]{} (223, 160)[$P_N$]{} (50, 130)[$P_1$]{} (45, 20)[$\partial_+^{\operatorname{red}}(M_o)$]{} (210, 45)[$(0, 0, 0)$]{}
Then the equivalence relations above only identify points on the lower-strata of $\mathcal{C}^o_{T^3}$, i.e. strata of dimension less than $3$ consisting of the faces, edges, and vertices of the cube. Then the image $\partial_-({\mathcal{M}}_2)$ is given by the quotient of the plane $$\label{pn}
P_N:= \{(0, y, z) : y \in [0, {1 \over 2}], z \in [-{1 \over 2}, {1 \over 2}]\}$$ whose quotient $[P_M] \subset \chi(T^3)$ is a pillowcase. The image $\partial_+({\mathcal{M}}^{\operatorname{red}}(M_o))$ of the unperturbed reducible locus on the manifold $M_o$ is given by the quotient of the plane $$\label{pm}
P_M:=\{(x, 0, z): x \in [-{1 \over 2}, 0], z \in [-{1 \over 2}, {1 \over 2}] \}$$ whose quotient $[P_M]$ is also a pillowcase. The image $\varphi_1^* \circ \partial_-({\mathcal{M}}_2)$ is given by the quotient of the plane $$P_1:=\{ (x, -x, z) : x \in [-{1 \over 2}, 0], z \in [-{1 \over 2}, {1 \over 2}] \}$$ whose quotient $[P_1] \subset \chi(T^3)$ is a cylinder $[0, 1] \times S^1$. Finally we consider a parallel copy of $P_M$ given by $$P_0:= \{ (x, {1 \over 2}, z) : x \in [-{1 \over 2}, 0], z \in [-{1 \over 2}, {1 \over 2}] \}$$ whose quotient $[P_0] \subset \chi(T^3)$ is again a pillowcase. We then orient the portion of the fundamental cube $\mathcal{C}^o_{T^3}$ by $dx \wedge dy \wedge dz$. It’s straightforward to see that the equivalence relations defined on the faces of $\mathcal{C}^o_{T^3}$ is orientation-preserving. Thus all the top strata of the quotient of the planes $P_N$, $P_M$, $P_1$, and $P_0$ are oriented by the orientation induced from that of $\mathcal{C}^o_{T^3}$. Let’s consider a solid $$V:=\{ (x, y, z) : x+y \geq 0, x \in [-{1 \over 2}, 0], y \in [0, {1 \over 2}], z \in [-{1 \over 2}, {1 \over 2}] \}.$$ Then the quotient $[V]$ is enclosed by $-[P_1]$, $[P_N]$, and $[P_0]$ in $\chi(T^3)$. Note that $$V \cap P_M =\{(0, 0, z): z \in [-{1 \over 2}, {1 \over 2}] \}.$$ Let $[A] \in {\mathcal{M}}^{\operatorname{red}}(M_o)$ be a non-central instanton such that $\partial_+([A]) \in [V]\subset \chi(T^3)$. The admissibility of $X$ implies that $H^1(M_o; \operatorname{ad}A_{{\mathbb{C}}}) =0$ since $[A]$ comes from a reducible instanton on $X$. Thus $[V]$ avoids the asymptotic values of the bifurcation points in ${\mathcal{M}}_{\sigma}(M_o)$ with respect to small perturbations. By choosing generic perturbations making $\partial_+$ transverse to $[V]$, we conclude that $$\begin{split}
\# {\mathcal{M}}_{\sigma}^*(X_1) - \# {\mathcal{M}}_{\sigma}^*(X) & = \# \big(\partial_+({\mathcal{M}}_1) \cap [P_1] \big) - \# \big(\partial_+({\mathcal{M}}_1) \cap [P_N] \big) \\
& = \# \big(\partial_+({\mathcal{M}}_1) \cap [P_0] \big) \\
&=\# \partial_+^{-1}([P_0]).
\end{split}$$ Now the proof has been reduced to the following result.
Continuing with notations above, one has $$\# \partial_+^{-1}([P_0]) = 2 D^0_{w_{\mathcal{T}}}(X_0),$$ where $D^0_{w_{\mathcal{T}}}(X_0)$ counts the irreducible anti-self-dual $SO(3)$-instantons on the ${\mathbb{R}}^3$-bundle $E_0 \to X_0$ characterized by $$p_1(E_0) = 0, \; w_2(E_0)=\operatorname{PD}[\mathcal{T}_0] \in H^2(X_0; {\mathbb{Z}}/2),$$ and $\mathcal{T}_0 \subset X_0$ is the core torus of the gluing $D^2 \times T^2$.
From its construction, $\# \partial_+^{-1}([P_0])$ counts the irreducible $SU(2)$-instantons on $M_o$ whose asymptotic holonomy around $\lambda$ is $\operatorname{Diag}(-1) \in SU(2)$. Through the isomorphism $\operatorname{ad}: {\mathfrak{su}}(2) \to \mathfrak{so}(3)$, every $SU(2)$-connection gives rise to an $SO(3)$-connection. Let $[A] \in \partial_+^{-1}([P_0])$. Then $A$ corresponds to a perturbed anti-self-dual $SO(3)$-connection $A'$ whose asymptotic holonomy around $\lambda$ is the identify. Since performing $0$-surgery amounts to gluing $D^2 \times T^2$ by sending the meridian $\partial D^2 \times \{pt.\} \times \{pt.\}$ to $\lambda$, the gluing theorem for $SO(3)$-instantons glues $[A']$ to an $SO(3)$-instanton $[A'_0]$ on $E_0$. The fact that $A'_0$ fails to lift to an $SU(2)$-connection forces $w_2(E_0) = \operatorname{PD}[\mathcal{T}_0]$.
To see why there is a factor of ‘2’ in the equation, we note the $SU(2)$-gauge group and $SO(3)$-gauge group fit into the exact sequence $$\operatorname{Map}(M_o, \pm 1) \to {\mathcal{G}}^{SU(2)}(M_o) \to {\mathcal{G}}^{SO(3)}(M_o).$$ Thus each $SU(2)$-gauge equivalence class corresponds to two $SO(3)$-gauge equivalence classes.
An Application to Finite Order Diffeomorphisms
----------------------------------------------
This subsection is devoted to the proof of Proposition \[fod\] using the surgery formula. We briefly recall the set-up. Let $\mathcal{K} \subset Y$ be a knot in an integral homology $S^1 \times S^3$. We write $\Sigma_n$ for the $n$-fold cyclic cover of $Y$ branched along $\mathcal{K}$, and $\tau_n: \Sigma_n \to \Sigma_n$ for the covering translation. We assume $\Sigma_n$ is a rational homology sphere, and denote by $X_n$ the mapping torus of $\Sigma_n$ under the covering translation $\tau_n$.
The argument is exactly the same as in the case of the Casson-Seiberg-Witten invariant [@M1 Proposition 1.2]. We denote by $\mathcal{T} \subset X_n$ the mapping torus of the branching set $\tilde{\mathcal{K}} \subset \Sigma_n$, and $X'_n$ the manifold resulted from performing $(1, 1)$-surgery of $X_n$ along $\mathcal{T}$. Lemma 7.1 in [@M1] tells us that the restriction of the covering translation $\tau_n$ on the knot complement extends to a free self-diffeomorphism $\tau_n': \Sigma'_n \to \Sigma'_n$ with $\Sigma'_n$ the manifold given by performing $1$-sugery of $\Sigma_n$ along $\tilde{\mathcal{K}}$. Moreover $X'_n$ is the mapping torus of $\Sigma'_n$ under $\tau'_n$. From Corollary 7.7 in [@RS1], we have $$\lambda_{FO}(X'_n) = n \lambda(Y) + {1 \over 8} \sum_{m=0}^{n-1} \operatorname{sign}^{m/n}(Y, \mathcal{K}) + {1 \over 2}\Delta''_{\mathcal{K}}(1).$$ Let’s denote by $X^0_n$ the manifold obtained by performing $(0, 1)$-surgery of $X_n$ along $\mathcal{T}$. In the proof of [@M1 Proposition 1.2], it has been shown that $X^0_n =S^1 \times Y_0(\mathcal{K})$. Combining the surgery formula and Corollary \[0d0s\], we get $$\lambda_{FO}(X_n) = n \lambda(Y) + {1 \over 8} \sum_{m=0}^{n-1} \operatorname{sign}^{m/n}(Y, \mathcal{K}).$$
The Excision Formula {#fsfo}
====================
The Set-Up
----------
We start with a more explicit description of the excision operation. Let $(X_1, \mathcal{T}_1)$ and $(X_2, \mathcal{T}_2)$ be two pairs of admissible homology $S^1 \times S^3$ with an essentially embedded torus. We choose an identification $\nu(\mathcal{T}_i) \cong D^2 \times T^2$ for a tubular neighborhood of $\mathcal{T}_i$ as in Section \[sffo\] so that we get a basis $\{\mu_i, \lambda_i, \gamma_i \}$ of $\partial \nu(\mathcal{T}_i)=-\partial M_i$ for each $i$. Let $\varphi: \partial M_2 \to \partial M_1$ be a diffeomorphism so that the manifold $$X_1 \#_{\varphi} X_2 := M_1 \cup_{\varphi} M_2$$ is an admissible homology $S^1 \times S^3$. Let $A_{\varphi}$ be the matrix representing the induced map $\varphi_*: H_1(\partial M_2; {\mathbb{Z}}) \to H_1(\partial M_1; {\mathbb{Z}})$ under the basis $\{[\mu_i], [\lambda_i], [\gamma_i]\}$. Over $D^2 \times T^2$, we write $$\mu'=\{pt.\} \times S^1 \times \{pt.\}, \lambda'=\partial D^2 \times \{pt.\} \times \{pt.\}, \gamma' = \{pt.\} \times \{pt.\} \times S^1.$$ We let $X_{1, \varphi}:=M_1 \cup_{\varphi_1} D^2 \times T^2$ and $X_{2, \varphi}:=D^2 \times T^2 \cup_{\varphi_2} M_2$ with the gluing map $\varphi_i$ inducing the matrix $A_{\varphi}$ on first homology groups with respect to the bases $\{[\mu_i], [\lambda_i], [\gamma_i]\}$ and $\{[\mu'], [\lambda'], [\gamma']\}$.
Since we require $X_1 \#_{\varphi} X_2$ to be an admissible homology $S^1 \times S^3$, the form of the matrix $A_{\varphi}$ can be described more explicitly. Due to the ambiguity of the choice of $\gamma_i$, one can find a framing of $\nu(\mathcal{T}_1)$ by adding to $\gamma_1$ certain multiples of $\mu_1$ and $\lambda_1$ so that $A_{\varphi}$ has the form $$A_{\varphi}=
\begin{pmatrix}
a & b & 0 \\
c & d & 0 \\
p & q & 1
\end{pmatrix}.$$ Since $A_{\varphi}$ is orientation-reversing, we have $\det A_{\varphi}=-1$. With the help of the Mayer-Vietoris sequence, one can show that $X_1 \#_{\varphi} X_2$ is an integral homology $S^1 \times S^3$ if and only if $\operatorname{gcd}(aq, b)=1$ and $(aq)^2 + b^2 \neq 0$. If we wish to repeat the argument in Section \[sffo\] to derive the admissibilily of $X_1 \#_{\varphi} X_2$ from that of $X_1$ and $X_2$ purely on the homology level, i.e. $\operatorname{im}r_{M_1} \cap \operatorname{im}r_{M_2} = \operatorname{im}r_{M_1} \cap \operatorname{im}\varphi^* \circ r_{M_2}$, we get $$b=\pm 1, q=0.$$ We note that the diffeomorphism type of $X_{1, \varphi}$ is determined by the image $\varphi_{1, *}([\lambda']) \in H_1(\partial M_1; {\mathbb{Z}})$. Thus in this case, $X_{1 ,\varphi}$ is obtained from $X_1$ via the $(1, d)$-surgery. The same can be derived for $X_{2, \varphi}$. In general we need to know more about the topology of $X_1$ and $X_2$ to determine whether $X_1 \#_{\varphi} X_2$ is admissible with a given matrix $A_{\varphi}$.
When we take the fiber sum of $(X_1, \mathcal{T}_1)$ and $(X_2, \mathcal{T}_2)$, the gluing map $\varphi_{\mathcal{T}}$ corresponds to the matrix: $$A_{\varphi_{\mathcal{T}}} =
\begin{pmatrix}
0 & 1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1
\end{pmatrix}.$$ We see that $X_{i, \varphi_{\mathcal{T}}} = X_i$, $i=1, 2$, and the fiber sum $X_1 \#_{\mathcal{T}} X_2$ is an admissible integral homology $S^1 \times S^3$ as $b=1, q=0$ in this case.
To apply the neck-stretching argument, we put metrics $g_1$ on $X_1$ and $g_2$ on $X_2$ so that when restricting to collar neighborhoods of $\partial M_1$ and $\partial M_2$ they are respectively of the form $$dt^2 + \varphi^*h \text{ and } dt^2 + h.$$ where $h$ is a flat metric on $T^3$. We also let $L$ be the length-parameter of the neck, and write various neck-stretched manifolds as in Section \[sffo\].
We say one more word about the choice of perturbations. The purpose of perturbing the anti-self-dual equation is to achieve various transversality properties, i.e. Proposition \[TSC\], Proposition \[TIS\], Proposition \[TRS\], and Proposition \[gph\], which correspond to the surjectivity of certain differential operators. Thus one can use perturbations of compact support since the transversality is an open condition. So given perturbations of compact support $\sigma_i$ on the end-cylindrical manifold $M_{i, o}$, we get a perturbation $\sigma_1 \# \sigma_2$ on $X_1 \#_{\varphi} X_2$ when the neck is stretched long enough.
The Proof of Theorem \[exif\]
-----------------------------
We shall omit the neck-lenghth parameter $L$ in the notations, which is chosen to be large enough to apply the gluing argument. To simplify the notation, we write $N=D^2 \times T^2$. Depending on the context, $N_o$ could mean either $N \cup [0, \infty) \times T^3$ or $(-\infty, 0] \times T^3 \cup N$. The gluing theorem tells us that $$\begin{split}
\# {\mathcal{M}}^*_{\sigma_1}(X_{1, \varphi}) & = \# \big( \partial_+({\mathcal{M}}_{\sigma_1}^*({M_{1, o}})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^{\operatorname{red}}(N_o)) \big) \\
\# {\mathcal{M}}^*_{\sigma_2}(X_{2, \varphi}) & = \# \big( \partial_+({\mathcal{M}}^{\operatorname{red}}(N_o)) \cap \varphi^* \circ \partial_-({\mathcal{M}}^*_{\sigma_2}(M_{2, o})) \big),
\end{split}$$ and moreover $$\begin{split}
\# {\mathcal{M}}^*_{\sigma_1 \# \sigma_2} (X_1 \#_{\varphi} X_2) & =
\# \big( \partial_+ ({\mathcal{M}}^{\operatorname{red}}_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^*_{\sigma_2}(M_{2, o})) \big) \\
&+ \# \big( \partial_+ ({\mathcal{M}}^*_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^*_{\sigma_2}(M_{2, o})) \big) \\
& + \# \big( \partial_+ ({\mathcal{M}}^*_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^{\operatorname{red}}_{\sigma_2}(M_{2, o})) \big).
\end{split}$$ The counting on right hand side of the third equation above makes sense because we can first fix a generic $\sigma_2$, then choose a generic $\sigma_1$ so that the asymptotic map $\partial_+: {\mathcal{M}}^*_{\sigma}(M_{1, o}) \to \chi(T^3)$ is transverse to $\varphi^* \circ \partial_-({\mathcal{M}}_{\sigma_2}(M_{2, o}))$.
As in the proof of Theorem \[sur1\], we regard the character variety $\chi(T^3)$ as the quotient the fundamental cuber $\mathcal{C}_{T^3}$ under appropriate relations. We identify the copy $T^3= \partial M_1$ with a basis given by $\{\mu_1, \lambda_1, \gamma_1\}$. Then $$\partial_+({\mathcal{M}}^{\operatorname{red}}(M_{1, o})) = [P_M] \text{ and }\varphi^* \circ \partial_- ({\mathcal{M}}^{\operatorname{red}}(M_{2, o})) =\varphi^* [P_M],$$ where $P_M$ is defined in (\[pm\]). The admissibility of $X_1 \#_{\varphi} X_2$ ensures there is no bifurcation points on neither ${\mathcal{M}}^{\operatorname{red}}(M_{1, o})$ nor ${\mathcal{M}}^{\operatorname{red}}(M_{2, o})$ asymptotic to $[P_M] \cap \varphi^*[P_M]$. Since $\dim {\mathcal{M}}^*_{\sigma_1}(M_{1, o}) = \dim {\mathcal{M}}^*_{\sigma_2}(M_{2, o}) = 1$, the transversality of the asymptotic maps implies that $$\label{fibr1}
\partial_+ ({\mathcal{M}}^*_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^*_{\sigma_2}(M_{2, o})) = \varnothing.$$
The proof of Proposition \[TSC\] gives us an isotopy from the unperturbed reducible locus ${\mathcal{M}}^{\operatorname{red}}(M_{2, o})$ to the perturbed one ${\mathcal{M}}^{\operatorname{red}}_{\sigma_2}(M_{2, o})$. Thus we conclude the counting $$\# \big( \partial_+ ({\mathcal{M}}^*_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^{\operatorname{red}}_{\sigma_2}(M_{2, o})) \big)$$ is equal to $$\# \big( \partial_+ ({\mathcal{M}}^*_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^{\operatorname{red}}(M_{2, o})) \big).$$ Note that $\varphi^* \circ \partial_-({\mathcal{M}}^{\operatorname{red}}(M_{2, o})) = \varphi^* \circ \partial_-{\mathcal{M}}^{\operatorname{red}}(N_o).$ Thus $$\label{fibr2}
\# \big( \partial_+ ({\mathcal{M}}^*_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^{\operatorname{red}}_{\sigma_2}(M_{2, o})) \big) = \# {\mathcal{M}}^*_{\sigma_1}(X_{1, \varphi}).$$ Similarly we have $$\label{fibr3}
\# \big( \partial_+ ({\mathcal{M}}^{\operatorname{red}}_{\sigma_1}(M_{1, o})) \cap \varphi^* \circ \partial_-({\mathcal{M}}^*_{\sigma_2}(M_{2, o})) \big) = \# {\mathcal{M}}^*_{\sigma_2}(X_{2, \varphi}).$$ Combining (\[fibr1\]), (\[fibr2\]), and (\[fibr3\]), we conclude that $$\# {\mathcal{M}}^*_{\sigma_1 \# \sigma_2} (X_1 \#_{\varphi} X_2) = \# {\mathcal{M}}^*_{\sigma_1}(X_{1, \varphi}) + \# {\mathcal{M}}^*_{\sigma_2}(X_{2, \varphi}),$$ which finishes the proof.
Examples
--------
In this subsection we compute the Furuta-Ohta invariants for two families of admissible integral homology $S^1 \times S^3$ arisen from mapping tori under diffeomorphisms of infinite order.
Let $(Y_1, \mathcal{K}_1)$ and $(Y_2, \mathcal{K}_2)$ be two pairs of integral homology sphere with an embedded knot. Fix two integers $n_1, n_2 > 1$. In what follows, $j=1$ or $j=2$. We denote by $\Sigma_j$ the cyclic $n_j$-fold cover of $Y_i$ branched along $\mathcal{K}_j$, and $\tilde{\mathcal{K}}_j$ the preimage of $\mathcal{K}_j$ in the cover $\Sigma_j$. Now we take $X_j$ to be the mapping torus of $\Sigma_j$ under the covering translation, and $\mathcal{T}_j$ the mapping torus of $\tilde{\mathcal{K}}_j$. Then the fiber sum formula tells us that the Furuta-Ohta invariant of $X_1 \#_{\mathcal{T}} X_2$ is given by $$\begin{aligned}
\lambda_{FO} (X_1 \#_{\mathcal{T}} X_2) & = n_1\lambda(Y_1) + {1 \over 8} \sum_{m_1=1}^{n_1-1}\operatorname{sign}^{m_1/n_1}(Y_1, \mathcal{K}_1) \\
& + n_2 \lambda(Y_2) + {1 \over 8} \sum_{m_2=1}^{n_2-1}\operatorname{sign}^{m_2/n_2}(Y_2, \mathcal{K}_2). \end{aligned}$$ We claim that the fiber $X_1 \#_{\mathcal{T}} X_2$ is the mapping torus of the knot splicing, denoted by $\Sigma_1 \#_{\mathcal{K}} \Sigma_2$, under certain self-diffeomorphism. We denote by $\tau_j$ the covering translation on $\Sigma_j$. A $\tau_j$-invariant neighborhood of $\tilde{K}_j$ is identified with $S^1 \times D^2$ where $\tau_j$ acts as $$\tau_j(e^{i\eta_j}, re^{i\theta_j}) = \big(e^{i\eta_j}, re^{i(\theta_j+{2\pi \over {n_j}})} \big).$$ A neighborhood of $\mathcal{T}_j$ is now identified with $[0, 1] \times S^1 \times D^2 / \sim$, for which we identify with $D^2 \times T^2$ as follows: $$\begin{aligned}
\nu(\mathcal{T}_j) & \longrightarrow S^1 \times S^1 \times D^2 \\
[t, e^{i\eta_j}, e^{i\theta_j}] & \longmapsto \big([t], e^{i\eta_j}, re^{i(\theta_j+t{2\pi \over {n_j}})} \big).\end{aligned}$$ Under the identifications above, along the mapping circle the knot complements $V_1:=\Sigma_1 \backslash \nu{\tilde{\mathcal{K}}_1}$ and $V_2:=\Sigma_2 \backslash \nu{\tilde{\mathcal{K}}_2}$ are glued at time $t$ via the map $$\begin{aligned}
\phi_t: \partial V_2 & \longrightarrow \partial V_1 \\
(e^{i\eta_2}, e^{i \theta_2}) & \longmapsto \big(e^{i(\theta_2+t{2\pi \over {n_2}})}, e^{i(\eta_2 - t {2\pi \over {n_1}})} \big). \end{aligned}$$ We write $F_t:=V_1 \cup_{\phi_t} [0, 1]_s \times T^2 \cup_{\operatorname{id}} V_2$ for the fiber at time $t$. We identify $F_t$ with $F_0$ by inserting the isotopy $\phi_0^{-1} \circ \phi_{(1-s)t}$ from $\phi^{-1}_0 \circ \phi_t$ to $\operatorname{id}$ along $[0, 1] \times T^2$, and denote by $f_t: F_t \to F_0$ this identification. From its construction, $F_0$ is the knot splicing $\Sigma_1 \#_{\mathcal{K}} \Sigma_2$. To see how the monodromy map looks like, we note that for $x \in \{0\} \times T^2 \subset F_1$, one has $$\phi_0 \circ \tau_2(x) = \tau_1 \circ \phi_1(x).$$ Thus $\tau_1$ and $\tau_2$ combine to a map $\tau_1 \# \tau_2: F_1 \to F_0$. So the monodromy map is given by $\tau_1 \# \tau_2 \circ f^{-1}_1: F_0 \to F_0$ whose restriction to the neck $[0, 1] \times T^2 \subset F_0$ has the form $$(s, e^{i\eta}, e^{i\theta}) \longmapsto \big(s, e^{i(\eta+(1-s){2\pi \over {n_1}})}, e^{i(\theta+ s{2\pi \over {n_2}})} \big).$$ In particular the monodromy map is of infinite order.
We consider the ‘Dehn twist’ along a torus in this example. Let $Y=Y_1 \#_{\mathcal{K}} Y_2$ be a splicing of two integral homology spheres along embedded knots. We denote by $V_i$ the knot complement in $Y_i$, and write $Y=V_1 \cup [0, 1]_s \times T^2 \cup V_2$. We use $(e^{i\eta}, e^{i\theta}) \in S^1 \times S^1$ to parametrize $T^2$ so that $S^1 \times \{pt.\}$ is null-homologous in $V_1$ and $\{pt.\} \times S^1$ is null-homologous in $V_2$. Let $p, q$ be a relatively prime pair and $c:[0, 1] \to T^2$ be a curve $$c(t):=(e^{i(\eta + t2\pi p)}, e^{i(\theta+ t2\pi q)}).$$ The Dehn twist along $c$ is a diffeomorphism $\tau_c: Y \to Y$ whose restriction on $V_1$ and $V_2$ is identity, and on the neck $[0, 1] \times T^2$ is given by $$\tau_c(s, e^{i\eta}, e^{i\theta}) = (s, e^{i (\eta + s2\pi p)}, e^{i(\theta + s2\pi q)}).$$ Then we see that $\tau_c$ has infinite order. Let $X_c$ be the mapping torus of $Y$ under $\tau_c$. Since $Y$ is an integral homology sphere, $X_c$ is an admissible homology $S^1 \times S^3$.
We claim that $X_c$ is given by torus excision. Let $M_1 = S^1 \times V_1$ and $M_2=S^1 \times V_2$. We regard $X_c=[0, 1]_t \times Y / \sim$, where $(0, \tau_c(y)) \sim (1, y)$. Since $\tau_c|_{[0, 1] \times T^2}$ is isotopic to identify, we can identify $[0,1]_t \times ([0,1]_s \times T^2 \cup V_2) / \sim$ with $[0, 1] \times T^3 \cup M_2$ by $$(t, s, e^{i\eta}, e^{i\theta}) \longmapsto \big( s, e^{2\pi t}, e^{i(\eta+(1-t)s2\pi p)}, e^{i(\theta+ (1-t)s2\pi q)} \big).$$ Then $[0, 1] \times T^3 \cup M_2$ is glued to $M_1$ by $$(e^{i\zeta}, e^{i\eta}, e^{i\theta}) \longmapsto (e^{i\zeta}, e^{i(\eta - p\xi )}, e^{i(\theta- p \xi)}).$$ In terms of the gluing matrix, the gluing map $\varphi$ is given by $$A_{\varphi}=
\begin{pmatrix}
0 & 1 & 0 \\
1 & 0 & 0 \\
-p & -q & 1
\end{pmatrix}.$$ If we write $X_{1, \varphi}=M_1 \cup_{\varphi} D^2 \times T^2$, $X_{2, \varphi}=D^2 \times T^2 \cup_{\varphi} M_2$, then $$\lambda_{FO}(X_c) = \lambda_{FO}(X_{1, \varphi}) + \lambda_{FO}(X_{2, \varphi}).$$ Finally, we note that $X_{i, \varphi}$ is obtained from $S^1 \times Y_i$ by a torus surgery. However, the gluing map $\varphi$ is not the type we condsidered in Theorem \[surq\], and we don’t know how to compare $\lambda_{FO}(X_{i, \varphi})$ with $\lambda_{FO}(S^1 \times Y_i)$ in this case.
| {
"pile_set_name": "ArXiv"
} |
23 cm 17.5 cm -0.30 cm -0.30 cm -1 cm
1 true cm
[**INFLUENCE OF PION-DELTA-HOLE CONFIGURATION ON THE**]{} 0.4 true cm [**RHO MESON MASS SPECTRUM IN DENSE HADRONIC MATTER**]{} 1 true cm
G. CHANFRAY 0.3 true cm [*IPN-Lyon, 43 Bd. du 11 Novembre 1918, F-69622 Villeurbanne Cédex, France.*]{} 0.6 true cm P. SCHUCK 0.3 true cm [*ISN-Grenoble, 53 Avenue des Martyrs, 38026 Grenoble-Cédex, France*]{}
2 true cm
[**Abstract**]{}
[*We discuss medium modification of the rho meson originating from its coupling to in medium quasi-pions and delta-hole states. This medium effect leads to a marked structure of the rho meson strength function in the $500\, MeV$ invariant mass region. We suggest that this effect may provide an explanation of the structure observed by the DLS collaboration in the $Ca+Ca/C+C$ ratio at $1\, GeV/u$.*]{}
1.5 true cm The modification of the rho meson mass spectrum at finite density and/or temperature is a very vividly debated subject. The interest for vector mesons is certainly related to the expected relationship between the evolution of their properties especially their masses with the evolution of the glue and scalar condensates and chiral symmetry restoration in a hot and dense environment. In addition, dilepton production in relativistic heavy ion collisions provides, through a vector dominance picture, a unique tool to study the rho meson in highly excited matter. Indeed, data obtained at CERN/SPS by the CERES collaboration in $S+Au$ at $200\, GeV/u$ and $Pb+Au$ at $158\, GeV/u$ [@CERES1; @CERES2] and the HELIOS collaboration [@HELIOS] have revealed a significant amount of strength below the free rho meson peak in the invariant mass region of $500\, MeV$ [@CERES1; @CERES2]. This led some authors to the conclusion that one is seing a downward shift of the rho meson mass induced by a QCD vacuum modification possibly intimately related to partial chiral symmetry restoration [@LI]. However it has been shown that more conventional many body mechanisms previously proposed in ref. [@BASIS; @ASA; @HERR] may explain the major part of the effect [@RAPP1]. The main mechanism at work is the coupling of the rho meson to states made of collective pion branch and transverse delta-hole states (ie those $\Delta-h$ states excited in photo-reactions) [@BASIS; @RAPP1]. This mechanism alone cannot explain the whole effect but inclusion of the p-wave coupling of the rho to some higher nucleon or delta resonances, building up the so-called rhosobar [@PIR], can give a reasonable agreement with data when incorporated in realistic calculation [@RAPP2]. More recently, new DLS [@PORTER] data of dielectron cross section in nucleus-nucleus collisions at $1\,GeV$ have shown some striking features. In particular, the $Ca+Ca$ to $C+C$ ratio exhibits a marked structure in the $500-600 MeV$ region. The purpose of this paper is to show that the original approach developped in ref. [@BASIS] for comparison with previous DLS data [@DLS0] (and subsenquently adapted to the CERN regime [@RAPP1]) may provide, at least qualitatively, an explanation of the effect.
We will briefly summarize the formalism for the calculation of the in medium rho meson propagator, incorporating coupling to collective pionic modes, longitudinal and transverse $\Delta-h$ states and also $2p-2h$ configurations. Then we will present some results for the ratio of the imaginary part of the rho meson propagator taken at two different densities and give a semi-quantitative discussion of the effect seen in the $Ca+Ca$ to $C+C$ ratio.
1 true cm The rho meson propagator at momentum $q(q^0,\bf{q})$ $$G(q)=q^2-m^2_\rho-\Sigma(q)$$ is fully known once the self-energy $\Sigma(q)$ is known. In the model described in [@BASIS] this self energy is obtained by substracting the standard two-pion loop at $q=0$ ensuring the proper normalization $F_\pi(0)=1$ of the pion electromagnetic form factor. In free space it reads : $$\Sigma(q)={4 g^2\over 3}\,\int {d^3t\over (2\pi)^3}\, v^2(t)\,
{q^2\over 4\omega_t^3}\,{1\over q^2-4\omega^2_t + i\eta}$$ with $\omega_t=(t^2+m^2_\pi)^{1/2}$. The coupling coupling constant $g=5$ and the parameter entering the form factor $v(t)$ have been fitted on the pion electromagnetic form factor and the $I=J=1$ $\pi\pi$ phase shifts.
In the medium the rho self energy is modified by the dressing of the pion propagator through p-wave coupling to $\Delta-h$ states (fig. 1a) and various vertex correction (fig. 1b, c) necessary to ensure gauge invariance. In a two-level model the final result is entirely expressible in terms of the energy $\Omega_{1,2}(k)$ of two longitudinal collective modes built on pion and $\Delta-h$ states and two transverse modes with energy ${\cal E}_{1,2}(k)$. Here, we limit ourselves to the case where ${\bf q}=0$ : $$\Sigma(q_0) =\Sigma_{LL}(q_0)\,+\,\Sigma_{LT}(q_0)$$ The first piece only depends on longitudinal modes : $$\Sigma_{LL}(q_0)={4 g^2\over 3}\,\int\,{d^3k\over (2\pi)^3}\,k^2 v^2(k)
\,\sum_{i=j=1}^2\,{q_0^2\over 2\big(\Omega_i(k)+\Omega_j(k)\big)\,
\Omega_i(k)\Omega_j(k)}$$ $$\times\,\bigg(1+{1\over 2}\left(\alpha_i(k) +\alpha_j(k)\right)\bigg)^2\,
{Z_i(k) \,Z_j(k)\over q^2_0- \big(\Omega_i(k) + \Omega_j(k)\big)^2\,+\,i\eta}
\label{SLL}$$ The eigenmodes $\Omega_1(k)$ are solution of the dispersion equation : $$\Omega_i^2(k)-\omega_k^2-k^2 \tilde\Pi^0\big(\Omega_i(k),{\bf k}\big)=0$$ where $\tilde \Pi^0 (k^0, {\bf k})$ is the pion self-energy dominated by the p-wave coupling to the $\Delta-h$ states screened by short range correlations associated to the $g'=0.5$ parameter. The explicit expression of $\Omega_i(k)$ and the associated weigth factors $Z_i(k)$ in the two-level model can be found in ref. [@AOUSS]. From eq. (\[SLL\]), it can be seen that the $\rho-\pi\pi$ vertex is modified by the presence of the $\alpha_i(k)
=\tilde \Pi^0(\Omega_i(k),{\bf k})$ factors. This vertex correction is represented by the diagram of fig. 1b and its net effect is to kill the structure at $q_0=2m_\pi$ originating from the dressing of the pions in the medium (fig. 1a) ; a detailed discussion can be found in [@BASIS].
The second piece of the in-medium self energy involves the coupling of the rho to the above longitudinal pionic modes and to a transverse mode, with energy ${\cal E}_i$ and strength $Y_i$, made of transverse delta-hole states and rho . In practice the collectivity of this states is very weak and the main contribution comes from the first state which is almost a pure transverse $\Delta-h$ state ($Y_1\simeq 1$). $\Sigma_{LT}$ reads : $$\Sigma_{LT}(q_0)={4 g^2\over 3}\,\int\,{d^3k\over (2\pi)^3}\,C(k)\, v^2(k)
\,\sum_{i=j=1}^2\,{q_0^2\over 2\big(\Omega_i(k)+{\cal E}_j(k)\big)\,
\Omega_i(k){\cal E}_j(k)}$$ $$\times\,
{Z_i(k) \bigg(1-Y_j(k)\bigg)
\over q^2_0- \big(\Omega_i(k) + {\cal E}_j(k)\big)^2\,+\,i\eta}
\label{SLT}$$ with $C(k)$ given by $C(k)=(8/9)(f^*_{\pi N\Delta}/m_\pi)^2{\cal E}_{\Delta k}
\Gamma^2_{\pi N\Delta(k)}\rho$ with notation of ref. [@BASIS]. The opening of this channel, depicted in fig. 1c, leads to a rather marked structure in the invariant mass region $q_0=M\ge m_\pi+\omega_\Delta\simeq 500\,MeV$ (see fig.2 and discussion below). This is the mechanism proposed in ref [@RAPP1] to explain an important part of the observed enhancement in the CERES data. We will discuss in the following to which extent it can provide an explanation of the structure seen in the $Ca+Ca/C+C$ ratio measured by the DLS collaboration.
In practice the above results for the self-energy has to be improved by the inclusion of the width of the states. For instance, the width of the quasi-pion states is incorporated through the replacement : $$\Omega_i(k)\to \Omega_i(k)\,+i\,
{{\bf k}^2 \,Im \tilde\Pi^0(\Omega_i(k), {\bf k})\over 2\,\Omega_i(k)}
\label{REPL}$$ and similar replacements for the transverse states. Here, the imaginary part of the irreducible pion self-energy $Im \tilde\Pi^0$, calculated along the dispersion line $i$, takes into account the delta width corrected from Pauli blocking [@OSET] together with extra $2p-2h$ contributions not reducible to a delta width [@BASIS]. The imaginary part of the self-energy $\Sigma_{LL}$ and $\Sigma_{LT}$ involving two-state propagators are now calculated using a spectral representation from which the real part is obtained by dispersion relation. All the details and phenomenological input are given in ref. [@BASIS].
0.5 true cm The dilepton production is directly propotionnal to the strength function $$S(q)=-{Im \Sigma(q)\over |q^2-m^2_\rho-\Sigma^(q)|^2}
\label{DILEP}$$ This imaginary part of the rho meson propagator is shown on fig. 2 for several densities. The bump between $500$ and $600\, MeV$ reflects the the pion-transverse delta-hole intermediate states (see fig 1.c). In heavy ion reactions around $1 GeV/u$ incident energies, one can expect compressions of nuclei ranging from $1$ to $2.5 \rho_0$ depending on the size of the system. To see how the structure at $ 500- 600\, MeV$ evolves as a function of density, we show in figure 3 the ratio of the strength function (\[DILEP\]) taken at two different densities typical of the $Ca+Ca$ and $C+C$ reactions measured by the DLS collaboration (this ratio is actually multiplied by the ratio of $A_p.A_t$ values corresponding to $Ca+Ca/C+C$). We also show on the same figure the DLS data. Of course this comparison can only give a first indication of the relevance of the mechanism mentionned just above. A realistic comparison requires a dynamical calculation incorporating precise experimental acceptance but we believe that the mechanism associated to $\pi-(\Delta-h)_T$ intermediate states, is certainly one important ingredient to account for the observed structure. 1 true cm Aknowledgement : Discussion with R.J. Porter, R. Rapp and J. Wambach are greatfully aknowledged.
[99]{} G. Agakichiev [*et.al*]{}, Phys. Rev. Lett. [**75**]{}, 1272 (1995); CERES Collaboration, J.P. Wurm, Nucl. Phys. [**A590**]{}, 103c (1995) A. Drees, Proceedings of the International Workshop on Gross Properties of Nuclei and Nuclear Excitations, Hirschegg, Austria January 1997, Editors H. Feldmeier, W. Nörenberg M. Masera [*et.al.*]{}, Nucl.Phys. [**A 590**]{}, 93c (1995) G.Q. Li, C.M. Ko, G.E. Brown, Phys. Rev. Lett. [**75**]{}, 4007 (1995) G. Chanfray, P. Schuck, Nucl. Phys. [**A545**]{}, 271c (1992); [**A555**]{}, 329 (1993) M. Asakawa, C.M. Ko, P. Levai, X.J. Qiu, Phys. Rev. [**C46**]{}, R1159 (1992) M. Herrmann, B.L. Friman, W. Nörenberg, Nucl. Phys. [**A545**]{}, 267c (1992) G. Chanfray, R. Rapp, J. Wambach, Phys. Rev. Lett. [**76**]{}, 368 (1996) R. Rapp, G. Chanfray, J. Wambach, Nucl. Phys. [**A617**]{}, 472 (1997) B. Friman, H. J. Pirner, Nucl. Phys. [**A617**]{}, 496 (1997) The DLS Collaboration, R. J. Porter [*et. al.*]{}, nucl-ex/9703001 G. Roche [*et.al*]{}, Phys. Rev. Lett. [**61**]{}, 1069, (1988); Phys. Lett. [**B226**]{}, 228 (1989) G. Chanfray, Z. Aouissat, P. Schuck, W. Norenberg, Phys. Lett. [**B256**]{}, 325 (1991) E. Oset, L.L , Nucl. Phys. [**A468**]{}, 631 (1987)
1 true cm [**Figure 1**]{} : Medium corrections to the rho meson propagator in the medium. (1a) : dressing of the pions in the medium. (1b) : vertex correction required by gauge invariance. (1c) : coupling of the rho to quasi-pion and transverse delta-hole states. 1 true cm [**Figure 2**]{} : Imaginary part of the rho meson propagator at values of $\rho / \rho_0=0,1,2$ with imaginary part from $\Delta$ resonance and $2p-2h$ included. 1 true cm [**Figure 3**]{} : Ratio of the strength function (multiplied by the ratio of $A_p.A_t$ values for $Ca+Ca/C+C$) for various couples of $\rho/\rho_0$. Dotted line : $1.8/ 1.2$; dashed line : $2/1$; full line : $2.4/1.2$. Experimental data obtained by the DLS collaboration [@PORTER] are also shown
| {
"pile_set_name": "ArXiv"
} |
\#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
****
Bose-Einstein correlations of neutral gauge bosons in $pp$ collisions
G.A. Kozlov
[ Bogolyubov Laboratory of Theoretical Physics\
Joint Institute for Nuclear Research,\
Joliot Curie st., 6, Dubna, Moscow region, 141980 Russia\
]{}
Abstract {#abstract .unnumbered}
========
The theory for Bose-Einstein correlations in case of neutral gauge bosons in $pp$ collisions at high energies is presented. Based on quantum field theory at finite temperature the two-particle Bose-Einstein correlations of neutral gauge bosons are carried out for the first time. As a result, the important parameters of the correlation functions can be obtained for the $Z^{0}Z^{0}$ pairs. The correlations of two bosons in 4-momentum space presented in this paper offer useful and instructive complimentary viewpoints to theoretical and experimental works in multiparticle femtoscopy and interferometry measurements at hadron colliders.
Introduction
============
An investigation of the space-time extension or even squeezing of particle sources via the multiparticle quantum-statistics correlation in high energy interactions is still attract the attention of physical society in both experiment and theory. Over the past few decades, a considerable number of successful studies have been done in this direction \[1\]. It is well understood that the studies of correlations between produced particles, the effects of coherence and chaoticity, an estimation of particle emitting source size play an important role in high energy physics.
By studying the Bose-Einstein correlations (BEC) of identical particles (we mean like-sign charge particles and the neutral charge ones), it is possible experimentally to determine the time scale and spatial region over which particles do not have the interactions. Such a surface is called as decoupling one. In fact, for an evolving system such as $p p$ collisions, it is not really a surface, since at each time there is a spread out surface due to fluctuations in the last interactions, and the shape of this surface evolve even in time. The particle source is not approximately constant because of energy-momentum conservation constraint.
More than half a century ago Hanbury-Brown and Twiss \[2\] used BEC between photons to measure the size of distant stars. In the works \[3,4\], the master equations for evolution of thermodynamic system that can be created at the final state of a high multiplicity process were established. The equations have the form of the field operator evolution equation (Langevin-like \[5\]) and allows one to gain the basic features of the emitting source space-time structure. In particular, it has been conjectured and further confirmed that the size of the emitting source through BEC is strongly affected by non-classical off-shell effect.
The shapes of BEC function were experimentally established in the LEP experiments ALEPH \[6\], DELPHI \[7\] and OPAL \[8\], and ZEUS Collaboration at HERA \[9\], which also indicated a dependence of the measured correlation radius on the hadron $(\pi,\ K)$ mass. The results for $\pi^{\pm}\pi^{\pm}$ and $\pi^{\pm}\pi^{\mp}$ correlations with $p\bar p$ collisions at $\sqrt {s}$ = 1.8 TeV were published by E735 Collaboration in \[10\].
The correlations between heavy particles (e.g., neutral gauge bosons $Z^{0}Z^{0}$) of Bose-Einstein type have not been carried out previously at hadron colliders. Such a study can be addressed to the Large Hadron Collider (LHC) which provide proton-proton interactions at $\sqrt{s} = 14$ TeV centre-of-mass system (c.m.s.) energy.
In this work, we make an attempt to demonstrate that the problem of properties of the genuine interactions can be explored using experimental data which can be collected by ATLAS and CMS Collaborations at the LHC. These data can be analyzed through the compared measures of some inclusive distributions and final state correlations.
One of the aims of this paper is to carry out the proposal for the experimental measurements of the $Z^{0}Z^{0}$ pair correlations.
This exploration will be theoretically supported by the quantum field theory at finite temperature $({QFT}_\beta)$ model approach \[3\]. It is known that the effective temperature of the vacuum or the ground state or even the thermalized state of particles distorted by external forces is occurring in models quantized in external fields. One of the main parameters of the model is the temperature of the particle source under the random source operator influence. The main channels are the di-lepton production $pp\rightarrow Z^{0}Z^{0}\rightarrow
2e^{-}2e^{+},~2\mu^{-}2\mu^{+},~e^{-}e^{+}\mu^{-}\mu^{+}$ in $pp$ collisions.
An efficient selection of leptons needs to be performed according to the following criteria. First, all leptons were required to lie in the pseudorapidity range covered by, e.g., the CMS muon system that is, $|\eta| \le$ 2.4. Second, the leptons were required to be unlikely charged in pairs. Note that the acceptances of another multipurpose detector ATLAS in the azimuthal angle and pseudorapidity are close to the respective parameters of CMS.
The dilepton channel is especially promising from the experimental point of view, since it is expected that the experimental facilities related for LHC (CMS and ATLAS detectors) will make it possible to record muons of energy in the TeV range with a resolution of about a few percent and an efficiency close to 100 %. Moreover, this channel is characterized by a maximum signal-to-background ratio in the energy region being considered.
BEC in case of two particles {#bec}
============================
A pair of identical bosons with the mass $m$ produced incoherently (in ideal nondisturbed, noninteracting cases) from an extended source will have an enhanced probability $C_{2}(p_{1},p_{2})=
N_{12}(p_{1},p_{2})/[N_{1}(p_{1})\cdot N_{2}(p_{2})]$ to be measured in terms of differential cross section $\sigma$, where $$\label{e31}
N_{12}(p_{1},p_{2})=\frac{1}{\sigma}\frac{d^{2}\sigma}{d\Omega_{1}\,d\Omega_{2}}$$ to be found close in 4-momentum space $\Re_{4}$ when detected simultaneously, as compared to if they are detected separately with $$\label{e32}
N_{i}(p_{i})=\frac{1}{\sigma}\frac{d\sigma}{d\Omega_{i}}, \,\,\,
d\Omega_{i}=\frac{d^{3}\vec p_{i}}{(2\pi)^{3}\,2E_{p_{i}}}, \,\,
E_{p_{i}}=\sqrt {\vec p_{i}^{2}+m^{2}},\,\,\,
i = 1, 2.$$
On the other hand, the following relation can be used to retrieve the BEC function $C_2(Q)$: $$\label{Konstrukcia C2}
C_2(Q) = \frac{N(Q)}{N^{ref}(Q)},$$ where $N(Q)$ in general case refer to the numbers for neutral gauge bosons (eg., $Z^{0} Z^{0}$) with $$\label{eq_01}
Q = \sqrt {-(p_1-p_2)_{\mu}\cdot (p_1-p_2)^{\mu}}= \sqrt{M^{2} - 4\,m^{2}}.$$ In definitions (\[Konstrukcia C2\]) and (\[eq\_01\]), $N^{ref}$ is the number of particle pairs without BEC and $p_{\mu_{i}}= (\omega_{i}, \vec p_{i})$ are four-momenta of produced bosons $(i = 1,\ 2)$; $M = \sqrt {(p_1+p_2)^{2}_{\mu}}$ is the invariant mass of the pair of bosons.
An essential problem in extracting the correlation is the estimate of the reference distribution $N^{ref}(Q)$ in Eq. (\[Konstrukcia C2\]). If there are other correlations beside the Bose-Einstein effect, the distribution $N^{ref}(Q)$ should be replaced by a reference distribution corresponding to the two-particle distribution in a geometry without BEC. Hence, the expression (\[Konstrukcia C2\]) represents the ratio between the number of $Z^{0}Z^{0}$ pairs $N(Q)$ in the real world and the reference sample $N^{ref}(Q)$ in the imaginary world. Note that the reference sample can not be directly observed in an experiment. Different methods are usually applied for the construction of reference samples \[1\], however all of them have strong restrictions. One of the preferable methods is to construct the reference samples directly from data. For our aim for reference sample $N^{ref}(Q)$, it is suitable to use the pairs $ Z^{0}Z^{0}$ from different (mixed) events.
It is commonly assumed that the maximum of two-particle BEC function $C_2(Q)$ is 2 for $\vec p_{1} =
\vec p_{2}$ if no any distortion and final state interactions are taking into account.
There are experimental difficulties in a determination of $ Z^{0}Z^{0}$ correlations, which are associated with acceptance limitations and limited statistics in the $ Z^{0}Z^{0}$ sample.
In general, the shape of the BEC function $C_2(Q)$ is model dependent. The most simple form of Goldhaber-like parameterization for $C_2(Q)$ \[11\] has been used for data fitting: $$\label{c2_aleph}
C_2(Q)=C_0\cdot (1+\lambda e^{-Q^2R^2})\cdot (1+\varepsilon Q) ,$$ where $C_0$ is the normalization factor, $\lambda$ is the so-called the chaoticity strength factor, meaning $\lambda =1$ for fully incoherent and $\lambda =0$ for fully coherent sources; the parameter $R$ is interpreted as a radius of the particle source, often called as the “correlation radius”, and assumed to be spherical in this parameterization. The linear term in (\[c2\_aleph\]) is supposed to be account within the long-range correlations outside the region of BEC. Note that distribution of bosons can be either far from isotropic, usually concentrated in some directions or almost isotropic, and what is important that in both cases the particles are under the random chaotic interactions caused by other fields in the thermal bath. In the parameterization (\[c2\_aleph\]) all of these problems are embedded in the random chaoticity parameter $\lambda$. To advocate the formula (\[c2\_aleph\]) it is assumed:\
a. incoherent average over particle source where $\lambda$ serve to account for:\
- partial coherence,\
- long-lived resonances associated with multiple distinguishable sources,\
- $Z^{0}Z^{0}$ purity;\
b. spherical Gaussian density of particle emission cell (with radius $R$);\
c. static source which means no time (energy) dependence.\
In order to save the quantum pattern of particle production process and to avoid the static and undistorted character of particle emitter source we also suggest to use the $C_2(Q)$ function within $QFT_\beta$ accompanying by quantum evolution approach in the form: $$\label{c2_Kozlov}
C_2(Q)=\xi(N)\cdot \Bigl[ 1+ \frac{1}{(1+\alpha)(1+\alpha^{\prime})}\
\tilde\Omega(Q)+
\frac{2{\sqrt{\alpha\alpha^{\prime}}}}
{(1+\alpha)(1+\alpha^{\prime})}\
\sqrt{\tilde\Omega(Q)} \Bigr]\cdot F(Q, \Delta x),$$ where $\xi(N)$ depends on the multiplicity $N$ as $$\label{eq_02}
\xi(N)= \frac{\langle{N (N-1)}\rangle}{\langle N\rangle^2} .$$ The function $F(Q, \Delta x)$ that expresses the correlation magnitude as a function of $Q$ and two-particle relative distance $\Delta x$ is a consequence of the Bogolyubov’s principle of correlations weakening at large distances \[12\] $$\label{e021}
F(Q, \Delta x) = \frac{f(Q,\Delta x)}{f(p_{1})\cdot
f(p_{2})} = 1 + r_{f}\,Q + \ldots$$ The function (\[e021\]) is normalized as $F(Q, \Delta x = \infty) = 1$, and $r_{f}$ is the measure of correlations weakening where $r_{f} \rightarrow 0$ as $\Delta
x\rightarrow \infty $; $f(Q,\Delta x)$ is the two-particle distribution function with $\Delta x$, while $f(p_{i})$ are one-particle probability functions with $i=1,2$. The important parameter $\alpha$ (as well as $\alpha^{\prime})$ in (\[c2\_Kozlov\]) summarizes our knowledge of other than space-time characteristics of the particle emitting source, and plays the role of a coherence parameter (see \[4\] for details).
The $\tilde\Omega(q)$ in (\[c2\_Kozlov\]) has the following structure in momentum space $$\label{e031}
\tilde\Omega (Q)=\Omega (Q)\cdot\gamma (n) ,
$$ where $$\label{e0310}
\Omega (Q) = \exp (-\Delta_{p\Re}) =
\exp \left [-(p_1-p_2)^{\mu}\,\Re_{\mu\nu}\, (p_1-p_2)^{\nu}\right ]$$ is the smearing smooth dimensionless generalized function, $\Re_{\mu\nu}$ is the (nonlocal) structure tensor of the space-time size (BEC formation domain), and it defines the spherically-like domain of emitted (produced) bosons.
To clarify with $\gamma (n)$ in (\[e031\]) let us emphasize that most of experiments dealing with elementary particles at high energies are of an inclusive as one measures quantum effect of BEC on limited samples of particles produced only. The unobserved part of the rest particle system acts then as a kind of thermal (heat) bath influencing measured samples of data (observables). Actually, the temperature $T$ being the most important parameter describing the influence of such a thermal bath is occurred in this model.
The function $\gamma (n)$ reflects the quantum thermal features of BEC pattern and is defined as $$\label{e032}
\gamma (n) = \frac{{n^2 (\bar \omega )}}{{n(\omega )\ n(\omega
')}} ,\ \
n(\omega ) \equiv n(\omega ,\beta ) =
\frac{1}{{e^{(\omega - \mu )\beta} - 1 }} ,\ \
\bar\omega = \frac{{\omega + \omega '}}{2} ,$$ where $n(\omega,\beta )$ is the mean value of quantum numbers for Bose-Einstein statistics particles with the energy $\omega$ and the chemical potential $\mu$ in the thermal bath with statistical equilibrium at the temperature $T= 1/\beta$. The following condition $\sum_{f} n_{f}(\omega,\beta) = N$ is evident, where the discrete index $f$ reflects the one-particle state $f$.
In terms of time-like $R_{0}$, longitudinal $R_{L}$ and transverse $R_{T}$ components of the space-time size $R_{\mu}$ the distribution $\Delta_{p\Re}$ looks like: $$\label{e33}
\Delta_{p\Re}\rightarrow\Delta_{pR} = (\Delta p^{0})^2 R^{2}_{0} + (\Delta p^{L})^2 R^{2}_{L} +
(\Delta p^{T})^2 R^{2}_{T} .$$ Seeking for simplicity one has ($R_{L}=R_{T}=R$) $$\label{e34}
\Delta_{pR} = (p^{0}_{1}-p^{0}_{2})^{2}R^{2}_{0} +
(\vec p_{1} - \vec p_{2})^{2} \vec R^{2}$$ for identical bosons. Hence, we have introduced a new parameter $R_{\mu}$, a 4-vector, which defines the region of nonvanishing particle density with the space-time extension of the particle emission source. Expression (\[e0310\]) must be understood in the sense that $\Omega(Q)$ is a function that in the limit $R\rightarrow\infty$, strictly becomes a $\delta$-function. For practical using with ignoring the energy-momentum dependence of $\alpha$, and assuming that $\alpha' = \alpha$ ($\alpha$ is related with $C_{2}(0)$ and $N$), we get the expression with $ \Omega(Q)\simeq\exp(-Q^{2}\,R^{2})$: $$C_2(Q) \simeq \xi(N) \left \{1 + \lambda_{new}(\beta)\,e^{-Q^2 R^{2}}
\left [1+\lambda_{corr}(\beta)\,e^{+Q^{2} R^{2}/2}\right ]
\right \},
\label{e58}$$ where the new intercept function becomes as $\lambda_{new} =
\gamma(\omega,\beta)/(1+\alpha)^{2}$, and the new coherence correction in the brackets of Eq. (\[e58\]) carries an additional intercept function $\lambda_{corr} = 2\,\alpha/\sqrt{\gamma(\omega,\beta)}$. In fact, since $\alpha \neq \alpha'$ (because $\omega \neq \omega'$ and, therefore, the number of states identified here with the number of particles $n(\omega)$ with given energy is also different), one can use the general precise form (\[c2\_Kozlov\]) for $C_2$ with details given by Eqs. (\[e031\]) and (\[e032\]) and with $\alpha$ coherence function depending on the particle mass, the energy of final leptons produced in pairs within the decays of $Z^{0}$’s and such characteristics of the emission process as the temperature $T$ and chemical potential $\mu$ occurring in the definition of $n(\omega)$ in (\[e032\]).
Since we did not follow special assumptions on the quantum operator level for $C_{2}(Q)$ from the initial stage, it may correspond to a physically real and observable effect at the LHC. This pattern may lead to a new squeezing state of correlation region.
Stochastic field and Green’s function
=====================================
Let us consider the stochastic field $B_{\mu}(x)= B_{{\mu}_{\tilde{s}}}(x,\tau)$ that depends on the arbitrary random source $\tilde{s}(x)$, and the fifth component $\tau$ means the “stochastic time”. The differential equation of an evolution of the field operator $B_{{\mu}_{\tilde{s}}}(x,\tau)= B_{\mu}(x,\tau)$ in the system under the associated stochastic process is
$$\begin{aligned}
\label{e1}
\partial_{\tau}B_{\mu}(x,\tau)= O[B_{\mu}(x,\tau)]\, ,\end{aligned}$$
where $O[B_{\mu}(x,\tau)]$ is the differential stochastic operator which has the form $$\begin{aligned}
\label{e2}
O[B_{\mu}(x,\tau)]= -\frac{1}{V}\frac{\delta J[B_{\mu}(x,\tau)]}{\delta B_{\mu}(x,\tau)} +
\tilde s_{\mu}(x,\tau)\end{aligned}$$ with a volume $V$ being introduced by dimensional reason. The r.h.s. of Eq. (\[e2\]) is the so-called stochastic operator where $J=\int d^{4} y L[B_{\mu}(y), \partial_{\nu} B_{\mu\nu}(y)]$ is the action defined by the Lagrangian density $L$; $\tilde s_{\mu}(x,\tau) = s_{\mu}(x,\tau) + n_{\mu} P $ carries the random stochastic history where the memory dissipation forces and the heat bath effects are included into $ s_{\mu}(x) = s_{\mu}(x,\tau)$, the constant $P$ emerges within the action of the stationary forces. Equation (\[e1\]) is nothing other but the evolution equation of the Langevin type applied already to stochastic processes on the quantum operator level in derivation of multiparticle Bose-Einstein correlations \[3\].
For simplicity, we assume that $\tilde{s}_{\mu}(x)$ varies stochastically with the Gaussian correlation function $$\begin{aligned}
\label{e3}
\langle\tilde{s}_{\mu}(x)\tilde{s}_{\nu}(y)\rangle = const\,\delta_{\mu\nu}
\exp (-z^{2}/l^{2}_{ch})\, ,\end{aligned}$$ where $z_{\nu}=(x-y)_{\nu}$, and $const$ is the strength of the noise described by the distribution function $\exp (-z^{2}/l^{2}_{ch})$ with $l_{ch}$ being the noise characteristic scale. Both $const$ and $\l_{ch}$ define the influence of the (Gaussian) noise on, e.g., correlations between particles that “feel” an action of an environment. Actually, Eq. (\[e1\]) can be transferred to the standard field equation of motion (in Euclidean space) $$\begin{aligned}
\label{e4}
\frac{1}{V}\frac{\delta J[B_{\mu}(x)]}{\delta B_{\mu}(x)} = \tilde s_{\mu}(x)\end{aligned}$$ with the source $\tilde s_{\mu}(x)$ if both $B_{\mu}$ and $\tilde s_{\mu}$ do not depend on “stochastic time” $\tau$. In classical theory, the random process given by $\tilde {s}_{\mu}(x)$ is nothing other but the white (Gaussian) noise.
In this paper, we focus on the role of particle masses and energies, effects of coherence and distortion, and the heat bath influences which are rather important to describe the correlations between particles. To solve this problem, especially to derive the memory term in evolution equation one can use the general properties of ${QFT}_{\beta}$. The model is defined by the following generating functional in four-dimensional space-time $$\begin{aligned}
\label{e5}
Z = \int D B_{\mu}\exp \left [ -i\int d^{4} x L(B_{\mu},B_{\mu\nu})\right ]\, ,\end{aligned}$$ where $$\begin{aligned}
\label{e6}
L = -\frac{1}{4} B_{\mu\nu}B^{\mu\nu} + \frac{1}{2}(m^{2} + U)B_{\mu}B^{\mu}\end{aligned}$$ with $B_{\mu\nu} = \partial_{\mu}B_{\nu} - \partial_{\nu}B_{\mu}$.
The direct calculations using the solution of Eq. (\[e1\]) with the Lagrangian density (\[e6\]) leads to the propagator of the field $B_{\mu}(x,\tau)$ distorted by $\tilde s_{\mu}(x,\tau)$. The transverse part of $B_{\mu}(x,\tau)$ will give the correct expression for the Euclidean vector field propagator at $\tau\rightarrow\infty$.
We are working with fields that correspond to a thermal field $B_{\mu}(x)$ with the standard definition of the Fourier transformed propagator $F[\tilde G_{\mu\nu}(p)]$ $$F[\tilde G_{\mu\nu}(p)]= G_{\mu\nu}(x-y) =
Tr\left\{T[B_{\mu}(x)B_{\nu}(y)]\rho_{\beta}\right\}
\label{e7}$$ with $\rho_{\beta}= e^{-\beta H}/Tr e^{-\beta H}$ being the density matrix of a local system in equilibrium at temperature $T$ under the Hamiltonian $H$ $$H = \int \frac{d^{3}\vec p}{(2\pi)^{3} 2p^{0}}\,p^{0}
\sum_{\lambda =1}^{3} b^{{\lambda}^{+}}(p)b^{\lambda}(p)
\label{e8}$$ with the operators of annihilation $b^{\lambda}(p)$ and creation $b^{{\lambda}^{+}}(p)$ to be defined later.
The interaction of $B_{\mu}(x)$ with the external field is given by the potential $U$. The equation of motion is $$(\nabla^2 + m^2)B_{\mu}(x) = -J_{\mu}(x),
\label{e9}$$ where $J_{\mu}(x) = U B_{\mu}(x)$ is the source density operator. A simple model like this allows one to investigate the origin of the unstable state of the thermalized equilibrium in a nonhomogeneous external field under the influence of source density operator $J_{\mu}(x) = U B_{\mu}(x)$. For example, the source can be considered as $\delta$-like generalized function, $J_{\mu}(x)=\tilde\mu\,\rho(x,\epsilon)B_{\mu}(x)$ in which $\rho(x,\epsilon)$ is a $\delta$-like succession giving the $\delta$-function as $\epsilon\rightarrow 0$ (where $\tilde\mu$ is some massive parameter). This model is useful because the $\delta$-like potential $U(x)$ provides the conditions for restricting the particle emission domain (or the deconfinement region). We suggest the following form: $$J_{\mu}(x) = - J_{sys}(x)\,B_{\mu}(x) + J_{R_{\mu}}(x),
\label{e10}$$ where the source $J_{\mu}(x)$ is a sum of a regular systematic motion part $J_{sys}(x)$ and the random source $J_{R_{\mu}}(x)$. The equation of motion (\[e9\]) becomes $$[\nabla^2 + m^2 - J_{sys}]B_{\mu}(x) = -J_{R_{\mu}}(x),
\label{e11}$$ and the propagator satisfies the following equation (in the Fourier transformed form labeled by tilde): $$[p^{2}_{\mu} - m^2 +\tilde J_{sys}]\tilde G_{\mu\nu}(p_{\mu}) = \tilde d_{\mu\nu} (p),
\label{e12}$$ where $$d_{\mu\nu}(x) = \left (g_{\mu\nu} + \frac{1}{m^{2}}\frac{\partial^{2}}
{\partial x_{\mu}\partial x_{\nu}}\right )\delta (x).
\label{e13}$$ As the standard point, the Green’s function of the vector field can be obtained from the one of the scalar field acting by the relevant operator $g_{\mu\nu} + m^{-2}\partial^{2} / (\partial x_{\mu}\partial x_{\nu})$. The solution of Eq. (\[e9\]) is $$B_{\mu}(x) = -\int dy\, G_{\mu\nu} (x,y)\,J_{R_{\nu}}(y),
\label{e14}$$ where the Green’s function obeys the Eq. (\[e12\]).
Green’s function and kernel operator
=====================================
Let us go to the thermal field operator $B_{\mu}(x)$ by means of the linear combination of the frequency parts $B_{\mu}^{1}(x)$ and $B_{\mu}^{{2}^{+}}(x)$ $$B_{\mu}(x) = B_{\mu}^{1}(x) + B_{\mu}^{{2}^{+}}(x)
\label{e24}$$ with $[B_{\mu}(x), B_{\nu}(y)] = i D_{\mu\nu}(x-y)$ and
$$B_{\mu}^{1} (x) = \int \frac{d^{3}\vec{p}}{(2\pi)^{3} 2 (\vec p^{2} +m^{2})^{1/2}}
\sum_{\lambda =1}^{3} \epsilon_{\mu}^{(\lambda)}(p)\tilde b^{(\lambda)}(p)\,e^{-ipx},$$ $$B_{\mu}^{{2}^{+}}(x) = \int \frac{d^{3}\vec{p}}{(2\pi)^{3} 2 (\vec p^{2} +m^{2})^{1/2}}
\sum_{\lambda =1}^{3} \epsilon_{\mu}^{(\lambda)}(p)\tilde b^{{(\lambda)}^{+}}(p)\,e^{ipx}.$$ The following properties of polarization vectors are the standard ones: $$\epsilon_{\mu}^{(\lambda)}(p)\epsilon_{\mu}^{(\lambda^{\prime})}(p) = g_{\lambda\lambda^{\prime}},$$ $$\sum_{\lambda =1}^{3} \epsilon_{\mu}^{(\lambda)}(p)\epsilon_{\nu}^{(\lambda)}(p) =
- g_{\mu\nu} + \frac{p_{\mu}p_{\nu}}{m^{2}}.$$ We assume that the deviation from the asymptotic free state given by the operator $a(\vec p,t)$ is provided by the random operator $r(\vec p,t):a(\vec p,t)\rightarrow
b(\vec p,t) = a(\vec p,t) + r(\vec p,t)$. The operators $\tilde b^{(\lambda)}(p)$ and $\tilde b^{{(\lambda)}^{+}}(p)$ obey the following equations in $\Re_{4}$ (see details in \[3\]): $$[\omega - \tilde K(p)]\tilde b^{(\lambda)}(p) = \tilde F(p) + \rho(\omega_{P},\epsilon),
\label{e27}$$ $$[\omega - \tilde K^{+}(p)] \tilde b^{{(\lambda)}^{+}}(p) =
\tilde F^{+}(p) + \rho^{*}(\omega_{P},\epsilon),
\label{e28}$$ where $p_{\mu} = (\omega = p^{0}, \vec p)$. Both equations (\[e27\]) and (\[e28\]) can be transformed into new equations for the frequency parts $ B_{\mu}^{1}(x)$ and $B_{\mu}^{{2}^{+}} (x)$ $$i\partial_{0} B_{\mu}^{1} (x) + \int_{\Re_{4}} K(x-y)\,B_{\mu}^{1}(y)dy = f_{\mu}(x),
\label{e29}$$ $$- i\partial_{0} B_{\mu}^{{2}^{+}} (x) + \int_{\Re_{4}} K^{+}(x-y)\,B_{\mu}^{{2}^{+}}(y)dy =
f_{\mu}^{+}(x),
\label{e30}$$ where $$f_{\mu}(x) = \int \frac{d^{3}\vec{p}}{(2\pi)^{3} 2 (\vec p^{2} +m^{2})^{1/2}}
\sum_{\lambda =1}^{3} \epsilon_{\mu}^{(\lambda)}(p)
[\tilde F(p) + \rho(\omega_{P},\epsilon)]e^{-ipx},
\label{e31}$$ $$f_{\mu}^{+}(x) = \int \frac{d^{3}\vec{p}}{(2\pi)^{3} 2 (\vec p^{2} +m^{2})^{1/2}}
\sum_{\lambda =1}^{3} \epsilon_{\mu}^{(\lambda)}(p)
[\tilde F^{+}(p) + \rho^{*}(\omega_{P},\epsilon)]e^{ipx}.
\label{e32}$$ The equations for field components $B_{\mu}^{1} (x)$ and $B_{\mu}^{{2}^{+}}(x)$ (\[e29\]) and (\[e30\]), respectively, are nonlocal within the presence of the formfactors $K(x-y)$ and $K^{+}(x-y)$, respectively. In principle, these formfactors can admit the description of locality for nonlocal interactions. At this stage, it must be stressed that we have new generalized evolution equations (\[e29\]) and (\[e30\]), which retain the general features of the propagating and interacting of the quantum vector fields with mass $m$ that are in the heat bath (thermal reservoir) and are chaotically distorted by other fields. For further analysis, let us rewrite the system of Eqs. (\[e29\]) and (\[e30\]) in the following form: $$i\partial_{0} B_{\mu}^{1} (x) + K(x)\star B_{\mu}^{1} (x) = f_{\mu}(x),
\label{e34}$$ $$- i\partial_{0} B_{\mu}^{{2}^{+}} (x) + K^{+}(x)\star B_{\mu}^{{2}^{+}} (x) = f_{\mu}^{+}(x),
\label{e35}$$ where $A(x)\star B(x)$ is the convoluted function of the generalized functions $A(x)$ and $B(x)$. Applying the direct Fourier transformation to both sides of Eqs. (\[e34\]) and (\[e35\]) with the following properties of the Fourier transformation $$F[K(x)\star B_{\mu}^{i} (x) ] = F[K(x)]\,F[B_{\mu}^{i} (x)] \,\,\, (i = 1, 2^{+}),$$ we get two equations $$[p^{0} + \tilde K(p)]\tilde B_{\mu}^{1}(p) = F[f_{\mu}(x)],
\label{e38}$$ $$[- p^{0} - \tilde K^{+}(p)]\tilde B_{\mu}^{{2}^{+}}(p) = F[f_{\mu}^{+}(x)].
\label{e39}$$ Finally, we have got the following equation for $\tilde B_{\mu}(p)$ field:
$$[- p^{0} + \tilde K^{+}(p)][p^{0} + \tilde K(p)]\tilde B_{\mu}(p) = \tilde T_{\mu}(p),
\label{e040}$$
where $$\tilde T_{\mu}(p) = [- p^{0} + \tilde K^{+}(p)]F[f_{\mu}(x)]+
[p^{0} + \tilde K(p)]F[f_{\mu}^{+}(x)].$$ We are now at the stage of the main strategy: one has to identify the field $B_{\mu} (x)$ and the random source operator $J_{{R_{\mu}}}(x)$, introduced in Eq. (\[e11\], with the Fourier transformed field $\tilde B_{\mu}(p)$ and $\tilde T(p)$ in (\[e040\]), respectively.
The next step is our requirement that Green’s function $\tilde G_{\mu\nu}(p)$ in Eq. (\[e12\]) and the function $ \Gamma_{\mu\nu}(p)$, satisfying the equation $$[- p^{0} + \tilde K^{+}(p)][p^{0} + \tilde K(p)]\tilde \Gamma_{\mu\nu}(p) = g_{\mu\nu}
\label{e42}$$ must be equal to each other, i.e. $$F[\tilde G_{\mu\nu}(p) - \tilde \Gamma_{\mu\nu}(p)] = 0.$$ The kernel operator $\tilde K(p)$ is $$\tilde K(p) \simeq \epsilon \sqrt {1 + \frac{m^{2}}{\epsilon^{2}}},
\label{e43}$$ where $\epsilon = 2 \sqrt {\vec k_{l}^{2} + m_{l}^{2}}$ is the total energy of the final lepton-antilepton pair (with momentum $\vec k_{l}$ and the mass $m_{l}$ for the lepton) produced within the decay of $Z^{0}$ boson being in the rest frame. To get $\tilde K (p)$ in the form( \[e43\]) we used the fact that the full Green’s function $\tilde G_{\mu\nu}(p)$ is given by the corresponding full Green’s function of the scalar field \[13\] under the action by the differential operator $(g_{\mu\nu} - m^{-2}\, p_{\mu}\, p_{\nu})$.
Source size
===========
It has been emphasized \[4\] that there are two different scale parameters in the model considered here. One of them is the so-called “correlation radius” $R$ introduced in (\[c2\_aleph\]) and (\[c2\_Kozlov\]) with (\[e33\]). In fact, this $R$-parameter gives the pure size of the particle emission source without the external distortion and interaction coming from other fields. The other (scale) parameter is the stochastic scale $L_{st}$ which carries the dependence of the particle mass, the $\alpha$-coherence degree and what is very important — the temperature $T$-dependence: $$\label{e36}
L_{st}\simeq {\left[\frac{1}{\alpha(N)\, {\vert p^{0}-\tilde K(p)\vert}^{2}\,
n(m,\beta)}\right ]}^{\frac{1}{2}}\rightarrow
{\left[\frac{1}{\alpha(N)\, 4\vec k^{2}_{l}\,
{\vert 1-\delta_{k}\vert}^{2}\,\bar n(m,\beta)}\right ]}^{\frac{1}{5}},$$ where $$\delta_{k} = \sqrt {1 + \frac{m^{2}}{4\vec k^{2}_{l}}}$$ and the lepton mass $m_{l}$ is neglected.
It turns out that the scale $L_{st}$ defines the range of stochastic forces. This effect is given by $\alpha
(N)$-coherence degree which can be estimated from the experiment within the two-particle BEC function $C_{2}(Q)$ when $Q$ close to zero, $C_{2}(0)$, at fixed value of mean multiplicity $\langle N\rangle$: $$\label{e37}
\alpha (N)\simeq \frac{2-\bar C_{2}(0) + \sqrt {2-\bar
C_{2}(0)}}{\bar C_{2}(0)-1},\,\,\, \bar C_{2}(0)= C_{2}(0)/\xi(N) .$$ In formula (\[e36\]), $\bar n(m,\beta)$ is the thermal relativistic particle number density $$\label{e38}
\bar n(m,\beta) = 3\int\frac{d^{3}\vec
p}{(2\,\pi)^{3}}\,n(\omega,\beta)=3\frac{\mu^2 +m^2}
{2\,\pi^2}\,T\,\sum_{l=1}^{\infty}\frac{1}{l}K_{2}\left
(\frac{l}{T}\sqrt{\mu^2 + m^2}\right) ,$$ where $K_{2}(...)$ is the modified Bessel function.
The coherence function $\alpha$ is another very important one that summarizes our knowledge of other than space-time characteristics of the particle emission source, and the prediction of $\alpha$ from an experiment is very instructive aim itself. For $\alpha =0$, one actually finds $$1 < C_{2}(Q) < \xi (N) (1 + \gamma e^{-Q^{2}R^{2}})$$ which is nothing other but the Goldhaber parameterization \[11\] with $0< \gamma < 1$ being a free parameter adjusting the observed value of $C_{2}(Q=0)$.
Within our aim to explore the correlation between $Z^{0}Z^{0}$ the scale $L_{st}$ has the form $$\label{e39}
L_{st}\simeq {\left [\frac{e^{\sqrt{m^{2}+\mu^{2}}/T}}{12\, \alpha(N)\,\vec k^{2}_{l}\,
{(m^{2}+\mu^{2})}^{3/4}{\left
(\frac{T}{2\,\pi}\right)}^{3/2}\, \left
(1+\frac{15}{8}\frac{T}{\sqrt{m^{2}+\mu^{2}}}\right )
{\vert 1-\delta_{k}\vert}^{2}}\right ]}^{\frac{1}{5}} ,$$ where the condition $l\,\beta\,\sqrt{m^{2}+\mu^{2}} >1$ for any integer $l$ in (\[e38\]) was taken into account. The only lower temperatures will drive $L_{st}$ within formula (\[e39\]) even if $\mu = 0$ and $l=1$ with the condition $T < m$.
Note that the condition $\mu < m$ is a general restriction in the relativistic “Bose-like gas”, and $\mu = m$ corresponds to the Bose-Einstein condensation.
For high enough $T$ no $\mu$ - dependence is found for $L_{st}$: $$\label{e40}
L_{st}\simeq {\left [\frac{\pi^{2}}{12\,\zeta (3)\, \alpha(N)\,\vec k^{2}_{l}\,
T^{3} {\vert 1-\delta_{k}\vert}^{2}}\right ]}^{\frac{1}{5}} ,$$ where the condition $T > l \sqrt{m^{2}+\mu^{2}}, l=1,2, ...$ is taken into account. The origin of formula (\[e40\]) comes from $$\label{e41}
\bar n(m,\beta)\rightarrow \bar n(\beta)\simeq \frac{3\,T^{3}}{\pi^{2}}\,\zeta (3)
$$ where neither a $Z^{0}$ boson mass nor the $\mu$ - dependence occurred; $\zeta (3) = \sum^{\infty}_{l=1}l^{-3} = 1.202$ is the zeta-function with the argument $3$.
To be close to the experiment there is necessary to include transverse momenta, where the $Z^{0}$ boson mass $m$, in Eqs. (\[e38\]), (\[e39\]), (\[e40\]) is replaced by the transverse mass $m_T = \sqrt{m^{2} + p^{2}_{T}}$.
Actually, the increasing of $T$ leads to squeezing of the domain of stochastic force influence, and $L_{st}(T=T_{0})= R$ at some effective temperature $T_{0}$. The higher temperatures, $T > T_{0}$, satisfy to more squeezing effect and at the critical temperature $T_{c}$ the scale $L_{st}(T=T_{c})$ takes its minimal value. Obviously $T_{c}\sim O(200~GeV)$ defines the phase transition where the chiral symmetry restoration will occur. Since in this phase all the masses tend to zero and $\alpha\rightarrow 0$ at $T>T_{c}$ one should expect the sharp expansion of the region with $L_{st}(T>T_{c})\rightarrow \infty$.
The qualitative relation between $R$ and $L_{st}$ above mentioned is the only one we can emphasize in order to explain the mass dependence of the source size.
Conclusions
===========
To summarize: the theoretical proposal for two-particle Bose-Einstein correlation function in case of $Z^{0}Z^{0}$ pairs in $pp$ collisions is carried out for the first time.
The correlations of two bosons in 4-momentum space presented in this paper offer useful and instructive complimentary viewpoints to theoretical and experimental works in multiparticle femtoscopy and interferometry measurements at hadron colliders.
We find the time dependence of correlation function calculated in time-dependent external field provided by the operator $r(\vec p, t)$ and the chaotic coherence function $\alpha(m,\beta)$. The result can be compared with the static correlation functions (see, e.g., \[14\] and the references therein mainly devoted to heavy-ion collisions) and also can be used for experimental data fitting.
The stochastic scale $L_{st}$ decreases with increasing temperatures slowly at low temperatures, and it decreases rather abruptly when the critical temperature is approached.
Our results first predicted for correlation radius $R$ are both$ Z^{0}$ boson mass and lepton energy dependent $$\label{e411}
R \sim \frac{e^{m/{5T_{0}}}}{\alpha ^{1/5}\,{\vert\vec k_{l}\vert}^{2/5}
\,m^{3/10}\, T_{0}^{3/10}}$$ for low values of $T_{0} < m$, while for higher temperatures, $T_{0} > \sqrt {m^{2} + \mu^{2}}$, one has $$\label{e412}
R \sim \frac{1}{\alpha ^{1/5}\, {\vert\vec k_{l}\vert}^{2/5}\, T_{0}^{3/5}}.$$ The theoretical correlation radius $R$ at temperature $T_{0}$ decreases as $Z^{0}$-boson momentum increases. Both estimations (\[e411\]) and (\[e412\]) serve as the first approximation to explain the experimental data at different $\sqrt s$ and hence at $T$. We claim that the experimental measuring of $R$ (in $fm$) can provide the precise estimation of the effective temperature $T_{0}$ which is the main thermal character in the $Z^{0}Z^{0}$ pair emitter source (given by the effective dimension $R$) in the proper leptonic decaying channel $Z^{0}Z^{0}\rightarrow l\bar ll\bar l$ with the final lepton energy $\sqrt{\vec k^{2}_{l} + m^{2}_{l}}$ at given $\alpha$ fixed by $C_{2}(Q=0)$ and $\langle N\rangle$. Actually, $T_{0}$ is the true temperature in the region of multiparticle production with dimension $R = L_{st}$, because at this temperature it is exactly the creation of two particles ($Z^{0}Z^{0}$) occurred, and these particles obey the criterion of BEC.
[30]{}
R.M.Weiner, Phys. Rep. [**327**]{} (2000) 249. R.Hanbury-Brown and R.Q.Twiss, Nature [**178**]{} (1956) 1046; G.A.Kozlov, O.V.Utyuzh and G.Wilk, Phys. Rev. [**C68**]{} (2003) 024901. G.A.Kozlov, Phys. Rev. [**C58**]{} (1998) 1188; J. Math. Phys. [**42**]{} (2001) 4749 and New J. of Physics [**4**]{} (2002) 23.1; G.A.Kozlov, “BEC and the particle mass”, hep-ph/0512184; G.A.Kozlov, J. Elem. Part. Phys. Atom. Nucl. [**36**]{} (2005) 108; G.A. Kozlov, O. Utyuzh, G. Wilk and Z. Wlodarczyk, “Some forgotten features of the Bose-Einstein correlations”, hep-ph/0710.3710. G.A. Kozlov, “BEC and the particle mass”, hep-ph/0512184. P.Langevin, C. R. Acad. Sci. [**146**]{} (1908) 530. The ALEPH Collab., Eur. Phys. J. [**C36**]{} (2004) 147; Phys. Lett. [**B611**]{} (2005) 66. The DELPHI Collab., Phys. Lett. [**B379**]{} (1996) 330. G.Abbiendi et al. (OPAL Collab.), Phys. Lett. [**B559**]{} (2003) 131. The ZEUS Collab., Phys. Lett. [**B583**]{} (2004) 231. T. Alexopoulos et al., \[E735 Collaboration\], Phys. Rev. [**D48**]{} (1993) 1931.
G.Goldhaber et al., Phys. Rev. Lett. [**3**]{} (1959) 181; G.Goldhaber et al., Phys. Rev. [**120**]{} (1960) 300. N.N. Bogolyubov, “Quasiaveragies in problems of statistical mechanics”, JINR report D-781, JINR, Dubna (1961). G.A. Kozlov, “Bose-Einstein correlations and the stochastic scale of light hadrons emitter source”, hep-ph/arXiv:0801.2072. C.Y. Wong, W.N. Zhang, Phys. Rev. [**C76**]{} (2007) 034905.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We investigate shear and buoyancy instabilities in radially stratified, magnetized, cylindrical flows, for application to magnetocentrifugally driven winds - such as those from protostars - and to magnetized accretion disks. Our motivation is to characterize the susceptibility of cold MHD disk winds to growing internal perturbations, and to understand the relation of wind instabilities to known accretion disk instabilities. Using four different linear analysis techniques, we identify and study nine principal types of unstable or overstable disturbances, providing numerical and analytic solutions for growth rates for a wide range of parameters.
When magnetic fields are predominantly [*toroidal*]{}, as in protostellar winds far from their source, we find the system is susceptible to growth of five different kinds of perturbations: axisymmetric fundamental and toroidal resonance modes, axisymmetric and non-axisymmetric toroidal buoyancy modes, and non-axisymmetric magnetorotational modes. Winds having a sufficiently steep field gradient ($d\ln{B}/d\ln{R} < -0.75$ for a purely toroidal-field case) are globally unstable to the long wavelength fundamental mode concentrated at small radii; these promote the establishment of narrow dense jets in the centers of wider winds. Long-wavelength outer-wind modes are all stable for power-law wind equilibria. The toroidal buoyancy instabilities promote small-scale radial mixing provided the equilibrium has nonzero magnetic forces. For low-temperature toroidal-${\bf B}$ winds, both axisymmetric and non-axisymmetric magnetorotational instabilities have very low growth rates. The stabilization of buoyancy instabilities by shear and of magnetorotational instabilities by compressibility may be important in allowing cold MHD winds to propagate over vast distances in space.
When magnetic fields are predominantly [*poloidal*]{}, as may occur in protostellar winds close to their source or in astrophysical disks, we find the system is susceptible to four additional growing modes: axisymmetric magnetorotational (Balbus-Hawley), axisymmetric poloidal buoyancy, non-axisymmetric geometric buoyancy, and poloidal resonance modes. The well-known axisymmetric Balbus-Hawley mode has the fastest growth rate. When the magnetic field is nonuniform, the axisymmetric poloidal buoyancy mode promotes radial mixing on small scales. The geometric poloidal buoyancy mode requires high $m$, thus is readily stabilized by shear.
Previous work on magnetorotational instabilities has concentrated on near-incompressible systems (accretion disks or stellar interiors). We extend this analysis to allow for compressibility (important in winds). We introduce a “coherent wavelet” technique (a WKB temporal approximation), and derive closed-form analytic expressions for instantaneous instability criteria, growth rates, and net amplification factors for generalized non-axisymmetric magnetorotational instabilities in compressible flows with both poloidal and toroidal fields. We confirm that these are in excellent agreement with the results of shearing-sheet temporal integrations, and that “locally-axisymmetric” perturbations have the largest amplifications only provided $(\bold{k}\cdot\bold{{v_{\rm A}}})/\Omega \lesssim 1$.
author:
- 'Woong-Tae Kim and Eve C. Ostriker'
title: 'Magnetohydrodynamic Instabilities in Shearing, Rotating, Stratified Winds and Disks'
---
Introduction
============
The ubiquity of energetic molecular outflows and atomic jets from young stellar objects (YSOs) ranging from deeply embedded infrared sources to classical T Tauri stars suggests that they are an inescapable by-product of star formation [e.g., @ric00 and references therein]. Winds from YSO disks play an important role in shedding angular momentum carried by inflowing material, thereby permitting further accretion in order for the central objects to attain stellar dimensions [@har82; @pud86; @shu88]. These winds may also have strong effects on the dynamical evolution of the parent cloud by providing a source of turbulent energy [@nor80], and may help to determine the final masses of stars by reversing the infall of surrounding gas [@shu87]. Therefore, understanding the physics of protostellar winds is of essential importance to the theory of star formation.
Among the various scenarios regarding the origin and nature of the protostellar winds, the most promising is magnetohydrodynamic (MHD) models in which winds are driven by the interaction of the centrifugal force with open magnetic fields threading rapidly rotating disks. These magnetocentrifugally driven wind models can account for observed high mass and momentum losses [@lad85; @bac96] and the kinematic and structural characteristics of bipolar molecular outflows in general [@li96], as well as in specific cases [e.g., HH111, cf. @nag97]. It is, however, still controversial whether the wind originates only from a small magnetosphere-disk interaction region near the central star [@shu94; @shu00], or whether it emanates from an extended region of the disk (following the seminal model of Blandford & Payne 1982; see, e.g., K$\ddot{\rm o}$nigl & Pudritz 2000).
Although the role of magnetic fields in driving protostellar winds is by now well established (at least theoretically), their complementary role in governing the stability properties of winds is less well explored. In addition, azimuthal and vertical shear within winds may also affect their stability properties. Questions of wind stability are potentially important for both large scale and small-scale phenomena. These include understanding the role of magnetic fields and velocity shear in (a) helping winds to propagate over enormous distances (up to a factor $\sim 10^6$ in dynamic range) through the ISM in parsec-scale giant HH flows [@rei97]; (b) creating bright HH “knots” spaced throughout optical jets [@har00]; (c) governing the angular extent of the emergent wind and establishing the momentum distribution for driving molecular outflows; and (d) converting large-scale ordered flow energy to jet heating through small-scale instabilities. In addition, it is important to study the dynamical properties of large classes of theoretical wind solutions to test whether they are stable equilibria which can represent real astronomical systems, or whether they are unstable equilibria which should rarely be observed in nature because they evolve rapidly into other configurations.
Previous studies of time-dependent behavior in steadily-input MHD winds have focussed primarily on the Kelvin-Helmholtz instabilities driven by the interaction of the wind surface layer with the ambient medium [see, e.g., the studies of @app92; @har92; @ros99 and references therein]. Generally, heavy jets containing strong toroidal fields are relatively resistant to these instabilities. In addition to “driven” instabilities resulting from boundary conditions, winds may also be subject to “free” instabilities in their interiors. Whether a given wind solution is internally unstable must depend on the velocity shear, magnetic geometry, and internal stratification. One route to studying internal wind instabilities is via time-dependent numerical simulations of winds. Although such studies have yielded intriguing results on the development of episodic knots in MHD winds, the computational demands in carrying out simulations precludes extensive exploration of parameter space, large spatial dynamic range, or very long-term integration. In addition, some of the time-dependent internal features found in simulations may be introduced by particular choices of inflow boundary conditions that are inconsistent with a steady-state flow, rather than occurring as a result of intrinsic instability of the wind.
Due to the importance of gaseous accretion disks in a wide variety of astrophysical systems, major attention has focussed on disk dynamics, and, in particular, the role of instability-driven turbulence in angular momentum transport . Saturated magnetorotational instabilities [hereafter MRIs; @bal91; @bal92; @haw91; @haw95] represent perhaps the most important local dynamical process affecting disk evolution. Magnetized disk winds share many generic properties with disks, so it is interesting to investigate the potential importance of MRIs in winds.
In this work, we investigate the internal stability of rotating, magnetized protostellar winds to (primarily) local shear and buoyancy modes. We also extend previous studies of local MRIs in accretion disks. The fundamental difference between “wind” and “disk” systems in our idealized models is in the absence or presence of gravity as a confining force. These systems may also be distinguished by the geometry of the magnetic field, with toroidally-dominant fields expected in the wind case, but either poloidal or toroidal fields possible for the disk case. Our most general analysis and results apply to cold flows, but we also perform separate calculations (see §7) including thermal effects which specialize to local analysis of MRIs.
Whether a wind originates from a narrow boundary layer or an extended radial region, the radial expansion of the flow will lead to shear in the azimuthal velocity field of the asymptotic state. The total specific angular momentum of the flow, $J = R(v_\phi - B_{\rm pol}B_\phi/(4\pi\rho v_{\rm pol}))$, is conserved along streamlines (where $B_{\rm pol}$, $B_\phi$ and $v_{\rm pol}$, $v_\phi$ are the poloidal, toroidal components of the magnetic field and the flow velocity). If the [Alfv$\acute{\rm e}$n]{} mach number $M_{\rm A}\equiv v_{\rm pol}/v_{\rm A,pol} \gg 1$, the kinetic part dominates the specific angular momentum and $v_\phi \approx J/R$. Thus, the angular velocity $\Omega =
v_\phi/R \approx J/R^2$ in the asymptotic wind will have a gradient $d\ln \Omega/d\ln R= d \ln J /d\ln R - 2$. If the wind comes from a boundary layer, then $d \ln J /d\ln R$ may be small; if the wind originates from a large region with self-similar scalings, then $d \ln J /d\ln R = 1/2$. In either case, $d\ln \Omega/d\ln R$ is expected to be a negative, order-unity quantity for wind systems. For thin disk systems (i.e., negligible pressure support), a Keplerian radial profile $d\ln \Omega/d\ln R = -3/2$ is expected if the central mass is dominant. For the analysis of this paper (except where noted otherwise), we adopt $d\ln \Omega/d\ln R=-3/2$ for both “wind” and “disk” systems, but our qualitative results are insensitive to this assumption. As discussed below, order-unity radial logarithmic gradients may also be expected in the magnetic field strengths; we allow for a range of magnetic gradients.
Significant shear may also exist in the poloidal velocities of jets if they originate from an extended radial region. The asymptotic outflow speed $v_{\rm 0z}$ generally scales linearly with the rotational speed at the footpoint $R_{\rm foot}$ of the streamline, so that $$\frac{\partial \ln v_{\rm 0z}}{\partial \ln R}=
\frac{\partial \ln v_{\rm 0z}}{\partial \ln v_{\rm foot}}
\frac{\partial \ln v_{\rm foot}}{\partial \ln R_{\rm foot}}
\frac{\partial \ln R_{\rm foot}}{\partial \ln R}
\sim -\frac{1}{2} \frac{\partial\ln R_{\rm foot}}{\partial \ln R}.$$ If the range of footpoint radii is small compared to the range of asymptotic radii (as for a wind from a boundary layer), then $|\partial\ln v_{\rm 0z}/\partial\ln R| \ll 1$; if the radial ranges are comparable, then $\partial\ln v_{\rm 0z}/\partial\ln R$ is negative and order unity. For disks, the vertical velocity shear is negligible. We allow for a range of vertical shear rates in the present analysis.
Our analysis consists of developing and solving sets of linearized MHD equations for several general classes of background flows. Both axisymmetric and non-axisymmetric disturbances are explored. Even in linearized form, MHD problems present considerable technical challenges. Thus, instead of attacking sophisticated sets of steady-state wind solutions, for many specific examples we will take as an unperturbed configuration one of the power-law cylindrical equilibrium solutions recently identified by @ost97 as asymptotic states of self-similar disk winds. These have density $\rho \propto R^{-q}$, $\bold{B}
\propto R^{-(1+q)/2}$, $\bold{v}
\propto R^{-1/2}$, and sound speed ${c_{\rm s}}=0$. Although these adopted initial configurations are relatively simple, they retain general asymptotic characteristics of MHD disk winds in the sense that they have both azimuthal and vertical magnetic field and velocity components with arbitrary ratios, and all the physical variables have radial gradients. These gradients may also be thought of as representing local scalings within a more complex overall stratification. We also include models without vertical motion but with significant equilibrium gravity to study stability in magnetized astrophysical disks. In our analysis of MRIs, we use equilibria with uniform $\bold{B}, \rho$, and ${c_{\rm s}}\neq 0$, and take $\Omega\propto R^{-a}$ with arbitrary $a$.
Because the systems we are studying contain significant azimuthal shear, an arbitrary initial spatial planform is not maintained indefinitely. When $\Omega'\neq 0$ and the azimuthal wavenumber $m\neq 0$ (and/or when $v_{\rm 0z}'\neq0$ and the vertical wavenumber ${k_{\rm z}}\neq0$; the prime represents a differentiation with respect to $R$), spatial wavefunctions may become increasingly radially corrugated in time due to the kinematic shearing of the wavefronts imposed in the initial conditions. If we describe the radial spatial wavefunction in terms of the amplitudes of Fourier coefficients with radial wavenumbers ${k_{\rm R}}$, this corresponds at late times to a secular increase in the amplitudes of large-${k_{\rm R}}$ terms and decrease in the amplitudes of small-${k_{\rm R}}$ terms. As we shall show, all the disturbances we identify are stabilized at sufficiently large ${k_{\rm R}}$. Thus, if $m\Omega'$ and/or ${k_{\rm z}}v_{\rm 0z}' \neq 0$, the net amplification factor for any arbitrary initial disturbance is limited by the rate of kinematic growth of radial corrugation compared to the growth rate of any dynamically-driven instabilities.
In previous work, two complementary analytical methods have been used to study small-amplitude disturbances in shearing astrophysical systems. One approach adopts the “shearing-sheet” formalism, and integrates the local, time-dependent, linearized wave equations directly to obtain the evolutionary behavior of shearing wavelets treated as an initial-value problem [@gol65; @jul66; @bal92]. An alternative analytical approach uses WKB techniques to derive dispersion relations for spatial Fourier modes, superpositions of which represent local wavefunctions [e.g., @shu74; @shu92; @ryu92; @ter96]. This paper includes analyses using both approaches, and also introduces a hybrid technique which we term a “coherent wavelet” analysis. We adopt the “modal” strategy in order to identify characteristic instantaneous growth rates and physical mechanisms for a wide variety of spatial disturbances. By considering the time over which a spatial pattern is altered by shear, we can estimate the net amplification factor of a given initial modal disturbance. We use the shearing-sheet formalism for studying magnetorotational instabilities, which are cut off at relatively small values of $R{k_{\rm R}}/m$ (whereas the modal analysis applies to large $R{k_{\rm R}}/m$), and also for studying buoyancy instabilities in the high-$m$ regime where modes are short-lived. We show that the results obtained from the shearing-sheet integrations in both cases are in excellent agreement with the predictions of a coherent wavelet analysis, in which time-dependent growth rates $\gamma(t)$ are computed by time-localizing the shearing-sheet equations and solving an analytic dispersion relation.
The organization of this paper is as follows: We begin by studying instabilities in cold, magnetized winds. In §2, the basic MHD equations and the specific adopted wind equilibria are described. In §3, we analyze the stability of winds to the simplest perturbation with ${k_{\rm z}}=m=0$, where ${k_{\rm z}}$ and $m$ are respectively the vertical and azimuthal wavenumbers of the perturbation. We term these the “fundamental modes”; we present solutions for stable and unstable global modes under the assumption of free Lagrangian boundary conditions. The modal analysis and general local dispersion relation for cold flows with arbitrary ${k_{\rm z}}$ and $m$ are presented in §4. We present numerical solutions of the dispersion relation for both axisymmetric and non-axisymmetric perturbations in §5. In §6, we classify the unstable or overstable modes and provide the physical interpretation for each mode. Next, we include (variable) thermal pressure terms to compare the susceptibility of cold winds vs. warm disks to shear instabilities. In §7 we analyze the axisymmetric Balbus-Hawley instability of poloidal fields and the non-axisymmetric MRI of toroidal fields, discuss the respective instability mechanisms, and provide the corresponding instability criteria. Here, we use the coherent wavelet technique to compute growth rates, and compare with direct shearing-sheet integrations. The generalized instability criteria and net amplification factors for the magnetorotational disturbances with both toroidal and poloidal background fields are also derived. Finally in §8, we summarize and discuss conclusions of the present work.
Basic Equations and Cylindrical Equilibrium for Cold Wind
=========================================================
We begin with the ideal MHD equations $$\frac{\partial\rho}{\partial t} + \nabla\cdot (\rho \bold{v} ) = 0,$$ $$\frac{\partial\bold{v}}{\partial t} + \bold{v}\cdot\nabla\bold{v}
= \frac{1}{4\pi\rho} (\nabla\times\bold{B})\times\bold{B}
-\frac{\nabla P}{\rho} - \nabla\Phi_{\rm G},$$ $$\frac{\partial\bold{B}}{\partial t} = \nabla\times(\bold{v}\times\bold{B}),$$ and $$\nabla\cdot\bold{B} = 0,$$ where $\rho$ is the density, $\bold{v}$ is the fluid velocity, $\bold{B}$ is the magnetic field, $P$ is the thermal pressure, and $-\nabla \Phi_{\rm G}
\equiv -\bold{g}$ is the gravitational force due to a central object. We ignore self-gravity in the flow.
We now consider cold, magnetized cylindrical flows. Since the flow velocity in disk winds is always supersonic except in the vicinity of the disk where the material is lifted by the thermal pressure [@bla82], the thermal pressure term in eq. (2) can generally be neglected compared to magnetic stress. Except for investigations of generalized MRIs (§7), we shall drop the thermal pressure term. We adopt standard cylindrical coordinates $(R,\phi,z)$.
By assuming that ${v_{\rm R}}={B_{\rm R}}=0$ and all variables are independent of $z$, we have a general equilibrium condition from eq. (2) $$\Omega^2R \equiv \frac{v_\phi^2}{R} = \frac{1}{4\pi\rho}
\left(\frac{B_\phi^2}{R} + \bold{B} \cdot
\bold{B}'\right) + g_{\rm R},$$ where a prime denotes differentiation with respect to $R$. At a large distance from the origin, the gravitational force due to the central source can also be ignored on the grounds that magnetic and centrifugal forces far exceed it. In this case, the magnetic hoop stress acting inward is the only force that balances the outward centrifugal force and outward magnetic pressure gradient force (under the assumption that the magnetic field strength decreases outward).
As an initial equilibrium configuration of the wind, for specific cases we will adopt the asymptotic solutions for cylindrically symmetric axial flows presented by @ost97. All variables have power-law dependences on $R$: $\rho \propto R^{-q}, B_{\phi} \propto B_{\rm z} \propto R^{-(1+q)/2}$, and $v_{\phi} \propto v_{\rm z} \propto R^{-1/2}$. We define the local pitch angle $i$ of the magnetic fields as $i \equiv \tan^{-1}{(B_{\rm z}/B_{\phi})}$. Neglecting $g_{\rm R}$ in eq. (5), radial momentum balance requires $$v_{\phi}^2 = \frac{v_{\rm A}^2}{2}\left(\cos{2i} - q \right),$$ where $v_{\rm A}$ is the local Alfv$\acute{\rm e}$n speed defined by $$v_{\rm A}^2 \equiv v_{\rm A\phi}^2 + v_{\rm Az}^2\;\;\;\;
{\rm with} \;\;\;\; v_{\rm A\phi}\equiv
\frac{B_\phi}{\sqrt{4\pi\rho}}\;\;\;\;
{\rm and}\;\;\;\;v_{\rm Az}\equiv\frac{B_{\rm z}}{\sqrt{4\pi\rho}}.$$
It is obvious from eq. (6) that there would be no such power-law solutions if $q>1$. This is because when $q>1$, the magnetic field has so steep a gradient that the corresponding pressure force always exceeds the tension. Therefore, to ensure force balance and cylindrical collimation in winds with power-law profiles, the magnetic field strength must decline with $R$ more slowly than $R^{-1}$.
We can define an angular velocity ${\Omega_{\rm f}}\equiv
\Omega-v_{\rm pol}B_\phi/(RB_{\rm pol})$ as that of a rotating frame in which the flow of winds is parallel to the local field line. ${\Omega_{\rm f}}$ is the rotation rate of the magnetic field lines thought of as rigid wires. In such a frame, the family of solutions can be completely described in terms of scaled values of the specific angular momentum $j$, the Bernoulli constant $e$, and $q$, where $$j \equiv \frac{\Omega}{{\Omega_{\rm f}}}\left( 1 -
\frac{({v_{\rm A\phi}}/R\Omega)^2}{1- {\Omega_{\rm f}}/\Omega}\right)\;\;\;\;{\rm and}
\;\;\;\;
e\equiv \frac{1}{(R{\Omega_{\rm f}})^2}\left(
\frac{1}{2}\bold{v}^2 + \Phi_{\rm G} - R^2{\Omega_{\rm f}}\Omega \right),$$ [@ost97]. The condition for a super-Alfv$\acute{\rm e}$nic outflow velocity requires $0 < j \leq 1$. Generally speaking, the pitch angle $i$ does not depend on $q$, although one can parameterize $i$ in terms of $q$, $e$, and $j$. However, for flows originating from a Kepler-rotating disk, angular momentum and energy conservation requirements limit the range of $i$ available to an equilibrium (asymptotic) magnetic field configuration. Utilizing eqs. (15) to (23) of @ost97 one can show that the maximum, over all permitted values of $e$ and $j$, pitch angle $i_{\rm max}$ is given by $$\tan^2 i_{\rm max} = \left(\frac{4}{3+q}\right)^2 - 1,$$ which is attained when $e=0$ and $j=1$. If $j>1$, the streamline never reaches the Alfv$\acute{\rm e}$n radius.
Fundamental Mode
================
Dynamical Equations
-------------------
We first consider the response of the equilibrium state when small, axisymmetric perturbations with an infinite wavelength along a vertical direction are imposed. We term the waves with ${k_{\rm z}}=0$ and $m=0$ the “fundamental modes” analogous to eigenfunctions of oscillations without any node. Let the subscripts 0 and 1 denote the equilibrium and perturbed states, respectively. Linearizing the set of the dynamical equations (1) to (4), we may write $$\frac{\partial \rho_1}{\partial t} =
-\frac{1}{R}\frac{\partial}{\partial R} \left(R {\rho_{\rm 0}}v_{\rm 1R}\right),$$ $$\frac{\partial v_{\rm 1R}}{\partial t}
= 2\Omega v_{1\phi} - \frac{1}{4\pi{\rho_{\rm 0}}}
\left\{
\frac{2{B_{\rm 0\phi}}B_{1\phi}}{R} + \frac{\partial}{\partial R}
\left( \bold{B}_{\rm 0} \cdot \bold{B}_1 \right)
-\frac{\rho_1}{{\rho_{\rm 0}}}
\left( \frac{{B_{\rm 0\phi}}^2}{R} + \bold{B}_{\rm 0} \cdot \bold{B}_{\rm 0}'
\right) \right\},$$ $$\frac{\partial v_{1\phi}}{\partial t} =
-\frac{\kappa^2}{2\Omega}v_{\rm 1R},$$ $$\frac{\partial v_{\rm 1z}}{\partial t} =
-v_{\rm 0z}' v_{\rm 1R},$$ $$\frac{\partial B_{1\phi}}{\partial t} =
-\frac{\partial}{\partial R}\left(B_{\rm 0 \phi} v_{\rm 1R}\right),$$ $$\frac{\partial B_{\rm 1z}}{\partial t} =
- \frac{1}{R}\frac{\partial}{\partial R}\left(
RB_{\rm 0z}v_{\rm 1R}\right),$$ and $B_{\rm 1R} = 0$. In eq. (10), $\kappa$ stands for the epicycle frequency $$\kappa^2 \equiv \frac{1}{R^3}\frac{d}{dR}(R^4\Omega^2).$$
Combining the perturbed equations (8)$-$(13) and eliminating all other variables in favor of the perturbed radial velocity ${v_{\rm 1R}}$, one obtains the wave equation
$$\begin{aligned}
\frac{1}{v_{\rm A}^2}\frac{\partial^2{v_{\rm 1R}}}{\partial t^2} &=&
\frac{\partial^2{v_{\rm 1R}}}{\partial R^2}
+ \frac{d\ln{(R{B_{\rm 0}}^2)}}{dR} \frac{\partial{v_{\rm 1R}}}{\partial R}
\nonumber \\
&-& \left[
\frac{\kappa^2}{v_{\rm A}^2} + \frac{1}{R^2}\left(
1 + \cos^2{i}\frac{d\ln{{\rho_{\rm 0}}}}{d\ln{R}}\right)
- \frac{{\rho_{\rm 0}}}{R{B_{\rm 0}}^2}\frac{d}{dR}
\left(\frac{R}{{\rho_{\rm 0}}}\bold{B}_0\cdot\bold{B}_0'\right)
\right]
{v_{\rm 1R}}.\end{aligned}$$
For power-law profiles, this becomes $$\frac{1}{v_{\rm A}^2}
\frac{\partial^2{v_{\rm 1R}}}{\partial t^2} =
\frac{\partial^2{v_{\rm 1R}}}{\partial R^2} -
\frac{q}{R}
\frac{\partial{v_{\rm 1R}}}{\partial R} -
\left[\frac{\kappa^2}{v_{\rm A}^2} + \frac{1}{R^2}
\left(\frac{1-q}{2} - q\cos^2i
\right)
\right]
{v_{\rm 1R}}.
\eqnum{14b}$$
To transform eqs. (14) to the ${\rm Schr\ddot{o}dinger}$ form, we define a new independent variable $\Psi$ through $${v_{\rm 1R}}= \frac{\Psi(R)}{\sqrt{R{B_{\rm 0}}^2}}e^{i\omega t};\;\;\;{\rm or}\;\;\;
{v_{\rm 1R}}= R^{q/2}\Psi(R)e^{i\omega t}$$ for the power-law case. Then, we have $$\frac{d^2\Psi}{dR^2} + K^2(R)\Psi = 0,$$ with $K(R)$ defined by $$K^2(R) \equiv \frac{\omega^2-\kappa^2}{{v_{\rm A}}^2}
-\frac{3}{4R^2}
-\frac{d\ln{{\rho_{\rm 0}}}}{dR}\left(\frac{\cos^2{i}}{R} +
\frac{\bold{B}_0\cdot\bold{B}_0'}{B_0^2}\right)
+ \left(\frac{\bold{B}_0\cdot\bold{B}_0'}{B_0^2}\right)^2,
\eqnum{17a}$$ or $$K^2(R) \equiv \frac{\omega^2-\kappa^2}{{v_{\rm A}}^2} - \frac{1}{R^2}
\left(
\frac{1}{2} + \frac{q^2}{4} - q\cos^2i
\right)
\eqnum{17b}$$ for the power-law case. Local (WKB) solutions to eq. (16) have $\Psi \sim e^{i\int {k_{\rm R}}dR}$ with $R{k_{\rm R}}\gg 1$ and $d{k_{\rm R}}/dR \ll {k_{\rm R}}^2$. In this case, $\Psi'' \rightarrow - {k_{\rm R}}^2\Psi$, and we can use $K^2(R) = {k_{\rm R}}^2 $ to write a local dispersion relation $$\omega^2 = {v_{\rm A}}^2 {k_{\rm R}}^2,$$ which corresponds to MHD fast modes propagating along the radial direction. When $|\omega|^2$ is comparable to or smaller than ${v_{\rm A}}^2/R^2$, however, modes are not localized, and solutions must be sought as a global problem subject to boundary conditions.
Global Analysis for the Fundamental Modes
-----------------------------------------
In the previous section we showed that there is no short wavelength (local) unstable fundamental mode with ${k_{\rm z}}=m=0$ in self-similar MHD disk winds. Here, we present the results of a global normal-mode analysis performed with carefully chosen boundary conditions, and adopting the power-law equilibrium. Define the dimensionless radial variable $r \equiv R/{R_{\rm e}}$, and dimensionless parameters $\alpha \equiv (q-1)\cos^2{i} + q(2-q)/4$ and $\sigma^2 \equiv \omega^2 {R_{\rm e}}^2/v_{\rm A}^2({R_{\rm e}})$, with ${R_{\rm e}}$ being the position of the unperturbed outer edge of the wind. Then eq. (16) can be cast into the form $$\frac{d^2\Psi}{dr^2} + \left(\frac{\alpha}{r^2} + \sigma^2r\right)\Psi = 0.$$ It is not difficult to show that the general solutions of eq. (19) are $$\hskip 2.8cm
\Psi = \left\{ \begin{array}{ll}
A\sqrt{r}J_{\nu}(2\sigma r^{3/2}/3) + B\sqrt{r}Y_{\nu}(2\sigma r^{3/2}/3),
&{\rm if} \quad \sigma^2 > 0,
\hskip 2.15cm
{\rm (20a)}
\\
& \\
C\sqrt{r}I_{\nu}(2|\sigma| r^{3/2}/3) + D\sqrt{r}K_{\nu}(2|\sigma| r^{3/2}/3),
&{\rm if} \quad \sigma^2 < 0,
\hskip 2.15cm
{\rm (20b)}
\end{array} \right.$$ with $3\nu \equiv \sqrt{1-4\alpha} = \sqrt{(1-q)^2 + 4(1-q)\cos^2{i}}$. In eqs. (20), $J_{\nu}$, $Y_{\nu}$, and $I_{\nu}$, $K_{\nu}$ are the ordinary and modified Bessel functions of the 1st and 2nd kinds, respectively, and the coefficients $A, B, C$, and $D$ are constants to be determined from imposed boundary conditions.
Let us consider the case of a free Lagrangian boundary at which the total pressure due to initial and perturbed fields balances with a fixed external pressure at both inner and outer edges, which is equal to the unperturbed magnetic pressure. If the total pressure at an edge of an outflow is different from the external pressure, the boundary itself will move until a new balance exists. To first order, this condition of constant pressure at the boundary is written $$\frac{1}{2}\frac{dB_{\rm 0}^2}{dR}\frac{\partial R_{\rm b}}
{\partial t} + \bold{B}_{\rm 0}\cdot \frac{\partial \bold{B}_{\rm 1}}
{\partial t} = 0,$$ where $R_{\rm b}$ is the location of the perturbed boundary. The first term of eq. (21) represents the change in the total pressure due to the boundary movement, while the second term arises from perturbed magnetic pressure itself. All quantities are evaluated at the unperturbed boundary position. Using eqs. (12), (13), and (15), and using $\partial R_{\rm b}/\partial
t= v_{\rm 1R}$ at boundaries, we find the desired boundary conditions are $$\frac{d\Psi}{dr} + \frac{q/2+\sin^2i}{r}\Psi = 0, \;\;\;{\rm at}\;\;\;
r = r_{\rm i}\;\;{\rm and}\;\;1,$$ where $r_{\rm i} \equiv {R_{\rm i}}/{R_{\rm e}}$ is the normalized distance of an inner boundary from the axis. Together with the boundary conditions (22), eq. (19) forms a Sturm-Liouville system. By employing the variational principle one can show that $\sigma^2$ is real and that $\sigma^2(\Psi)$ is stationary subject to an arbitrary variation of $\Psi$.
When $\sigma^2 > 0$ (stable modes), the oscillatory properties of $J_{\nu}$ and $Y_{\nu}$ guarantee the existence of discrete eigenvalues $\sigma_{\rm n}$ with $n$ denoting the number of nodes in the corresponding eigenfunction $\Psi_{\rm n}$. The resulting eigenvalues for $r_{\rm i} = 10^{-1}$ and $10^{-4}$, and $0< i < i_{\rm max}(q)$ are plotted in Fig. 1. Only a few cases with small $n$ are shown. Eigenvalues associated with different $i$’s fill each shaded area completely. When $r_{\rm i}= 10^{-4}$, eigenfunctions which link inner and outer boundaries have to extend across enormous changes in density and magnetic field strengths. In this case, $B/A \ll 1$ and eigenvalues become rather insensitive to the local properties such as $q$ and $i$. When $r_{\rm i}=10^{-1}$, however, the wind mimics a slender hollow cylinder. The variation in density and field strengths over radius is slight, causing eigenvalues to be sensitive to $i$ and $q$. In addition, the narrow width of the wind changes the number of nodes in eigenfunctions. When $q>0.5$, for example, the eigenfunctions with $r_{\rm i}=10^{-1}$ have almost the same eigenvalues as, but one more node than, the $r_{\rm i}=10^{-4}$ case, as seen in Fig. 1.
When $r_{\rm i} \ll 1$ and $\sigma^2 \gg 1$, the asymptotic solutions to eqs. (20a) and (22) gives $ \sigma_{\rm n} = 3\pi n/2 + 3\pi(2\nu+1)/8$. These are plotted with dotted lines in Fig. 1b, and show good agreement with the values calculated without any assumption (even for $n=0$). The case with $q=0.5$ and $i=0$ is special, because the slope of the eigenfunction at the inner boundary is $-1/4$, which automatically satisfies the boundary condition (22). In this case the asymptotic eigenvalues are $\sigma_{\rm n}=3\pi n/2$, drawn as filled circles in Fig. 1b. Eigenvalues have no upper bound as $n\rightarrow \infty$, which is a general property of solutions to a Sturm-Liouville equation [@mor53].
Now consider the unstable global solutions with $\sigma^2<0$. Let $\psi_1$ and $\psi_2$ be the two linearly independent solutions of $\Psi$: $\psi_1 \equiv \sqrt{r}I_{\nu}(2|\sigma| r^{3/2}/3)$ and $\psi_2 \equiv \sqrt{r}K_{\nu}(2|\sigma| r^{3/2}/3)$ such that $\Psi = C\psi_1 + D\psi_2$. Because $\psi_1$ is a monotonically increasing function of $r$ (i.e., $\psi_1, \psi_1' > 0$ always) and $\Psi$ must have a negative logarithmic slope at the inner and outer boundaries (cf.eq. \[22\]), $\psi_1$ alone can not constitute eigenfunctions for global modes. In addition, since $\psi_1$ increases exponentially for a large value of $|\sigma|r^{3/2}$, while $\psi_2, \psi_2' \rightarrow 0$, the outer free Lagrangian boundary condition requires $C/D \rightarrow 0$. Although $C$ is not strictly zero when $|\sigma|$ has a relatively small value, the contribution of $\psi_1$ to global solutions near the inner boundary is negligibly small. Thus, unstable eigenvalues, if they exist, are essentially determined by the inner boundary condition imposed on $\psi_2$.
As we move inward from the outer boundary, $\psi_2$ rapidly increases asymptoting to $$\psi_2 \sim r^{(1-3\nu)/2}
\left[ 1 - \frac{\pi\nu |\sigma|^{2\nu}}{3^{2\nu}\sin(\pi\nu)\Gamma^2(\nu+1)}
r^{3\nu} + \cdot\cdot\cdot \right],$$ for $r \ll 1$ [cf. @abr65], where $\Gamma(\nu+1)$ is a Gamma function. In fact, $(1-3\nu)/2$ is the maximum logarithmic slope $\psi_2(r)$ can ever attain. From the inner boundary constraint (22), the existence of unstable global solutions is guaranteed if $(1-3\nu)/2 > -(q/2+\sin^2{i})$, or, equivalently $$q > 1 - \frac{(1+\sin^2{i})^2}{2},$$ is satisfied. Eq. (24) is the global instability criterion for the fundamental modes of self-similar, cold, magnetized winds, subject to the free boundary conditions expressed by eq. (22). By putting $C=0$ and neglecting higher order terms in $\psi_2$, we derive from eqs. (22) and (23) the approximate, analytic expression for the eigenvalues of the global instability $$|\sigma|r_{\rm i}^{3/2} =
\frac{|\omega| {R_{\rm i}}}{{v_{\rm A}}({R_{\rm i}})} =
3\left[\frac{\sin(\pi\nu)\Gamma^2(\nu+1)}{\pi\nu}
\frac{(1-3\nu+q+2\sin^2i)}{(1+3\nu+q+2\sin^2i)}\right]^{1/2\nu},$$ which again shows that $|\sigma|$ has a positive real value if eq. (24) holds.
In Fig. 2 we plot the approximate growth rates for unstable modes from eq. (25) as dotted lines, as well as the exact growth rates numerically computed for $r_{\rm i}=10^{-4}$ (thin solid lines) and for $r_{\rm i}=0.1$ (dashed lines). The curves shown are for $i=0^{\rm o}, 5^{\rm o},\cdot\cdot\cdot,35^{\rm o},
40^{\rm o}$ from right to left, and the uppermost thick lines are for $i_{\rm max}$ calculated from eq. (7). Note that varying the width of outflow via $r_{\rm i}$ yields very little change in the plotted solutions: $r_{\rm i}$-dependence of the growth rates appears mainly through the product $|\sigma|r_{\rm i}^{3/2}$. Eq. (25) gives accurate growth rates for relatively small values of $|\sigma| r_{\rm i}^{3/2}$, while its estimates deviate up to $\sim 16\%$ from the exact values as $|\sigma| r_{\rm i}^{3/2}$ becomes larger. In this case we need to include next order terms in $\psi_2$ (cf. eq. \[23\]) to obtain more accurate results. For a given set of equilibrium parameters, we note that whereas there exist an infinite set of stable eigenmodes, there is (at most) a unique unstable eigenmode. Noting $|\omega| \equiv |\sigma|v_{\rm A}({R_{\rm e}})/{R_{\rm e}}= |\sigma|r_{\rm i}^{3/2}
v_{\rm A}(R_{\rm i})/R_{\rm i}$, we expect from Fig. 2 that the system is typically globally unstable within $\sim$5 crossing times of [Alfv$\acute{\rm e}$n]{} waves at the inner boundary.
Once the ratios of the coefficients and the eigenfrequencies are found for the fundamental modes, one can easily construct radial solutions for the perturbed variables: $\rho_1$, ${v_{\rm 1R}}$, $v_{1\phi}$, and $B_{1\phi}$. These are plotted in Fig. 3 for $i=0^{\rm o}$ and $r_{\rm i}=10^{-3}$. Fig. 3a corresponds to a stable case with $q=0.4$ and $\sigma_0=2.42$, while Fig. 3b depicts an unstable case with $q=0.6$ and $|\sigma|r_{\rm i}^{3/2}=0.11$. Although normalization is arbitrary, we note that for the unstable modes, the negative radial velocity case drives the entire system into a more stable configuration (with lower magnetic energy) when the equilibrium magnetic field is predominantly toroidal. This can be shown as follows: Let $\delta M$, $\delta E_{\rm B}$, and $\delta \Phi_{\rm B}$ denote the mass, magnetic energy, and toroidal magnetic flux per unit height in a local flux tube. Then we have $\delta E_{\rm B} = (\pi/2)(\delta \Phi_{\rm B}/\delta M)^2 \delta M
\rho R^2$. For a given flux tube, $\delta M$ and $\delta \Phi_{\rm B}/\delta M$ are constant in time and $\delta M>0$. Thus, ${\rm sgn}\; d(\delta E_{\rm B}) /dt = {\rm sgn} \; d(\rho R^2)/dt =
{\rm sgn}\; [\rho_0 R^2 ({v_{\rm 1R}}/R - \partial {v_{\rm 1R}}/\partial R)]$ from the equation of continuity. If ${v_{\rm 1R}}/R$ dominates $\partial {v_{\rm 1R}}/\partial R$ and ${v_{\rm 1R}}<0$, then ${\rm sgn}\; d(\delta E_{\rm B}) /dt <0$; magnetic energy decreases with time, meaning that the system evolves into a more stable state. Thus we scale $v_{\rm 1R}/v_{0\phi} = 1$ at $r=1$ for Fig. 3a and $v_{\rm 1R}/v_{0\phi} = -1$ at $r=r_{\rm i}$ for Fig. 3b, respectively. Note that stable eigenfunctions have their largest amplitude near the outer boundary, while the inner, high density region is nearly static during oscillation. Unstable eigenfunctions, on the other hand, are almost zero except in the region close to the inner boundary. The respective inner-region vs. outer-region predominance of unstable vs. stable eigenfunctions reflects the respective characteristic frequencies as well: the inner-wind unstable modes grow at large rates comparable to [Alfv$\acute{\rm e}$n]{} frequencies in the interior, whereas outer-wind stable modes oscillate at low frequencies comparable to the [Alfv$\acute{\rm e}$n]{} frequencies in the exterior.
We remark that there is no globally unstable fundamental mode when one adopts rigid boundaries with $\Psi(r)=0$ at both $r=r_{\rm i}$ and $r=1$, instead of the free Lagrangian boundaries, since both $\psi_1$ and $\psi_2$ are monotonic functions of $R$. If $\Psi'(r)=0$ is imposed at both boundaries [cf. @dub93], however, we still have unstable fundamental modes with a different instability criterion[^1] and different growth rates.
We discuss the significance of fundamental modes to protostellar outflows in §8.3.
Local Analysis for Cold Winds
=============================
We now consider general non-axisymmetric Eulerian perturbations with small amplitudes. Neglecting the effects of thermal pressure and external gravity due to a central object, we linearize eqs. (1)$\sim$(4) $$\frac{d\rho_1}{dt} = - \frac{1}{R}\frac{\partial}{\partial R}(R\rho_0 {v_{\rm 1R}})
- \frac{1}{R}\frac{\partial}{\partial\phi}(\rho_0 v_{1\phi})
- \frac{\partial}{\partial z}(\rho_0 v_{1z}),$$ $$\frac{d{v_{\rm 1R}}}{dt} = 2\Omega v_{1\phi}
+\frac{1}{4\pi\rho_0}\left[
-\frac{2{B_{\rm 0\phi}}B_{1\phi}}{R}
+ \frac{{B_{\rm 0\phi}}}{R}\frac{\partial B_{\rm 1R}}{\partial \phi}
+ {B_{\rm 0z}}\frac{\partial B_{\rm 1R}}{\partial z}
-\frac{\partial}{\partial R} (\bold{B}_0\cdot\bold{B}_1)
+\frac{\rho_1}{\rho_0}
\left(\frac{{B_{\rm 0\phi}}^2}{R} + \bold{B}_0 \cdot \bold{B}_0' \right)
\right],$$ $$\frac{dv_{1\phi}}{dt} = - \frac{\kappa^2}{2\Omega}{v_{\rm 1R}}+\frac{1}{4\pi\rho_0}\left[
\left({B_{\rm 0\phi}}'+\frac{{B_{\rm 0\phi}}}{R}\right)B_{\rm 1R}
+{B_{\rm 0z}}\left(\frac{\partial B_{1\phi}}{\partial z}
- \frac{1}{R}\frac{\partial B_{\rm 1z}}{\partial \phi}\right)
\right],$$ $$\frac{dv_{\rm 1z}}{dt} = - v_{\rm 0z}' {v_{\rm 1R}}+\frac{1}{4\pi\rho_0}\left[
{B_{\rm 0z}}'B_{\rm 1R} + {B_{\rm 0\phi}}\left(
\frac{1}{R}\frac{\partial B_{\rm 1z}}{\partial \phi}
-\frac{\partial B_{1\phi}}{\partial z}\right)
\right],$$ $$\frac{dB_{\rm 1R}}{dt} = \frac{{B_{\rm 0\phi}}}{R}\frac{\partial {v_{\rm 1R}}}{\partial \phi}
+{B_{\rm 0z}}\frac{\partial {v_{\rm 1R}}}{\partial z},$$ $$\frac{dB_{1\phi}}{dt} = R\Omega'B_{\rm 1R}
-\frac{\partial}{\partial R}({B_{\rm 0\phi}}{v_{\rm 1R}})
+{B_{\rm 0z}}\frac{\partial v_{1\phi}}{\partial z}
-{B_{\rm 0\phi}}\frac{\partial v_{\rm 1z}}{\partial z},$$ $$\frac{dB_{\rm 1z}}{dt} = v_{\rm 0z}' B_{\rm 1R}
-\frac{1}{R}\frac{\partial}{\partial R}(R{B_{\rm 0z}}{v_{\rm 1R}})
+\frac{{B_{\rm 0\phi}}}{R}\frac{\partial v_{\rm 1z}}{\partial \phi}
-\frac{{B_{\rm 0z}}}{R}\frac{\partial v_{1\phi}}{\partial \phi},$$ where the Lagrangian time derivative is denoted by $$\frac{d}{dt} \equiv \frac{\partial}{\partial t}
+ \Omega(R)\frac{\partial}{\partial \phi}
+ v_{\rm 0z}\frac{\partial}{\partial z},$$ and again the subscripts 0 and 1 indicate the equilibrium and perturbation variables, respectively.
Since all the coefficients of the perturbed variables in eqs. (26)$\sim$(32) do not depend on the coordinates $\phi$ and $z$, we may look for solutions having sinusoidal dependence on $\phi$ and $z$. Furthermore, if there exist any normal modes, we can write eigenfunctions in the form $$\chi_1(R,\phi,z,t) = \chi_1(R)e^{i(m\phi + {k_{\rm z}}z - \omega t)},$$ where $\chi_1$ refers to any physical variable of perturbations. Substituting eq. (34) into the set of eqs. (26)$\sim$(32) and eliminating all other variables in terms of the radial Lagrangian displacement ${\xi_{\rm R}}\equiv - {v_{\rm 1R}}/i{\tilde{\omega}}$ with a Doppler shifted frequency $${\tilde{\omega}}\equiv \omega - m\Omega - {k_{\rm z}}v_{\rm 0z},$$ we obtain the second order differential equation $$\frac{d^2{\xi_{\rm R}}}{dR^2} + \frac{d}{dR}\ln\left(RB_{\rm 0}^2\frac{{\tilde{\omega}_{\rm A}}^2}
{{\tilde{\omega}_{\rm F}}^2}\right)\frac{d{\xi_{\rm R}}}{dR} + \frac{H(R)}{{\tilde{\omega}_{\rm A}}^2}{\xi_{\rm R}}= 0,$$ where
$$\begin{aligned}
H(R) &\equiv &
{\tilde{\omega}_{\rm F}}^2 \left\{ \frac{{\tilde{\omega}_{\rm A}}^2-\kappa^2}{{v_{\rm A}}^2} -
{F_{\rm B}}\frac{d\ln{\rho_{\rm 0}}}{dR}
- \frac{1}{R^2} + \frac{1}{R{B_{\rm 0}}^2} \frac{d}{dR}
\left(R \bold{B}_0 \cdot \bold{B}_0' \right)
\right\} \nonumber
\\
&-& 4\Omega\left\{
{\tilde{\omega}}\left( \frac{m}{R}{F_{\rm B}}+ \frac{{B_{\rm 0z}}}{R{B_{\rm 0}}}G_+ \right)
+\Omega k^2 \frac{{B_{\rm 0z}}^2}{{B_{\rm 0}}^2}
- {v_{\rm A}}^2 k^2 {F_{\rm B}}\frac{{B_{\rm 0\phi}}(\bold{k}\cdot\bold{B}_0)}{{\tilde{\omega}}{B_{\rm 0}}^2}
\right\}
\\
& - & \left(\frac{d}{dR}\ln{\frac{{\tilde{\omega}_{\rm F}}^2}{R{B_{\rm 0}}^2}}
- \frac{d}{dR} \right)
\left\{
{v_{\rm A}}^2 {F_{\rm B}}G_-^2 + 2\Omega G_-{\tilde{\omega}}\frac{{B_{\rm 0z}}}{{B_{\rm 0}}} + \frac{{v_{\rm A}}^2}{R}G_+G_-
\right\}
\nonumber
\\
& - &{v_{\rm A}}^2 \left\{
k^2{F_{\rm B}}^2 \frac{{\tilde{\omega}_{\rm A}}^2}{{\tilde{\omega}}^2} + \frac{2}{R} {F_{\rm B}}G_+ G_-
+ \frac{G_+^2}{R^2}
\right\},
\nonumber
\eqnum{36a}\end{aligned}$$
with $${F_{\rm B}}\equiv
\frac{1}{{B_{\rm 0}}^2}\left(\frac{{B_{\rm 0\phi}}^2}{R} + \bold{{B_{\rm 0}}} \cdot \bold{{B_{\rm 0}}}'\right),
\quad
G_{\pm} \equiv \frac{1}{{B_{\rm 0}}}\left(\frac{m}{R}{B_{\rm 0z}}\pm {k_{\rm z}}{B_{\rm 0\phi}}\right),
\eqnum{36b}$$ $${\tilde{\omega}_{\rm A}}^2 \equiv {\tilde{\omega}}^2 - v_{\rm A}^2(\bold{k}\cdot\bold{b}_0)^2,
\quad
{\tilde{\omega}_{\rm F}}^2 \equiv {\tilde{\omega}}^2 - v_{\rm A}^2k^2,
\quad {\rm and} \quad
\bold{k} \equiv (0, \frac{m}{R}, {k_{\rm z}}).$$ Here, ${F_{\rm B}}$ represents the equilibrium magnetic force, and ${\tilde{\omega}_{\rm A}}$ and ${\tilde{\omega}_{\rm F}}$ are frequencies connected to the Alfv${\acute{\rm e}}$nic and fast magnetosonic modes in cold MHD fluids, respectively. $\bold{b}_0\equiv
\bold{{B_{\rm 0}}}/|\bold{{B_{\rm 0}}}|$ is the unit vector along an equilibrium field direction, and finally $\bold{k}$ is a vector wavenumber. When $m={k_{\rm z}}=0$, only the terms in the first bracket in the definition of $H(R)$ do not vanish, recovering the radial wave equation (eq. \[16\], \[17a\]) for the fundamental mode.
The second order differential equation (35) has a singularity at ${\tilde{\omega}_{\rm A}}^2=0$, but if we treat a fully general problem including the thermal effects of compressible gas, we will find another singularity (a so called cusp singularity) at the positions where Doppler shifted frequencies of traveling waves match with slow MHD wave frequencies of the medium [cf. @rob85]. For an incompressible medium, ${\tilde{\omega}_{\rm A}}^2=0$ singularities are often referred to as shear [Alfv$\acute{\rm e}$n]{} singularities where because of resonances the characteristics of waves propagating radially would be modified to be either absorbed into or amplified by background medium, if considered as a boundary value problem [@ros82; @cur96]. As pointed out by @app74, the locations with ${\tilde{\omega}_{\rm F}}^2=0$ in eq. (35) are not singularities; these cut-off points in our local analyses appear as resonance waves with frequencies having relatively small imaginary parts, suggesting potential attenuation or amplification of amplitudes.
To remove the second term in eq. (35) we further define $${\xi_{\rm R}}\equiv \Psi
\left(RB_{\rm 0}^2\frac{{\tilde{\omega}_{\rm A}}^2}{{\tilde{\omega}_{\rm F}}^2}\right)^{-1/2}.$$ Then eq. (35) is reduced to the standard ${\rm Schr\ddot{o}dinger}$ form of eq. (16), with generalized $K^2(R)$ defined by $$K^2(R) \equiv \frac{H(R)}{{\tilde{\omega}_{\rm A}}^2} - \frac{1}{2}\frac{d^2}{dR^2}
\ln{\left(RB_{\rm 0}^2\frac{{\tilde{\omega}_{\rm A}}^2}{{\tilde{\omega}_{\rm F}}^2}\right)}
-\frac{1}{4}\left[
\frac{d}{dR}\ln{\left(RB_{\rm 0}^2\frac{{\tilde{\omega}_{\rm A}}^2}{{\tilde{\omega}_{\rm F}}^2}\right)}\right]^2.$$
Generally speaking, $K(R)$ is a function of $R$ for fixed values of $m$, ${k_{\rm z}}$, and $\omega$. However, we can still consider the behavior in a local sense near some fixed ${R_{\rm o}}$, such that $K$ is close to $K({R_{\rm o}})$. This is mathematically formalized as described in @lin93 and @ter96. Let us consider in the nonuniform background the spatially localized wave packet of the form $$\Psi = \psi(R-R_{\rm o})e^{i{k_{\rm R}}(R-R_{\rm o})} + O(\frac{1}{{k_{\rm R}}}),$$ where $\psi(r)$ is a function which is non-zero only in a small neighborhood of $r \equiv R-{R_{\rm o}}=0$. The scale over which $\psi(r)$ varies significantly must tend to zero as ${k_{\rm R}}\rightarrow \infty$, but no faster than ${k_{\rm R}}^{-1}$. Then, to leading order, $d^2\Psi/dR^2 \approx -{k_{\rm R}}^2\Psi$, and the solution ${k_{\rm R}}^2 = K^2(R,{\tilde{\omega}},m,{k_{\rm z}})$ of the ${\rm Schr\ddot{o}dinger}$ equation yields a local dispersion relation with the right hand side evaluated at a reference point ${R_{\rm o}}$, provided ${k_{\rm R}}$ is limited to a sufficiently large value (i.e., ${R_{\rm o}}{k_{\rm R}}\gg 1$). We may invert this dispersion relation to find ${\tilde{\omega}}= W(m,{k_{\rm z}},{k_{\rm R}},R)$, so that ${k_{\rm R}}$ now plays the role of an independent parameter and the dispersion relation yields the Doppler-shifted frequency ${\tilde{\omega}}$ of a wave near a position ${R_{\rm o}}$ having local wavevector $\bold{k}$. This is equivalent to a standard WKB approximation in the radial direction.
Solution of the local dispersion relation near ${R_{\rm o}}$ yields $$\omega = W(m,{k_{\rm z}},{k_{\rm R}},{R_{\rm o}}) + m\Omega({R_{\rm o}}) + {k_{\rm z}}v_{\rm 0z}({R_{\rm o}})
+ {\cal O}(r),$$ where the ${\cal O}(r)$ term is $(m\Omega'({R_{\rm o}}) + {k_{\rm z}}v_{\rm 0z}'({R_{\rm o}}))r$. Defining ${d{k_{\rm R}}}/{dR} \equiv - ({\partial W}/{\partial R}) /
({\partial W}/{\partial {k_{\rm R}}}) \sim {k_{\rm R}}/R$, the WKB condition $|d{k_{\rm R}}/dR| \ll {k_{\rm R}}^2$ will be satisfied for ${k_{\rm R}}R\gg 1$. For normal mode solutions, the ${\cal O}(r)$ term in $\omega$ must provide a negligible contribution to the phase; this requires that we must have $|m\Omega'({R_{\rm o}}) + {k_{\rm z}}v_{\rm 0z}'({R_{\rm o}})|t \ll {k_{\rm R}}$. For axisymmetric modes with negligible vertical shear, this is always satisfied. However, for $m\neq 0$ disturbances, or flows with non-negligible ${k_{\rm z}}v_{\rm 0z}'$, spatially localized wavepackets maintain a characteristic radial wavenumber for only a limited time, altering their spatial pattern because of the background shear. For a wavepacket with initial wavenumber ${k_{\rm R}}(0)$, the radial wavenumber at time $t$ becomes, upon inclusion of $-t$ times the ${\cal O}(r)$ term in $\omega$ in the phase, $${k_{\rm R}}(t) = {k_{\rm R}}(0) - (m\Omega'+{k_{\rm z}}v_{\rm 0z}')t.$$ Thus, for example, with ${k_{\rm z}}=0$, the pitch $\tan p \equiv m/R{k_{\rm R}}$ of a spiral pattern changes by a fraction $\epsilon = |d{k_{\rm R}}|/{k_{\rm R}}$ over time $t = \epsilon {k_{\rm R}}/|m\Omega'|$. If $R{k_{\rm R}}\gg m$, the pattern changes slowly compared to the orbit time. Among nonaxisymmetric disturbances, the wavepackets with low $m/R{k_{\rm R}}$ have the largest temporal range for which they remain close to normal modes of the system. In the following two sections, we present solutions for the growth rates of unstable disturbances determined from a local modal analysis (i.e., producing solutions ${\tilde{\omega}}= W({\bold k},{R_{\rm o}}))$, with the understanding that when $|m\Omega' + {k_{\rm z}}v_{\rm 0z}'| \neq 0$, the modal growth (i.e., $\sim e^{|{\tilde{\omega}}|t}$) with fixed pattern holds only for a limited time. In assessing the potential of the non-axisymmetric instabilities we shall identify to affect flow dynamics, we will consider their total amplification over times $< {k_{\rm R}}R/m\Omega$, for which the spatial pattern changes little.
For simplicity let us define dimensionless variables $$\sigma \equiv {\tilde{\omega}}{R_{\rm o}}/ {{v_{\rm A}}({R_{\rm o}})},\; {x_{\rm z}}\equiv {k_{\rm z}}{R_{\rm o}},\;
{x_{\rm R}}\equiv {k_{\rm R}}{R_{\rm o}},\;
\kappa \equiv {{R_{\rm o}}\kappa_{\rm o}}/{{v_{\rm A}}({R_{\rm o}})},$$ $$\Omega \equiv {{R_{\rm o}}\Omega_{\rm o}}/{{v_{\rm A}}({R_{\rm o}})}, \;{\rm and}\;
\zeta \equiv {{R_{\rm o}}v_{\rm 0z}'({R_{\rm o}})}/{{v_{\rm A}}({R_{\rm o}})}.$$ Here, $\zeta$ measures the amount of shear in the vertical velocity of the winds, and the “o” subscripts in the equilibrium epicyclic and rotation frequencies denote evaluation at the reference point ${R_{\rm o}}$. We adopt the power-law equilibria of §2. We now organize the terms in eq. (38) finally to get a 12th-degree polynomial $$0 = \sigma^{12} +
\sum_{j=0}^{10} f_j(q,i,{x_{\rm R}},{x_{\rm z}},m,\Omega,\zeta) \sigma^j.$$ The functional dependences of the coefficient $f_j$’s on the parameters are so complicated that it is not illuminating to write down the whole expression here. We may obtain more simplified forms for $f_j$’s by sorting out terms and taking the limit of ${x_{\rm R}}\gg 1$. Because there are various interesting modes which demand different regimes of parameters, however, we keep all the terms as the general local dispersion relation (with ${x_{\rm R}}$ large) and compute numerical growth rates by solving eq. (41) for $\sigma$ as a function of the other variables. We present these numerical results in §5. In §6, we will classify individual (either unstable or overstable) modes and provide their limiting dispersion relations.
@ter96 considered only incompressible modes, for disk applications where thermal pressure is considerable, by taking a divergence-free displacement vector as a perturbation eigenvector, thus obtaining a 4th-order polynomial. Our dispersion relation, for applications to supersonic MHD winds in which thermal pressure is negligible hence motions are compressible, contains information about all possible, either oscillating or unstable, modes of cold MHD winds.
Numerical Solutions of Modal Dispersion Relation
================================================
The derived dispersion relation (41) is a 12th-order polynomial with real coefficients, indicating that solutions appear as complex conjugate pairs. Solutions having non-zero real and imaginary parts are overstable modes, and solutions with vanishing real parts are unstable modes. In this section, we present both types of solutions which are consistent with the local analysis by fixing ${x_{\rm R}}=10$. As we shall show, both higher values of ${x_{\rm R}}$ (more spatially localized) and lower values of ${x_{\rm R}}$ (less spatially localized) give qualitatively the same family of solutions as with ${x_{\rm R}}=10$.
Axisymmetric Modes of Instabilities
-----------------------------------
First, we consider the axisymmetric case with $m=0$. For 4 selected sets of parameters, we plot the real and imaginary parts of the unstable and overstable modes in Fig. 4. A Keplerian rotation gradient with $\kappa^2 = \Omega^2$ is assumed and vertical shear is neglected except in ${\tilde{\omega}}$. We take $\Omega$ as arbitrary rather than using the relation (6), by allowing that the gravitational force from a central object also contributes to the equilibrium rotation velocity. Then, from eq. (5), the normalized angular velocity becomes $$\Omega^2 = \kappa^2 = \cos^2{i} - \frac{1+q}{2} + G_{\rm R},$$ where ${G_{\rm R}}\equiv g_{\rm R}({R_{\rm o}}){R_{\rm o}}/{v_{\rm A}}^2({R_{\rm o}}) > 0$ is the normalized gravitational acceleration. Note that for $1+q>0$ (i.e., magnetic fields decreasing outward), equilibrium solutions with $i$ approaching 90$^{\rm o}$ require non-zero gravity (because hoop stresses do not confine a primarily-poloidal flow). Also note that as ${G_{\rm R}}$ strengthens, the initial equilibrium is maintained by the balance between centrifugal and gravitational forces, implying that the magnetic force is negligible.
The behavior of the solutions shown in Fig. 4 (and similar behavior for other parameters) allows us to identify 4 different axisymmetric mode families: a toroidal resonance mode (TR), an axisymmetric toroidal buoyancy mode (ATB), a poloidal buoyancy mode (PB), and a Balbus-Hawley (BH) mode. One (TR) of these is an overstable mode and the others (ATB, PB, and BH) are purely growing modes.
Fig. 4a and 4b correspond to a disk wind at large distance from the source, where magnetic fields are dominantly toroidal (small $i$) and centrifugal force balances magnetic force (small ${G_{\rm R}}$), while Fig. 4c and 4d correspond to an accretion disk or inner part of a wind where magnetic fields are poloidal (large $i$) and centrifugal force is balanced by the gravity from a central object (relatively large ${G_{\rm R}}$). In each frame, solid and dotted lines represent the imaginary and real parts of the frequencies, respectively. Fig. 4a shows the TR mode which splits into two branches in the presence of (arbitrarily small) poloidal magnetic fields (Fig. 4b and 4c). This TR mode is not a generic instability mode because it has a far larger real part (associated with ordinary MHD oscillations), indicating an overstability. With the presence of poloidal field components, there exist two different types of buoyancy modes, namely ATB and PB modes. When the pitch angle of the magnetic field is relatively small, the buoyancy instabilities are driven by the interplay of the centrifugal force with the hoop stress of toroidal fields, so we call these ATB modes. Since, as explained in §6.1.2, the ATB modes need non-zero poloidal fields as well to be unstable, they disappear when $i=0$. On the other hand, the PB instability modes arise when the fields are predominantly poloidal so that the pressure gradient forces of poloidal fields and the gravity from a central object are main driving forces, similar in character to the Parker instability. ATB and PB are pure instability modes with Re($\sigma$)=0, as shown in Fig. 4. These instability modes operate even in the arbitrarily high-${k_{\rm z}}$ limit because of our cold MHD assumption; otherwise sound waves would stabilize short wavelength perturbations, as they do in the Parker instability. Fig. 4d shows that the BH instability mode appears when ${G_{\rm R}}\gg 1$, corresponding to dynamically weak magnetic fields in the equilibrium; with reduced ${G_{\rm R}}$ (also shown in Fig. 4d), BH is stabilized by radial MHD wave motions when ${x_{\rm R}}$ is large. As discussed in §6.1.3 and §7.1, one interesting finding in our work is that the [*compressible*]{} axisymmetric BH mode is strongly suppressed even for ${G_{\rm R}}$ large when the toroidal field is sufficiently strong; in a cold, Kepler-rotating MHD flow, it is fully stabilized when the pitch angle $i<30^{\rm o}$ (see §6.1.3).
Fig. 5 shows how the characteristics of unstable/overstable modes change as $i$ and ${G_{\rm R}}$ vary for the fixed values of $x_{\rm z}=4$, $x_{\rm R}=10$, and $q=\zeta =0$. For a pure toroidal field configuration with $i=0^{\rm o}$, we observe only overstable TR modes that are almost independent of ${G_{\rm R}}$. As $i$ increases, ATB emerges but is stabilized by rotation with ${G_{\rm R}}$ large. When $i=45^{\rm o}$ and $q=0$, the buoyancy mode disappears because with these parameters the net force from the background magnetic fields vanishes (cf. eqs. \[5\], \[6\], and \[42\]). When $i>30^{\rm o}$ the BH mode strengthens as ${G_{\rm R}}$ increases. This is because in our normalization higher values of ${G_{\rm R}}$ correspond to weaker equilibrium magnetic fields, with which the BH instability operates efficiently. At a pure poloidal configuration of magnetic fields, BH and PB modes remain unstable (Fig. 5d). Dotted lines at very small ${G_{\rm R}}$ in Fig. 5c and 5d mark the minimum value of ${G_{\rm R}}$, available for given values of $q$ and $i$, below which no initial equilibrium exists (cf. eq. \[42\]).
Non-Axisymmetric Modes of Instabilities
---------------------------------------
When non-axisymmetric perturbations are applied, the cold MHD system responds with 3 more modes which are either unstable or overstable, in addition to the axisymmetric modes. We shall refer to these as non-axisymmetric toroidal buoyancy (NTB), geometric poloidal buoyancy (GPB), and poloidal resonance (PR) modes. The PR modes are MHD waves which have non-zero azimuthal wavenumbers and become overstable when there is a radial gradient of the axial field, analogous to the TR modes. In addition to the above modes, systems with toroidal magnetic field configurations and non-zero sound speed are also subject to significant non-axisymmetric magnetorotational instability (NMRI) modes. Unlike the other three non-axisymmetric modes, the NMRI mode arises due to a differential rotation with $d\Omega/dR <0$, where $\Omega$ is an angular velocity of the rotation. More than anything else, the fact that the NMRI in $B_\phi$-dominated systems needs a finite sound speed to be unstable distinguishes it from the axisymmetric Balbus-Hawley instability, which can be unstable regardless of temperature for purely axial fields. As we shall discuss later, the basic mechanism for the onset of NMRI is quite different from that of axisymmetric BH instability. We reserve the discussion of NMRI modes for §7, concentrating here on numerical results for our basic cold MHD system.
Fig. 6 shows the unstable and overstable solutions of the dispersion relations for two combinations of selected parameters: Fig. 6a corresponds to a disk wind with small $i$ and small ${G_{\rm R}}$, while Fig. 6b is for the near-disk case with large $i$ and large ${G_{\rm R}}$. We assume a Keplerian rotation and take an arbitrary $\Omega$ once again using eq. (42). For all cases, we chose ${x_{\rm R}}=10$, $q=\zeta=0$, and $m=1$, and confirmed that changes to these parameters do not appreciably affect the qualitative results. Solid and dotted lines in Fig. 6 represent imaginary and real parts of the normalized wave frequencies, respectively. When $i={G_{\rm R}}=0$, Fig. 6a shows the presence of unstable NTB and overstable TR modes which are split by the non-axisymmetry. NTB modes are nearly like PB modes in their physical basis and have an almost constant growth rate over a wide range of ${x_{\rm z}}$. But they depend sensitively on the logarithmic gradients of the density and magnetic structures (i.e., $q$; see Fig. 7). For an intermediate value of $i$, both TR and PR modes coexist. At some wavenumber ${x_{\rm z}}$, they combine to simply vanish, but overall they give rise to complicated behavior of Im($\sigma$). When $i=90^{\rm o}$, we observe three unstable GPB, BH, and PB modes, and overstable PR modes (Fig. 6b). GPB modes are driven by a buoyancy force together with the geometrical effect. Note that the real parts of PR modes are linearly proportional to the vertical wavenumber ${x_{\rm z}}$, as TR modes are, indicating that they are really overstable modes. Since ${x_{\rm R}}\gg m$, however, there exists only a small contribution from non-axisymmetric effects to the axisymmetric BH and PB modes (cf. Fig. 4d). Remember that when $m\gg{x_{\rm R}}$, the normal mode assumption rapidly breaks down because such high-$m$ modes lose their spatial pattern very quickly; we investigate $m\gg {x_{\rm R}}$ cases using different methods in §§7 and 8.
Fig. 7 shows how the characteristics of the buoyancy modes change as ${x_{\rm z}}$, $i$, and $q$ vary. When ${G_{\rm R}}$=0, an initial equilibrium exists only for a limited range of $i < i_{\rm crit}\equiv \cos^{-1} \sqrt{(1+q)/2}$ from eq. (42), with toroidal field components dominating over poloidal field components. In Fig. 7, therefore, the unstable modes with $i<i_{\rm crit}$ correspond to toroidal buoyancy modes, while poloidal buoyancy modes have $i>i_{\rm crit}$. Generally speaking, with the assumption of extremely cold medium, smaller-scale buoyancy modes with high ${x_{\rm z}}$ have larger growth rates. When $i$ is very small, as seen in Fig. 7, ATB modes are stable because they need the aid of poloidal fields to be unstable, while NTB modes become unstable for [*all*]{} $i<i_{\rm crit}$. This reflects the physically different driving mechanisms between ATB and NTB instabilities. PB modes become more unstable with higher $q$ (steeper background gradients), while ATB/NTB modes are more efficient with smaller $q$. Greater instability is simply associated with higher background magnetic force in the respective cases (cf. the initial equilibrium condition \[42\]).
The ${k_{\rm R}}$-dependence of the unstable/overstable modes are summarized in Fig. 8. Here we fix ${x_{\rm z}}=2$ for all cases and choose $q=0.8$, ${G_{\rm R}}=0$, and small $i$ for Fig. 8a and 8c, corresponding to disk wind-like systems, and $q=0$, ${G_{\rm R}}=5$, and large $i$ for Fig. 8b and 8d, corresponding to accretion disks or disk winds near their sources. The BH instability modes are completely suppressed by MHD waves when ${x_{\rm R}}\gtrsim 3$; we will show that this is consistent with the prediction of the asymptotic dispersion relation. All the other modes extend with smaller growth rates to larger ${x_{\rm R}}$, with Im($\sigma)\sim {x_{\rm R}}^{-1}$, which we will show agrees well with the asymptotic dispersion relations (43), (45), (50), and (52) for the PB, ATB, TR, and NTB modes, respectively. For the PR modes the asymptotic dispersion relation (56), showing Im($\sigma)\sim {x_{\rm R}}^{-2/3}$, is valid only when $R{k_{\rm z}}\gg m$, which is not consistent with the parameters adopted in Fig. 8d. When ${k_{\rm R}}\gg {k_{\rm z}}\sim m/R$, one can confirm analytically that the PR modes also behave as Im($\sigma)\sim {x_{\rm R}}^{-1}$. In the shearing-wavelet point of view with eq. (40), Fig. 8 shows that kinematic shear arising from the background flows ultimately stabilizes both unstable and overstable modes, as ${k_{\rm R}}$ grows secularly increases in time. Although the local approximation breaks down if ${x_{\rm R}}$ is not large, Fig. 8 indicates that the BH mode exists and may show interesting behavior for small ${x_{\rm R}}$. In addition, Fig. 8 also suggests larger growth rates when ${x_{\rm R}}$ is small for other modes, although the assumptions of this section of a radially-local, slowly-changing pattern are not self-consistent when ${x_{\rm R}}$ is small. To study dynamical growth of disturbances which occurs when ${x_{\rm R}}\ll m$, we use direct integrations of the shearing-sheet equations. We present these results in §8.2 (for the NTB modes) and §§7.2 and 7.3 (for the NMRI modes and generalized MRIs).
Mode Classification
===================
The cold MHD system we are investigating has 8 distinct local modes with Im$(\sigma)>0$. Some of them (TR and PR) have larger Re$(\sigma)$ corresponding to overstability, while the others (PB, ATB, BH, NTB, GPB, and NMRI) have negligible Re$(\sigma)$, indicating pure instability. The NMRI modes do not appear in the numerical solutions because of the cold MHD assumption we made. Detailed discussion of the NMRI modes will be separately given in §7.2. In this section we describe the physical nature of the individual cold-fluid modes and present the respective dispersion relations under some limiting approximations.
Axisymmetric Modes
------------------
### Poloidal Buoyancy Mode
Consider a system with pure axial fields. If gravitational forces are large, then they may balance the combined outward radial centrifugal force and pressure gradient force of outward-decreasing $B_{\rm 0z}(R)$; otherwise, if $g_{\rm R}=0$, then the strength in the magnetic fields must increase outward for an equilibrium to exist. In an initial state, at any point in the system the magnetic pressure force acting outward is balanced by the difference between gravity and the centrifugal force acting inward. If perturbed, a denser fluid element experiences reduced magnetic forces but unchanged centrifugal and gravitational forces per unit mass, and thus it would tend to sink radially inward dragging the field line with it; a lighter fluid element would correspondingly tend to float outward. Then, in a frozen-in-field condition, the neighboring gas finds itself on sloping lines of force and thus slides inward to add its weight and to cause field lines to bend more, expediting the instability. This poloidal buoyancy mode is analogous to the Parker instability [@par66], with the driving force role of external gravity in Parker’s instability replaced by combination of gravity and the centrifugal force in the PB. The PB mode can occur for both axisymmetric and non-axisymmetric disturbances.
Putting $B_{0\phi}=m=0$ and considering short wavelength perturbations with ${v_{\rm Az}}^2{k_{\rm z}}^2 \gg {\tilde{\omega}}^2$ in eq. (38), one can find the dispersion relation for the poloidal buoyancy mode is $${\tilde{\omega}}^2 = -\left(\frac{1+q}{2}\right)^2
\frac{{v_{\rm Az}}^2{k_{\rm z}}^2}{R^2({k_{\rm z}}^2+{k_{\rm R}}^2)}
= - \left(\frac{|g_{\rm R} -R\Omega^2|^2}
{{v_{\rm Az}}^2}\right)\frac{{k_{\rm z}}^2}{{k_{\rm z}}^2+{k_{\rm R}}^2}.$$
Eq. (43) states that there is no preferred length scale as long as ${k_{\rm z}}$ is large. However, the inclusion of thermal effects would stabilize the PB mode with shorter wavelengths, as in the Parker instability[^2] Also, sufficiently large ${k_{\rm R}}\gg {k_{\rm z}}$ stabilizes the mode.
### Axisymmetric Toroidal Buoyancy Mode
Now consider a system with weak poloidal but strong toroidal field components and negligible gravity. When magnetic fields are predominantly toroidal (i.e., when $i< \cos^{-1} \sqrt{(1+q)/2}$ from eq. \[42\]), an initial equilibrium state is maintained by the balance mainly between the centrifugal force acting outward, and magnetic hoop stresses which act inward. With sinusoidal density perturbations with ${k_{\rm z}}$ imposed on the equilibrium, a heavier blob of material would tend to float radially outward under the action of unchanged centrifugal forces per unit mass but reduced specific magnetic forces; a lighter element would correspondingly tend to sink. The radial motions of the heavier and lighter blobs are in opposite directions and thus cause the poloidal field lines to bend, creating radial perturbed fields.
The azimuthal fluid motion is slightly accelerated by the tension force exerted by the initial toroidal and the perturbed radial fields (cf. ${B_{\rm 0\phi}}B_{\rm 1R}/R$ term in eq. \[28\], associated with spiral magnetic field line projections in the $z$=constant plane). This causes the initial poloidal component of field lines to bend now in the azimuthal direction, creating bands of perturbed azimuthal fields with signs alternating in the $\hat{\bold{z}}$ direction. The resulting total azimuthal fields are distributed in such a way that the heavier (lighter) blob in the initial perturbation has a lower (higher) toroidal field strength. Induced motions due to the vertical magnetic pressure gradient force carry the matter from under dense to over dense regions, closing the loop and amplifying the initial perturbation.
By setting $m=v_{\rm 0z}=0$ and taking the ${v_{\rm A}}^2{k_{\rm z}}^2 \gg {\tilde{\omega}}^2$ limit, we obtain from eq. (38) the following dispersion relation $$0 = {\tilde{\omega}}^4 - \left[ {v_{\rm Az}}^2\left({k_{\rm z}}^2+{k_{\rm R}}^2\right) + \kappa^2
-4\Omega^2\frac{{B_{\rm 0z}}^2}{{B_{\rm 0}}^2}
\right]{\tilde{\omega}}^2
-4\Omega
{v_{\rm A\phi}}{v_{\rm Az}}{F_{\rm B}}{k_{\rm z}}{\tilde{\omega}}- {v_{\rm A}}^2{v_{\rm Az}}^2{F_{\rm B}}^2{k_{\rm z}}^2,$$ with ${F_{\rm B}}=(\cos^2 i - (1+q)/2)/R$. When ${v_{\rm A\phi}}=0$ (i.e., with pure poloidal fields), eq. (44) immediately recovers eq. (43), the limiting dispersion relation for PB modes. On the other hand, if ${v_{\rm Az}}=0$, there is no unstable ATB mode, clearly demonstrating that ATB modes operate by bending poloidal field lines. From Fig. 5b, we note that ATB modes are stabilized by rotation (larger ${G_{\rm R}}$ corresponds to stronger rotation). It can be shown from eq. (44) that when $R{k_{\rm z}}, R{k_{\rm R}}\gg {\tilde{\omega}}$, the critical wavenumbers are $({k_{\rm z}}^2 + {k_{\rm R}}^2)_{\rm crit} = -(d\Omega^2/d\ln{R})/({v_{\rm A}}^2\sin^2{i})$, below which the system is stable against the ATB modes. For ${k_{\rm z}}\gg k_{\rm z,crit}$, eq. (44) is further reduced to $${\tilde{\omega}}^2 = - \left(\cos^2 i - \frac{1+q}{2}\right)^2
\frac{{v_{\rm A}}^2{k_{\rm z}}^2}{R^2({k_{\rm z}}^2+{k_{\rm R}}^2)}
= - \left(\frac{|g_{\rm R} -R\Omega^2|^2}{{v_{\rm A}}^2}\right)
\frac{{k_{\rm z}}^2}{{k_{\rm z}}^2+{k_{\rm R}}^2}.$$ Since ATB instabilities are axisymmetric modes, they can persist without being disturbed by the kinematic growth of ${k_{\rm R}}$ due to shear, if $v_{\rm 0z}'=0$.
Among well-known plasma modes, the pinch or sausage mode of a plasma column is most similar to the ATB in overall geometry and effect. Both are axisymmetric and require the radial tension force from predominantly toroidal magnetic fields to drive the instability. For both the plasma pinch mode and the ATB of cold cylindrical winds, the net effect is that matter tends to be ejected radially in bands alternating with contracting magnetic field loops. However, in pinch modes, the plasma is generally unmagnetized and surrounded by [*external*]{} toroidal fields, and axial fields tend to suppress the instability. In the ATB, on the other hand, [*internal*]{} toroidal magnetic fields permeate the fluid, and non-zero axial fields are required for instability.
### Compressible Balbus-Hawley Mode
In the presence of axial magnetic fields, a differentially rotating disk is unstable to an axisymmetric incompressible perturbation (Balbus & Hawley 1991; see also Velikhov 1959 and Chandrasekhar 1960). Because this Balbus-Hawley instability[^3] has a rapid growth time (comparable to the local rate of rotation) and exists for arbitrarily weak magnetic field strength, it is believed to provide a powerful mechanism for the generation of the effective viscosity in astrophysical accretion disks. Through numerical simulations, @haw91 argued that the roles of compressibility and toroidal fields are not significant as long as the total field strength is subthermal. Also, @bla94 studied the effect of toroidal fields on the compressible axisymmetric BH instability and showed that toroidal fields do not modify the instability criterion, while reducing growth rates slightly if ${v_{\rm A\phi}}<{c_{\rm s}}$. We find the striking result that under extremely cold conditions (i.e., ${v_{\rm A}}\gg{c_{\rm s}}$), compressibility prohibits the axisymmetric BH instability from occurring if the toroidal fields are as strong as the poloidal fields.
By taking the weak magnetic field limit ($\Omega \gg {v_{\rm A}}{k_{\rm R}}, {v_{\rm A}}{k_{\rm z}}$) and $m=0$, we obtain from eq. (38) the following dispersion relation for the compressible axisymmetric BH instability in a cold MHD flow $$\omega^4 - \omega^2[{v_{\rm A}}^2({k_{\rm z}}^2+{k_{\rm R}}^2) + {v_{\rm Az}}^2{k_{\rm z}}^2 + \kappa^2]
+ {v_{\rm A}}^2{k_{\rm z}}^2[{v_{\rm Az}}^2({k_{\rm z}}^2+{k_{\rm R}}^2) + \kappa^2 - 4\Omega^2\sin^2 i] = 0,$$ and thus from the last term in eq. (46) we obtain the instability criterion $${v_{\rm Az}}^2({k_{\rm z}}^2+{k_{\rm R}}^2) + \kappa^2 - 4\Omega^2\sin^2 i < 0.$$ With an $\Omega \propto R^{-a}$ rotation profile, eq. (47) implies that if $\sin^2 i < 1-a/2$, we anticipate no BH instability in a cold flow. For a Keplerian rotation law with $a=3/2$, for instance, no axisymmetric BH instability occurs if $i < 30^{\rm o}$!; [*when the magnetic field strength is superthermal, the inclusion of toroidal fields tends to suppress the growth of the BH instability*]{}. With a steeper rotation profile (as would occur, for example, in winds from boundary layers), there is an increase in the range of $i$ for which a system is BH-unstable.
We defer the full discussion on the BH instability until §7.1, where we explicitly include pressure terms in the dynamical equations.
### Toroidal Resonance Mode
Consider a system having pure toroidal fields without rotation. If the initial fields are homogeneous in space, magnetosonic waves, driven solely by magnetic pressure (with the assumption of the cold medium), would propagate without any interruption in the plane whose normal is perpendicular to the magnetic field direction. In an inhomogeneous medium, however, MHD waves no longer maintain a sinusoidal planform, and the characteristics of the waves change through the interaction with the background medium. The amplitudes of the waves may sometimes increase as they propagate, or sometimes they may become evanescent and decay at a resonance, or even may be trapped between two resonance points [cf. @rae82]. In such a strongly structured medium, the classification of MHD waves is not in general possible.
Our local treatment of MHD waves can provide some insight on the amplification or evanescence of propagating MHD waves in a structured medium. For ${\tilde{\omega}_{\rm F}}^2 \rightarrow 0$ and $B_{0\rm z}=\Omega=0$, the local wave equation (35) can be simplified as $$\frac{d^2{\xi_{\rm R}}}{dR^2} - \frac{d\ln{{\tilde{\omega}_{\rm F}}^2}}{dR}
\frac{d{\xi_{\rm R}}}{dR} + \frac{{\tilde{\omega}_{\rm F}}^2}{{v_{\rm A\phi}}^2} {\xi_{\rm R}}= 0.$$ Again we define ${\xi_{\rm R}}\equiv ({\tilde{\omega}_{\rm F}}^2)^{1/2}\Psi$, then eq. (48) takes the form of eq. (16) with ${k_{\rm R}}\equiv K(R)$ defined by $${k_{\rm R}}^2 \equiv \frac{{\tilde{\omega}_{\rm F}}^2}{{v_{\rm A\phi}}^2}
- \frac{3}{4}\left(\frac{{v_{\rm A\phi}}^2{k_{\rm z}}^2}{R{\tilde{\omega}_{\rm F}}^2}\right)^2,$$ where ${v_{\rm A\phi}}\propto R^{-1/2}$ was assumed. Thus, considering limiting cases of ${k_{\rm z}}$ and ${k_{\rm R}}$, one can find the dispersion relation for this mode near the resonance frequencies (i.e., ${\tilde{\omega}}\approx {v_{\rm A\phi}}{k_{\rm z}}$) $${\tilde{\omega}}^2 = \left\{ \begin{array}{ll}
{v_{\rm A\phi}}^2{k_{\rm z}}^2\left[1 + (3/4)^{1/2}e^{\pm\pi i/2}(R{k_{\rm R}})^{-1}\right],
&{\rm for}\;\; R{k_{\rm R}}\gg R{k_{\rm z}}\gg 1, \\
& \\
{v_{\rm A\phi}}^2{k_{\rm z}}^2\left[1 + (3/4)^{1/3}e^{\pm2\pi i/3}(R{k_{\rm z}})^{-2/3}\right],
&{\rm for}\;\; R{k_{\rm z}}\gg R{k_{\rm R}}\gg 1, \\
\end{array} \right.$$ showing that the imaginary part of toroidal resonance mode vanishes quickly as ${k_{\rm R}}\rightarrow \infty$, while its real part gets bigger as ${k_{\rm z}}$ increases. Therefore, it is not adequate to regard TR modes as a true local instability mode. Though the TR mode is not a local instability, it suggests potential for waves to have global instabilities in which the magnetosonic resonance (${\tilde{\omega}}= {v_{\rm A\phi}}{k_{\rm z}}$) plays a similar role to the Lindblad resonance in rotating disks. Thus, waves of fixed frequency propagating with a radial component of $\bold{k}$ into their magnetosonic resonances may be amplified or reflected.
The modification of traveling waves due to the inhomogeneity of the medium is mediated through the magnetic pressure. A similar effect would occur when hydrodynamic waves propagate into an inhomogeneous medium.
Non-Axisymmetric Modes
----------------------
### Non-Axisymmetric Toroidal Buoyancy Mode
The non-axisymmetric toroidal buoyancy mode is very similar to the PB mode in its physical mechanism, in spite of the different field geometry. For toroidal-field dominated cases, an equilibrium can exist with the net inward magnetic stresses balancing the outward centrifugal force. When the system is perturbed non-axisymmetrically, the instability would develop similarly to PB modes, as described in §6.1.1. For the $B_\phi$-dominated case, however, over-dense regions float outward and under-dense regions sink, because the inward magnetic forces are enhanced when the density drops.
Setting ${B_{\rm 0z}}=v_{\rm 0z}=0$ and assuming $R{k_{\rm z}}\ll m, R{k_{\rm R}}$, one can show that the general dispersion relation (38) is reduced to the following quartic equation in terms of ${\tilde{\omega}}$ $$0 = {\tilde{\omega}}^4 - \left[ {v_{\rm A\phi}}^2\left(\frac{m^2}{R^2}+{k_{\rm R}}^2\right) + \kappa^2
\right]{\tilde{\omega}}^2
-4\Omega\frac{m}{R}
{F_{\rm B}}{v_{\rm A\phi}}^2{\tilde{\omega}}- {v_{\rm A\phi}}^4{F_{\rm B}}^2\left(\frac{m}{R}\right)^2,$$ with ${F_{\rm B}}=(1-q)/2R$, The negative last term in eq. (51) guarantees the existence of unstable NTB modes. The third term (caused by the coupling of the rotation with the background fields) tends to stabilize NTB modes. Thus, if $$\frac{m^2}{R^2} + {k_{\rm R}}^2 \; < \; \frac{(4\Omega^2-\kappa^2)}{{v_{\rm A\phi}}^2},$$ there is no unstable NTB mode. Note that for wind equilibria with ${B_{\rm 0z}}=0$, from eq. (6) the RHS of the above equals $3(1-q)/2R^2$; thus the NTB instability will be present at all $m$ when $1/3<q<1$.
In the limit of large $m$, we obtain the asymptotic dispersion relation for the NTB mode $${\tilde{\omega}}^2 = - \left(\frac{1-q}{2}\right)^2 \frac{{v_{\rm A\phi}}^2m^2}
{R^2(m^2+R^2{k_{\rm R}}^2)} =
- \left(\frac{|g_{\rm R}-R\Omega^2|^2}{{v_{\rm A\phi}}^2}\right)\frac{m^2}{m^2+R^2{k_{\rm R}}^2}.$$ Here for the second equality, eq. (6) with $i=0^{\rm o}$ is used. Eq. (52) is akin to eq. (43), the dispersion relation for the PB mode, and to eq. (45), the dispersion relation for the ATB mode, reflecting the common origin in buoyancy forces of all three. In fact, Fig. 7 clearly shows how the various buoyancy modes extend and smoothly join at intermediate pitch angles.
### Geometric Poloidal Buoyancy Mode
Now suppose a system with pure vertical fields. When perturbed azimuthally, a fluid element becomes over dense and tends to move inward due to the decreased background magnetic pressure force per unit mass if $0<q<1$. This geometrically converging motion of fluid increases density and field strength by factors of $(1-q)$ and $(1-q)/2$, respectively. On the other hand, the magnetic field enhancement induces diverging motions of the fluid in the azimuthal direction by building up a pressure gradient, tending to lower the density. The net effect of these two processes is a density increase by a factor of $(1-q)/2$, accelerating the inward motion of the heavier element. When $m \gg R{k_{\rm z}}$, the dispersion relation for this GPB mode is found to be $${\tilde{\omega}}^2 = - \left(\frac{1-q^2}{4}\right) \frac{{v_{\rm Az}}^2m^2}{R^2(m^2+R^2{k_{\rm R}}^2)}.$$ When $q=1$, there is no instability. This is because the initial configuration of the density and the field is such that the mass and magnetic flux contained in a thin ring with the thickness $dR$ and the radius $R$ are constant over $R$, and no gain from the geometrical effect is possible.
### Poloidal Resonance Mode
The physical basis for the poloidal resonance mode is quite similar to that of the toroidal resonance mode. The only difference between them is the background field geometry. In the presence of pure axial fields, MHD waves with non-zero $m$ are easily influenced by radial magnetic pressure gradients.
To derive the dispersion relation near the resonance frequencies (i.e., ${\tilde{\omega}_{\rm F}}\approx 0$), let us suppose a system with pure axial fields and neglect the vertical velocity shear. The system is also assumed to rotate slowly enough that the effects of rotation may not be important in the wave dynamics (i.e., $m{v_{\rm Az}}\gg R\Omega$). For ${\tilde{\omega}}^2 \rightarrow {v_{\rm Az}}^2({k_{\rm z}}^2+m^2/R^2)$, we are left from eq. (35) with $$\frac{d^2{\xi_{\rm R}}}{dR^2} + \frac{d}{dR}
\ln{\left(\frac{{\tilde{\omega}_{\rm A}}^2}{{\tilde{\omega}_{\rm F}}^2}\right)}
\frac{d{\xi_{\rm R}}}{dR} + \frac{{\tilde{\omega}_{\rm F}}^2}{{v_{\rm Az}}^2} {\xi_{\rm R}}= 0.$$ We now define ${\xi_{\rm R}}\equiv ({\tilde{\omega}_{\rm F}}^2/{\tilde{\omega}_{\rm A}}^2)^{1/2}\Psi$ to simplify eq. (54) into eq. (16) with ${k_{\rm R}}\equiv K(R)$ defined by $${k_{\rm R}}^2 \equiv \frac{{\tilde{\omega}_{\rm F}}^2}{{v_{\rm Az}}^2}
- \frac{m^2{v_{\rm Az}}^2}{R^2{\tilde{\omega}_{\rm F}}^2}\left(\frac{d\ln{{\tilde{\omega}_{\rm F}}^2}}{dR}\right)^2,$$ where we took the limit of $R{k_{\rm z}}\gg m$ and assumed ${v_{\rm Az}}\propto R^{-1/2}$. Solving eq. (55) for two limits of ${k_{\rm R}}$, we obtain the dispersion relation near the resonance frequencies $${\tilde{\omega}}^2 = \left\{ \begin{array}{ll}
{v_{\rm Az}}^2{k_{\rm z}}^2\left[1 + e^{\pi i/3}
(ma_{\rm sh})^{2/3}(R^2{k_{\rm R}}{k_{\rm z}})^{-2/3}\right],
&{\rm for}\;\; R{k_{\rm R}}\gg R{k_{\rm z}}\gg m, \\
& \\
{v_{\rm Az}}^2{k_{\rm z}}^2\left[1 \pm e^{\pi i/2} (ma_{\rm sh})^{1/2}
(R{k_{\rm z}})^{-1}\right],
&{\rm for}\;\; R{k_{\rm z}}\gg R{k_{\rm R}}\gg m, \\
\end{array} \right.$$ where $a_{\rm sh} \equiv 1\pm 3m\Omega/{v_{\rm Az}}{k_{\rm z}}$, showing again a rapidly declining imaginary part as ${k_{\rm R}}$ increases, at which the local approximation is valid. Thus, just like TR modes, PR modes are not strictly local instability modes.
Magnetorotational Instability
=============================
Axisymmetric BH Instability
----------------------------
In the preceding section, we briefly discussed the axisymmetric BH instability in a cold, differentially rotating medium and found that the BH instability can be suppressed by the azimuthal component of magnetic fields, if the medium is cold enough. Incompressibility has generally been adopted in the study of the BH instability in an accretion disk on the grounds that in such a system the magnetic fields are subthermal and thus acoustic waves can maintain the incompressible condition over many rotation periods. For magnetocentrifugally driven winds, however, sound waves play a minor role in controlling the dynamics and thus the incompressible approximation is inapplicable. In addition, since an initial equilibrium is attained through the balance between the centrifugal and magnetic forces (cf. eq. \[5\]), the [Alfv$\acute{\rm e}$n]{} crossing time scale is comparable to the rotation time scale (${v_{\rm A}}\sim R\Omega$); in this case, the fields are not weak and the unstable range of wavenumbers becomes narrow.
We generalize the previous discussion of the axisymmetric compressible BH instability by explicitly including the thermal pressure terms in the momentum equation and exploring the role of compressibility to the development of the Balbus-Hawley instability. We consider a cylindrical flow threaded by both vertical and azimuthal magnetic fields, ignoring the radial variations in the initial configuration except $\Omega=\Omega(R)$ and neglecting the vertical velocity. We assume the medium is isothermal and take the WKB ($R {k_{\rm z}}\gg 1$) approximation. Through the standard approach to linear analyses, we arrive at the dispersion relation for the compressible version of the BH instability $$({\omega_{\rm A}}^2 - \kappa^2)f(\omega^2) = {k_{\rm R}}^2{\omega_{\rm A}}^2(({c_{\rm s}}^2 + {v_{\rm A}}^2)\omega^2
-{c_{\rm s}}^2{v_{\rm Az}}^2{k_{\rm z}}^2) + 4\Omega^2{v_{\rm Az}}^2{k_{\rm z}}^2(\omega^2-{c_{\rm s}}^2{k_{\rm z}}^2),$$ where ${c_{\rm s}}$ is the isothermal sound speed of the medium, ${\omega_{\rm A}}^2 \equiv {\omega}^2 - {v_{\rm Az}}^2{k_{\rm z}}^2$, and $f(\omega^2)$ is defined by $$f(\omega^2) \equiv
\omega^4 - \omega^2 ({c_{\rm s}}^2+{v_{\rm A}}^2){k_{\rm z}}^2 + {c_{\rm s}}^2{v_{\rm Az}}^2{k_{\rm z}}^4.$$ Eq. (57) is a sixth-order equation for $\omega$ with only even terms. When ${k_{\rm R}}=0$, eq. (57) is identical to eq. (64) of [@bla94] or eq. (99) of @bal98. Now let us take the two opposite limits of ${c_{\rm s}}$ to obtain the following dispersion relations
$$\begin{aligned}
{\omega_{\rm A}}^4 - (\kappa^2 + {v_{\rm A}}^2{k_{\rm R}}^2 + {v_{\rm A\phi}}^2{k_{\rm z}}^2){\omega_{\rm A}}^2
+(\kappa^2{v_{\rm A\phi}}^2 - 4\Omega^2{v_{\rm Az}}^2){k_{\rm z}}^2 &=& 0,
\quad{\rm for}\quad {c_{\rm s}}\rightarrow 0, \\
(1 + {k_{\rm R}}^2/{k_{\rm z}}^2){\omega_{\rm A}}^4 - \kappa^2{\omega_{\rm A}}^2 - 4\Omega^2{v_{\rm Az}}^2{k_{\rm z}}^2 &=& 0,
\quad{\rm for}\quad {c_{\rm s}}\rightarrow \infty,\end{aligned}$$
and the corresponding instability criteria[^4]
$$\begin{aligned}
{v_{\rm Az}}^2({k_{\rm z}}^2 + {k_{\rm R}}^2) + \kappa^2 - 4\Omega^2\sin^2i &<& 0,
\quad{\rm for} \quad {c_{\rm s}}\rightarrow 0, \\
{v_{\rm Az}}^2({k_{\rm z}}^2 + {k_{\rm R}}^2) + d\Omega^2/d\ln R &<& 0,
\quad{\rm for} \quad {c_{\rm s}}\rightarrow \infty.\end{aligned}$$
Note that eq. (58b) is the same as the original dispersion relation of the incompressible BH instability (eq. \[2.9\] of Balbus & Hawley 1991 without the Brunt-${\rm V\ddot{a}is\ddot{a}l\ddot{a}}$ frequency). The instability criterion (59a) in the extremely compressible limit depends explicitly on the local pitch angle, showing that as $i$ departs from $90^{\rm o}$ the instability becomes gradually confined to smaller values of ${k_{\rm z}}$. For a cold Keplerian flow, no instability occurs when $i$ is smaller than $30^{\rm o}$.
To examine what role thermal pressure plays to the growth of the BH instability and why the instability criterion depends on $i$, we plot the unstable solutions of eq. (57) as functions of ${q_{\rm A}}\equiv (\bold{k}\cdot\bold{v}_{\rm A})/\Omega$ $(= {k_{\rm z}}{v_{\rm Az}}/\Omega$ for the axisymmetric case) and $\beta \equiv {c_{\rm s}}^2/{v_{\rm A}}^2$ in Fig. 9. For the time being, we confine our discussion to the ${k_{\rm R}}=0$ case. When $i=90^{\rm o}$, the instability criterion from eqs. (59a,b) is ${v_{\rm A}}^2{k_{\rm z}}^2 < - d\Omega^2/d\ln{R}$ and the growth rate is independent of $\beta$, implying that the compressibility does not alter the instability (Fig. 9a). This can be understood as follows: when magnetic fields are mainly axial, sound waves propagating along a vertical direction decouple completely from the magnetic fields and are undisturbed by rotation. But transverse MHD waves which are intrinsically incompressible are influenced by rotation to become ultimately unstable for a range of ${k_{\rm z}}$ when $d\Omega /dR <0$. Therefore, $i=90^{\rm o}$ is a very special case.
On the other hand, when both vertical and azimuthal fields are present, toroidal perturbed fields generated by an initial azimuthal displacement or by sheared motion following a radial perturbation of the initial axial fields tend to cause vertical oscillations, but in a cold assumption, mainly due to the magnetic field gradient terms, $-{B_{\rm 0\phi}}(\partial B_{1\phi}/\partial z)$. This oscillatory vertical motion tries to distribute the perturbed fields as uniformly as possible, thereby tending to suppress the growth of the disturbances. However, the vertical magnetic pressure gradients are not strong enough to create significant vertical motions if thermal pressure is large: a compressed region tends to expand vertically but with little change in the strength of the toroidal fields, thus providing a favorable condition for the development of the BH instability. This explains why higher $\beta$ cases have higher growth rates at fixed $i$, and why the growth rate decreases as $i$ decreases at fixed $\beta$ (Fig. 9b$\sim$9d).
Although the instability criterion (59b) is completely independent of the strength of the azimuthal fields provided that $\beta\neq 0$, indicating as noted by @bla94 that to all orders, azimuthal fields do not modify the stability criterion, the corresponding growth rates drop progressively as $\beta$ decreases if $i\neq 90^{\rm o}$. When $\beta \gtrsim 1$, any change of an inclination angle $i$ from $90^{\rm o}$ does not bring significant reductions in growth rates, implying that the characteristics of the instability are essentially the same as the pure poloidal case. If $\beta \ll 1$, however, we observe dramatic stabilizing effects from toroidal fields, as illustrated in Fig. 9.
A few comments should be devoted to the effect of ${k_{\rm R}}$. Radial wave motions do nothing but add another restoring force to perturbations. This in turn means that thermal pressure has a stabilizing influence on the growth of the BH instability. Thus there are two competing processes of thermal pressure: thermal pressure associated with vertical wave motion promotes the BH instability, while thermal pressure controlling radial motion opposes it. It turns out that for $i \neq 90^{\rm o}$ the former process always dominates. For $i=90^{\rm o}$, only the latter effect exists, giving higher growth rates for smaller $\beta$, when ${k_{\rm R}}\neq0$.
Notice the stabilizing effect of ${k_{\rm R}}$ in eqs. (59). If the background vertical flow has significant shear, the local radial wavenumber would increase with time (cf. eq. \[40\]), suppressing the instability. Thus when $v_{\rm 0z}' \neq 0$, the compressible BH instability will exhibit a transient growth, as must happen to all modes if ${k_{\rm z}}v_{\rm 0z}' \neq 0$ and/or $m\Omega' \neq 0$.
In conclusion, we have found that compressibility has a stabilizing effect on the axisymmetric BH instability. Even though its effect is small if the sound speed is super-Alfv$\acute{\rm e}$nic, compressibility must be considered whenever the Alfv$\acute{\rm e}$n speed is comparable to or even exceeds the thermal sound speed, as is expected in winds and also in disk coronae [cf. @mil99].
The above discussion applies only for axisymmetric perturbations. It was also found that an accretion disk with purely toroidal fields is subject to non-axisymmetric instability [@bal92; @ter96], but we will show in the following section that the physical role of compressibility in that case is completely different, in spite of the same quantitative instability criteria.
Non-Axisymmetric MRI: Coherent Wavelet Analysis
-----------------------------------------------
@bal92 found that a differentially rotating disk of incompressible fluid with embedded toroidal magnetic field is unstable to non-axisymmetric perturbations. Adopting shearing sheet coordinates (see below), they integrated a set of the perturbed equations and showed that perturbations with an intermediate azimuthal wavenumber $m$ can exhibit transient, but enormous growth over a time scale of several percent of $\Omega^{-1}$.
An alternative approach was taken by @ter96 to study a similar instability to that identified by @bal92. They solved the problem using the local WKB approximation. They started from a general compressible equation of state, but subsequently they supposed divergence-free poloidal Lagrangian displacements, which made their treatment essentially incompressible. They derived a sufficient condition for the instability which is exactly the same form as that of axisymmetric BH instability (i.e., $d\Omega^2/d\ln R < 0$). Noting that azimuthal shear is the main driving mechanism and bending of the field lines provides a stabilizing restoring force, they suggested the non-axisymmetric instability of toroidal magnetic fields might resemble the original BH instability.
We argue in this work that the underlying physical mechanisms for non-axisymmetric toroidal-${\bf B}$ MRI (which we refer to as “NMRI”) and axisymmetric poloidal-${\bf B}$ MRI (which we refer to as “BH”) are in fact quite different from each other. In this section, we analyze the NMRI by looking at “coherent wavelet” solutions in which every physical variable, localized in both space and time, oscillates or grows with the same space-time dependence, and provide quantitative results in detail.
### Localization in Space and Time
We begin by considering a shearing, rotating disk with uniform density, and magnetic fields with only an azimuthal component. We ignore any unperturbed vertical motion in the medium. We include thermal pressure effects with an isothermal equation of state to obtain the explicit dependence of the NMRI on the temperature, but neglect effects of cylindrical geometry. This configuration is the same as Balbus & Hawley’s (1992), except that they considered only the incompressible case with the Boussinesq approximation, and allowed for vertical equilibrium gradients yielding buoyant oscillations. Adopting the shearing sheet coordinates $(\tilde{R},\tilde{\phi},\tilde{z})$ such that $\tilde{R} = R$, $\tilde{\phi} = \phi - \Omega(R) (t-t_{\rm o})$, and $\tilde{z}=z$ [@gol65; @jul66; @bal92], we consider the time development of an initial plane-wave disturbance which preserves sinusoidal variation in the local rest frame of the equilibrium shearing, rotating flow $$\chi_1(R,\phi,z,t)
= \chi_1(t)e^{im\tilde{\phi} + i{k_{\rm z}}\tilde{z} +i{k_{\rm R}}(t_{\rm o})\tilde{R}},$$ where ${k_{\rm R}}(t_{\rm o})$ is a radial wavenumber at a fiducial time $t=t_{\rm o}$. The linearized form of the MHD eqs. (1)$\sim$(4), can be written in dimensionless form as $$\frac{d \alpha}{d\tau}
= -{q_{\rm R}}u_{\rm 1R} - {q_{\rm m}}u_{1\phi} -{q_{\rm z}}u_{\rm 1z},$$ $$\frac{d u_{\rm 1R}}{d\tau} = 2 u_{1\phi}
- {q_{\rm m}}b_{\rm R} + {q_{\rm R}}(\beta\alpha + b_{\phi}),$$ $$\frac{d u_{1\phi}}{d\tau}
= -\frac{\kappa^2}{2\Omega^2}u_{\rm 1R} + \beta {q_{\rm m}}\alpha,$$ $$\frac{d u_{\rm 1z}}{d\tau}=
-{q_{\rm m}}b_{\rm z} + {q_{\rm z}}(\beta\alpha + b_{\phi}),$$ $$\frac{d b_{\rm R}}{d\tau}={q_{\rm m}}u_{\rm 1R},$$ $$\frac{d b_{\phi}}{d\tau}
= \frac{d\ln{\Omega}}{d\ln{R}} b_{\rm R} - {q_{\rm z}}u_{\rm 1z} - {q_{\rm R}}u_{\rm 1R},$$ $$\frac{d b_{\rm z}}{d\tau} ={q_{\rm m}}u_{\rm 1z},$$ where the dimensionless Lagrangian derivative is denoted by $$\frac{d}{d\tau} = \frac{1}{\Omega}
\frac{\partial}{\partial t} + \frac{\partial}{\partial \phi}.$$ In eqs. (61)$\sim$(68), all perturbed variables are dimensionless and defined by $\alpha \equiv \rho_1/\rho_0$, $\bold{u_1} \equiv i \bold{v_1}/{v_{\rm A\phi}}$, $\bold{b}\equiv \bold{B_1}/{B_{\rm 0\phi}}$, and $\tau \equiv t\Omega $, and dimensionless parameters are $\beta \equiv {c_{\rm s}}^2/{v_{\rm A\phi}}^2$, ${q_{\rm m}}\equiv {v_{\rm A\phi}}m/R\Omega$, ${q_{\rm z}}\equiv {v_{\rm A\phi}}{k_{\rm z}}/\Omega$, and $${q_{\rm R}}(\tau) \equiv \frac{{v_{\rm A\phi}}{k_{\rm R}}(t)}{\Omega} =
\frac{{v_{\rm A\phi}}}{\Omega}
\left[{k_{\rm R}}(t_{\rm o}) - m(t-t_{\rm o})\frac{d\Omega}{dR}\right]
= -mt\frac{{v_{\rm A\phi}}}{\Omega}\frac{d\Omega}{dR}
= -\tau{q_{\rm m}}\frac{d\ln\Omega}{d\ln R},$$ where the third equality holds when $t_{\rm o}\equiv -{k_{\rm R}}(t_{\rm o})/m\Omega'$. Eqs. (65)$\sim$(67) yield the divergence free condition for the perturbed magnetic fields.
Since ${q_{\rm R}}$ has a $\tau$-dependence, the linear system of eqs. (61)$\sim$(67) does not form an eigenvalue problem; kinematics of shear wrap a given disturbance by increasing its radial wavenumber linearly with time. In the original shearing sheet formalism, the fate of a system exposed to perturbations is analyzed through direct integrations of linearized equations over time. In doing so, one may observe transient amplification or decay of applied disturbances depending on their stability. One can say that a system is unstable if some physical variables grow sufficiently over certain time scales. The efficiency of instability for a system is identified by computing the response of the system to variation of parameters input to temporal integrations. This approach was adopted by @bal92 in their identification of the NMRI.
Here, we instead analyze the NMRI by proceeding one more step from the original shearing sheet formalism to find solutions which are localized in time as well as in space. First, we note that there exist two distinct time scales: the growth time of instabilities determined by the inverse of the dimensionless instantaneous growth rate, $$\gamma(\tau) \equiv \frac{d}{d\tau}\ln{\chi_1(\tau)} =
\frac{1}{\Omega}\frac{d}{dt}\ln{\chi_1(t)},$$ and the dimensionless shearing time, $(d\ln{q_{\rm R}}/d\tau)^{-1}={q_{\rm R}}|{q_{\rm m}}d\ln \Omega/d\ln R|^{-1}$, as a typical time scale of the linear growth of the radial wavenumber. If the shearing time is much longer than the growth time, that is, if ${q_{\rm m}}|d\ln \Omega/d\ln R|/(\gamma{q_{\rm R}}) \ll 1$ (the “weak shear limit”), the time dependence of ${q_{\rm R}}$ in eq. (69) can be neglected, and thus normal mode solutions having an exponential or oscillatory behavior can be sought. @shu74 applied this technique to investigate the effects of a differential rotation on the Parker instability. Also, @ryu92 obtained an algebraic dispersion relation for the convective instability in differentially rotating disks, by assuming that ${q_{\rm R}}$ is time-independent.
Because the convective and the Parker instabilities arise from hydrodynamic and magnetic buoyancy effects, respectively, independent of the rotation of a disk, one can always find a regime in which the weak shear limit is applicable. In some cases, however, as for example in the axisymmetric poloidal or the non-axisymmetric toroidal MRIs with weak magnetic fields, the instabilities result directly from a differential rotation with $\Omega' <0$. In such cases, peak growth rates are of the same order as rotational frequencies [@bal98], and thus the weak shear is not a good approximation for these non-axisymmetric instabilities.
However, we can still look for coherent solutions in which [*all*]{} perturbed variables vary as $e^{\gamma \tau}$ with time, provided the variation of the instantaneous growth rate $\gamma(\tau)$ over the growth time $\gamma^{-1}$ is relatively small, i.e., $$\left|\frac{d\ln \gamma(\tau)}{d\tau}\right| \ll \gamma(\tau).$$ We refer to the solutions under this approximation as “coherent wavelet solutions” because all physical quantities localized in both space and time grow at the same instantaneous rate. If the condition (71) holds, the changes in $\gamma(\tau)$ can be neglected over a short time interval, and the set of dynamical equations (61)$\sim$(67) constitutes an eigensystem instantaneously. This is equivalent to the WKB method in the time dimension. Since $\gamma^{-1}d\ln \gamma(\tau)/d\tau =
-{q_{\rm m}}(d\ln\Omega/d\ln R) (d\ln \gamma/d\ln{q_{\rm R}}) /(\gamma{q_{\rm R}})$, eq. (71) is satisfied if either ${q_{\rm m}}|d\ln\Omega/d\ln R|/(\gamma{q_{\rm R}}) \ll 1$ (the weak shear limit), or $|d\ln\gamma/d\ln{q_{\rm R}}| \ll 1$ (instantaneous growth rates are relatively insensitive to the radial wavenumber); the condition (71) is less restrictive and in fact is the generalization of the weak shear limit. Of course, we need to check the self-consistency of this coherent wavelet approximation by examining [*a posteriori*]{} whether resulting solutions satisfy the condition (71). For incompressible media, @bal92 mapped the regime of instability in \[$({\bf k\cdot{v_{\rm A}}})^2$, $|{\bf k}|/{k_{\rm z}}$\] space using WKB methods similar to those we adopt.
### Coherent Wavelet Dispersion Relation
Upon substituting eq. (70) into eqs. (61)$\sim$(67) and applying the approximation (71) so that $d\chi_1/d\tau \rightarrow \gamma\chi_1$, one can form a matrix equation $\gamma Q = {\cal M}Q$, where $Q=(\alpha,u_{\rm 1R},u_{1\phi},u_{\rm 1z},b_{\rm R},
b_\phi,b_{\rm z})^{\rm T}$ is a column vector and ${\cal M}$ is a $7\times$7 matrix whose components are determined by the coefficients of $Q$ in the right hand sides of eqs. (61)$\sim$(67). By solving the condition det(${\cal M} - \gamma{\cal I})=0$, where ${\cal I}$ is the identity matrix, we obtain a seventh order polynomial in $\gamma$. As a further approximation, however, if at least one of the conditions, ${q_{\rm z}}\gg {q_{\rm R}}$, $\gamma \gg \tau$, or $\gamma\tau \gg 1$, is satisfied, all even order terms that depend linearly on $\tau$ and ${q_{\rm m}}$ but are independent of ${q_{\rm z}}$, can be neglected compared to the remaining terms. The first two conditions apply when the radial wavenumber is not significant, either because disturbances are highly localized in the vertical direction (${q_{\rm z}}\gg {q_{\rm R}}$) or simply because we are looking at modal behaviors at the time $\tau \sim 0$, while the third condition holds when net amplification of perturbations is large. This simplification yields a trivial solution $\gamma=0$ (this arises from the fact that perturbed magnetic fields, $b_{\rm R}$, $b_\phi$, and $b_{\rm z}$, are linearly dependent via the divergence-free condition) and a third-order polynomial in $\gamma^2$ which is the resulting instantaneous dispersion relation for NMRI
$$\begin{aligned}
0 = \gamma^6 &+& \gamma^4
\left[(1+\beta)q^2 + {q_{\rm m}}^2 + \frac{\kappa^2}{\Omega^2} \right]
\nonumber \\
&+& \gamma^2 \left[(1+2\beta)q^2{q_{\rm m}}^2 +
\frac{\kappa^2}{\Omega^2}({q_{\rm m}}^2 + (1+\beta){q_{\rm z}}^2)\right]
+\beta {q_{\rm m}}^2 \left[q^2{q_{\rm m}}^2 + \frac{d\ln{\Omega^2}}{d\ln R}{q_{\rm z}}^2\right],
\eqnum{72}\end{aligned}$$
where the amplitude of the total wavenumber defined by $q^2(\tau)
\equiv {q_{\rm R}}(\tau)^2 + {q_{\rm m}}^2 + {q_{\rm z}}^2$ is a function of $\tau$ through eq. (69). Combining eqs. (69) and (72), one can evaluate a local, instantaneous growth rate at a given time.
With vanishing magnetic fields and thermal pressure, we would obtain from eq. (72) stable epicyclic oscillations. In the limit of strong magnetic fields and no rotation, eq. (72) is immediately reduced to $(\gamma^2+{q_{\rm m}}^2)(\gamma^4 + (1+\beta)q^2\gamma^2 + \beta{q_{\rm m}}^2q^2)=0$, the usual dispersion relations for the [Alfv$\acute{\rm e}$n]{} waves and the fast and slow MHD waves in a medium embedded with toroidal magnetic fields. In the presence of rotation with non-vanishing but weak fields, however, these [Alfv$\acute{\rm e}$n]{} and MHD modes are coupled to exhibit generally complex modal behaviors. They can be stable or unstable depending on the parameters, but it is always a slow MHD wave that becomes unstable because it has the lowest frequency so that there is a plenty of time during which destabilizing forces (centrifugal forces for NMRI) act on it. We have instantaneously growing solutions with real positive values of $\gamma$ provided that the last term in eq. (72) is negative. Thus, when $\beta{q_{\rm m}}\neq 0$, the local, instantaneous instability criterion in terms of the dimensionless variables for the NMRI can be written $${q_{\rm m}}^2\left(1 + \frac{{q_{\rm R}}(\tau)^2 + {q_{\rm m}}^2}{{q_{\rm z}}^2}\right)
+ \frac{d\ln\Omega^2}{d\ln R} < 0,$$ demonstrating that $d\Omega^2/d\ln{R} = \kappa^2 - 4\Omega^2<0$ is indeed a sufficient condition for the instability to arise when the magnetic field strength is negligible [@bal92; @fog95; @ter96]. Eq. (73) recovers the results of @bal92 for the instability regime for toroidal-field NMRI. The result of eq. (73) can be compared to the poloidal-field BH instability criterion of eq. (59b). Notice that the NMRI (with $B_{\rm 0z}=0$) vanishes completely as $\beta\rightarrow 0$, while the axisymmetric BH instability still exists even when $\beta=0$ (see §7.1).
Although they share the same instability criterion, the operating mechanisms for the NMRI of toroidal fields are quite different from the axisymmetric BH instability of poloidal fields. Both arise via destabilization of the slow mode. The NMRI mode, just as the poloidal BH mode, depends on shear to generate azimuthal fields from radial perturbations of the background fields. But there is more to the story. The key mechanism for the NMRI instability lies in the vertical MHD wave motions driven by the gradient of the total (initial plus perturbed) azimuthal fields, as schematically illustrated in Fig. 10. Since we suppose perturbations which are sinusoidal in both vertical and azimuthal directions, the perturbed azimuthal fields are also periodic in both directions. Rapid vertical motions with high ${k_{\rm z}}$, generated by $-{B_{\rm 0\phi}}(\partial B_{1\phi}/\partial z)$ stress, would produce over- and under-dense regions which regularly alternate along the azimuthal direction (Fig. 10a). And then, azimuthal fluid motions are induced, according to the equation of continuity, from over-dense regions to under-dense regions (Fig. 10b). Depending on the direction ($\mp \hat{\phi}$) of these induced motions, the coriolis and/or centrifugal force would alter the paths, radially inward or outward (Fig. 10c). Under the condition of field freezing, these radial motions would produce radial magnetic fields with a small amplitude from the background toroidal fields (Fig. 10d). These radial fields would in turn be sheared out to generate (positive or negative) perturbed azimuthal fields, due to the differential rotation of the background flows. When $d\Omega/dR <0$, the resulting azimuthal fields from initial and perturbed ones would be distributed (Fig. 10e) such that they reinforce the applied initial vertical perturbations (Figs. 10a and 10f), implying the MRI; the entire system would just oscillate with rotation-modified MHD frequencies if $d\Omega/dR >0$. This explains how the NMRI operates.
When ${k_{\rm z}}$ is large, the stabilization of the NMRI occurs when a magnetic tension from radially bent field lines exceeds the centrifugal or coriolis force (Figs. 10b and 10c). Shear [Alfv$\acute{\rm e}$n]{} waves with radial polarization can suppress the instability if the field lines are sufficiently strong or if the azimuthal wavenumber is large enough, as expressed by the dimensionless parameter ${q_{\rm m}}^2$ outside the parentheses in eq. (73). When ${q_{\rm z}}\ll {q_{\rm R}}$, on the other hand, MHD waves propagating along the radial direction stabilize the NMRI, as indicated by the terms inside parentheses in eq. (73); ${q_{\rm R}}(\tau)$ clearly reflects the stabilizing effect of the background shear.
The maximum instantaneous growth rate is achieved when ${q_{\rm R}}\approx 0$. In this case, eq. (73) implies instability if $$q^2_{\rm m} < q^2_{\rm m,crit} \equiv \frac{1}{2}
\left[-{q_{\rm z}}^2\! +\! \sqrt{{q_{\rm z}}^4-
4{q_{\rm z}}^2\frac{d\ln{\Omega^2}}{d\ln{R}}}\;\right].$$ Note that $q^2_{\rm m,crit} \rightarrow -d\ln{\Omega^2}/d\ln{R}$, for ${q_{\rm z}}\gg 1$, while $q^2_{\rm m,crit} \rightarrow {q_{\rm z}}\sqrt{-d\ln \Omega^2/
d\ln R}$, for ${q_{\rm z}}\ll 1$. Numerical solutions of eq. (72) with ${q_{\rm R}}=0$ are presented in Fig. 11. As both eq. (73) and Fig. 11 show, the maximum growth rates for toroidal-field background states are attained when ${q_{\rm z}}\rightarrow
\infty$, which is a sharp contrast to the axisymmetric poloidal-field BH instability that has fastest growing mode at moderate ${q_{\rm z}}$’s [cf. Fig. 9 and see also discussion in @bal98]. But both forms of the MRI have the same maximum growth rates at the same ${q_{\rm A}}\equiv
(\bold{k}\!\cdot\!\bold{v}_{\rm A})/\Omega$. Fig. 11b shows how growth rates depend on the sound speed. As the sound speed increases, the medium becomes more unstable. This reflects the incompressible nature of the NMRI. Even though the marginally critical wavenumber is independent of temperature (for $\beta \neq 0$), the virulence of the instability is greatly inhibited as $\beta$ decreases. For ${q_{\rm m}}\ll 1$, one can find from eq. (72) the temperature dependence of the limiting growth rate $$\gamma = {q_{\rm m}}\sqrt{-\frac{\beta}{\kappa^2(1+\beta)}\frac{d\Omega^2}{d\ln R}},$$ or $\gamma = {q_{\rm m}}\sqrt{3\beta/(1+\beta)}$ for a Keplerian rotation. Eq. (74) gives slopes of the growth rates for small ${q_{\rm m}}$ (Fig. 11). For magnetocentrifugally driven winds which are as cold as $\beta < 0.01$, the NMRI is not expected to play a significant role; the growth rate in dimensional units is $\sqrt{3}{c_{\rm s}}m/R$.
When the medium is incompressible ($\beta \rightarrow \infty$), eq. (72) allows the analytic expression for the instantaneous growth rate for pure toroidal-field background states, $$\gamma^2 = \left\{ \begin{array}{ll}
\frac{{q_{\rm z}}^2}{2q^2}\frac{\kappa^2}{\Omega^2}\left[\sqrt{
1 + 16\frac{q^2{q_{\rm m}}^2}{{q_{\rm z}}^2}\frac{\Omega^4}{\kappa^4}} -1 \right] - {q_{\rm m}}^2,
&{\rm if\;\;\;} \frac{q^2{q_{\rm m}}^2}{{q_{\rm z}}^2} + \frac{d\ln\Omega^2}{d\ln R} < 0, \\
& \\
0.
&{\rm otherwise}, \\
\end{array} \right.$$ When ${q_{\rm z}}\gg 1$, one can derive the maximum growth rate $\gamma_{\rm max} = |d\ln\Omega/d\ln R|/2$, which is achieved when $q^2_{\rm m,max} = -(d\ln\Omega^2/d\ln R) /2 - \gamma^2_{\rm max}$. It can be shown from eq. (72) or (75) that $d\gamma^2/d\tau \sim {q_{\rm m}}^3{q_{\rm R}}/q^2 \rightarrow 0$ as ${q_{\rm z}}\rightarrow
\infty$. This proves that the coherent wavelet approximation is self-consistent for the NMRI with high ${q_{\rm z}}$.
### Comparison With the Shearing Sheet Formalism
In order to compare the coherent wavelet solutions for toroidal-field NMRI with the results from the shearing sheet approximation, we directly integrate eqs. (61)$\sim$(67) over time, with given sets of initial conditions. In Fig. 12, we display the time evolution of all perturbed variables for ${q_{\rm m}}=0.1$, ${q_{\rm z}}=1$, and $\beta=100$, which are the same parameters as chosen for Fig. 3 of @bal92. We adopt a Keplerian rotation profile in what follows. The initial amplitudes are 0.1 for every variable except $b_{\rm R}=0.01$ and $b_{\rm z}=0.4$, and the initial ${\tilde{\tau}}$, where the orbit number ${\tilde{\tau}}\equiv \tau/2\pi = t\Omega/2\pi$, is allowed to be determined from the divergence free condition of the initial, perturbed magnetic fields. When ${\tilde{\tau}}< -20 $, the system responds with MHD wave motions before they start to grow. During this relaxation stage, fast MHD modes having large $|{q_{\rm R}}|$ are nearly longitudinal acoustic waves, affecting $u_{\rm 1R}$ and $\alpha$, while $u_{1\phi}$ and $b_\phi$ are mostly influenced by transverse slow modes. As time increases, $|{q_{\rm R}}|$ gradually decreases, permitting rotational shear to affect the overall dynamics. Once the condition (73) is satisfied, shear drives the slow modes to be unstable, following the process illustrated in Fig. 10. Even though the growth of disturbances shows a transient nature due to the kinematic growth of ${q_{\rm R}}$, the net amplification is about 9 orders of magnitude, a bit higher than Balbus & Hawley’s result. This is because the integration interval in Balbus & Hawley covered a slightly smaller part of the unstable time range. At later time when ${q_{\rm R}}$ has a large value, the system exhibits stable oscillations with the slow MHD wave frequency. Fig. 12 also shows the predicted amplification magnitude (thick solid line) from the coherent wavelet approximation (see below).
In Fig. 13, we plot the numerical growth rates for each variable calculated from Fig. 12 based on the direct numerical integrations in the shearing sheet formalism, together with the growth rate of the corresponding coherent wavelet solution. Here, a dimensionless instantaneous growth rate $\tilde{\gamma}$ as a function of ${\tilde{\tau}}$ is defined through $10^{\tilde{\gamma}{\tilde{\tau}}}=e^{\gamma \tau}$, (or $\tilde{\gamma}({\tilde{\tau}}) = 2\pi \gamma(\tau) \log e)$. Note that the heavy solid line for $\tilde{\gamma}$ drawn from eq. (72) fits well with various curves computed from the direct numerical integrations. The instantaneous growth rates are almost symmetric with respect to their maxima near ${\tilde{\tau}}=0$, as expected. Growth of the modes occurs only when $|{\tilde{\tau}}| < 18.3$, which is in good agreement with the results of the direct integrations, demonstrating the validity of the coherent wavelet approximation.
We define a dimensionless amplification magnitude as $$\Gamma({\tilde{\tau}}) \equiv \int_{-\infty}^{{\tilde{\tau}}}
\tilde{\gamma}({\tilde{\tau}}')d{\tilde{\tau}}'
= \log e \int_{-\infty}^{\tau} \gamma(\tau')d\tau'.$$ Then, $\Gamma({\tilde{\tau}})$ is an order of magnitude measurement of the amplification of an unstable mode during the time interval ($-\infty, {\tilde{\tau}}$). The total amplification is given by $10^{\Gamma(\infty)}$. When eq. (75) is substituted, the analytic evaluation of the integral in eq. (76) is not an easy task. In view of a shape of $\tilde{\gamma}({\tilde{\tau}})$ (Fig. 13), we further approximate $\tilde{\gamma}$ with a simple form $$\tilde{\gamma}= \left\{ \begin{array}{ll}
\tilde{\gamma}_{\rm o}(1- |{\tilde{\tau}}|/{\tilde{\tau}}_{\rm c})^{1-{q_{\rm m}}/2},
&{\rm if\;\;\;} |{\tilde{\tau}}|<{\tilde{\tau}}_{\rm c}, \\
0,
&{\rm otherwise}, \\
\end{array} \right.$$ where $$\eqnum{78a}
\tilde{\gamma}_{\rm o} \equiv
\sqrt{2} \pi\log e \left\{\frac{{q_{\rm z}}^2}{{q_{\rm m}}^2+{q_{\rm z}}^2}\frac{\kappa^2}{\Omega^2}
\left[\sqrt{
1 + 16\frac{({q_{\rm m}}^2+{q_{\rm z}}^2){q_{\rm m}}^2}{{q_{\rm z}}^2}
\frac{\Omega^4}{\kappa^4}} - 1 \right] - 2{q_{\rm m}}^2\right\}^{1/2},$$ and the termination epoch of growth ${\tilde{\tau}}_{\rm c}$ is defined by $$\eqnum{78b}
{\tilde{\tau}}_{\rm c} \equiv
\frac{1}{2\pi{q_{\rm m}}}\left|\frac{d\ln\Omega}{d\ln R}\right|^{-1}
\sqrt{-\left(\frac{d\ln \Omega^2}{d\ln R} \right)
\frac{{q_{\rm z}}^2}{{q_{\rm m}}^2}-{q_{\rm m}}^2-{q_{\rm z}}^2}.$$ Notice that eq. (77) is valid only if ${\tilde{\tau}}_{\rm c}$ is real, that is, only if the condition (73) is satisfied. From eqs. (76) and (77), the total amplification magnitude is easily found to be $$\Gamma(\infty) = \frac{4\tilde{\gamma}_{\rm o}{\tilde{\tau}}_{\rm c}}{4-{q_{\rm m}}},$$ which is illustrated with solid contours in Fig. 14. Also shown with dotted contours are the direct results from numeral integration of eq. (75), which are in excellent agreement with $\Gamma(\infty)$. The thick contour is the locus of ${\tilde{\tau}}_{\rm c}=0$, demarcating the stable and unstable regions: the system is stable at the right hand side of the thick contour. In the ${q_{\rm m}}-{q_{\rm z}}$ plane, the total amplification tends to be greater as ${q_{\rm z}}$ becomes larger and as ${q_{\rm m}}$ becomes smaller. This is because the NMRI with $B_{\rm 0z}=0$ acquires maximum instantaneous growth rates at ${q_{\rm z}}=\infty$ (Fig. 11a) and because the shearing time is longer with smaller ${q_{\rm m}}$ (cf. eq. \[78b\]). For comparison, we also include in Fig. 14 the results from the shearing sheet equations for four parameter sets: (${q_{\rm m}}$, ${q_{\rm z}}$) = (0.03, 0.1), (0.1, 1), (1, 10), and ($\sqrt{2}$, 100$\sqrt{2}$), and $\beta$=100 for all cases: these are marked with dots on ${q_{\rm m}}-{q_{\rm z}}$ plane, labeled by the respective exact and estimated (in parentheses) amplification magnitudes. Note that all of the estimated amplification magnitudes are within 5% of the results of direct shearing sheet integrations. This indicates that eq. (79) is an excellent analytic estimate for the amplifications of incompressible NMRI modes.
Generalized MRI
---------------
Motivated by the success of the coherent wavelet method in finding the solutions of the NMRI with purely toroidal background fields, we now generalize both the axisymmetric BH and NMRI instabilities by considering non-axisymmetric perturbations applied to the rotating medium threaded by both vertical and azimuthal magnetic fields. We include the effect of thermal pressure and allow the angular velocity $\Omega$ to be a function of $R$, but ignore any other radial variations in the initial state. We adopt the shearing sheet coordinates as before, and linearize eqs. (1)$\sim$(4). After applying perturbations in the form of eq. (60), we assume that the perturbations evolve with time as $e^{\gamma (t) t}$ with the coherent wavelet condition (i.e., $d\ln \gamma(t)/dt \ll \gamma(t)$). Following the same procedure as §7.2, we obtain the general instantaneous dispersion relation for the MRI (now written in dimensional form)
$$\begin{aligned}
0 = \gamma^6 &+& \gamma^4
\left[({c_{\rm s}}^2+{v_{\rm A}}^2)k^2 + \kappa^2 + (\bold{k}\cdot\bold{{v_{\rm A}}})^2 \right]
\nonumber \\
&+& \gamma^2 \left[(2{c_{\rm s}}^2 + {v_{\rm A}}^2)(\bold{k}\cdot\bold{{v_{\rm A}}})^2k^2 +
\kappa^2\left({c_{\rm s}}^2{k_{\rm z}}^2 + {v_{\rm A\phi}}^2(\frac{m^2}{R^2}+{k_{\rm z}}^2) \right)
+(\bold{k}\cdot\bold{{v_{\rm A}}}){k_{\rm z}}{v_{\rm Az}}\frac{d\Omega^2}{d\ln R}
\right]
\nonumber \\
&+& {c_{\rm s}}^2 (\bold{k}\cdot\bold{{v_{\rm A}}})^2
\left[(\bold{k}\cdot\bold{{v_{\rm A}}})^2 k^2 +
\frac{d{\Omega^2}}{d\ln R}{k_{\rm z}}^2\right],
\eqnum{80} \end{aligned}$$
where $k^2\equiv k_{\rm R}^2(t) + m^2/R^2 + {k_{\rm z}}^2$ with the radial wavenumber defined by $k_{\rm R}(t) = - mtd\Omega/dR$ when we choose $t_{\rm o} = -{k_{\rm R}}(t_{\rm o})/m\Omega'$. When either $m=0$ (axisymmetric case) or ${v_{\rm Az}}=0$ (pure toroidal field case), eq. (80) becomes identical respectively with eq. (57) for the BH modes or eq. (72) for the NMRI modes.
From eq. (80), we obtain the instantaneous instability criteria for the generalized MRI modes
$$\begin{aligned}
{v_{\rm A}}^2k^2(t)(\bold{k}\cdot\bold{v}_{\rm A})^2 +
\kappa^2{v_{\rm A\phi}}^2\left(\frac{m^2}{R^2}+{k_{\rm z}}^2 \right)
+ (\bold{k}\cdot\bold{v}_{\rm A})
{v_{\rm Az}}{k_{\rm z}}\frac{d\Omega^2}{d\ln R}\; & < \;0,
\quad{\rm for} \quad {c_{\rm s}}= 0,
\\
(\bold{k}\cdot\bold{v}_{\rm A})^2
\left(1 + \frac{{k_{\rm R}}^2(t) + m^2/R^2}{{k_{\rm z}}^2} \right) +
\frac{d\Omega^2}{d\ln R}\; & <\; 0,
\quad{\rm for} \quad {c_{\rm s}}\neq 0,\end{aligned}$$
which are obviously the generalizations of eqs. (59) and (73). It can also be shown that when $i=90^{\rm o}$, both equations (81a) and (81b) become identically ${v_{\rm Az}}^2({k_{\rm R}}^2(t)+m^2/R^2 + {k_{\rm z}}^2) + d\Omega^2/d\ln R <0$. With an $\Omega \propto R^{-a}$ rotation profile, eq. (81a) gives the sufficient condition for the instability in an extremely cold medium: $\sin i > (4-2a)/(4-a)$. Keplerian flows for example become unstable only if $i\gtrsim24^{\rm o}$, indicating that cold, $B_\phi$-dominated media are not subject to the generalized non-axisymmetric MRI disturbances, just as we found earlier that axisymmetric BH modes are also stable in cold flows for small $i$. We remark that the case with $a\rightarrow 2$, as potentially possible in MHD winds from boundary layers, can just barely satisfy the cold medium instability criterion for $i\rightarrow 0$. Note that unlike the NMRI mode with ${v_{\rm Az}}=0$, maximum growth rates in the ${c_{\rm s}}\neq 0$ case are not achieved as ${k_{\rm z}}\rightarrow \infty$. In fact, high-${k_{\rm z}}$ or high-$m$ disturbances are efficiently stabilized by [Alfv$\acute{\rm e}$n]{} and/or MHD waves whenever [*both*]{} poloidal and toroidal fields are present. However small they may be, therefore, inclusion of poloidal fields would yield a different result from the case with pure toroidal fields (this point was previously noted by Balbus & Hawley 1998). Again the stabilizing effect of kinematic shear appears through the time dependence of ${k_{\rm R}}^2(t)$, when $m\neq0$.
Fig. 15a shows how the compressible BH modes are stabilized by azimuthal magnetosonic waves. Here, we confine consideration to the radial wavenumber ${k_{\rm R}}=0$. As ${q_{\rm m}}(= {v_{\rm A}}m/R\Omega)$ increases, both the growth rates and the ranges in ${q_{\rm z}}(={v_{\rm A}}{k_{\rm z}}/\Omega)$ of unstable modes decrease. This is because if ${q_{\rm m}}\neq0$, azimuthally displaced material feels relatively stronger restoring forces due to both thermal and magnetic pressures of the medium as well as stronger tension forces from bent field lines. Non-axisymmetric poloidal-field BH instability modes become stabilized with increasing values of $m$. When ${q_{\rm m}}> \sqrt{-d\ln \Omega^2/d\ln R}$ ($=\sqrt{3}$ for a Keplerian rotation), the instability is strictly cut off, even when the effect of kinematic shear is not taken into account.
We remark that the role of temperature of the medium to the BH instability is different between axisymmetric (with ${q_{\rm m}}=0$ and $i\neq 90^{\rm o}$; see Fig. 9) and non-axisymmetric (with ${q_{\rm m}}\neq 0$ and $i=90^{\rm o}$; see Fig. 15) cases. When $i=90^{\rm o}$, as already explained in §7.1, the axisymmetric BH instability with ${q_{\rm m}}=0$ is independent of $\beta$, because only Alfv$\acute{\rm e}$n and sound waves exist and they do not interact with each other. When ${q_{\rm m}}=0$ and $i\neq 90^{\rm o}$, magnetic pressure induces vertical MHD wave motions which tend to stabilize the system when $\beta$ is small. If $\beta \gg 1$, however, the vertical wave motions become nearly acoustic, leaving the toroidal component of perturbed fields unaffected and permitting higher growth rates. When ${q_{\rm m}}\neq 0$ and $i=90^{\rm o}$, on the other hand, the coupling of thermal pressure with magnetic pressure occurs through azimuthal MHD motions, and the growth rate depends only weakly on $\beta$.
Fig. 15b shows loci of equi-growth rate on the ${q_{\rm z}}\!-\!{q_{\rm m}}$ plane for $i=10^{\rm o}$ and $\beta=0.01$ (dotted contours) and $\beta=100$ (thin solid contours). For $0<{q_{\rm m}}<1.17$, there exist upper and lower critical vertical wavenumbers, $q_{\rm z,u}$ and $q_{\rm z,l}$ such that the system is unstable with $q_{\rm z,l}<{q_{\rm z}}<q_{\rm z,u}$. When ${q_{\rm z}}> q_{\rm z,u}$, disturbances are stabilized by MHD waves propagating mainly along vertical direction, while perturbations with ${q_{\rm z}}<q_{\rm z,l}$ approach stable Alfv$\acute{\rm e}$n waves. Each contour has a slope of $\sim -\tan i$ ($= -0.18$ for $i=10^{\rm o}$) at both ends. Note that dotted contours with lower $\beta$ are labeled with much smaller growth rates than solid ones with higher $\beta$, even though they are similar in shape. Compared to Fig. 9 or Fig. 15a, this implies that the NMRI instability with a toroidal field configuration is more sensitive to temperature than the axisymmetric/non-axisymmetric BH instability with poloidal fields.
When $\beta \rightarrow \infty$, from eq. (80) we have the instantaneous growth rates for the generalized MRI modes $$\gamma^2 =
\frac{\kappa^2{k_{\rm z}}^2}{2k^2}\left[\sqrt{
1 + 16({\bf k\cdot{v_{\rm A}}})^2\frac{k^2\Omega^2}{{k_{\rm z}}^2\kappa^4}} -1 \right] -
({\bf k\cdot{v_{\rm A}}})^2,$$ for $({\bf k\cdot{v_{\rm A}}})^2k^2/{k_{\rm z}}^2 + d\Omega^2/d\ln R < 0$, which is also a generalization of eq. (75). With weak magnetic field strength, one can show from eq. (82) that $\gamma^{-3}d\gamma^2/dt \sim -m{k_{\rm R}}({\bf k\cdot{v_{\rm A}}})/(\Omega{k_{\rm z}}^2)
\rightarrow 0$ as $({\bf k\cdot{v_{\rm A}}})/\Omega \rightarrow 0$. Thus, we see that for the generalized MRIs, the coherent wavelet approach is self-consistent in the weak field limit. When the field strength is moderate, on the other hand, we obtain $d\gamma^2/dt \sim m\Omega'{k_{\rm R}}({\bf k\cdot{v_{\rm A}}})^2/k^2$. Since $\gamma\sim\Omega\sim ({\bf k\cdot{v_{\rm A}}})$ in this case, the coherent wavelet condition is met only when $m|\Omega'| \ll {k_{\rm R}}\gamma$ (the weak shear limit). Of course, the predominantly toroidal-field case that becomes unstable with ${k_{\rm z}}\gg 1$ also satisfies the coherent wavelet condition, since $k^2$ becomes arbitrarily large without increasing the $({\bf k\cdot{v_{\rm A}}})$-term, as discussed in §7.2.2.
Comparing eq. (75) with eq. (82), we note that the incompressible MRI can be generalized simply by replacing the dimensionless azimuthal wavenumber ${q_{\rm m}}$ with $({\bf k\cdot{v_{\rm A}}})/\Omega$. Therefore, we can write the net amplification magnitude for the generalized incompressible MRI as $$\Gamma(\infty) = \frac{ 4\log e}
{4-{q_{\rm A}}}\;\gamma_{\rm o}t_{\rm c}$$ where the dimensional peak growth rate $\gamma_{\rm o}$ and the cut-off time of the instability $t_{\rm c}$ are defined by $$\frac{\gamma_{\rm o}^2}{\Omega^2} \equiv
\frac{\kappa^2\cos^2\theta}{2\Omega^2}\left[
\sqrt{1 + 16\frac{{q_{\rm A}}^2}{\cos^2\theta}\frac{\Omega^4}{\kappa^4}}
-1 \right] - {q_{\rm A}}^2$$ and $$t_{\rm c}\equiv
\frac{1}{\sin\theta}\left|\frac{d\Omega}{d\ln R}\right|^{-1}
\sqrt{-\frac{d\ln\Omega^2}{d\ln R}\frac{\cos^2\theta}{{q_{\rm A}}^2}
-1},$$ respectively. The net amplification can thus be completely determined by the two parameters: the dimensionless wavenumber ${q_{\rm A}}\equiv ({\bf k \cdot {v_{\rm A}}})/\Omega$ projected in the direction of initial equilibrium magnetic fields, and the angle $\theta \equiv \tan^{-1}(m/R{k_{\rm z}})$ of the wavenumber vector with respect to the vertical axis. Note that $\gamma_{\rm o}\rightarrow 0$ with vanishing ${q_{\rm A}}$, while $t_{\rm c}\rightarrow \infty$ as $\theta\rightarrow 0$, indicating that low-$m$ instabilities show higher net amplifications than high-$m$ disturbances as long as ${q_{\rm A}}\neq 0$. The total amplification magnitudes, eq. (83), are plotted in Fig. 16 with thin solid contours. For comparison, we also plot the numerical results from eqs. (76) and (82) with dotted contours. We assume a Keplerian rotation profile. The heavy curve with ${q_{\rm A}}^2=3\cos^2\theta$ draws the locus of the marginal stability. In the limit of a weak magnetic field strength (i.e., ${q_{\rm A}}\rightarrow 0$), it can be shown from eqs. (83)$\sim$(85) that $\Gamma(\infty) = (2\log e)\Omega/(\kappa\tan\theta)$, inversely proportional to $\theta$ (for $\theta \ll 1$) but independent of ${q_{\rm A}}$, as illustrated in Fig. 16. Also shown in Fig. 16 are the results from the direct temporal integrations of shearing-sheet equations with $\beta=100$ as filled circles (for $i=90^{\rm o}$), filled triangles (for $i=30^{\rm o}$), and open circles (for $i=0^{\rm o}$), labeled by the respective exact and estimated (in parentheses) amplification magnitudes. These two results agree very well, implying that the coherent wavelet approach indeed provides excellent approximations to the solutions for amplification of generalized MRIs.
From Fig. 16, it is apparent that it is the locally near-axisymmetric (in the sense $m/R{k_{\rm z}}= \tan\theta \ll 1$) disturbances that experience maximum amplification, with the amplification magnitude only weakly dependent on ${q_{\rm A}}= ({\bf k \cdot {v_{\rm A}}})/\Omega$ within the unstable regime (${q_{\rm A}}\lesssim 1$). The increase in amplification factor with $R{k_{\rm z}}/m$ predicted from linear theory may in part explain the larger amplitudes of power spectra for modes with larger ${\bf \hat{k}\cdot\hat{z}}$ measured from nonlinear simulations of the saturated MRI [cf. @haw95]. In addition, for the case of pure toroidal fields, Fig. 16 suggests only low amplification factors unless ${k_{\rm z}}$ is very large, which may help explain why @haw95 found lower magnetic field saturation amplitudes in cases with initial $B_{\rm z} =0$.
Summary and Discussion
======================
General Conclusions Based on Linearized Analysis
------------------------------------------------
Through linear analyses of the ideal MHD equations, we have explored the stability of shearing, rotating flows to a wide range of (primarily local) disturbances. The chief motivation for this study is to characterize the internal instabilities that could develop in disk winds that emanate from an extended region of a differentially rotating protostellar disk around a young star. The dynamics of such winds has inspired intensive theoretical effort because they may be responsible for observed YSO jets and outflows. In our analysis, we include both results based on generic density, magnetic field, and flow profiles, and results which adopt as initial equilibrium configurations the power-law asymptotic solutions of self-confined cylindrically symmetric winds presented by Ostriker (1997): $\rho \propto R^{-q}, B_{\phi} \propto B_{\rm z} \propto R^{-(1+q)/2}$, and $v_{\phi} \propto v_{\rm z} \propto R^{-1/2}$. For most of our analysis (§§2-6), the flows were assumed to be cold enough that thermal effects can be ignored compared with magnetic forces. To make contact with other studies of shear-induced MHD instabilities in rotating disks, we also consider stability of specific models which include non-zero thermal pressure (§7). For the lowest-order “fundamental” modes, we employ a normal-mode analysis with free Lagrangian boundary conditions to find eigenvalues and eigenfunctions of both stable and unstable modes (§3). For higher-order modes, we employ three different local techniques to study growth of unstable disturbances: In §§4-6, we present numerical and analytic solutions of dispersion relations obtained from normal mode analyses in the $R{k_{\rm R}}\gg 1$ WKB limit. These are exact for $m\Omega'+{k_{\rm z}}v_{\rm 0z}'=0$ disturbances and are valid for a limited time for weakly-shearing circumstances where ${k_{\rm R}}\gg m|\Omega'/\Omega|, {k_{\rm z}}|Rv_{\rm 0z}'/v_{\rm 0z}|$ (see also §8.2 below). In §7, we employ temporal integrations of the shearing-sheet equations to study MRI modes (which are cut off for $R{k_{\rm R}}\gg 1$). We also introduce, in §7.2, a “coherent wavelet” formalism which adapts modal analyses for situations where shear is considerable (i.e., small $R{k_{\rm R}}/m$); the coherent wavelet analysis is equivalent to a WKB approach in the temporal domain. We include a comparison of results from the shearing-sheet and coherent-wavelet techniques applied to MRIs, in §7.2.3.
Applying these techniques we have identified a total of nine different unstable or overstable families of disturbances that occur for a wide range of flow parameters: five (FM, BH, ATB, PB, and TR) of them are axisymmetric and the other four (NTB, GPB, PR, and NMRI) are non-axisymmetric. Table 1 summarizes the properties of these modes. The main general conclusions drawn from the analysis in this work can be summarized as follows:
\(1) Systems having a primarily azimuthal magnetic field, for example, disk winds far from their source, are susceptible to the fundamental (FM), axisymmetric (ATB) and non-axisymmetric (NTB) toroidal buoyancy, non-axisymmetric magnetorotational instability (NMRI) and toroidal resonance (TR) modes. Unstable fundamental modes (see §3.2) are concentrated in the central parts of jets, and occur in $B_\phi$-dominated flows when the logarithmic gradient of the magnetic field is steeper than $\approx -0.75$ (cf. eqs. \[24\] and \[25\]). Growth rates of unstable FM are comparable to inner-wind [Alfv$\acute{\rm e}$n]{} frequencies. Long wavelength modes with large amplitudes at large radii are all stable, for power-law wind profiles. The TR mode (see §6.1.4) is an overstability, with growth suppressed when ${k_{\rm R}}$ increases through shear of the vertical velocity, simply becoming oscillatory MHD waves. The axisymmetric toroidal buoyancy mode (ATB; §6.1.2) is activated initially by the buoyancy force and subsequently by bending poloidal magnetic fields. In geometrical form, it is locally similar to the sausage mode of a plasma column confined by toroidal fields, and leads to radial mixing. Because growth rates are larger on smaller scales, ATB can contribute to the generation of local turbulence in disk winds. The non-axisymmetric toroidal buoyancy mode (NTB; see §6.2.1) is much like the Parker instability, but with the centrifugal force replacing the role of external gravity. Although the normal-mode analysis for NTB has the largest temporal validity at small $m/R{k_{\rm R}}$, the instantaneous growth rate increases with increasing $m/R{k_{\rm R}}$ (cf. eq. \[52\] and Fig. 8c). We thus return to the NTB in §8.2, below, applying time-dependent techniques to study the $R{k_{\rm R}}/m \ll 1$ limit. Because the NTB is present whenever radial magnetic forces are non-zero, it may be important in promoting radial mixing. Both of the toroidal buoyancy instabilities require non-zero magnetic forces in the equilibrium state. The rarefied and cold conditions of disk winds do not favor the development of the NMRI (see §7.2.2 and §7.3). Like the original (poloidal field) magnetorotational (BH) instability, NMRI requires $d\ln{\Omega^2}/d\ln{R} <0$, but also requires a relatively incompressible medium, as is provided by the relatively dense and warm (${c_{\rm s}}\gtrsim {v_{\rm A}}$) conditions in an accretion disk. We show the NMRI vanishes in the limit of ${c_{\rm s}}/{v_{\rm A}}\rightarrow 0$. We further discuss perturbations in cold, $B_\phi$-dominated flows in §§8.2 and 8.3, below.
\(2) Systems having primarily axial magnetic fields, for example, accretion disks or winds very near their origin, are susceptible to the Balbus-Hawley (BH), poloidal buoyancy (PB and GPB), and poloidal resonance (PR) modes. The well-known axisymmetric Balbus-Hawley instability (BH; see §6.1.3 and §7.1) is the most efficient member of the family of magnetorotational instabilities (MRIs; see §7.3). It will work to produce channel flows, eventually generating fully-developed MHD turbulence through coupling to non-axisymmetric disturbances in the non-linear stage. Driven by background magnetic pressure and the centrifugal force, the axisymmetric poloidal buoyancy mode (PB; see §6.1.1) requires a gradient in the magnetic field strength to be unstable. If the field distribution is steep enough, the poloidal buoyancy modes would also work effectively to generate radial mixing and turbulence over much smaller scales than the BH instability. Because of their overstable characteristics, the impact on the system of poloidal resonance modes (PR; see §6.2.3) would be best evaluated with a global rather than local formalism. Configurations with shallower background magnetic gradients $(q<1)$ are also subject to a non-axisymmetric poloidal buoyancy instability (GPB; see §6.2.2) which arises in part from geometric effects.
\(3) In distinction to the original, incompressible, axisymmetric BH instability, we found that the compressible axisymmetric BH mode is strongly stabilized by the presence of an azimuthal magnetic field if the medium has substantially sub-Alfv$\acute{\rm e}$nic sound speeds. For example, in a cold rotating flow with $\Omega \propto R^{-3/2}$, the axisymmetric BH instability would be completely suppressed if the local pitch angle $i \equiv \tan^{-1}(B_{\rm z}/B_\phi)$ is less than 30$^{\rm o}$ (cf. eq. \[47\]). In an incompressible medium (as provided by a disk with ${c_{\rm s}}\gtrsim {v_{\rm A}}$), faster sound waves preserve perturbed toroidal fields from being dispersed by MHD wave motions, thereby providing a favorable condition for the BH instability. When the field configuration is purely poloidal, the compressible BH instability is identical with its incompressible counterpart, independent of temperature (cf. eqs. \[58\] and \[59\]).
\(4) Even though they share the same instability criterion (cf. eqs. \[59b\] and \[73\]), the operating mechanisms for the NMRI of purely toroidal $\bold{B}$-fields is entirely different from the axisymmetric BH instability of primarily poloidal $\bold{B}$-fields. In the NMRI (see §7.2), vertical MHD wave motions driven by magnetic pressure play an essential role in the feedback loop for induced radial disturbances, while the axisymmetric BH instability tends to be stabilized by vertical wave motions. Faster sound speeds produce higher growth rates in both instabilities, but for different reasons: in the NMRI by activating azimuthal fluid motions preceded by the vertical MHD wave motions; in the BH instability by maintaining the perturbed azimuthal fields generated by shear (when ${B_{\rm 0\phi}}\neq 0$). Because of their non-axisymmetric nature, the NMRI has a transient growth, stabilized by the growth of ${k_{\rm R}}$ from kinematic azimuthal shear. For the NMRI mode, we show explicitly by comparison to direct temporal integrations of the shearing sheet equations that the growth rate at ${k_{\rm R}}=0$ can be used to provide a good estimate of the net amplification magnitude (see §7.2.3).
\(5) The coherent wavelet formalism we develop (§7.2) may be used to compute instability criteria and net amplification factors for generalized MRI disturbances with arbitrary magnetic field and wavevector orientations (§7.3). Eq. (80) gives the instantaneous dispersion relation for generalized MRIs. For strongly compressible flows (${c_{\rm s}}/{v_{\rm A}}\rightarrow 0$), instability does not occur in $B_\phi$-dominated configurations (cf. eq. \[81a\]); in this case, flows with an $\Omega\propto R^{-3/2}$ rotation law can be unstable only when the magnetic pitch angle $i>24^{\rm o}$. Because MHD disk winds generally have very small pitch angles, this result has the important implication that such [*winds will not be subject to the development of strong internal turbulence that occurs as a consequence of nonlinear MRIs in disks*]{}. The absence of MRIs in cold, $B_\phi$-dominated winds may be crucial in enabling them to propagate over large distances from their sources. High-${k_{\rm z}}$ and/or high-$m$ modes of the generalized MRI are stabilized by MHD waves, which is a sharp contrast with the NMRI of [*purely toroidal*]{} fields in which maximum growth rates are attained at ${k_{\rm z}}\rightarrow \infty$. For incompressible flows, the amplification factor for all MRIs can be written analytically in terms of ${q_{\rm A}}= ({\bf k \cdot {v_{\rm A}}})/\Omega$ and $\theta = \tan^{-1}(m/R{k_{\rm z}})$ (eqs. \[83\]$\sim$\[85\]); within the unstable regime (${q_{\rm A}}<
|d\ln\Omega/d\ln R|^{1/2}$, from eq. \[81b\]), the amplification is $\sim \exp[2\Omega/(\kappa\tan\theta)]$, favoring “locally-axisymmetric” disturbances.
Effect of Shear on Dynamical Growth of Buoyancy Instabilities
-------------------------------------------------------------
Apart from the results of §7 where we adopted the shearing sheet formulation of the dynamical equations to study MRIs, the results in this work have been elicited on the basis of the local normal mode analyses. As described in §4, these modes may have a limited range of temporal validity, due to the effects of background shear. For axisymmetric disturbances with negligible vertical shear (i.e., $m\Omega'+{k_{\rm z}}v_{\rm 0z}'\rightarrow 0$), the results presented in §5 and §6 are acceptable for all time; the modes with pure imaginary $\omega$ will show an exponential growth without interruption over arbitrarily long time until nonlinearity sets in. However, for non-axisymmetric disturbances, or for flows with non-negligible vertical shear, unstable modes identified in §5 and §6 are not purely growing. As time evolves, the differential velocities build up the radial wavenumber through the kinematic shear (cf.eq. \[40\]), which in turn tends to stop the further growth of disturbances. This can be seen directly through the suppression of instabilities in the local analysis when ${k_{\rm R}}$ is large (cf. Fig. 8). The characteristic time for the wave pattern to change by a fraction $\epsilon$ is $t=\epsilon{k_{\rm R}}/|m\Omega'+{k_{\rm z}}v_{\rm 0z}'|$; over this interval, the disturbance will be amplified by a factor $\exp(\epsilon{k_{\rm R}}{\rm Im}({\tilde{\omega}})/|m\Omega'+{k_{\rm z}}v_{\rm 0z}'|)$. When ${k_{\rm R}}\gg m/R, {k_{\rm z}}$, Fig. 8 shows that Im(${\tilde{\omega}}) \propto {k_{\rm R}}^{-1}$, so that the net amplification factor is nearly independent of ${k_{\rm R}}$. Since, however, Im(${\tilde{\omega}}$) is not larger than $\sim|m\Omega'+{k_{\rm z}}v_{\rm 0z}'|{k_{\rm R}}^{-1}$, only order-unity amplification can be expected for disturbances which are consistent with the requirements for quasi-steady normal mode analysis.
Because the normal-mode dispersion relations indicate larger values of growth rates when $R{k_{\rm R}}/m$ is small (which is however not self-consistent with the WKB treatment), it is desirable to extend investigation to allow for $R{k_{\rm R}}/m$ small. The coherent wavelet formalism used for the MRI in §7 suggests that when $m\gg 1$ (or $R{k_{\rm z}}\gg 1$ for $v_{\rm 0z}'\neq 0$ cases), this can be done by regarding ${k_{\rm R}}$ as a time-dependent variable according to eq. (40) and using the asymptotic dispersion relations of §6 (i.e., eq. \[43\] for PB, eq. \[45\] for ATB, eq. \[52\] for NTB, and eq. \[53\] for GPB). To verify this argument, we specifically consider the NTB modes (which are one of the chief instabilities in $B_\phi$-dominated winds) and compare the results with the shearing sheet temporal integrations. For the latter, we set $B_{\rm z}=0$ and integrate eqs. (26)$\sim$(31) in time, setting all of the coefficients to constant values. The resulting instantaneous growth rates and time evolutions of variables are plotted in Fig. 17 as functions of the normalized time $\tau = t\Omega$. We omitted the ${k_{\rm R}}$-dependent term in eq. (26) in order to remove rapid oscillations arising from a phase mismatch between the density and radial velocity; the amplitude evolution is independent of this term. We also neglected the vertical velocity shear and and selected $q={k_{\rm R}}(0)={k_{\rm z}}=0$, $R\Omega = 0.1 {v_{\rm A\phi}}$, and $m=100$. As initial conditions, we chose 0.1 for every variable except $b_{\rm R}=0.01$ and integrated the system of the linearized equations. Various curves are computed from direct numerical integrations of shearing sheet equations, while the heavy solid lines are drawn from the normal mode solution, eqs. (52) and (86) (see below), after taking allowance for the time dependence of ${k_{\rm R}}$. The rapid fluctuations of the perturbed variables for $\tau<0$ are due to MHD waves with high $|{k_{\rm R}}|$, disappearing after variables grow substantially. Again, most of growth occurs over a relatively short period of time near ${k_{\rm R}}\sim 0$. Note an excellent agreement between the results from two different approaches; we have also obtained similar results with integration from other initial conditions. This confirms that our normal mode results can also be applied to high-$m$ disturbances if ${k_{\rm R}}$ is allowed to vary with time.
Using eq. (52) with ${k_{\rm R}}(t)$ from eq. (40), we integrate ${\tilde{\omega}}$ over time to estimate the net amplification for the NTB modes $$\frac{\chi_1(t)}{\chi_1(0)} = \exp{\int_0^t {\rm Im}({\tilde{\omega}})\, dt} =
\left[ \frac{3}{2}\Omega t + \sqrt{1+\left(\frac{3}{2}\Omega t\right)^2}
\right]^{(1-q){v_{\rm A\phi}}/3R\Omega},$$ where we put $v_{\rm 0z}'={k_{\rm R}}(0)=0$ and assume a Keplerian rotation. Thus instead of an exponential growth, at later time of evolution we have the power-law growth due to the kinematic shear. This behavior is distinct from the MRI modes, which are strictly stable for large enough ${k_{\rm R}}$. However, the continued local growth of buoyancy perturbations is offset by the role of kinematic shear in mixing phases of disturbances. Considering waves on $z$=constant plane in a square with sides $L$, the maximum averaged contrast in any variable relative to the mean value is $(\lambda/L)\chi_1(t)$, where $\lambda$ is the local wavenumber of the waves. With $\lambda \sim {k_{\rm R}}(t)^{-1}$ and using eqs. (6) and (86), the average contrast for the NTB modes evolve with time as $\sim (\Omega t)^{\sqrt{2(1-q)}/3-1}$, vanishing as $t \rightarrow \infty$ for $0<q<1$. From eq. (86), amplification factors are essentially scale-free. Although there may be a significant growth of the NTB modes on large scales, their dynamical effect on small scales is limited by the phase mixing due to shear.
In the presence of vertical shear, the evolution of ATB modes is also affected by the kinematic growth of ${k_{\rm R}}$ although they are axisymmetric modes. The net amplification of the ATB modes follows a power-law growth as that of the NTB mods does. In fact, eq. (86) with $\Omega$ replaced by $2v_{\rm 0z}'/3$ gives the temporal behavior of the net amplification for the ATB modes.
Discussion of Applications to Protostellar Winds
------------------------------------------------
In order for a disk wind to overcome the gravitational barrier due to a central object and to be centrifugally launched from the surface of a Keplerian disk, the poloidal components of field lines should thread the disk at an angle of 30$^{\rm o}$ or more from the axis [@bla82]. Once material starts to flow outward along such field lines, it is accelerated primarily in the radial direction by the centrifugal force or by the pressure gradient in the toroidal field. Beyond the [Alfv$\acute{\rm e}$n]{} surface where the local, poloidal component of the flow velocity is equal to that of the [Alfv$\acute{\rm e}$n]{}wave velocity, the magnetic field is not strong enough to play a role of “a rigid wire", and the inertia of gas becomes important, winding up the field lines to be progressively more toroidal. In this process, the azimuthal flow velocity decreases below the corotation value. With the increase in the azimuthal component of the magnetic field, the associated hoop stress provides the collimation of the outflow and causes the streamlines to bend upward. The radial flow velocity of the outflow is still positive, although it decreases gradually, eventually becoming zero at the cylindrical asymptotic limit. The power-law solutions (with ${v_{\rm R}}={B_{\rm R}}=0$) we adopted for many specific cases represent the asymptotic limit of each streamline.
@ost97 presented self-similar steady solutions for disk winds with cylindrical asymptotics and gave the asymptotic fluid and [Alfv$\acute{\rm e}$n]{}speeds and the location of the asymptotic streamlines, characterized by $q$ together with ${R_{\rm A}}/R_1$ or $R_0/R_1$, where $R_0, {R_{\rm A}}$, and $R_1$ denote the radii of the footpoint, the [Alfv$\acute{\rm e}$n]{} surface, and the asymptote of each given streamline, respectively. Typical numerical values for those solutions are $\Omega = 0.2\Omega_0$, ${v_{\rm A\phi}}/R = 0.42\Omega_0$, and ${v_{\rm A\phi}}/v_{\rm z}=6$ for $q=0.5$, and $\Omega = 0.1\Omega_0$ and ${v_{\rm A\phi}}/R = 0.45\Omega_0$ for $q=0.9$, where $\Omega_0$ is the Keplerian rotation rate at the streamline footpoint.
Using these values we can estimate the growth times of the global fundamental mode and the fastest growing (with ${k_{\rm R}}$ near 0) toroidal buoyancy modes. The foregoing analysis suggests that these disturbances will play the most significant dynamical role, given the ineffectiveness of MRIs in cold, $B_\phi$-dominated flows. We define the time to grow by $\Gamma$ orders of magnitude as $t_\Gamma$. For FM, $t_{\Gamma,\rm FM} = \Gamma/(|\omega|\log e)$, so from Fig. 2 and eq. (25) we have for $q=0.9$ $$t_{\Gamma,\rm FM} \sim
\frac{14\Gamma}{\Omega_{\rm 0,i}} =
25\; \Gamma \;{\rm days} \;
\left(\frac{M}{M_\odot}\right)^{-1/2}
\left(\frac{R_{\rm 0,i}}{0.1\;\rm AU}\right)^{3/2}.$$ For ATB and NTB, from eqs. (45) and (52), we have for $i\rightarrow 0$ and $q=0.5$ $$t_{\Gamma,\rm ATB} \sim
\frac{10^\Gamma}{0.21\Omega_0} =
24\times 10^\Gamma\; {\rm yrs} \;
\left(\frac{M}{M_\odot}\right)^{-1/2}
\left(\frac{R_0}{10\;\rm AU}\right)^{3/2}\;\;\;{\rm and}$$ $$t_{\Gamma,\rm NTB} \sim
\frac{10^{1.4\Gamma}}{3\Omega_0} =
1.7\times 10^{1.4\Gamma} \;{\rm yrs} \;
\left(\frac{M}{M_\odot}\right)^{-1/2}
\left(\frac{R_0}{10\;\rm AU}\right)^{3/2},$$ respectively. Here, $M$ is the mass of the central star and $\Omega_{\rm 0,i}$ is the angular speed of a disk at the footpoint $R_{\rm 0,i}$ of the innermost streamline of winds. The fact that the growth of the FM by a factor $10^\Gamma$ occurs within $\sim \Gamma$ times the rotation period of the disk at the inner radius, far shorter than the lifetime of winds ($\sim 10^4-10^5$ yrs), suggests that the FM mode is dynamically important in the evolution of the disk winds. When $q$ is small, the radial turbulent mixing of the wind, caused by both axisymmetric and non-axisymmetric toroidal buoyancy modes over a relatively short time, is likely to cascade down into arbitrarily smaller scales to dissipate when the microscopic processes such as magnetic reconnection are included. The released energy in the dissipation processes may heat up the flow, potentially making a significant contribution to the heating of protostellar winds and jets. Because the growth rates of buoyancy modes are proportional to the equilibrium magnetic forces (cf. eqs. \[45\] and \[52\]), winds that have approached a force-free magnetic configuration will not be subject to the ATB and NTB instabilities.
The global fundamental mode affects only the inner region of the disk winds (e.g., the central tenth for the model shown in Fig. 3). The logarithmic density gradient $\partial\ln\rho/\partial\ln R$ changes relative to the equilibrium value by $\partial(\rho_1/\rho_0)/
\partial\ln R$. Fig. 3 shows that as a consequence of the fundamental mode, the very central region becomes more steeply stratified, a surrounding concentric region less steeply stratified, and the balance (most of the wind) remains nearly unchanged. Thus, the FM tends to enhance jetlike structure in the central parts of winds. In addition, because of their tendency to compress interior gas via the FM mechanism, disk winds may help to collimate any interior flows into narrow, fast jets, even when the disk winds themselves have relatively slow motion [cf. @ost97].
What do the present results imply about the likely radial extent of protostellar winds? First, we note that observed optical jets are unlikely to be isolated structures, because if so they would be significantly overpressured relative to the ambient medium: Since magnetocentrifugal jet models typically predict internal [Alfv$\acute{\rm e}$n]{} speeds comparable to their flow speeds in the range $150\sim400$ km s$^{-1}$, they have strong internal magnetic pressure $P_{\rm wind} = B_\phi^2/8\pi \sim
\rho_{\rm w}{v_{\rm A}}^2/2 \sim 2.8\times 10^{-7}$ ergs cm$^{-3}$, which is about 6 orders of magnitude greater than the gas pressure of ambient medium, $P_{\rm ext} = {c_{\rm s}}^2\rho_{\rm ext}
\sim 1.3\times 10^{-13}$ ergs cm$^{-3}$. Here as reference values we adopted $\rho_{\rm w} = 350\; m_{\rm H}$ cm$^{-3}$, $\rho_{\rm ext} = 200\; m_{\rm H}$ cm$^{-3}$, ${v_{\rm A}}= 310$ km s$^{-1}$, and ${c_{\rm s}}= 0.2$ km s$^{-1}$ with $m_{\rm H}$ being the mass of a hydrogen atom [@har99]. Thus the pressure imbalance at the outer boundary of the jet would cause either the wind as a whole or only its surface layer to expand until a new balance is attained. Based on the results of this paper, if the magnetic field at the base of the wind is stratified less steeply than $R^{-1}$, then perturbations of the outer parts of the wind are stable. As a consequence, only the surface layers of such winds would expand in order to achieve a pressure-balanced condition with the ambient medium. If, on the other hand, the wind’s magnetic field is stratified more steeply than $R^{-1}$ at its base, no equilibrium is even possible; the wind would expand as a whole to fill the entire $4\pi$ steradians, with the inner parts having higher density observable as a narrow optical jet [cf. @shu94; @sha98]. Numerical simulations presently underway [@lee00] support previous work indicating that protostellar winds with a wide-angle component are better able to produce observed molecular outflow structures than purely jetlike winds [see also @li96; @ost97; @ost98; @mat99] but further studies are required to determine just how distributed in angle the wind momentum should be - i.e., to discriminate between “fully-expanded” and “surface-expanded” models. Recent observations [see, e.g., @ric00] showing a correlation in molecular outflow kinematics with age - with extremely high velocity, highly-collimated flows seen only in the youngest sources - may indicate an underlying temporal evolution from more-collimated to more-expanded protostellar winds.
We acknowledge a stimulating report from an anonymous referee, and helpful comments from N. Turner and S. Balbus.
Abramowitz, M., & Stegun, I. A. 1965, Handbook of Mathematical Functions (New York: Dover) Appert, K., Gruber, R., & Vaclavik, J. 1974, Phys. Fluids, 17, 1471 Appl, S., & Camenzind, M. 1992, A&A, 256, 354 Bachiller, R. 1996, ARAA, 34, 111 Balbus, S. A., & Hawley, J. F. 1991, ApJ, 376, 214 Balbus, S. A., & Hawley, J. F. 1992, ApJ, 400, 610 Balbus, S. A., & Hawley, J. F. 1998, Rev. Mod. Phys., 70, 1 Blaes, O. M., & Balbus, S. A. 1994, ApJ, 421, 163 Blandford, R. D., & Payne, D. G. 1982, MNRAS, 199, 883 Chandrasekhar, S. 1960, Proc. Natl. Acad. Sci., 46, 253 Curry, C., & Pudritz, R. E. 1996, MNRAS, 281, 119 Dubrulle, B., & Knobloch, E. 1993, A&A, 274, 667 Foglizzo, T., & Tagger, M. 1995, A&A, 301, 293 Goldreich, P., & Lynden-Bell, D. 1965, MNRAS, 130, 125 Hardee, P. E., Cooper, M. A., Norman, M. L., & Stone, J. M. 1992, ApJ, 399, 478 Hartigan, P., Morse, J. A., Tumlinson, J., Raymond, J., & Heathcote, S. 1999, ApJ, 512, 901 Hartigan, P., Bally, J., Reipurth, B., and Morse, J. A. 2000, in Protostars and Planets IV, ed. V. Mannings, A. P. Boss, & S. S. Russell (Tucson: University of Arizona Press), in press Hartmann, L., & MacGregor, K. B. 1982, ApJ, 259, 180 Hawley, J. F., & Balbus, S. A. 1991, ApJ, 376, 223 Hawley, J. F., Gammie, C. F., & Balbus, S. A. 1995, ApJ, 440, 742 Julian, W. H., & Toomre, A. 1966, ApJ, 146, 810 K$\ddot{\rm o}$nigl, A., & Pudritz, R. E. 2000, in Protostars and Planets IV, ed. V. Mannings, A. P. Boss, & S. S. Russell (Tucson: University of Arizona Press), in press Lada, C. J. 1985, ARAA, 23, 267 Li, Z. Y., & Shu, F. H. 1996, ApJ, 472, 211 Lee, C. F., Stone, J. M., Ostriker, E. C., & Mundy, L. G. 2000, in preparation Lin, D. C. N., Papaloizou, J. C. B., & Kley, W. 1993, ApJ, 416, 689 Matzner, C. D., & McKee, C. F. 1999, ApJ, 526, 109 Miller, K. A., & Stone, J. H. 1999, ApJ, submitted Morse, P. M., & Feshbach, H. 1953, Methods of Theoretical Physics (New York: McGraw-Hill) pp 719-725 Nagar, N. M., Vogel, S. N., Stone, J. M., & Ostriker, E. C. 1997, ApJ, 482, L195 Norman, C., & Silk, J. 1980, , 238, 158 Ostriker, E. C. 1997, ApJ, 486, 291 Ostriker, E. C. 1998, in Accretion Processes in Astrophysical Systems, ed. S. Holt, & T. Kallman (Woodbury NY:AIP press) p484 Ouyed, R., & Pudritz, R. E. 1997, ApJ, 484, 794 Parker, E. N. 1966, ApJ, 145, 811 Pudritz, R. E., & Norman, C. A. 1986, ApJ, 301, 571 Rae, I. C., & Roberts, B. 1982, MNRAS, 201, 1171 Richer, J. S., Shepherd, D. S., Cabrit, S., Bachiller, R., and Churchwell, E. 2000, in Protostars and Planets IV, ed. V. Mannings, A. P. Boss, & S. S. Russell (Tucson: University of Arizona Press), in press Reipurth, B., Bally, J., & Devine, D. 1997, AJ, 114, 2708 Roberts, B. 1985, in Solar System Magnetic Fields, ed. E. R. Priest (Dordrecht: D. Reidel Publishing Company) p37 Rosen, A., Hardee, P. E., Clarke, D. A., Johnson, A. 1999, ApJ, 510, 136 Ross, D. W., Chen, G. L., & Mahajan, S. M. 1982, Phys. Fluids, 25, 652 Ryu, D., & Goodman, J. 1992, ApJ, 388, 438 Shang, E., Shu, F. H., & Glassgold, A. E. 1998, ApJ, 493, L91 Shu, F. H. 1974, A&A, 33, 55 Shu, F. H. 1992, The Physics of Astrophysics. II. Gas Dynamics (Mill Valley: Univ. Science Books) Shu, F. H., Adams, F. C., & Lizano, S. 1987, ARAA, 25, 23 Shu, F. H., Lizano, S., Ruden, S., & Najita, J.. 1988, ApJ, 328, L19 Shu, F. H., Najita, J. R., Ostriker, E. C., Wilkin, F., Ruden, S., & Lizano, S. 1994, ApJ, 429, 781 Shu, F. H., Najita, J. R., Shang, H., & Li, Z. Y. 2000, in Protostars and Planets IV, ed. V. Mannings, A. P. Boss, & S. S. Russell (Tucson: University of Arizona Press), in press Stone, J. M., Gammie, C. F., Balbus, S. A., & Hawley, J. F. 2000, in Protostars and Planets IV, ed. V. Mannings, A. P. Boss, & S. S. Russell (Tucson: University of Arizona Press), in press Terquem, C., & Papaloizou, J. C. B. 1996, MNRAS, 279, 767 Velikhov, E. 1959, Sov. Phys.–JETP, 36, 995
[ccccc]{} FM &$m={k_{\rm z}}=0$ & global mode & toroidal & unstable\
& & & &\
ATB & & buoyancy & toroidal & unstable\
PB & $m=0$ & Parker & poloidal & unstable\
BH &${k_{\rm z}}\neq 0$& MRI & poloidal & unstable\
TR & & resonance & toroidal & overstable\
& & & &\
NTB & & Parker & toroidal & unstable\
PR & $m\neq 0$ & resonance & poloidal & overstable\
GPB & ${k_{\rm z}}\neq0$&geometric & poloidal & unstable\
NMRI& & MRI & toroidal & unstable\
[^1]: In this case, eq. (24) would become $(1-3\nu)/4=\alpha > 0$, corresponding to $R^2(K^2(R) - \omega^2/{v_{\rm A}}^2) > 0$.
[^2]: Also by taking a local approximation and by neglecting density stratification and the effects of thermal and cosmic ray pressures, one can simplify eq. (III.12) of @par66 to get the asymptotic (${\bold k} \rightarrow \infty$) dispersion relation $$\omega^2 = - \left(\frac{g^2}{v_{\rm A\parallel}^2}\right)
\frac{k_{\parallel}^2}{k_{\perp}^2+k_{\parallel}^2},$$ where $g$ is the gravity perpendicular to the galactic plane, $v_{\rm A\parallel}$ is the Alfv$\acute{\rm e}$n speed of initial fields parallel to the galactic plane, and $k_{\parallel}$ and $k_{\perp}$ are perturbation wavenumbers in the respective directions parallel and perpendicular to the galactic disk and magnetic field. Comparing the above with eq. (43), we may write $g_{\rm eff}
\equiv g_{\rm R} - R\Omega^2$ for the PB modes, with wavenumber correspondence ${k_{\rm z}}\leftrightarrow k_\parallel$ and ${k_{\rm R}}\leftrightarrow k_\perp$.
[^3]: Often referred to magnetorotational instability, or briefly, MRI.
[^4]: In fact, from eq. (57) the formal instability criterion (59b) is generic for any value of ${c_{\rm s}}\neq0$; it may be written as ${v_{\rm Az}}^2({k_{\rm z}}^2+{k_{\rm R}}^2) + \kappa^2 - 4 \Omega^2 <0$. However, when ${c_{\rm s}}/{v_{\rm A}}\ll 1$, growth rates for small $i$ are very low.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present two types of numerical prescriptions that accelerate the radiative transfer calculation around point sources within a three-dimensional Cartesian grid by using the oct-tree structure for the distribution of radiation sources. In one prescription, distant radiation sources are grouped as a bright extended source when the group’s angular size, $\theta_{\rm s}$, is smaller than a critical value, $\theta_{\rm crit}$, and radiative transfer is solved on supermeshes whose angular size is similar to that of the group of sources. The supermesh structure is constructed by coarse-graining the mesh structure. With this method, the computational time scales with $N_{\rm m} \log(N_{\rm m}) \log(N_{\rm s})$ where $N_{\rm m}$ and $N_{\rm s}$ are the number of meshes and that of radiation sources, respectively. While this method is very efficient, it inevitably overestimates the optical depth when a group of sources acts as an extended powerful radiation source and affects distant meshes. In the other prescription, a distant group of sources is treated as a bright point source ignoring the spatial extent of the group and the radiative transfer is solved on the meshes rather than the supermeshes. This prescription is simply a grid-based version of [START]{} by Hasegawa & Umemura and yields better results in general with slightly more computational cost ($\propto N_{\rm m}^{4/3} \log(N_{\rm s})$) than the supermesh prescription. Our methods can easily be implemented to any grid-based hydrodynamic codes and are well-suited to adaptive mesh refinement methods.'
author:
- |
Takashi Okamoto$^{1}$[^1], Kohji Yoshikawa$^{1}$, Masayuki Umemura$^{1}$\
$^{1}$ Center for Computational Sciences, University of Tsukuba, 1-1-1 Tennodai, Tsukuba 305-8577 Ibaraki, Japan
date: 'Accepted . Received ; in original form '
title: 'ARGOT: Accelerated radiative transfer on grids using oct-tree'
---
\[firstpage\]
methods: numerical – radiative transfer.
Introduction
============
Radiative transfer (RT) of photons has fundamental importance for formation of astronomical objects, such as galaxies, stars, and blackholes. Unfortunately, the nature of RT, [in which we have to solve the time evolution of the six-dimensional phase-space information of photons]{} (three spatial dimensions, two angular dimensions, and one frequency dimension; or equivalently three spatial and three momentum dimensions), makes it difficult to solve RT accurately and to couple it with hydrodynamics. To date, various RT schemes has been proposed [@iliev06], some of which are coupled with hydrodynamics [@iliev09]. A wide range of approximation have been used to deal with multi-dimensional nature of the transfer equation and they have their own pros and cons.
When radiation sources are embedded in media on meshes, RT calculations can be categorised into two types; one premises that the source functions are assigned on meshes and the other does that radiation sources are treated as point sources independent of meshes. In the former type, the RT equations are integrated along long or short characteristics between meshes. The latter is advantageous when the number of the point sources, $N_{\rm s}$, is smaller than that of the boundary meshes, $\sim N_{\rm m}^{2/3}$, where $N_{\rm m}$ is the total number of the meshes. The latter type of the RT schemes is often called ‘ray-tracing’ that we deal with in this paper.
The most accurate and straight-forward RT scheme is the long characteristics method in which all source meshes are connected to all other relevant meshes [@abel99; @sokasian01; @RSPH]. This method is however very expensive computationally. The computational costs scales with $N_{\rm m}^2$ in general and with $N_{\rm m}^{4/3} N_{\rm s}$ for the transfer from point sources.
The short characteristics method [@kunasz88; @stone92; @mellema98; @nakamoto01] reduces the computational cost by integrating the equation of RT only along lines that connect nearby cells. It scales with $N_{\rm m}^{5/3}$ and with $N_{\rm m} N_{\rm s}$ for the transfer from point sources. Its known disadvantage is the inability to track collimated radiation fields and hence the inability to cast sharp shadows owing to the numerical diffusion.
The methods whose computational cost is similar to that of the short characteristics method with small loss of accuracy compared to the long characteristics method have also been developed (@razoumov05 and ‘authentic RT’ by Nakamoto et al. in @iliev06). Adaptive ray tracing [@abel02] has been widely used for RT around point sources [@wise11].
Mote Carlo transport [@ciardi01] is also straight forward. The advantage of this approach is that comparatively few approximations to the RT equations need to be made. The resulting radiation field however inevitably becomes noisy [see @iliev06] due to its stochastic nature unless a huge number of photon packets are transported. This method is computationally very expensive in the optically thick regime.
The methods, which consider the moments of the RT equations and consist in choosing a closure relation to solve them, can lead to substantial simplifications that can drastically speed up the calculations because its computational cost scales with $\sim N_{\rm m}$. The most common of these methods is the flux-limited diffusion, which solves the evolution of the first moment and uses a closure relation valid in the diffusion limit, which is an isotropic radiative pressure tensor. The equation is modified with an ad-hoc function (the flux limiter) in order to ensure that the radiative flux is valid in the free-streaming limit. This method is very useful in diffusive regions and have been used to study accretion discs [@ohsuga05] and star formation [@krumholz06]. Another method of closing the system is the variable Eddington tensor formalism. It gives better results than the flux-limited diffusion but are much more complex and costly because it requires the local resolution of the transfer equation at each timestep. The methods which employ the optically thin variable Eddington tensor approximation [@ga01] have been used to study cosmic reionization [@ga01; @ricotti02; @petkova09]. A locally evaluated Eddington tensor, called the M$_1$ model, has also been used to close the system [@heracles] and has applied to study cosmic reionization [@aubert08]. The accuracy of the moments methods is problem-dependent and is hard to judge in general situation. @petkova11 have developed a method that employs a direct discretisation of the RT equation in Boltzmann form with finite angular resolution on moving meshes. This method is advantageous in solving problems in which time-dependent solution of the RT equation is important. The timestep however has to be very short because photons propagate at the speed of light unless a reduced speed of light approximation is employed.
In many astrophysical problems, for example cosmic reionization and galaxy formation, we have to deal with numerous radiation sources. [ @TRAPHIC introduced source merging procedure in order to avoid computationally expensive scaling with the number of sources and implemented it on Smoothed Particle Hydrodynamics (SPH). ]{} @START utilised the oct-tree algorithm [@tree] in order to accelerate the RT around point sources and they coupled the RT with SPH. In their method, distant sources from a target gas particle are grouped and regarded as a single point source when the angular size of the group of the sources is smaller than a critical value. Consequently, the effective number of radiation sources is largely reduced to $\log(N_{\rm s})$ when there are $N_{\rm s}$ sources.
The methods we explore in this paper are parallel to this approach except that we implement this grouping algorithm to grid-based codes. In one of our methods, we introduce supermeshes; a supermesh consists of $8^n$ meshes and it is characterised by the mean density of each chemical species of the meshes within the supermesh. Solving the RT on supermeshes whose angular size is similar to that of the group of the sources in question results in further reduction of computational time in principle. Another approach we take is the point source approximation, in which a group of sources sufficiently distant from a target mesh is treated as a point source. The latter can be regarded as a grid-based version of [START]{} [@START].
Unlike gravitational interactions to which the tree-algorithm has been widely applied, RT is affected by the medium between a source and a target. It is therefore very important to test these tree-based approaches in cases where an extended group of sources works as a powerful source in inhomogeneous medium and affects (e.g. ionizes) distant meshes. In this paper, we extensively investigate such cases in order to clarify advantages and disadvantages of the methods using tree-based algorithm.
This paper is organised as follows. In section 2, we describe the algorithm in detail. In section 3, we present several test problems and compare our methods to each other. We summarise and discuss the results in section 4.
Radiative transfer with tree-algorithm
======================================
In this section, we describe our ray-tracing algorithm that we use to solve the steady RT equation for a given frequency, $\nu$: $$\frac{{{\rm d}}I_\nu}{{{\rm d}}\tau_\nu} = - I_\nu + S_\nu,
\label{eq:rt}$$ where $I_\nu$, $\tau_\nu$, and $S_\nu$ are the specific intensity, the optical depth, and the source function, respectively. This equation is adequate for problems in which the absorption and emission coefficients change on timescales much longer than the light crossing time. This will always be the case in the volumes we will simulate by using our methods. Eqn. (\[eq:rt\]) has a formal solution: $$I_\nu(\tau_\nu) = I_{\nu, 0} \mathrm{e}^{- \tau_\nu}
+ \int_0^{\tau_\nu} S_\nu(\tau'_\nu) \mathrm{e}^{-\tau_\nu + \tau'_\nu} {{\rm d}}\tau'_\nu,
\label{eq:formal}$$ where $I_{\nu, 0}$ is the specific intensity at $\tau_\nu = 0$ and $\tau'_\nu$ is the optical depth at a position along the ray. Throughout this paper we employ so-called on-the-spot approximation [@agnagn] in which recombination photons are assumed to be absorbed where they were emitted. Using the on-the-spot approximation, the formal solution given by equation (\[eq:formal\]) is reduced to $$I_\nu(\tau_\nu) = I_{\nu, 0} \mathrm{e}^{- \tau_\nu}.
\label{eq:formal2}$$ To solve this equation numerically, one needs to calculate optical depth between each pair of a source and a target mesh. The computational cost is hence proportional to the number of sources. In the next subsection, we will describe the method to decrease the effective number of radiation sources by using the oct-tree structure.
Source grouping algorithm
-------------------------
As in @START, we construct the oct-tree structure for the distribution of radiation sources. A cubic computational domain is hierarchically subdivided into 8 cubic cells until each cell contains only one radiation source or the size of a cell becomes sufficiently small compared to that of the computational domain. We call these sub-volumes ‘tree nodes’. When the side length of the cubic computational domain is $L$, the width of a level $l$ tree node is given by $w^{(l)} = L/2^{l}$. Each tree node records the centre of the luminosity of the radiation sources contained in the node, $${\boldsymbol r } = \frac{\sum_m {\boldsymbol r }_m L_m}{\sum_m L_m},
\label{eq:centre}$$ and the total luminosity, $$L = \sum_m L_m,
\label{eq:totallum}$$ where ${\boldsymbol r }_m$ and $L_m$ indicate the position vector and the luminosity of a radiation source, respectively, and subscript $m$ runs over all sources within the tree node.
Once we have constructed the tree structure, we loop over all meshes. RT from all the radiation sources to each target mesh is performed by a simple recursive calculation as done in $N$-body calculation. We start at the root node (level 0 tree node), which covers entire computational domain. Let $w$ be the width of the node currently being processed and $D$ the distance between the closest edges of the tree node and the target mesh. If the angular size of the node is smaller than a fixed value of accuracy parameter, i.e. $$\frac{w}{D} < \theta_{\rm crit},
\label{eq:opening}$$ then we perform the RT calculation between the group of sources in the current node and the target mesh and move on to the next node. Otherwise, we examine the child nodes (subnodes) recursively. The effective number of sources is thus proportional to $\log(N_{\rm s})$. In the following subsections we will explain how we perform the RT calculation between a group of sources and a target mesh.
Supermesh approximation {#sec:smart}
-----------------------
![Schematic illustration of the supermesh structure for $8\times8$ two-dimensional meshes. In this case, the maximum level, $l_{\rm max}$, is 3 and the meshes themselves can be used as the highest level supermeshes. A level $l$ supermesh contains $2^{2(l_{\rm max} - l)}$ meshes. For three-dimensional meshes, a level $l$ supermesh consists of $2^{3(l_{\rm max} - l)}$ meshes. []{data-label="fig:supermesh"}](fig1.eps){width="8.0cm"}
We first introduce the supermesh approximation. In Fig. \[fig:supermesh\], we show a schematic illustration of the supermesh structure. If a three-dimensional computational domain is discretised by $2^{3 l_{\rm max}}$ meshes, a level $l$ supermesh consists of $2^{3 (l_{\rm max} -l)}$ meshes. We can calculate the mean density of each chemical species for every supermesh by using the meshes contained in it. The meshes can be used as the highest level supermeshes. The supermesh structure is resembling to an adaptive mesh refinement (AMR) structure and thus this method is well-suited to couple with the hydrodynamics by AMR codes.
Let us consider the case in which plane-parallel radiation with the specific intensity $I_0$ enters a supermesh that consists of $N_x \times N_y$ meshes. What we want to know is the mean intensity of the ray emerging from the other side of the supermesh, $\langle I_{\rm out} \rangle$ (see Fig. \[fig:supermeshapprox\]).
![Plane-parallel radiation with specific intensity $I_0$ entering to a supermesh that consists of $N_x \times N_y$ meshes. The $(i, j)$-th mesh has the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ number density, $n_{i, j}$. []{data-label="fig:supermeshapprox"}](fig2.eps){width="8.0cm"}
For simplicity, we here only consider the absorption by ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ atoms and drop the frequency dependence. The side length of each mesh is $\Delta L$ and the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ number density of the $(i, j)$-th mesh in the supermesh is $n_{i, j}$. The mean intensity of the emerging radiation is given by $$\langle I_{\rm out} \rangle = \frac{I_0}{N_y} \sum_j^{N_y} \exp[- \sigma_{\rm HI} {\cal N}_j],
\label{eq:exact}$$ where $\sigma_{\rm HI}$ is the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ cross-section and ${\cal N}_j$ is the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ column density of the $j$-th line, i.e. ${\cal N}_j = \sum_i^{N_x} n_{i,j} \Delta L$. [ In our supermesh approximation, we use the mean ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ number density, $\langle n \rangle = \sum_{i,j} n_{i, j} / (N_x N_y)$, to estimate the mean intensity of the emerging radiation $\langle I_{\rm out} \rangle$. Doing this introduces some error as we will show below. In order to understand the accuracy and the nature of the supermesh approximation, we compare the mean intensity of the emerging radiation by the supermesh approximation to that calculated by using the meshes. We first consider the Taylor series expansion of the mean intensity of the emerging radiation when we solve the RT on the supermesh: ]{} $$\begin{aligned}
\langle I_{\rm out} \rangle_{\rm mean} &=& I_0
\exp(-\sigma_{\rm HI}\langle {\cal N} \rangle ) \nonumber \\
& = & I_0 \left[
1 - \sigma_{\rm HI} \langle {\cal N} \rangle
+ \frac{\sigma_{\rm HI}^2}{2} \langle {\cal N} \rangle^2 + \cdots \right],
\label{eq:mean}\end{aligned}$$ where $\langle {\cal N} \rangle$ is the mean ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ column density given by $$\begin{aligned}
\langle {\cal N} \rangle = \frac{\sum^{N_y}_j {\cal N}_j}{N_y}
= \frac{\Delta L \sum_j^{N_y} \sum_i^{N_x} n_{i, j}}{N_y} = \Delta L N_x \langle n \rangle.
\label{eq:column}\end{aligned}$$ On the other hand, the Taylor series expansion of Eqn. (\[eq:exact\]) is $$\begin{aligned}
\langle I_{\rm out} \rangle &=& \frac{I_0}{N_y}
\sum_j^{N_y} \left[1 - \sigma_{\rm HI} {\cal N}_j + \frac{1}{2} (\sigma_{\rm HI} {\cal N}_j)^2
+ \cdots \right] \nonumber \\
&=& \frac{I_0}{N_y} \left[N_y - \sigma_{\rm HI} \sum^{N_y}_j {\cal N}_j
+ \frac{\sigma_{\rm HI}^2}{2} \sum^{N_y}_j {\cal N}_j^2 + \cdots \right] \nonumber \\
&=& I_0 \left[1 - \sigma_{\rm HI} \langle {\cal N} \rangle
+ \frac{\sigma_{\rm HI}^2}{2} \langle {\cal N}^2 \rangle + \cdots \right].
\label{eq:tayler}\end{aligned}$$ The difference between $\langle I_{\rm out} \rangle$ and $\langle I_{\rm out} \rangle_{\rm mean}$ is the second order in $\tau$. From Eqn. (\[eq:exact\]) and (\[eq:mean\]), the leading error in $\langle I_{\rm out} \rangle_{\rm mean}$ is $$\begin{aligned}
\langle I_{\rm out} \rangle - \langle I_{\rm out} \rangle_{\rm mean} &=& I_0 \frac{\sigma_{\rm HI}^2}{2}
\left(\langle {\cal N}^2 \rangle - \langle {\cal N} \rangle^2 \right).
\label{eq:error}\end{aligned}$$ Since the variance of the column density, $\langle {\cal N}^2 \rangle - \langle {\cal N} \rangle^2$, could be very large in the inhomogeneous medium, we substantially overestimate the optical depth if we use Eqn. (\[eq:mean\]).
We can therefore in principle improve the approximation by estimating the variance of the column density. According to the central limit theorem, the variance of the column density for large $N_x$ can be expressed by using the variance of the density, if the density, $n_{i, j}$, is a sequence of independent and identically distributed random variables: $$\langle {\cal N}^2 \rangle - \langle {\cal N} \rangle^2 = N_x \left[\langle n^2 \rangle - \langle n \rangle^2\right].
\label{eq:centrallimit}$$ Using this relation, the mean intensity of the emerging radiation can be approximated as: $$\begin{aligned}
\langle I_{\rm out} \rangle_{\rm variance} &=& I_0 \biggl[\exp\left(-\sigma_{\rm HI} \langle {\cal N} \rangle \right) \nonumber \\
&& \hspace{1cm} + \frac{\sigma_{\rm HI}^2 N_x}{2}\left(\langle n^2 \rangle - \langle n \rangle^2\right)\biggl].
\label{eq:smapprox}\end{aligned}$$ The effective column density for a ray segment that intersects the supermesh is hence $${\cal N}_{\rm eff} = - \frac{\ln\left[\exp\left(-\sigma_{\rm HI} \langle n \rangle h\right)
+ \frac{\sigma_{\rm HI}^2 h}{\Delta L} \left(\langle n^2 \rangle - \langle n \rangle^2\right)\right]}{\sigma_{\rm HI}},
\label{eq:effectivecolumn}$$ where $h$ is the length of a ray segment. We however do not employ this approximation because Eqn. (\[eq:centrallimit\]) is only valid for large $N_x$ and $N_x$ always becomes small near the target mesh. We thus only use the mean density in our supermesh approximation which is described by Eqn. (\[eq:mean\]). We will investigate the accuracy of this approximation in Section \[sec:tests\].
Now we have to determine on which supermeshes we perform the RT calculation. We chose to use the lowest level supermeshes whose angular size, $\theta$, is equal to or smaller than the angular size of a group of the sources, $\theta_{\rm s}$, since we assume plane-parallel radiation to construct the approximation. We define the luminosity-weighted rms projected radius as the effective projected size of the group of the sources[^2], i.e. if the target mesh is located along the $z$-direction from the centre of the luminosity, the projected size of the group is defined as $$r_{\rm rms}^2 = \frac{\sum_m L_m \left\{(x_m - \bar{x})^2 + (y_m - \bar{y})^2\right\}}{\sum_m L_m},
\label{eq:rms}$$ where $\bar{x}$ and $\bar{y}$ are, respectively, the $x$ and $y$ components of the position vector of the luminosity centre and the subscript $m$ runs over all sources in the tree node in question. Practically, we calculate the following tensor for each tree node: $${\cal I}_{ij} = \sum_m L_m ({\boldsymbol r }_{m, i} - \bar{{\boldsymbol r }}_i)({\boldsymbol r }_{m, j} - \bar{{\boldsymbol r }}_j),
\label{eq:inertia}$$ where the subscripts $i$ and $j$, respectively, indicate $i$-th and $j$-th components of the position vector, i.e. $i$ and $j$ are either $x$, $y$, or $z$; and the subscript $m$ has the same meaning as in Eqn. (\[eq:rms\]). By using $(0, 0)$ and $(1, 1)$ components of the tensor ${\cal I}'$ which is the tensor ${\cal I}$ in the rotated frame so that the target mesh is placed along the $z$-direction from the luminosity centre, we can estimate the angular size of the group of the source in the tree node as $$\theta_{\rm s} = \frac{2 r_{\rm rms}}{D} = \frac{2}{D} \left(\frac{{\cal I}'_{00} + {\cal I}'_{11}}{\sum_m L_m} \right)^\frac{1}{2},
\label{eq:angularsize}$$ where $D$ is the distance between the luminosity centre and the closest edge of the target mesh. In Fig. \[fig:supermeshrt\], we illustrate the procedure of the RT calculation using the supermeshes.
![A schematic illustration of the RT calculation using supermeshes. The radiation sources in the tree node indicated by red square are regarded as a single bright extended source. The target mesh is coloured by light blue. The RT is solved on the supermeshes at the lowest level, whose angular size is equal to or smaller than the angular size of the source group, $\theta_{\rm s}$. The supermeshes used are indicated by purple colour and their sizes are represented by the orange squares. []{data-label="fig:supermeshrt"}](fig3.eps){width="8.0cm"}
The computational cost by this method is expected to scale with $N_{\rm m} \log(N_{\rm m}) \log(N_{\rm s}) $.
Point source approximation {#sec:start}
--------------------------
Here we introduce another way of accelerating the RT calculation by using the oct-tree structure of the distribution of radiation sources. As in @START, we treat a group of sources in a tree node which satisfies the condition described by Eqn. (\[eq:opening\]) as a bright point source. Since we ignore the size of the source group, we solve the RT not on the supermeshes but on the meshes. Consequently, the computational cost scales with $N_{\rm m}^\frac{4}{3} \log(N_{\rm s})$. Although this is slightly more expensive computationally than the supermesh approximation in which the cost is proportional to $N_{\rm m} \log(N_{\rm m}) \log(N_{\rm s})$, this method may faster than the supermesh approximation for small $N_{\rm m}$ because we do not have to calculate ${\cal I}'_{00}$ and ${\cal I}'_{11}$ in the point source approximation[^3]. [Since the surface area of a Strömgren sphere is proportional to $\dot{N}^{2/3}$ where $\dot{N}$ is photoionization rate (see Section \[stroemgren\]), treating a source group as a point source underestimates the surface area of ionized regions. We will explore this effect in our tests. ]{}
RRB DRR CIR RCRB DRCR CICR CECR BCR CCR CS
--------------- ----- --------------- --------------- ------ --------------- --------------- ----- ----- -----
(4), (5), (4) (2) (7), (7), (1) (4), (5), (4) (3) (3), (3), (3) (3), (3), (3) (4) (6) (8)
\
\(1) @abel97; (2) @aldrovandi73; (3) @cen92; (4) @hummer94; (5) @hummer98; (6) @ikeuchi86; (7) @janev87; (8) @agnagn
Non-equilibrium chemistry {#sec:chem}
-------------------------
We solve the non-equilibrium chemistry for e, ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$, ${\mbox{{H$\,${\sc ii}}}}$, ${\mbox{{He$\,${\sc i}}}}$, ${\mbox{{He$\,${\sc ii}}}}$, and ${\mbox{{He$\,${\sc iii}}}}$ implicitly. Note that since we employ the on-the-spot approximation, we use ‘Case B’ recombination coefficients to calculate recombination rates of ${\mbox{{H$\,${\sc ii}}}}$, ${\mbox{{He$\,${\sc ii}}}}$, and ${\mbox{{He$\,${\sc iii}}}}$ throughout this paper.
Using the optical depth obtained by the methods described in Section \[sec:smart\] or \[sec:start\], the photoionization rates of ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$, ${\mbox{{He$\,${\sc i}}}}$, and ${\mbox{{He$\,${\sc ii}}}}$ in each mesh are given by $$\Gamma_i = \sum_\alpha \Gamma_{i, \alpha},$$ where $\Gamma_{i, \alpha}$ denotes the radiative contribution from a radiation source (or a group of radiation sources), $\alpha$, and $i = {\ensuremath{\mbox{{H$\,${\sc i}}}}}$, ${\mbox{{He$\,${\sc i}}}}$, and ${\mbox{{He$\,${\sc ii}}}}$. The contribution from a point-like radiation source, $\alpha$, is represented by $$\Gamma_{i, \alpha} = \frac{1}{4 \pi h r_\alpha^2} \int_{\nu_i}^\infty
\frac{{{\rm d}}\nu}{\nu} \sigma_i(\nu) L_\alpha(\nu)
\exp\left[-\sum_j {\cal N}_{j, \alpha} \sigma_j (\nu)\right],
\label{eq:photoionization}$$ where $\nu_{i}$ is the threshold frequency for the $i$-th species, $\sigma_i(\nu)$ is the cross-section of the $i$-th species, and $r_\alpha$, $L_\alpha(\nu)$, and ${\cal N}_{i, \alpha}$ are respectively the distance between the luminosity centre and the target mesh, the intrinsic luminosity of the radiation source (or the group of the sources), and the column density of $i$-th species. The sum in the exponent runs over all three chemical species. When all sources have the same spectral shape, i.e. $L_\alpha(\nu) = C_\alpha f(\nu)$, we generate a look-up table for each species as a function of column densities: $$g_i({\cal N}_k) = \int_{\nu_i}^{\infty}
\frac{{{\rm d}}\nu}{\nu} \sigma_i(\nu) f(\nu)
\exp\left[-\sum_j {\cal N}_j \sigma_j (\nu)\right].
\label{eq:tables}$$ In our case, a look-up table for each chemical species becomes three-dimensional table. We have confirmed that 20 logarithmic bins for each column density is sufficient. By using the look-up tables, the RT calculation is reduced to evaluating the column densities. Following @anninos97, we update the densities of each chemical species implicitly by using a backward difference formula (BDF). The equations to evolve the density of each species can be generally written as $$\frac{{{\rm d}}n_i}{{{\rm d}}t} = C_i (T, n_j) - D_i (T, n_j) n_i,
\label{eq:chemistry}$$ where $n_i = \rho_i / (A_i m_{\rm H})$, $A_i$ is the atomic mass number of the $i$-th species, and $m_{\rm H}$ is the proton mass. This time $i$ is either e, ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$, ${\mbox{{H$\,${\sc ii}}}}$, ${\mbox{{He$\,${\sc i}}}}$, ${\mbox{{He$\,${\sc ii}}}}$ or ${\mbox{{He$\,${\sc iii}}}}$. The first term of the right-hand side, $C_i$, is the collective source term responsible for the creation of the $i$-th species. The second term involving $D_i$ represents the destruction mechanisms for the $i$-th species and are thus proportional to $n_i$.
Since the timescales for the ionization and recombination differ by many orders of magnitude depending on chemical species, Eqn. (\[eq:chemistry\]) is a stiff set of differential equations. In numerically solving a stiff set of equations, implicit schemes are required unless an unreasonably small timestep is employed. As in @anninos97 we adopt a BDF. Discretisation of Eqn. (\[eq:chemistry\]) yields $$n_i^{t + \Delta t} = \frac{C_i^{t + \Delta t}\Delta t + n_i^t}{1 + D_i^{t + \Delta t} \Delta t},
\label{eq:bdf}$$ where all source terms are evaluated at the advanced timestep. However, not all source terms can be evaluated at the advanced timestep due to the intrinsic nonlinearity of Eqn. (\[eq:chemistry\]). We hence sequentially update densities of all species in the order of increasing ionization states rather than updating them simultaneously; We evaluate the source terms contributed by the ionization from and recombination to the lower states at the advanced timesteps. This method has been found to be very efficient and accurate [e.g. @anninos97; @yoshikawa06].
Further improvements in accuracy and stability can be made by subcycling the rate solver over a single timestep with which the RT is solved. The subcycle timestep, which we call the ‘chemical timestep’, is determined so that the maximum fractional change in the electron density is limited to 10% per timestep: $$\Delta t_{\rm chem} = 0.1 \left|\frac{n_{\rm e}}{\dot{n}_{\rm e}}\right|.
\label{eq:chemicaltimespte}$$
Photo-heating and radiative cooling {#sec:heatcool}
-----------------------------------
Similarly to the photoionization, photo-heating rate for each mesh due to the photoionization of the $i$-th species is given by $${\cal H}_i = \sum_\alpha {\cal H}_{i, \alpha},$$ where ${\cal H}_{i, \alpha}$ indicates the contribution from a radiation source (or a group of sources), $\alpha$, and $i = {\ensuremath{\mbox{{H$\,${\sc i}}}}}$, ${\mbox{{He$\,${\sc i}}}}$, and ${\mbox{{He$\,${\sc ii}}}}$. The total photo-heating rate is defined by ${\cal H} = \sum_i {\cal H}_i n_i$. The contribution from a point-like source, $\alpha$, is written as $${\cal H}_{i, \alpha} = \frac{1}{4 \pi r_\alpha^2} \int_{\nu_i}^\infty
\frac{{{\rm d}}\nu}{\nu} \sigma_i(\nu) L_\alpha(\nu) (\nu - \nu_i)
\exp\left[- \sum_j \sigma_j(\nu) {\cal N}_{j, \alpha} \right].
\label{eq:photoheat}$$ As for the photoionization, we generate a look-up table for each species when all sources have the identical spectral shape.
We solve the energy equation for each mesh implicitly as $$u^{t + \Delta t} = u^{t} + \frac{{\cal H}^{t + \Delta t} - \Lambda(n_i^{t + \Delta t}, T^{t+\Delta t})}{\rho^t} \Delta t,
\label{eq:heatcool}$$ where $u$ and $T^t = T(n_i^t, u^t)$ are respectively the specific internal energy and temperature of the gas and $\Lambda$ is the cooling function. Although this implicit integration is always stable, we need to subcycle the energy solver with $\Delta t_{\rm chem}$ because both $C_i$ and $D_i$ in Eqn. (\[eq:bdf\]) are functions of the temperature. We thus perform the rate solver and the energy solver alternately. The chemical timestep $\Delta t_{\rm chem}$ is recalculated before every subcycle.
Chemical reaction and cooling rates
-----------------------------------
We try to use the chemical reaction and cooling rates as up-to-date as possible. The sources of these rates are summarised in Table \[table:rates\]. Note that there are notable differences in the recombination cooling rates between literatures [see @iliev06].
Time stepping
-------------
Since the optical depth $\tau(\nu)$ at $t + \Delta t$ depends on densities of all species at $t + \Delta t$, we have to solve the static RT equation (Eqn. (\[eq:formal2\])), the chemical reactions (Eqn. (\[eq:bdf\])), and the energy equation (Eqn. (\[eq:heatcool\])) iteratively. We iterate these steps until the relative difference in the electron number density becomes sufficiently small: $|n_{\rm e}^{(n)} - n_{\rm e}^{(n - 1)}|/n_{\rm e}^{(n)} < \epsilon$, where superscripts indicate the number of iterations and we set $\epsilon$ to $10^{-4}$ throughout this paper. The timestep $\Delta t$, with which we solve the RT equation to obtain $\Gamma_i^{t + \Delta t}$ and ${\cal H}_i^{t + \Delta t}$, could be much larger than the chemical timestep $\Delta t_{\rm chem}$, with which we subcycle the rate and energy solvers.
We however choose to employ a timestep that is defined by the timescale of the chemical reactions: $$\Delta t_i = \epsilon_{\rm e} \left|\frac{n_{\rm e}}{\dot{n}_{\rm e}}\right|_i +
\epsilon_{\rm HI} \left|\frac{n_{\rm HI}}{\dot{n}_{\rm HI}}\right|_i,
\label{eq:timestep_i}$$ where the second term in the right-hand side prevent the timestep from becoming too short when the medium is almost neutral. Our choice for $\epsilon_{\rm e}$ and $\epsilon_{\rm HI}$ are 0.2 and 0.002, respectively. We follow the evolution of the system with the minimum of the timestep defined by Eqn. (\[eq:timestep\_i\]), i.e. $$\Delta t = \Delta t_{i, {\rm min}}.
\label{eq:timestep}$$ The timestep $\Delta t$ is therefore only about twice as long as the shortest chemical timestep, $\Delta t_{{\rm chem}, {\rm min}}$. With this timestep, we find the solutions typically within 3 to 6 iteration steps. While we can of course use a longer timestep, a longer timestep requires more iterations and the total number of steps becomes similar or even larger than the case we employ the timestep defined by Eqn. (\[eq:timestep\]). With a longer timestep, the solutions sometimes never converge. This timestep is in general much shorter than the timestep defined by the Courant timestep criterion and therefore we have to subcycle the hydrodynamical timestep with this timestep when we couple the RT with the hydrodynamics.
When optically thick meshes exist, the solutions converge very slowly. We thus use smoothed photoionization rates, $\tilde{\Gamma}_i$, and heating rates, $\tilde{\cal H}_i$, instead of $\Gamma_i$ and ${\cal H}_i$. The smoothed rates for the $i$-th mesh is calculated by using adjacent 26 meshes, i.e. 27 meshes in total, with a Gaussian kernel of the smoothing length $\Delta L$. Doing this drastically reduces the number of iterations required to find the solutions. [ This smoothing may introduce the smearing of the I-fronts especially when the spacial resolution is poor. While we do not find such an effect in our test simulations as we will show later, this can be avoided by applying the smoothing only to optically thick meshes as done by @RSPH. ]{}
Test simulations {#sec:tests}
================
In this section, we describe the tests we perform. In order to evaluate the accuracy of our tree-based RT algorithms, problems should involve many sources. Therefore some of the tests presented are neither simplest nor cleanest. All test problems are solved in three dimensions, with $128^3$ meshes unless otherwise stated.
Test 1 – Pure hydrogen isothermal ${\mbox{{H$\,${\sc ii}}}}$ region expansion {#stroemgren}
-----------------------------------------------------------------------------
![Test 1 – Images of the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ fraction, cut through at the mid plane of the simulation box at $t = 500~{\rm Myr}$.[]{data-label="fig:strmap"}](fig4.eps){width="8.0cm"}
The first test is the classical problem of a ${\mbox{{H$\,${\sc ii}}}}$ region expansion in a static, homogeneous, and isothermal gas, which consists of only hydrogen, around a single ionizing source. This problem has a known analytic solution and is therefore the most widely used test. Note however that since there is only a single radiation source, our RT schemes described in Sections \[sec:smart\] and \[sec:start\] have no difference and both methods become the long characteristics method. The aim of this test is hence to test our chemical reaction solver and time stepping procedure.
We adopt a monochromatic radiation source that steadily emits $\dot{N}_\gamma$ photons per second, whose frequency is the Lyman limit frequency ($h\nu_{\rm L} = 13.6$ eV). The density of the initially neutral gas is $n_{\rm H}$. Assuming the ionization equilibrium, the Strömgren radius is given by $$r_{\rm S} = \left(\frac{3\dot{N}_\gamma}{4 \pi \alpha_{\rm B}(T) n_{\rm H}^2}\right)^{1/3},
\label{eq:stroemgren}$$ where $\alpha_{\rm B}$ is the Case B recombination coefficient. If we assume that the ionization front (I-front) is infinitely thin, the evolution of the I-front radius is analytically given by $$r_{\rm I} = r_{\rm S} \left[1 - \exp(-t/t_{\rm rec})\right]^{1/3},
\label{eq:irad}$$ where $$t_{\rm rec} = \left(n_{\rm H} \alpha_{\rm B}\right)^{-1}
\label{eq:recombinationtime}$$ is the recombination time.
The analytical solution for the profile of the neutral and ionized fractions ($X_{\rm HI}(r) = n_{\rm HI}(r)/n_{\rm H}$ and $X_{\rm HII}(r) = n_{\rm HII}(r)/n_{\rm H}$) can also be calculated [e.g. @agnagn] from the equation of the ionization balance at radius $r$: $$\frac{n_{\rm HI}(r)}{4 \pi r^2} \dot{N}_\gamma e^{-\tau(r)}\sigma_{\rm HI}(\nu_{\rm L}) = n_{\rm HII}(r)^2 \alpha_{\rm B}(T),$$ where $$\tau(r) = \sigma_{\rm HI}\int_0^r n_{\rm HI}(r') {{\rm d}}r'.
\label{eq:taur}$$ The profile of the neutral fraction is thus given by $$X_{\rm HI}(r) = \frac {2 + \frac{\dot{N}_\gamma e^{-\tau(r)} \sigma_{\rm HI}}{4 \pi r^2 n_{\rm H} \alpha_{\rm B}}
- \sqrt{\left(2 + \frac{\dot{N}_\gamma e^{-\tau(r)}\sigma_{\rm HI}}{4 \pi r^2 n_{\rm H} \alpha_{\rm B}}\right)^2 - 4}}{2}.
\label{eq:analytic}$$ To derive this profile, we ignore the collisional ionization, which is included in our simulations.
The initial physical parameters of this test are the same as those of Test 1 in Cosmological Radiative Transfer Comparison Project [@iliev06], where the hydrogen number density, $n_{\rm H}$, is $10^{-3}~{\rm cm}^{-3}$, the temperature of the isothermal gas is $10^4$ K, and ionization rate, $\dot{N}_\gamma$, is $5 \times 10^{48}$ photons s$^{-1}$. Given these parameters and the recombination rate we use, $\alpha_{\rm B}(10^4~{\rm K}) = 2.58 \times 10^{-13}~{\rm cm}^3~{\rm s}^{-1}$, the recombination time and the Ströemgren radius are $t_{\rm rec} = 122.6~{\rm Myr}$ and $r_{\rm S} = 5.4~{\rm kpc}$, respectively.
We employ identical numerical parameters to those in [@iliev06]: The side length of the simulation box is 6.6 kpc, initial ionization fraction is set to $1.2 \times 10^{-3}$, and a radiation source is placed at the corner of the box, $(0, 0, 0)$. We compare our simulation results to the analytical solution given by Eqn. (\[eq:analytic\]) which represents the solution at $t = \infty$.
In Fig. \[fig:strmap\], we show the neutral fraction in the $z = 0.5 \Delta L$ plane at $t = 500~{\rm Myr}$, at which point the I-front is close to to the maximum radius, i.e. the Strömgren radius. The ${\mbox{{H$\,${\sc ii}}}}$ region is nicely spherical, though this is not surprising because, with a single source, our method is identical to the long characteristics method.
![Test 1 – The profiles of ionized and neutral fractions. The radius is in units of the Strömgren radius. The dot-dot-dot-dashed, dotted, dot-dashed, and dashed lines represent simulated results at $t = 120$, 250, 500, and 1000 Myr, respectively. The solid line indicates the analytical solution at $t = \infty$ given by Eqn. (\[eq:analytic\]). The minimum ionized fraction in the numerical results is set by the collisional ionization which is not included in the analytical solution. []{data-label="fig:strprofile"}](fig5.eps){width="8.0cm"}
In Fig. \[fig:strprofile\], we show the profiles of ionized and neutral fractions at $t = 120$, 250, 500, and 1000 Myr. The results asymptotically approach to the analytical solution at $t = \infty$. There is a minimum neutral fraction in the simulation results, which is set by the collisional ionization that is not included in the analytical solution.
Test 2 – Pure hydrogen ${\mbox{{H$\,${\sc ii}}}}$ region expansion with thermal evolution
-----------------------------------------------------------------------------------------
![Test 2 – [*Upper panel*]{}: Spherically averaged ionized and neutral fraction profiles. The dot-dashed, dashed, and solid lines indicate indicate the profile at $t = 10$, 100, and 500 Myr, respectively. The results from a high-resolution spherically symmetric one-dimensional simulation are shown by the dotted lines, which almost perfectly overlap with those by the three-dimensional simulation. The radius is in units of the Strömgren radius for the uniform isothermal gas with $n_{\rm H} = 10^{-3}~{\rm cm}^{-1}$ and $T = 10^4$ K. [*Lower panel*]{}: Spherically averaged temperature profiles. The meaning of the lines is the same as in the upper panel. []{data-label="fig:strprofilebb"}](fig6.eps){width="8.0cm"}
Test 2 solves essentially the same problem as Test 1, but the ionizing source is assumed to have a $10^5$ K blackbody spectrum and we allow the gas temperature to vary owing to heating and cooling processes. The initial gas temperature and ionized fraction are set to $10^2$ K and $1.2 \times 10^{-3}$, respectively.
In Fig. \[fig:strprofilebb\], we show the neutral and ionized fraction profiles (upper panel) and the temperature profiles (lower panel) at $t = 10$, 100, and 500 Myr. We also show the results from a high-resolution spherically symmetric one-dimensional simulation by the dotted line. For the one-dimensional simulation, we use 1024 meshes for a sphere of radius of $1.5 \times r_{\rm S}$ and we do not employ the smoothed ionization and heating rates whereas smoothed rates are employed in the tree-dimensional simulation. The results by the three-dimensional simulation are almost indistinguishable from those obtained by the one-dimensional one. The use of the smoothed rates to accelerate the convergence has thus no evident side-effects such as smearing of the I-front.
[ For this test, our results are most resembling to those obtained by [RSPH]{} for Test 2 in Cosmological Radiative Transfer Comparison Project [@iliev06][^4]. The agreement with [RSPH]{} is natural because both methods are essentially the long characteristics method. Small differences are probably caused by different adopted rates. ]{}
![Test 3 – Images of the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ fraction and temperature, cut through at the mid plane of the simulation box at $t = 30$, 100, and 500 Myr from left to right. The side length of the simulation box is 132 kpc in which we randomly distribute 1000 radiation sources and 1000 optically thick meshes. Upper two rows show results by the supermesh approximation with $\theta_{\rm crit} = 0.5$ and lower two rows by the point source approximation with $\theta_{\rm crit} = 0.5$. []{data-label="fig:multiplesources"}](fig7.eps){width="8.0cm"}
Test 3 – Multiple radiation sources in a clumpy medium
------------------------------------------------------
In order to test the validity of the RT solver based-on the source grouping, we have to solve problems that involve multiple sources. Moreover, the error in the supermesh approximation becomes large when the inhomogeneity of the medium is large (see Eqn. (\[eq:error\])). In this test, we therefore solve the RT from multiple sources in the clumpy medium. The side length of the simulation box is 132 kpc. We randomly select 1000 optically thick meshes whose hydrogen number density is $n_{\rm H} = 0.2~{\rm cm}^{-3}$ and optical depth at the Lyman limit frequency is $\sim 4 \times 10^3$ for the mesh size. The hydrogen number density of other meshes is set to $n_{\rm H} = 10^{-3}~{\rm cm}^{-3}$. We also randomly distribute 1000 radiation sources in the simulation box. Each source has a $10^5$ K blackbody spectrum and steadily emits $\dot{N}_\gamma = 5 \times 10^{48}$ ionizing photons per second. The initial gas temperature and ionization fraction are set to $10^2$ K and $1.2 \times 10^{-3}$, respectively.
![Test 3 – Dependence on the accuracy parameter $\theta_{\rm crit}$. [*Upper panels*]{}: The volume fractions of the neutral fraction at $t = 500$ Myr. The results by the supermesh approximation are presented in the left panel. The solid (black), dotted (red), and dashed (blue) lines indicate the results with $\theta_{\rm crit} = 1.0$, $0.5$, and $0.0$, respectively. The relative difference to the long characteristics method ($\theta_{\rm crit} = 0$), $\Delta$, is also shown. The right panel shows the results obtained by the point source approximation. [*Lower panels*]{}: The volume fractions of the gas temperature at $t = 500$ Myr. The meaning of the lines are the same as in the upper panels. []{data-label="fig:multiple_hist"}](fig8.eps){width="8cm"}
![Test 3 – Relative difference in the temperature, cut through at the mid plane of the simulation box at $t = 500~{\rm Myr}$. This figure compares temperature obtained by the supermesh approximation with $\theta_{\rm crit} = 1$ to that by the long characteristics method ($\theta_{\rm crit} = 0$). The relative difference in temperature is defined as $\Delta_T = (\left.T\right|_{\theta_{\rm crit} = 1}^{\rm supermesh} - \left.T\right|^{\rm long}) / \left.T\right|^{\rm long}$. []{data-label="fig:temp_comp"}](fig9.eps){width="8cm"}
In Fig. \[fig:multiplesources\], we show the neutral fraction and temperature maps at the mid plane of the simulation box at $t = 30$, 100, and 500 Myr. We show the results by the supermesh approximation and by the point source approximation with $\theta_{\rm crit} = 0.5$. The results by two methods are virtually identical to each other including the shape of shadows by the optically thick meshes.
In order to investigate the dependence on the accuracy parameter $\theta_{\rm crit}$, we compare the simulations with $\theta_{\rm crit} = 1$, 0.5, and 0. In Fig. \[fig:multiple\_hist\], we show the volume fractions of the neutral fraction and the volume fractions of the gas temperature respectively in the upper panels and lower panels. [We also show difference in the volume fractions relative to those obtained by the long characteristics method ($\theta_{\rm crit} = 0$). For example, the relative difference in the volume fraction of the neutral fraction by the supermesh approximation with $\theta_{\rm crit} = x$ is defined as ]{} $$\Delta = \frac{\left.p(X_{{\ensuremath{\mbox{{H$\,${\sc i}}}}}})\right|_{\theta_{\rm crit} = x}^{\rm supermesh} -
\left.p(X_{{\ensuremath{\mbox{{H$\,${\sc i}}}}}})\right.|_{\rm long}}{\left.p(X_{{\ensuremath{\mbox{{H$\,${\sc i}}}}}})\right|_{\rm long}}.$$ [ The volume fractions of the neutral fraction with $\theta_{\rm crit} = 1$ and $0.5$ agree quite well with those by the long characteristics method ($\theta_{\rm crit} = 0$). The relative differences are typically less than 1 % even with $\theta_{\rm crit} = 1$. For a given value of the accuracy parameter, the point source approximation shows slightly better agreement with the long characteristics method. On the other hand, agreement in the volume fraction of the gas temperature is not as excellent as that for the neutral fraction. In particular, both the supermesh and point source approximation predict much more low temperature gas around $10^3$ K. This is because treating a source group as a point source underestimates the surface are of the ionized regions as we stated in Section \[sec:start\] and the low temperature gas is primarily heated by high energy photons that permeate beyond the surfaces of highly ionized regions. Except for this disagreement for the low temperature gas ($ \lesssim 2 \times 10^3$ K), typical difference is less than 10 %. ]{}
[ To study how serious the deviation from the long characteristics method at low temperature, we compare the temperature map obtained by the supermesh approximation ($\theta_{\rm crit} = 1$), which shows the worst agreement with the long characteristics method, and that by the long characteristics method in Fig. \[fig:temp\_comp\]. We find that the temperature difference is largest for the low temperature gas with $T \sim 10^3~{\rm K}$ (see also Fig. \[fig:multiplesources\]). The difference in temperature is however very small, only 10 % at most. We therefore conclude that the results with $\theta_{\rm crit} = 1$ are almost converged to the result obtained by the long characteristics method. ]{}
This test proves that both tree-based methods produce equally good results even with a large value of the accuracy parameter, $\theta_{\rm crit} = 1$, in the situation where a local ${\mbox{{H$\,${\sc ii}}}}$ region is driven primarily by one or a few sources. This situation is resembling to the early stage of cosmic reionization. Only at very late stage of the reionization, the ${\mbox{{H$\,${\sc ii}}}}$ regions overlap each other and multiple sources become visible each other; at this stage, the reionization has largely completed already. We thus expect that our tree-based methods, in particular the supermesh approximation, are well suited to this type of problems.
Test 4 – Clustered radiation sources in a clumpy medium
-------------------------------------------------------
{width="15cm"}
Unlike Test 3, here we explore the problem in which groups of sources act like bright extended sources and they ionize distant meshes. This would be one of the toughest problems for the methods accelerated by source grouping. The side length of the simulation box is the same as in Test 3, i.e. $L_{\rm box} = 132$ kpc. In order to construct clustered distribution of radiation sources, we put a sphere of radius $r = L_{\rm box}/4$, whose centre is randomly placed in the simulation box. We uniformly distribute 1000 radiation sources in the sphere. We then put a new sphere whose radius is 20% smaller than the previous one and again we distribute 1000 sources in the sphere. We continue this procedure until we put 10 spheres, each of which contains 1000 sources. Consequently, there are $10^4$ radiation sources in the simulation box. Each source has a $10^5$ K blackbody spectrum and emits $\dot{N}_\gamma = 5 \times 10^{48}$ ionizing photons per second. We also randomly select $10^4$ optically thick meshes whose hydrogen number density is $n_{\rm H} = 0.2~{\rm cm}^{-3}$. The hydrogen number density of the remaining meshes is set to $n_{\rm H} = 10^{-3}~{\rm cm}^{-3}$. The initial gas temperature and ionization fraction are set to $10^2$ K and $1.2 \times 10^{-3}$, respectively.
In Fig. \[fig:multiple\_clustered\], we show the neutral fraction maps, cut through at the mid plane of the simulation box. The size and shape of the ionized regions by the supermesh approximation strongly depend on the value of the accuracy parameter; The larger the value is, the smaller the size of the ionized regions is. This is due to the very nature of the supermesh approximation, which significantly overestimates the optical depth when a size of supermesh is large and the variance of the ${\ensuremath{\mbox{{H$\,${\sc i}}}}}$ density is large (see Eqn. (\[eq:error\]) and (\[eq:centrallimit\])). On the other hand, the results by the point source approximation are relatively insensitive to the value of the accuracy parameter. The size of the ionized regions is almost same between $\theta_{\rm crit} = 1$ and 0 while small difference is seen in the shapes.
In Fig. \[fig:clustered\_hist\], we show the volume fractions of the neutral fraction and gas temperature at $t = 100$ Myr varying the value of the accuracy parameter, $\theta_{\rm crit}$, from 1 to 0. We also show the relative difference to the long characteristics method ($\theta_{\rm crit} = 0$). The volume fraction of the neutral fraction confirms the dependence of the supermesh approximation on the value of the accuracy parameter, i.e. the larger the value of $\theta_{\rm crit}$ is, the smaller the ionized fraction is. This dependence is more evident in the volume fraction of the gas temperature. There is more low temperature gas in the simulation with a larger value of the accuracy parameter. Importantly, the results by the supermesh approximation with $\theta_{\rm crit} = 0.2$ still significantly deviate from those by the long characteristics methods, and therefore we cannot trust the result even with $\theta_{\rm crit} = 0.2$.
On the other hand, the result by the point source approximation with $\theta_{\rm crit} = 1$ shows an excellent agreement with that with the long characteristics method, in spite of the fact that this approximation ignores the spatial extent of source groups. This result proves that the point source approximation is very efficient and accurate for this type of problems.
[ The relative difference to the long characteristics method indicates that both approximations overestimates the volume fraction of the almost fully-ionized gas ($X_{{\ensuremath{\mbox{{H$\,${\sc i}}}}}} \simeq 2 \times 10^{-6}$). This ionized fraction corresponds to the central regions of each source spheres. The volume of these regions are however very small and the neutral fraction is very low anyway; this overestimation of the ionization fraction at the central regions of the source spheres does not affect the evolution of the whole simulation box. In fact, by the point source approximation, the relative difference to the long characteristics method in the volume fraction of the neutral fraction is typically 1 % and $\sim 10$ % at most except for the highly ionized gas with $X_{{\ensuremath{\mbox{{H$\,${\sc i}}}}}} \lesssim 10^{-5}$. ]{}
[ Even by the point source approximation, the relative difference in the volume fraction of the gas temperature to the long characteristics method is rather large for the low temperature gas. The gas temperature however agrees very well with that by the long characteristics method just as we showed for Test 3. Except for the low temperature gas, the typical difference is $\sim 10$ %. Interestingly, decreasing the value of the accuracy parameter in the point source approximation from 1 to 0.2 does not improve the agreement with the long characteristics method very much in spite of the fact that the simulations with a smaller value of the accuracy parameter is much more computationally expensive as we will show in the next subsection. Since the point source approximation with $\theta_{\rm crit} = 1$ seems to be sufficiently accurate, we expect that this approximation with $\theta_{\rm crit} = 0.5$ would be a safe choice for most types of problems. ]{}
![Test 4 – Dependence on the accuracy parameter $\theta_{\rm crit}$. Results at $t = 100$ Myr are presented. The results are displayed in the same manner as Fig. \[fig:multiple\_hist\]. The volume fractions for the simulation with $\theta_{\rm crit} = 1$, 0.5, 0.2, and 0 are indicated by the solid (black), dotted (red), dot-dashed (green), and dashed (blue) lines, respectively. []{data-label="fig:clustered_hist"}](fig11.eps){width="8cm"}
Code performance
----------------
We here investigate how the computation time scales with the number of meshes and that of the sources. For this purpose, we measure the wall-clock time taken for one step of the RT calculation. The computation time for solving chemistry etc. is not included. We use 8 cores of 2.13 GHz Xeon E5506 processors for these simulations.
In order to study the scaling with the number of the meshes, we randomly place 1000 radiation sources in the simulation box. Each source and the simulation box is the same as used in Test 1 except that there are 1000 sources and we vary the number of the meshes. We show the result in the upper panel of Fig. \[fig:timing\]. We find that the supermesh approximation is slightly faster than the point source approximation for a given set of $N_{\rm m}$ and $\theta_{\rm crit}$. The computation time by the point source approximation is a slightly steeper function of the number of the meshes than that by the supermesh approximation. The computation time by the point source approximation scales with $N_{\rm m}^{4/3}$ as expected. The scaling of the computation time by the supermesh approximation is somewhere between $\propto N_{\rm m} \log(N_{\rm m})$ and $\propto N_{\rm m}^{4/3}$. Since the RT is solved on the supermeshes whose angular size is similar to the angular size of the source group, $\theta_{\rm s}$, which can be much smaller than $\theta_{\rm crit}$, the computation time becomes steeper function of $N_{\rm m}$ than the expected scaling, $\propto N_{\rm m} \log(N_{\rm m})$. In the lower panel of Fig. \[fig:timing\], we plot the computation time as a function of the number of the sources. The number of the meshes is fixed to $128^3$. The computation time scales with $\log(N_{\rm s})$ for $N_{\rm s} > 1000$ in all cases. This result proves that the tree-based source grouping is quite efficient to deal with a large number of radiation sources. For a given set of $N_{\rm s}$ and $\theta_{\rm crit}$, a simulation by the supermesh approximation is always faster than that by the point source approximation. It should be however noted that even with the same value of the accuracy parameter, simulations by the point source approximation are sometimes much more accurate than those by the supermesh approximation as we showed by Test 4.
![Computation time taken for one step of the RT calculation. [*Upper panel*]{}: Computation time as a function of the number of meshes, $N_{\rm m}$. The number of radiation sources, $N_{\rm s}$ is fixed to 1000. The solid (red) and dashed (green) lines show the results by the supermesh approximation with $\theta_{\rm crit} =
1$ and 0.5, respectively. The dot-dashed (blue) and dotted (light blue) lines indicate the point source approximation with $\theta_{\rm crit} = 1$ and 0.5, respectively. The thin dot-dot-dot-dashed lines show the scaling with $N_{\rm m}^{4/3}$ and $N_{\rm m} \log(N_{\rm m})$. [*Lower panel*]{}: Same as the upper panel but the number of the source, $N_{\rm s}$, is varied. The number of meshes, $N_{\rm m}$ is fixed to $128^3$. The thin dot-dot-dot-dashed line indicates the scaling with $\log(N_{\rm s})$. []{data-label="fig:timing"}](fig12.eps){width="8.0cm"}
Summary and discussion
======================
We have presented a code to solve radiative transfer around point sources within a three-dimensional Cartesian grid, [ARGOT]{}, which accelerates the RT calculation by utilising the oct-tree structure in order to reduce the effective number of radiation sources. We have explored two methods: one is the supermesh approximation and the other is the point source approximation. In both methods, sources in a tree node whose angular size is smaller than the accuracy parameter $\theta_{\rm crit}$ are treated as a single bright source. As a result, computation time only scales with $\log(N_{\rm s})$. The main difference between these two method is that while the former takes the spatial extent of a source group into account, the latter ignores the size of the source group and treat it as a point source. In the supermesh approximation, the RT is solved using supermeshes whose angular size is similar to the angular size of the source group in question. Doing this results in the further acceleration of the RT calculation.
One might thus see that the supermesh approximation is superior to the point source approximation. We have however shown that the point source approximation is always equally or more accurate than the supermesh approximation for a given value of the accuracy parameter. This is because RT in a inhomogeneous medium on a supermesh inevitably overestimates the optical depth. This approximation can be in principle improved by including higher order moments, such as variance, although we do not take such an approach. This method hence only applicable to the problems in which a local ${\mbox{{H$\,${\sc ii}}}}$ region is driven primarily by one or a few sources such as Test 3 in this paper. When one applies the supermesh method to the simulation of cosmic reionization, it could be combined with the ‘local clumping factor’ approach proposed by @raicevic11, although exploring such a method is beyond the scope of this paper.
The point source approximation, which can be regarded as a mesh version of [START]{} [@START], produces sufficiently accurate results with $\theta_{\rm crit} = 1$ for all test simulations presented in this paper. This approximation requires slightly more computational cost than the supermesh approximation and it scales with $N_{\rm m}^{4/3} \log(N_{\rm s}) $. The performance can be improved if we choose the angular resolution so that at least one ray from a radiation source (or a group of sources) crosses all target meshes instead of solving RT to all target meshes. Doing this reduces the total number of rays from $\propto N_{\rm m}$ to $\propto N_{\rm m}^{2/3}$. Such an algorithm has been applied for RT from point sources [@yajima09] and can be extended to our tree-based algorithm. The expected scaling is $ N_{\rm m} \log{N_{\rm s}}$, which is even faster than the supermesh approximation and the same scaling by [START]{}.
For parallel implementation, if the entire meshes and sources can fit into memory of one computer node, parallelisation via angle decomposition is preferable to volume decomposition. We implement the angle decomposition by using both [MPI]{} and [OpenMP]{}. If a simulation size becomes too large to fit into the memory of one computer node, we have to employ the volume decomposition. The volume decomposition for RT around point sources was introduce by @RSPH and the algorithm can be applied to our methods. We leave the volume decomposition to future work.
The method presented in this paper can be easily combined with any grid-based hydrodynamic code, even with codes based on AMR [@FLASH; @RAMSES; @ENZO] and will be useful for various astrophysical problems in which a large number of radiation sources are required such as cosmic reionization and galaxy formation. We will apply our code for these issues in a forth coming paper.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank Kenji Hasegawa and Hideki Yajima for stimulating discussion. We are also grateful to the anonymous referee for helpful comments. The simulations were performed with FIRST and T2K Tsukuba at Centre for Computational Sciences in University of Tsukub and with the Cray XT4 at CfCA of NAOJ. This work was supported in part by the [*FIRST*]{} project based on Grants-in-Aid for Specially Promoted Research by MEXT (16002003), Grant-in-Aid for Scientific Research (S) by JSPS (20224002). TO acknowledges financial support by Grant-in-Aid for Young Scientists (start-up: 21840015).
[39]{} natexlab\#1[\#1]{}
T., [Anninos]{} P., [Zhang]{} Y., [Norman]{} M. L., 1997, , 2, 181
T., [Norman]{} M. L., [Madau]{} P., 1999, , 523, 66
T., [Wandelt]{} B. D., 2002, , 330, L53
S. M. V., [Pequignot]{} D., 1973, , 25, 137
P., [Zhang]{} Y., [Abel]{} T., [Norman]{} M. L., 1997, , 2, 209
D., [Teyssier]{} R., 2008, , 387, 295
J., [Hut]{} P., 1986, , 324, 446
R., 1992, , 78, 341
B., [Ferrara]{} A., [Marri]{} S., [Raimondo]{} G., 2001, , 324, 381
B., [et al.]{}, 2000, , 131, 273
N. Y., [Abel]{} T., 2001, , 6, 437
M., [Audit]{} E., [Huynh]{} P., 2007, , 464, 429
K., [Umemura]{} M., 2010, , 407, 2632
D. G., 1994, , 268, 109
D. G., [Storey]{} P. J., 1998, , 297, 1073
S., [Ostriker]{} J. P., 1986, , 301, 522
I. T., [et al.]{}, 2006, , 371, 1057
—, 2009, , 400, 1283
R. K., [Langer]{} W. D., [Evans]{} K., 1987, [Elementary processes in Hydrogen-Helium plasmas - Cross sections and reaction rate coefficients]{}, [Janev, R. K., Langer, W. D., & Evans, K.]{}, ed. Springer
M. R., 2006, , 641, L45
P., [Auer]{} L. H., 1988, , 39, 67
G., [Raga]{} A. C., [Canto]{} J., [Lundqvist]{} P., [Balick]{} B., [Steffen]{} W., [Noriega-Crespo]{} A., 1998, , 331, 335
T., [Umemura]{} M., [Susa]{} H., 2001, , 321, 593
K., [Mori]{} M., [Nakamoto]{} T., [Mineshige]{} S., 2005, , 628, 368
B. W., [Bryan]{} G., [Bordner]{} J., [Norman]{} M. L., [Abel]{} T., [Harkness]{} R., [Kritsuk]{} A., 2004, ArXiv Astrophysics e-prints:astro-ph/0403044
D. E., [Ferland]{} G. J., 2006, [Astrophysics of gaseous nebulae and active galactic nuclei]{}, 2nd edn., [Osterbrock, D. E. & Ferland, G. J.]{}, ed. University Science Books
A. H., [Schaye]{} J., 2008, , 389, 651
M., [Springel]{} V., 2009, , 396, 1383
—, 2011, , 415, 3731
Rai[č]{}evi[ć]{} M., Theuns T., 2011, , 412, L16
A. O., [Cardall]{} C. Y., 2005, , 362, 1413
M., [Gnedin]{} N. Y., [Shull]{} J. M., 2002, , 575, 33
A., [Abel]{} T., [Hernquist]{} L. E., 2001, , 6, 359
J. M., [Mihalas]{} D., [Norman]{} M. L., 1992, , 80, 819
H., 2006, , 58, 445
R., 2002, , 385, 337
J. H., [Abel]{} T., 2011, , 414, 3458
Yajima H., Umemura M., Mori M., Nakamoto T., 2009, , 398, 715
K., [Sasaki]{} S., 2006, , 58, 641
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: This choice may somewhat underestimate the effective projected size as for the case of a disc with a constant surface brightness. We have confirmed that simulation results are not sensitive to such a level of difference (a factor of $\sqrt{2}$).
[^3]: It should be noted that, in [START]{} [@START], the computational time scales with $N_{\rm p} \log(N_{\rm s})$, where $N_{\rm p}$ is the number of the SPH particles, by utilising the optical depths for SPH particles in the order of distance from the radiation source (see @RSPH and @START for more details). This scaling is better than our point source approximation.
[^4]: We note that not all codes in Cosmological Radiative Transfer Comparison Project were capable of dealing with multifrequency RT.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We extend the time-dependent variational principle to the setting of dissipative dynamics. This provides a locally optimal (in time) approximation to the dynamics of any Lindblad equation within a given variational manifold of mixed states. In contrast to the pure-state setting there is no canonical information geometry for mixed states and this leads to a family of possible trajectories — one for each information metric. We focus on the case of the operationally motivated family of *monotone riemannian metrics* and show further, that in the particular case where the variational manifold is given by the set of *fermionic gaussian states* all of these possible trajectories coincide. We illustrate our results in the case of the Hubbard model subject to spin decoherence.'
author:
- 'Christina V. Kraus'
- 'Tobias J. Osborne'
title: 'A time-dependent variational principle for dissipative dynamics'
---
Introduction
============
One of the main challenges in a quantum mechanical experiment is to overcome the interaction of a system with its environment. Such interactions lead to *decoherence* and often obscure coherent quantum phenomena. Recently it has been shown that this vice can be turned into a virtue: dissipative processes can be exploited as a possible resource for quantum state engineering [@Wolf_DSE; @Diehl_Kantian; @Kraus_Kantian]. Several evolutions leading to non-trivial fixed points have now been proposed, including states with non-trivial topological properties [@Diehl_Rico]. Such *dissipative engineering* has opened up a completely new world for us.
Motivated by the new possibilities offered by dissipative engineering there has been renewed interest in understanding dissipative processes in more detail. However, this task is complicated by the fact that, as for ground states, we can only hope for analytic solutions in very special cases. Therefore, in general, we must take recourse to numerical approximation techniques in order to gain insight into the physics of a dissipative system. Typically the method of choice here is a *Monte Carlo* sampling algorithm. Such methods have led to many insights into the dissipative systems occurring in quantum optics, but have faced limitations when applied to strongly interacting many particle systems, particularly fermions, due to the inevitable *sign problem*.
There is, however, another general approach available to us, namely the *variational method*. This method has been very successfully applied in the pure-state case leading to unparalleled insights into the equilibrium physics of strongly interacting many body systems. Further, the elegant *time-dependent variational principle* (TDVP) [@Dirac; @TDVP] allows the *locally optimal* study of nonequilibrium *unitary* dynamics. The power of this method is well known in the field of quantum chemistry, where its application to the class of Hartree-Fock states is known as *time-dependent Hartree Fock theory* [@HFT_excitations]. This technique has also been exploited to great effect in the context of one-dimensional quantum spin systems in conjunction with powerful expressive variational classes such as matrix product states, a method synonymous with the *density matrix renormalisation group* (DMRG) [@Jutho_dispersion].
In contrast to the pure-state case, there is no operationally unique way to formulate the variational method for mixed states because there is no distinguished measure of information distance: there are infinite families of inequivalent distance measures, including examples such as the *fidelity* and the *trace distance*. This has complicated the formulation of a mixed-state TDVP, which requires knowledge of the *geometry* of state space. However, recent results in our understanding of the information geometry of mixed states allow us to revisit this problem. (See, however, [@Jackiw; @Rajagopal] for related variational approaches to the von Neumann equation.)
Thus, in this Article, we formulate the TDVP for mixed states in the general case of distance measures arising from *monotone riemannian metrics*. We then show that this method, when applied to the variational class of fermionic Gaussian states evolving according to an arbitrary Markovian CPT (completely positive trace-preserving) map $\rho_t = \mathcal{E}_t(\rho_0)$ are all equivalent to the application of Wick’s theorem (via *Gaussification*). Finally, we apply this method to the one-dimensional (1d) spinful Hubbard model subject to a decoherence process.
A review of the TDVP for pure states
====================================
In this section we review of the TDVP for pure quantum states, and explain why this approach cannot be immediately applied to the mixed case. First, we present the necessary notation. We denote by $\mathcal{M}_n(\mathds{C})$ the set of all complex $n\times n$ matrices with entries in $\mathds{C}$. The state space of an $n$-dimensional quantum system is given by the set $\mathcal{D}_n$ (here viewed as a differentiable manifold) of all density operators defined by $\mathcal{D}_n =\{\rho \in \mathcal{M}_n(\mathds{C})\vert \rho^{\dagger} = \rho, \rho \geq 0, {\mbox{tr}}(\rho)=1 \}$. In order to formulate a time-evolution within this manifold, we have to introduce the notion of tangent space. The tangent space $T_{\rho}\mathcal{D}_n$ to $\mathcal{D}_n$ at any interior point $\rho \in \mathcal{D}_n$ can be identified with the set $\{A \in \mathcal{M}_n(\mathds{C})\,\vert\, A^{\dagger} = A, {\mbox{tr}}(A) = 0\}$ of traceless hermitian matrices (we assume that our processes never include any part of the boundary of the manifold). The set $\mathcal{D}_n$ can be given the structure of a Riemannian manifold by choosing a positive bilinear form $M_{\rho}(A,B)$ on $T_{\rho}\mathcal{D}_n$ for all $\rho \in \mathcal{D}_n$. This supplements us with the notion of a distance. Throughout we define a *variational class* to be a submanifold $\mathcal{V}$ of $\mathcal{D}_n$ parametrized according to $\mathcal{V} = \{\rho(\mathbf{x})\,\vert\, \mathbf{x} \in \mathds{R}^D\}$, where the dependence on the parameters $x^j$ is assumed to be analytic.
We consider the time evolution of a quantum state $\rho_t = \rho(\mathbf{x}(t))\in \mathcal{D}_n$ in its most general form, i.e. $\rho_t \equiv \mathcal{E}_t(\rho_0)$, where $\mathcal{E}_t$ is a completely positive trace-preserving (CPT) map. Assuming that $\mathcal{E}_t$ is differentiable with respect to $t$ allows us to write the equation of motion as $\partial_t\rho_t = \mathcal{L}(\rho_t)$, where $\mathcal{L}$ is the infinitesimal generator of the dynamics. For example, $\mathcal{L}$ could describe a Hamiltonian or dissipative evolution of a quantum system. An exact solution to this evolution is in general hard to find, and so we aim at finding an optimal approximation to this evolution within the variational class $\mathcal{V}$.
To this end, we first review the case of Hamiltonian time evolution within the set of pure quantum states. Here, the variational class of state vectors is represented as the set $\{\vert \psi(\mathbf{x})\rangle\vert \mathbf{x}\in \mathds{R}^D\}$. The time-dependent Schroedinger equation then reads $\dot x^{j}\partial_j |\psi(\mathbf{x})\rangle = -iH |\psi(\mathbf{x})\rangle$, where $\partial_j = \partial/\partial x^j$. Note that, in general, the vector $H |\psi(\mathbf{x})\rangle$ is not an element of the tangent space to $\mathcal{V}$ at $|\psi(\mathbf{x})\rangle$, whereas the left side is a linear combination of vectors that span the tangent space $T_{|\psi(\mathbf{x})\rangle}\mathcal{V}$. Thus, in general, there is no exact solution for $\dot x^j$. The best approximation is given by the solution to the minimisation of the *information distance* (here measured using the *fidelity*) between the left- and right-hand sides: $\min_{\dot{x}^j}\| \dot x^{j}\partial_j |\psi(\mathbf{x})\rangle + iH |\psi(\mathbf{x})\rangle\|$. The minimum can be found by applying an orthogonal projection of $iH |\psi(\mathbf{x})\rangle$ onto the tangent space, as depicted in Fig.\[fig:TDVP\].
![Time-dependent variational principle with respect to a variational manifold of mixed states $\mathcal{V} = \{\rho(\mathbf{x})| \mathbf{x}\in \mathds{R}\}$. A mixed quantum state $\rho (\mathbf{x})$ within a variational manifold $\mathcal{V}$ evolves in time according to an arbitrary physical process described by a CPT map $\rho_t = \mathcal{E}_t(\rho_0)$. In general, such a process leads out of the tangent space of the variational manifold at the point $\rho (\mathbf{x})$. Hence, we have to “project back” into $\mathcal{V}$ using an appropriate measure of distance (color online). \[fig:TDVP\] ](Cone_TDVP.pdf){width="0.5\columnwidth"}
This discussion immediately reveals why the TDVP cannot be directly applied to the mixed-state setting: the approximation of the RHS of $\dot x^j \partial_j \rho = \mathcal{L}_t(\rho)$ by a vector in the tangent space requires a unique notion of information distance. In the mixed-state case there is no operationally unique answer, since there exist infinite families of inequivalent measures, and hence, there is no canonical choice of Riemannian metric on $\mathcal{D}_n$. However, we explain in the next section a possible solution to this problem.
Formulation of the TDVP for mixed states
========================================
As we have explained above, there exists no canonical choice of Riemannian metrics in on the set of mixed quantum states. However, it turns out that there are several *families* of Riemannian metrics which naturally arise from information-theoretic considerations. Here the natural condition is that the metric is *monotone*, meaning that the norm induced by the bilinear form $M_{(\rho)}$ cannot increase under any CPT-map $\mathcal{E}$, i.e. $M_{\mathcal{E}(\rho)}(\mathcal{E}(A), \mathcal{E}(A))\leq M_\rho(A,A)$. The reasoning here is that the distinguishability of two states infinitesimally close to $\rho$ can never be increased under the action of a channel. Remarkably, Petz showed there is a one-to-one correspondence between the set of monotone metrics and a special class of *superoperators* $\Omega_\rho$ (built in terms of convex operator functions) [@Petz]: these lead to monotone metrics according to $M_{\rho}(A,B) = \langle A, \Omega_{\rho}(B)\rangle \equiv {\mbox{tr}}[A^{\dagger} \Omega_{\rho}(B)]$.
These monotone metrics now allow an operationally motivated formulation of the TDVP for the dissipative dynamics generated by $\partial_t \rho_t = \mathcal{L}(\rho_t)$ within a given variational class $\mathcal{V}$. The setup is identical to the pure-state case: we aim to find the optimal trajectory $\rho_t \in \mathcal{V}$ generated by the vector field coming from the optimal element $A \in T_{\rho_t}\mathcal{V}$ which is closest to $\mathcal{L}(\rho_t)$, where we use the quadratic form $M_{\rho_t}(A,B)$ to measure the distance. That is, we solve $\inf_{A\in T_{\rho_t}\mathcal{V}} M_{\rho_t}(A- \mathcal{L}(\rho_t), A- \mathcal{L}(\rho_t))$. An intuitive picture of this last equation is given in Fig. \[fig:TDVP\]: The evolution under $\mathcal{E}$ takes us out of the variational manifold $\mathcal{V}$, and we want to ”project back” into $\mathcal{V}$ to find the state in the variational manifold that is the best approximation to this evolution. Using the definition of $M_{\rho_t}$ we find that this is equivalent to solving $$\begin{aligned}
\inf_{A\in T_{\rho_t}\mathcal{V}} \langle A- \mathcal{L}(\rho_t), \Omega_{\rho_t}(A- \mathcal{L}(\rho_t))\rangle.\end{aligned}$$ Parametrizing $A = v^j \partial_j \rho_t$ we can rewrite this infimum as $\inf_{\mathbf{v}\in\mathds{R}^D} \mathbf{v}^T \mathbf{G}_{\rho}\mathbf{v} - \mathbf{v}^T\mathbf{l}_{\rho} - \mathbf{l}_{\rho}^T \mathbf{v} + c_0$. The solution is given by $$\begin{aligned}
\label{eq:TDVP}
\mathbf{v} &= \mathbf{G}^{-1}_{\rho}\mathbf{l}_{\rho},\\
(\mathbf{G}_{\rho})_{jk} &= \langle \partial_j \rho(\mathbf{x}(t)), \Omega_{\rho}(\partial_k \rho(\mathbf{x}(t)))\rangle,\\
(\mathbf{l}_{\rho})_j &= \langle \partial_j \rho(\mathbf{x}(t)), \Omega_{\rho}(\mathcal{L}(\rho(\mathbf{x}(t))))\rangle, \end{aligned}$$ where $\mathbf{G}_{\rho}$ is the *pullback metric* or *Gram matrix*. This solution gives us the optimal trajectory (locally in time) within $\mathcal{V}$ via integration of the equation of motion $$\partial_t\rho_t = v^j(t, \rho_t)\partial_j \rho_t.$$ This equation of motion is the first main contribution of our Article. Eqs. (2)–(5) can be applied to any variational manifold subject to any physical process. In the next Section we apply this framework to a concrete example that is relevant in many-body physics.
The TDVP for fermionic Gaussian States
======================================
The understanding of fermionic quantum systems is of central interest many fields of physics. Fermions are building blocks of matter and thus central to some of the most fascinating effects known in the theory of many-body physics, like superconductivity or the quantum Hall effect. However, most problems of interest do not have a closed analytic solution, and we have to use numerical approximation techniques and we have to use appropriate variational wave functions to obtain insight into these systems. In the following we introduce the class of *fermionic Gaussian states* (fGS) that has been successfully applied to solve fermionic many-body problems in the pure state setting. We show then how the TDVP can be applied using this class of states as our variational class $\mathcal{V}_G$ and give a numerical example in the last Section.
In the following we describe fermionic systems in terms of $N$ fermionic mode operators $a_j$ obeying the canonical anti-commutation relations $\{{a^{\dagger}}_k, a_l\} = \delta_{kl}$. We use the equivalent representation in terms of $2N$ hermitian *Majorana* operators $c_{2j-1} = {a^{\dagger}}_j + a_j$ and $c_{2j} = (-i)({a^{\dagger}}_j - a_j)$ which obey $\{c_k, c_l\} = 2\delta_{kl}$. We take as our variational manifold the set of *fermionic Gaussian states* (fGS). fGS are those states whose density operator can be expressed as an exponential of a quadratic function of the Majorana operators, $\rho = \kappa \exp[-\tfrac{i}{2}c^TKc]$, where $K = -K^T \in \mathds{R}^{2N \times 2N}$. All information about the state is encoded in the real and anti-symmetric covariance matrix (CM), $\Gamma_{kl} = \tfrac{i}{2}{\mbox{tr}}([c_k, c_l]\rho)$, due to Wick’s theorem: $i^p {\mbox{tr}}[\rho c_{j_1}\ldots c_{j_{2p}}] = \mbox{Pf}(\Gamma_{j_1,
\ldots, j_{2p}})$ where $1 \leq j_1 < \ldots < j_{2p} \leq 2M$ and $\Gamma_{j_1,
\ldots, j_{2p}}$ is the corresponding $2p \times 2p$ submatrix of $\Gamma$. $\mbox{Pf}(\Gamma_{j_1, \ldots, j_{2p}})^2= \mbox{det}
(\Gamma_{j_1, \ldots, j_{2p}})$ is called the Pfaffian (see, e.g., [@LinearOptics] for further details).
The class of fGS is a natural generalization of the variational classes used in Hartree-Fock and BCS-theory. Thus, it is combining and extending the most successful tools in the description of fermionic many-body systems and hence allows for a description of a wide class of fermionic phases of matter, like superfluids, Mott and spin ordered phases. Recently it has also been shown that fGS with topological order can be engineered in a cold-atom implementation via a local dissipative process [@Diehl_Rico]. Thus, fGS have proven to be a powerful class capable of capturing fermionic phases with highly non-trivial properties. Every pure fGS is the ground state of a quadratic Hamiltonian. Further, fGS remain Gaussian under the evolution according to a quadratic Hamiltonian or a dissipative process with linear Lindblad operators. Using this, fGS have allowed the approximation of the ground and thermal states of, the time-evolution [@gHFTnum] of, as well as the excitation spectra [@KrausOsborne] of interacting fermionic systems.
The main ingredient used in all studies exploiting fGS is a process known as *Gaussification*, i.e., the approximation of any $N$-body correlation function in terms of a product of single-particle correlation functions via Wick’s theorem (see above):
Let $\sigma$ be a fermionic quantum state. Then its *Gaussification*, $\rho_G = \mathcal{G}(\sigma) \in \mathcal{V}_G$, is defined via the relation $\Gamma(\rho_G) = \Gamma(\sigma)$, which is equivalent to the application of Wick’s theorem to $\sigma$.
In the following, we show that the process of Gaussification is locally optimal in time within the variational class of fGS for all $\alpha$ metrics, since every monotone metric can be written as a convex combination of them [@Petz].
Let $\mathcal{E}_t$ be an arbitrary (differentiable) Markovian CPT map defining a time evolution on the space of density matrices via $\rho_t = \mathcal{E}_t(\rho_0)$. Then, the optimal approximation of this time evolution within the variational manifold of Gaussian states $\mathcal{V}_G$ with respect to any (convex combination of) monotone metrics of *$\alpha$-norm type*: $$\begin{aligned}
\label{eq:omega}
\Omega_{\rho}^{\alpha}(\sigma) = \tfrac{1}{2}(\rho^{-\alpha}\sigma \rho^{\alpha-1} + \rho^{\alpha-1} \sigma \rho^{-\alpha}),\end{aligned}$$ is obtained via Gaussification.
Theorem 1 is the second main result of this Article. To prove it we obtain a Gaussification of the time evolution of $\rho$ according to the generator $\mathcal{L}$ of the CPT map $\mathcal{E}_t$ as follows. Let $\rho(t+\delta t) = \rho(t) + \delta t \mathcal{L}(\rho)$, where $\delta t$ is an infinitesimal time step. The operator $\rho(t + \delta t) -\rho(t)$ is not necessarily a member of the tangent space $T_{\rho}\mathcal{V}_G$. However, the Gaussified operator $\mathcal{G}(\rho(t+\delta t)) -\rho(t)$ is, and it is therefore a linear combination of the tangent vectors $\partial_j \rho$, i.e. $\mathcal{G}(\rho(t+\delta t)) = \rho(t) + \delta t \sum_j v^j \partial_j \rho$, where $v^j \in \mathds{R}$. Since, by definition, Gaussification implies $\Gamma(\mathcal{G}(\rho)) = \Gamma(\rho)$, we obtain from the linearity of the map $\Gamma$, the following defining condition for the expansion parameters $v^j$: $$\begin{aligned}
\label{eq:Gaussification}
{\mbox{tr}}[c_{k_1} c_{k_2} \mathcal{L}(\rho)] = \sum_jv^j {\mbox{tr}}[c_{k_1} c_{k_2} \partial_j \rho].\end{aligned}$$ Now we show that an application of the TDVP projection with respect to any $\alpha$-norm also leads to Eq. . For the proof of this statement we need the following
Let $\rho \in \mathcal{V}_G$ be a fGS of $N$ modes, and let $T_{\rho}\mathcal{V}_G$ denote the tangent space of $\rho$. Then, $\mathcal{B}=\{\partial_{(j_1,j_2)} \rho = (\Omega_{\rho}^{\alpha})^{-1}(ic_{j_1}c_{j_2})\}_{ 1\leq j_1< j_2 \leq 2N}$ is a hermitian basis for the tangent space $T_{\rho}\mathcal{V}_G$. (Here $(j_1,j_2)$ denotes a *multi index*.)
We first claim that if $\{Q_a\}_{a=1}^{2N}$ is a set of linearly independent operators then so is $\{\Omega_{\rho}^{\alpha}(Q_a)\}_{a=1}^{2N}$. This follows immediately from the fact that $\Omega_\rho$ is a *positive* superoperator so that $\mathcal{B}$ is a set of linearly independent vectors.
Next, we show that $\dim(\mathcal{B}) = \dim(T_{\rho}\mathcal{V}_G)$. To this end, we determine a basis of $T_{\rho}\mathcal{V}_G$ by applying the most general infinitesimal Gaussian transformation on $\rho$ [@LinearOptics]. These are of the from $\rho \mapsto W \rho W^{\dagger} / {\mbox{tr}}[W \rho W^{\dagger}]$, where $W = e^{i \varepsilon Z_{kl}c_kc_l}$ with $Z_{kl} = X_{kl} + iY_{kl}$, $X_{kl}, Y_{kl} \in \mathds{R}$, and $\varepsilon \ll 1$. This leads directly to the tangent vectors $A^{(R)}_{kl} = [\rho, c_kc_l]$ and $A^{(I)}_{kl} = i\{\rho, c_kc_l\} - 2\Gamma_{kl}\rho$. In order to determine the number of linearly independent tangent vectors, we work in the basis $\tilde c_k =\sum_l O_{kl} c_l$, where $OO^T = \mathds{1}$, so that $\rho$ is in its standard form, $\rho = \prod_{j=1}^N\tfrac{1}{2}(\mathds{1} + i\lambda_j\tilde c_{2j-1}\tilde c_{2j})$, where $\lambda_j \in (-1, 1)$ (i.e. $\rho$ has no pure subspace). Then, the tangent vectors are readily calculated. For all $1 \leq k < l \leq N$ we obtain $A^{(R)}_{2k,2l} =$ $ -2i(\lambda_l \tilde c_{2k} \tilde c_{2l-1} + \lambda_k \tilde c_{2k-1} \tilde c_{2l}) \hat \rho_{kl}$, $A^{(R)}_{2k,2l-1} =$ $ 2i(\lambda_l \tilde c_{2k} \tilde c_{2l} - \lambda_k \tilde c_{2k-1} \tilde c_{2l-1})\hat \rho_{kl}$, $A^{(R)}_{2k-1,2l-1} =$ $ 2i(\lambda_l \tilde c_{2k-1} \tilde c_{2l} + \lambda_k \tilde c_{2k} \tilde c_{2l-1})\hat \rho_{kl}$ and $A^{(R)}_{2k-1,2l} =$ $ 2i(-\lambda_l \tilde c_{2k-1} \tilde c_{2l-1} + \lambda_k \tilde c_{2k} \tilde c_{2l})\hat \rho_{kl}$, where $\hat \rho_{kl} =$ $\prod_{j\neq k,l}\tfrac{1}{2}(\mathds{1} + i\lambda_j\tilde c_{2j-1}\tilde c_{2j})$. Further, we find that $A^{(I)}_{2k,2l} = $ $ 2i(\tilde c_{2k} \tilde c_{2l} - \lambda_k \lambda_l \tilde c_{2k-1} \tilde c_{2l-1})\hat \rho_{kl}$, $A^{(I)}_{2k,2l-1} =$ $ 2i(\tilde c_{2k} \tilde c_{2l-1} + \lambda_k \lambda_l \tilde c_{2k-1} \tilde c_{2l})\hat \rho_{kl}$, $A^{(I)}_{2k-1,2l-1} =$ $ 2i(\tilde c_{2k-1} \tilde c_{2l-1} - \lambda_k \lambda_l \tilde c_{2k} \tilde c_{2l})\hat \rho_{kl}$ and $A^{(I)}_{2k-1,2l} =$ $ 2i(\tilde c_{2k-1} \tilde c_{2l} + \lambda_k \lambda_l \tilde c_{2k} \tilde c_{2l-1})\hat \rho_{kl}$. Thus we obtain, for all $\lambda_{k,l} \in (-1,1)$, the four linearly independent basis vectors $i\tilde c_{2k-1}\tilde c_{2l-1}\hat \rho_{kl}$, $i\tilde c_{2k-1}\tilde c_{2l}\hat \rho_{kl}$, $i\tilde c_{2k}\tilde c_{2l-1}\hat \rho_{kl}$ and $i\tilde c_{2k-}\tilde c_{2l}\hat \rho_{kl}$. If $k=l$, we obtain $A^{(I)}_{2k-1,2k} =2 i(1-\lambda_k^2)\tilde c_{2k-1}\tilde c_{2k}\hat \rho_{k}$, while $A^{(R)}_{2k-1,2k}=0$. Thus the dimension of the tangent space is $\dim(T_{\rho}\mathcal{V}_G)=4 N(N-1)/2 + N = 2N(2N-1)= \dim(\mathcal{B})$.
Finally, we show that the operators $\partial_{(j_1, j_2)}\rho$ are hermitian. This follows immediately from the fact that $\Omega_{\rho}^{\alpha}((\partial_{(j_1, j_2)}\rho)^{\dagger}) = \Omega_{\rho}^{\alpha}(\partial_{(j_1, j_2)}\rho)$, and that $\Omega_{\rho}^{\alpha}$ is invertible
With Corollary 1 in hand we now directly apply the TDVP projection in our special basis, following Eq. . We see that $$\begin{aligned}
(G_{\rho})_{(j_1,j_2),(k_1,k_2)} &= \langle \partial_{(j_1,j_2)}\rho, \Omega_{\rho}^{\alpha}(\partial_{(k_1,k_2)}\rho)\rangle \nonumber\\
&= {\mbox{tr}}[\partial_{(j_1,j_2)}\rho c_{k_1}c_{k_2}],\\
(l_{\rho})_{(j_1,j_2)} &= \langle \partial_{(j_1,j_2)}\rho, \Omega_{\rho}^{\alpha}(\mathcal{L}(\rho))\rangle ={\mbox{tr}}[\mathcal{L}(\rho)c_{k_1}c_{k_2}],
\end{aligned}$$ and we arrive at Eq. . This proves the equivalence of Gaussification and the application of the TDVP to the variational class of Gaussian states.
![a) Anti-ferromagnetic order (AF) in the ground state $\rho_0$ of the Hubbard model for $u=4, \mu = -2$. The steady state $\rho_s$ shows a ferromagnetic order (FM). b) Comparison of the real-time evolution (time $t$ in units of $1/\kappa$) of the dissipative process with the Gaussified version. We present the difference of the CM and the two states $d\Gamma(t) = ||\Gamma_{\rho_G(t)} - \Gamma_{\rho(t)}||_2$, $d\rho(t) = || \rho_G(t) - \rho(t)||_2$. c) Purity, $p_{\rho} = {\mbox{tr}}[\rho^2]$, for the exact (solid) and the Gaussified (dashed) evolution, $\rho_G$. d) Evolution of the particle number $n$ for the two spin states $\sigma_s = \uparrow, \downarrow$ for the real ($\rho$) and Gaussified ($\rho_G$) process (color online).\[fig:example\]](TVDP_numerics.pdf){width="0.95\columnwidth"}
Example
=======
In the following we apply our approach to the $1d$ spinful Hubbard model with repulsive interactions subject to a magnetic field. The evolution of the system is described by a Lindblad equation $\partial_t \rho = -i[H,\rho] + \kappa\sum_{x}j_x \rho j_x^{\dagger} - \tfrac{1}{2}\{ j_x^{\dagger} j_x, \rho\}$, where $$\begin{aligned}
H &= J\sum_{x, \sigma_s} {a^{\dagger}}_{x, \sigma_s}a_{x+1,\sigma_s} + u\sum_{x}n_{x,\uparrow}n_{x,\downarrow} + \mu \sum_{x,\sigma_s} n_{x,\sigma_s},\\
j_x &= {a^{\dagger}}_{x,\uparrow}a_{x,\downarrow}. \end{aligned}$$ Starting from the ground state of the Hubbard Hamiltonian we expect that the external noisy magnetic field modeled by the operators $j_{x}$ drives the system to a completely spin-polarized state. The time scale of this process depends on the ratio between the decoherence strength $\kappa$ and the parameters of the Hubbard Hamiltonian.
We consider a system of $L=4$ sites with periodic boundary conditions at half filling and consider an interaction $u = 4$, a chemical potential $\mu = -2$ and $\kappa = 1$, where we take the hopping $J$ to be the energy scale. The unique ground state $\rho_0$ has anti-ferromagnetic (AF) order (see Fig.\[fig:example\]a (squares)). We implement a real-time evolution $\rho(t)$ of the dissipative process, arriving at a unique steady state $\rho_s$ which is completely spin polarized (ferromagnetic order (FM)) (Fig. \[fig:example\]a (circles)) and given by $\prod_{x=1}^4{a^{\dagger}}_{x\uparrow}{|0\rangle }$. In order to measure how well the Gaussified evolution given by $\rho_G(t)$ approximates the exact dynamics, we present the deviation $d\Gamma(t) = ||\Gamma_{\rho_G(t)} - \Gamma_{\rho(t)}||_2$ of the CM of the real and Gaussified process and the distance between the two states $d\rho(t) = || \rho_G(t) - \rho(t)||_2$ in Fig. \[fig:example\]b. We find that, as one might expect, on short time scales the Gaussified dynamics takes a different path from the exact evolution, but quickly coincides with the exact evolution for intermediate and long time scales. This can be explained by the fact that the ground state of the Hubbard model is not a Gaussian state, so that the initial state for the Gaussified evolution is a large distance from the ground state.
In order to obtain more insight into the Gaussified evolution we compare the time dependence of some physical quantities with the exact evolution. In Fig. \[fig:example\] c we present the purity $p_{\rho} = {\mbox{tr}}[\rho^2]$ for the exact evolution (solid) and the Gaussified process (dashed). We see that we end up in a pure state with only spin up particles (Fig. \[fig:example\]d) in the limit $t\rightarrow \infty$ in both cases.
Summary
=======
In summary, we have extended the time-dependent variational principle to the mixed-state setting, providing locally optimal (in time) equations that allow for an approximation of any Lindblad dynamics, given a variational manifold of mixed states and some information metric. In the case of fermionic Gaussian states we have proven that all $\alpha$-metrics lead to the same dynamics in the space of density matrices, which can equivalently be obtained via an application of Wick’s theorem at each time step (Gaussification). Thus, this method can easily be applied to large systems in any dimension and geometry, providing a powerful numerical tool for a variational study of dissipative dynamics for mixed quantum states.
*Acknowledgments.—*We thank J. Haegeman and F. Verstraete for useful discussion. This work was supported, in part, by the cluster of excellence EXC 201 “Quantum Engineering and Space-Time Research”, by the Deutsche Forschungsgemeinschaft (DFG), the EU grant QFTCMPS and by the Austrian Ministry of Science BMWF as part of the UniInfrastrukturprogramm of the Research Platform Scientific Computing at the University of Innsbruck.
[14]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , ****, ().
, , , , , , ****, ().
, , , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this brief communication we show why superclusters would naturally arise in the universe.'
author:
- |
B.G. Sidharth$^*$\
Centre for Applicable Mathematics & Computer Sciences\
B.M. Birla Science Centre, Adarsh Nagar, Hyderabad - 500063 (India)
title: LARGE SCALE STRUCTURES IN THE UNIVERSE
---
It is well known in the theory of the Random Walk or Brownian motion that $$R = \sqrt{N}l\label{e1}$$ holds where $R$ denotes the dimension of the system, $N$ the number of steps or events and $l$ represents a mean free path[@r1]. It has already been argued[@r2; @r3] that in the context of the universe as a whole, $N$ representing the total number of particles, (\[e1\]) gives the Eddington relation with $l$ being the pion Compton wavelength.\
Let us now consider $N \sim 10^6$ constituents in the universe. Then (\[e1\]) gives $$l \sim 10^{25} cm\label{e2}$$ Indeed there is observational evidence for (\[e2\]): We can easily see that (\[e2\]) holds for superclusters, both in terms of their size $l$ and their number $N$[@r4]. Further, (\[e1\]) shows that these superclusters must have a two dimension character, which is also true.\
It is ofcourse well known that one cannot apply the theory of Brownian motion to stars or even galaxies because they are gravitationally bound. However for superclusters with the huge separating voids, Brownian motion would be a reasonable approximation, as can be seen by the fact that (\[e2\]) is valid. So it is natural that such superclusters should arise.\
It is interesting to note also that recently, in a completely different context, it was suggested[@r5] that there could be a large scale quantization, giving precisely (\[e2\]) as the quantized length.
[99]{} F. Rief, “Fundamental of Statistical and Thermal Physics, McGraw Hill, Tokyo, 1965. B.G. Sidharth, ”The Chaotic Universe“, to appear in Chaos, Solitons and Fractals. (xxx.lanl.gov/quant-ph/abs/9811045). B.G. Sidharth, ”The Universe of Chaos and Quanta“, to appear in Chaos, Solitons and Fractals. (xxx.lanl.gov/quant-ph/abs/9902028). J.V. Narlikar,”Introduction to Cosmology", Foundation Books, New Delhi, 1993. S. Carneiro, Foundations of Physics Letters, (1) 1998, pp95-102.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this proceeding we present the results from a study of very high–redshift galaxies with the newly commissioned Wide Field Camera 3 on the [*Hubble Space Telescope*]{}. With the deepest near-infrared data ever taken, we discovered 31 galaxies at 6.3 $<$ z $\leq$ 8.6. The rest–frame ultraviolet (UV) colors of these galaxies are extremely blue, showing significant ($>$ 4 $\sigma$) evolution from z $\sim$ 3, over only 1 Gyr of cosmic time. While we cannot yet diagnose the exact cause of the bluer colors, it appears a low dust content is the primary factor. The stellar masses of these galaxies are less than comparably selected galaxies at 3 $<$ z $<$ 6, highlighting evolution in the stellar mass of characteristic (L$^{\ast}$) galaxies with redshift. Lastly, the measured rest–UV luminosity density of galaxies in our sample seems sufficient to sustain reionization at z $\sim$ 7 when we account for the likely contribution from galaxies below our magnitude limit.'
author:
- 'Steven L. Finkelstein'
title: 'Probing Stellar Populations at z $\sim$ 7 – 8'
---
[ address=[Texas A&M University\
George P. and Cynthia Woods Mitchell Institute for Fundamental Physics and Astronomy\
Department of Physics and Astronomy\
4242 TAMU, College Station, TX 77845\
[email protected]]{} ]{}
Introduction
============
Until recently, the robust identification of galaxies at z $\geq$ 7, corresponding to a time when the Universe was only $\sim$ 800 Myr old, was extremely difficult due to existing near–infrared instrumentation. The recent installation of the Wide Field Camera 3 (WFC3) on the [*Hubble Space Telescope*]{} ([*HST*]{}) has opened up the distant universe for detailed exploration for the first time. WFC3 has obtained the deepest near–infrared (NIR) images ever taken, in the [*Hubble*]{} Ultra Deep Field (HUDF), obtaining data in three filters, Y$_{105}$, J$_{125}$, and H$_{160}$ (PID 11563, PI G. Illingworth). In conjunction with existing optical Advanced Camera for Surveys (ACS) data, these new data allowed selection of galaxies at z $>$ 6.5 by searching for the Lyman break, due to intervening intergalactic medium (IGM) H[i]{} absorption. This is typically done by choosing a set of color criteria which select galaxies at the desired redshift, while excluding interlopers such as low–redshift passively evolving galaxies and galactic brown dwarfs, which can mimic high–redshift galaxies in a single color. Soon after these data were released, samples of z $>$ 6.5 galaxies were published, mostly using color cuts [e.g., @bouwens10; @oesch10; @mclure10; @bunker09; @yan09]. In these proceedings, we present the results of an independent study, using photometric redshift techniques to select a high–redshift galaxy sample, as well as rest–frame ultraviolet (UV) color analysis and stellar population modeling to study the physical characteristics of these galaxies. This full study is published in [@finkelstein10].
High–Redshift Galaxy Sample
===========================
We fit photometric redshifts to objects in the WFC3 data using the photometric redshift code EAZY [@brammer08]. This takes into account all available data, including the ACS photometry, when computing the redshift, and computes the full redshift probability distribution function, which is a sum over all available templates. This also provides a better redshift estimate, $\Delta$z $\pm$ 0.15–0.3, versus $\Delta$z $\pm$ 0.5 for the Lyman break technique. We found 31 objects which met our three selection criteria: 1) $>$ 3.5 $\sigma$ significance in the J$_{125}$ [*and*]{} H$_{160}$ bands; 2) Best–fit photometric redshift of 6.3 $<$ z$_\mathrm{phot}$ $\leq$ 8.6; and 3) $\geq$ 60% of the integrated probability distribution function ($\mathcal{P}_6$ ) at z $>$ 6.
{height=".28\textheight"}
Rest–Frame UV Colors of z $\sim$ 7 Galaxies
===========================================
Figure 1 shows the 23 objects in our z $\sim$ 7 sub–sample (6.3 $<$ z$_\mathrm{phot}$ $<$ 7.5) on a color–magnitude diagram. Individual galaxies have blue rest–UV colors, $-$0.5 $<$ $J-H$ $<$ 0.2, corresponding to a UV spectral slope of $-$4 $<$ $\beta$ $<$ $-1$. However, the faint nature of these galaxies results in a high uncertainty on their colors. Thus, to learn about the z $\sim$ 7 galaxy population as a whole, we computed the mean color of all objects in the z $\sim$ 7 sample, as well as the mean color in bright and faint bins (split at $H$ = 28.3). We ran 10$^{7}$ bootstrap Monte Carlo simulations to assess the uncertainty on these mean colors, where each simulation accounts for Poisson noise in the sample size as well as photometric scatter. The means and their corresopnding uncertainties are shown in Figure 1. We compare these means to the colors of local starburst galaxies [@kinney96], as well as the local extremely blue galaxy NGC 1705 (which is thought to be dust–free), and find that z $\sim$ 7 galaxies are bluer than local starbursts at $>$ 4 $\sigma$ significance (i.e., 99.9987% confidence). This is in stark contrast to the properties of Lyman break galaxies at z $\sim$ 3, which, only a billion years later, have rest–frame UV colors consistent with local starbursts [i.e., @papovich01; @reddy09; @bouwens09]. Thus there is significant evolution in the rest–frame UV properties of star–forming galaxies over only $\sim$ 1 Gyr of cosmic time. The faint galaxies in our z $\sim$ 7 sample have $\beta$ = $-$3.07, which is difficult for “normal” stellar populations to form, thus suggesting the presence of a top–heavy initial mass function (IMF), or zero metallicity (Population III) stars [e.g., @bouwens10b]. However, the photometric scatter at H $>$ 28.3 results in a large uncertainty of $\sigma_\beta$ = $\pm$ 0.5. Thus, while exotic stellar populations [*could*]{} be present, these data do not provide strong evidence to support their existence.
![The stellar masses of L$^{\ast}$ LBGs versus redshift (dark gray–[@reddy06]; black–[@shapley05]; gray–[@stark09] at z = 4, 5 and 6). The mass at L$^{\ast}$ at $z = $7–8 from our results are shown by the large stars. The error bars represent the central 80% range of the data. We show our individual galaxies as small stars with their uncertainties, which are typically a factor of +2/$-$4.5. The background gray circles denote stellar masses of Ly$\alpha$ emitting galaxies at 3.1 $\leq$ z $\leq$ 6.5 from the literature [e.g., @gawiser06b; @pirzkal07; @finkelstein07; @finkelstein09a]. The gray hatched region denotes the interquartile range of the LAE masses. The masses of the z $>$ 6.3 LBGs studied here are more similar to those of LAEs at all redshifts than LBGs at any redshift $<$ 6.](f2){height=".28\textheight"}
Stellar Mass Evolution
======================
We compared the spectral energy distributions (SEDs) of our sample to updated stellar population models of S. Charlot & G. Bruzual [@bruzual03]. Unfortunately, with only NIR detections (i.e. upper limits in the optical and in the infrared with [*Spitzer*]{}) we are unable to robustly constrain the ages and dust properties of these galaxies, though their very blue colors imply low levels of dust extinction. However, these blue colors do place limits on the metallicities, with a 68% confidence of Z $<$ 0.05 Z[$_{\odot}$]{}. Even though these galaxies are not detected in their rest–frame optical (observed infrared) we place constraints on their stellar masses as the young age of the Universe limits the amount of mass which can exist in old stars, unseen in the WFC3 bands. We find best–fit masses in the range of $\sim$ 10$^{8}$ – 10$^{9}$ M[$_{\odot}$]{}. Figure 2 shows the masses of our galaxies (interpolated to L$^{\ast}$ using the luminosity functions of [@oesch10] and [@bouwens10]) compared to the stellar masses of lower–redshift LBGs (also converted to the mass at L$^{\ast}$). Comparing the masses of L$^{\ast}$ galaxies is roughly equivalent to studying them at a constant number density, thus we trace the stellar masses of the direct progenitors and descendents of a given galaxy. Figure 2 also shows the masses of galaxies selected on the basis of their Ly$\alpha$ emission (LAEs), which are typically less evolved than LBGs [i.e., @gawiser06b; @finkelstein09a]. The masses of z $\geq$ 7 galaxies are more comparable to LAEs at all redshifts than LBGs at any lower redshift. The low masses and likely low–dust content of z $\geq$ 7 galaxies implies that they are physically similar to lower–redshift LAEs rather than scaled–down versions of LBGs. Lower redshift LAEs are the likely building blocks of more evolved galaxies [i.e., @gawiser07], thus at z $\sim$ 7 it appears we are observing an epoch dominated by these building blocks, and that the more evolved galaxies common at lower–redshifts are rare.
Summary
=======
We have analyzed the properties of 31 galaxies at z $>$ 6.3 in the extremely deep HUDF. These objects appear very similar to lower–redshift LAEs, implying that at z $\geq$ 7 we have reached the era of “baby” galaxies. However, it does not appear we have yet reached the era of “infant” galaxies, where one would expect Population III stars or top heavy IMFs, but these may come to light in the near future with the [*James Webb Space Telescope*]{}.
I would like to thank my collaborators on this work, Casey Papovich, Mauro Giavalisco, Naveen Reddy, Harry Ferguson, Anton Koekemoer and Mark Dickinson.
[21]{} natexlab\#1[\#1]{}\[1\][“\#1”]{} url \#1[`#1`]{}urlprefix\[2\]\[\][[\#2](#2)]{}
R. J. [Bouwens]{} et al., *ApJL*, **709**, L133–L137 (2010). P. A. [Oesch]{} et al., *ApJL*, **709**, L16–L20 (2010), R. J. [McLure]{} et al., *MNRAS*, **403**, 960–983 (2010), A. [Bunker]{} et al., *ArXiv e-prints* (2009), .
H. [Yan]{} et al. *ArXiv e-prints* (2009), .
S. L. [Finkelstein]{} et al., *ArXiv e-prints* (2009), .
G. B. [Brammer]{}, P. G. [van Dokkum]{}, and P. [Coppi]{}, *ApJ*, **686**, 1503–1513 (2008)
A. L. [Kinney]{} et al. *ApJ*, **467**, 38 (1996).
C. [Papovich]{}, M. [Dickinson]{}, and H. C. [Ferguson]{}, *ApJ*, **559**, 620–653 (2001)
N. A. [Reddy]{}, and C. C. [Steidel]{}, *ApJ*, **692**, 778–803 (2009),
R. J. [Bouwens]{} et al., *ApJ*, **705**, 936–961 (2009)
R. J. [Bouwens]{} et al., *ApJL*, **708**, L69–L73 (2010)
N. A. [Reddy]{}, C. C. [Steidel]{}, D. K. [Erb]{}, A. E. [Shapley]{}, and M. [Pettini]{}, *ApJ* **653**, 1004–1026 (2006)
A. E. [Shapley]{} et al., *ApJ*, **626**, 698–722 (2005)
D. P. [Stark]{} et al., *ApJ*, **697**, 1493–1511 (2009)
E. [Gawiser]{} et al., *ApJL*, **642**, L13–L16 (2006).
N. [Pirzkal]{}, S. [Malhotra]{}, J. E. [Rhoads]{}, and C. [Xu]{}, *ApJ*, **667**, 49–59 (2007).
S. L. [Finkelstein]{}, J. E. [Rhoads]{}, S. [Malhotra]{}, N. [Pirzkal]{}, and J. [Wang]{}, *ApJ*, **660**, 1023–1029 (2007).
S. L. [Finkelstein]{}, J. E. [Rhoads]{}, S. [Malhotra]{}, and N. [Grogin]{}, *ApJ*, **691**, 465–481 (2009).
G. [Bruzual]{}, and S. [Charlot]{}, *MNRAS*, **344**, 1000–1028 (2003).
E. [Gawiser]{} et al., *ApJ*, **671**, 278–284 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Enhancement of the sensitivities of optical magnetometers, atomic clocks and atom interferometers and other quantum metrology devices requires introducing new physical processes to improve on their present achievements. Many body collective correlations among the atoms, spins or, in general, quantum systems may prove to be a suitable method. As these correlations introduce interference terms in the intensity of the scattering amplitudes, they may enhance the signal as $N(N-1)$ for N correlated quantum systems. These correlations enhance the signal to noise ratio by a factor of $N^2$ and contribute to better sensitivity in quantum metrology. Moreover atomic correlation may provide quantum noise limit, Heisenberg limit. In the present communication excitation exchange induced by photons in a cavity between two atoms is calculated and clearly exhibits correlation and collective effects. A novel operator is introduced that expresses photon-induced excitation exchange that takes in account energy conservation, $V_{ij}=\hat{a}^\dag\sigma_i\sigma_j^\dag\hat{a}$, $\sigma_i=\left|g\right\rangle_{i}\left\langle e\right|_{i}$ is lowering operator of *i-th* atom, and $\hat{a}^\dag,\hat{a}$ are photon creation and annihilation operators. Here i and j stand for two atoms. This operator describes real or virtual photon assisted dipole-dipole interaction. Moreover, it conserves the total number of excitations in the joint em field and the quantum system. Experimental challenges are suggested.'
author:
- Reuben Shuker
- 'Gennady A. Koganov'
bibliography:
- 'Koganov.bib'
title: Collective Photon Assisted Dressing of Atomic Levels by the number $N$ of Correlated Atoms
---
Impressive achievements have been obtained in sensitivity of atomic clocks and optical magnetometers that practically arrive at the theoretical limit [@Romalis-subfemto; @Clocks-McGrew; @SERF-Review]. In both the signals are linear in the number $N$ of active atoms. Enhancement of these sensitivities requires introducing new physical processes that may improve on these achievements. Many body collective correlations among the atoms or in general quantum systems, may prove the right method [@Kimble2016; @Ritsch; @Zoller; @photonic; @Javanainen]. When $N$ quantum systems (atoms) interact with electromagnetic field the response intensity of the system is $\left|\sum_{i=1}^{N}{a_i}\right|^2$, where $a_i$ is the response amplitude of a single atom. This can be written as $\sum_{i=1}^{N}{\left|a_i\right|^2}+\sum_{i\neq j}^{N}{\left|a_i a_j\right|}$. The second term stems from pair correlations. If the atoms were statistically independent, this correlation function is equal to zero, and the system’s response intensity is merely equal to response intensity of an individual atom multiplied by $N$, as described by the first term. As these correlations introduce interference terms in the intensity of the amplitudes, they may enhance the signal up to $N^2$ for N correlated quantum systems, in case of constructive interference, as in super-radiance [@Haroche]. In the case of destructive interference the interference term may quench spontaneous emission and quantum noise [@Kimble2016]. Both of these cases, if correlation is obtained, result in better signal to noise ratio and may contribute to better sensitivity in atomic clocks, optical magnetometers, atomic interferometry [@Narducci] and other quantum processes, such as measurements, computing and communications. Atomic correlations may also reduce phase quantum noise and achieve the Heisenberg limit [@Mitchel-subprojection; @HisenLimit].\
In this communication we introduce a physics of generating correlations among quantum systems, for example two atoms, by cavity photonic excitation exchange. We present a case study of $N$ two-levels systems interacting with cavity photons in quest of quantum correlations. To take account for pair atomic correlations we define a collective excitation exchange operator $V$as follows:
$$\label{V1}
\hat{V}=\Omega_c\sum_{i\neq j}{\hat{a}^\dag\sigma_i\sigma_j^\dag\hat{a}},$$
where $\sigma_i=\left|g\right\rangle_{i}\left\langle e\right|_{i}$ is lowering operator of *i-th* atom, and $\hat{a}^\dag$ and $\hat{a}$ are photon creation and annihilation operators, $\Omega_c$ is the strength of direct interatomic interactions. The main virtue of this operator is that it conserves energy and the total number of excitations in the matter and photonic systems, as a whole. Excitation exchange operators, such as $\sigma_i\sigma_j^\dag$ [@antiresonance] or Stokes operator [@Mitchell-2014], are energy non-conserving. This is crucial to make the total Hamiltonian energy conserving.
Our novel operator explicitly presents fully quantum mechanically photon assisted atom correlation while it is absent in other works [@antiresonance; @Mitchell-2014; @superrad-dipole-int]. As will be seen later, we are able to obtain dressing by the number of atoms $N$. Indeed, the operator (\[V1\]) also correctly generates the necessary spin-spin correlation and their collective behavior treated in Refs.[@Mitchell-2014; @Mitchel-subprojection]. In fact, spin-spin correlation is generated due to Fermi-Dirac statistics of spins as presented in [@Mitchel-subprojection; @Mitchell-2014]. However, dressing by the number of atoms $N$ is not generated in these works, in particular in the case where photons are treated semiclassically [@antiresonance; @superrad-dipole-int]. Also, the virtue of atom correlations may provide the possibility of achieving Heisenberg limit [@Mitchel-subprojection; @HisenLimit] of quantum noise $1/N$, rather than standard quantum noise limit of $1/\sqrt{N}$.\
We introduce a general system of $N$ two-level atoms interacting with a resonant single mode of electromagnetic field. The system Hamiltonian in the interaction picture and the rotating wave approximation reads
$$\label{hamilt}
H=\sum_{i=1}^{N}{(g\sigma_i}\hat{a}^\dag+H.c.)+V,$$
where g is a coupling constant, and the collective operator $V$ is defined in Eq. (\[V1\]). The first term in (\[hamilt\]) describes the standard interaction of the atoms with the field, while the second term is responsible for field assisted pair atomic correlations, such as dipole-dipole and spin-spin interaction. It generates collective term proportional to $N(N-1)$ as a result of the interference term and the related correlation.\
In the following we present the collective effect and dependence of eigenvalues and dressed states on the number of atoms $N$. Calculation of the eigenvalues of the Hamiltonian (\[hamilt\]) results in
$$\label{eigen}
E=\frac{1}{2}N[\Omega_c(N-1)n\pm\sqrt{4ng^2+\Omega_c^2n^2(N-1)^2} ],$$
where $n$ is the number of photons inside the cavity. Note that the eigenvalues (\[eigen\]) manifest collective effect, as seen by asymptotic $N^2$ dependence at large values of $N$:
$$\begin{aligned}
E_1\approx \Omega_c n N^2,\label{asymptup}\\
E_2\approx -\frac{g^2}{\Omega_c}+O(\frac{1}{N^2})\label{asymptlow}\end{aligned}$$
Obviously no collective effect when $N=1$. Indeed, in this case the usual photon splitting effect leading to Autler-Towns doublet [@AutlerTownes] is found. For $n=0$ no cavity effect and no splitting. However, even for $n < 1$ cavity and correlation effects are possible [@Thompson]. In Fig. \[eigvalnoCorr\] the eigenvalues (\[eigen\]) are plotted as a function of the number of atoms $N$ in the absence of correlations $\Omega_c=0$. Both upper and lower eigenvalues change with constant step equal $g\sqrt{n}$ and are symmetric (with respect zero).
![(Color online) Upper and lower Eigenvalues of Hamiltonian (\[hamilt\]) without correlation term.[]{data-label="eigvalnoCorr"}](Fig1.eps)
When correlations are present, the eigenvalues change with $N$ in a different way (see Fig. \[eigvalCorr\]): the upper eigenvalue grows with the number of atoms with the rate proportional to $N$, i.e. it scales as $N^2$, whereas the lower eigenvalue does not decrease, as in the previous case, but grows slowly with the rate proportional to $1/N^2$ and asymptotically tends to negative limit value. These can be seen from the asymptotic expressions (\[asymptup\]) and (\[asymptlow\]). Both eigenvalues change with varying stair step, which can also be seen from the asymptotic expressions for the stair step values $\Delta E=E(N+1)-E(N)$ at large $N$.
$$\begin{aligned}
\Delta E^{(+)}=2 \Omega_cnN+\textit{O}\left(\frac{1}{N^2}\right) \label{E+}\\
\Delta E^{(-)}=\frac{g^2}{\Omega_c}\frac{1}{N^2}+\textit{O}\left(\frac{1}{N^3}\right)\label{E-}\end{aligned}$$
As seen from comparison of Figs. \[eigvalnoCorr\] and \[eigvalCorr\] the atomic correlations introduce symmetry breaking.
![(Color online) Upper and lower eigenvalues of Hamiltonian (\[hamilt\]) with correlation term.[]{data-label="eigvalCorr"}](Fig2.eps)
Figures \[eigval-Corr-NoCorr\] and \[eigval-AntiCorr-NoCorr\] show the symmetry breaking for two kinds of correlation, positive correlation with $\Omega_c>0$ (Fig. \[eigval-Corr-NoCorr\]), and negative correlation with $\Omega_c<0$ (Fig. \[eigval-AntiCorr-NoCorr\]). In the first case the upper eigenvalue grows as $N^2$ at large $N$, while the lower one asymptotically approaches negative limit value $-g^2/\Omega_C$, as described by the approximate expressions (\[asymptup\]) and (\[asymptlow\]), respectively.
In the case of negative correlation, $\Omega_c<0$, symmetry is broken in a different way, somewhat opposite to the case of positive correlation, as seen in Fig. \[eigval-AntiCorr-NoCorr\]. Here the upper eigenvalue at large number of atoms, $N>>1$, asymptotically approaches positive limit value $g^2/\left|\Omega_c\right|$, while the lower eigenvalue being negative scales as $N^2$.
![(Color online) Eigenvalues of the Hamiltonian (\[hamilt\]) with positive correlations ($\Omega_c>0$) and without correlations (black lines).[]{data-label="eigval-Corr-NoCorr"}](Fig3.eps)
![(Color online) Eigenvalues of the Hamiltonian (\[hamilt\]) with negative correlations ($\Omega_c<0$) and without (black lines) correlations.[]{data-label="eigval-AntiCorr-NoCorr"}](Fig4.eps)
In Fig. \[eigval-Corr-AntiCorr\] eigenvalues in two cases of positive and negative correlations are presented, elucidating symmetry breaking. This is a kind of restoring symmetry, namely, there is an inversion symmetry between positive and negative correlations.
![(Color online) Eigenvalues of the Hamiltonian (\[hamilt\]) with positive ($\Omega_c>0$) and negative correlations ($\Omega_c<0$).[]{data-label="eigval-Corr-AntiCorr"}](Fig5.eps)
Finally, we present in Fig. \[eigval-per-Atom\] eigenvalues (actually, these are frequencies of the transitions between the dressed states) per single atom. In the case of non-correlated atoms this frequency does not depend on the number of atoms $N$, as expected. When the atoms are correlated, the dressed transition frequency per atom grows linearly with $N$, both at positive and negative correlation.
![(Color online) Dressed states transition frequency (eigenvalues of the Hamiltonian (\[hamilt\])) per single atom as a function of the number of atoms $N$ at positive, negative and no correlation cases.\[eigval-per-Atom\]](Fig6.eps)
In summary, the novel operator describing field-assisted pair correlations among quantum systems, was introduced. A case study of $N$ two-levels systems interacting with cavity photons in quest of quantum correlations has been presented. Calculation of the eigenvalues of the Hamiltonian containing the new term responsible for quantum correlations shows that accounting for such correlations gives rise to breaking the system symmetry. The upper eigenvalue is no longer linear with respect to the number of atoms $N$, as it is when the atoms are independent, but scales as $N^2$ for $N>>1$. It has also been found that when the correlation is negative, the system symmetry breaks in a different way, namely, the lower (negative) eigenvalue of the Hamiltonian scales as $N^2$, while the upper one asymptotically drops to small limit value. The symmetry is restored taking both positive and negative correlation constant.\
Our novel operator explicitly presents in the Hamiltonian quantum mechanically photon assisted atom correlation while it is absent in other works [@antiresonance; @Mitchell-2014; @superrad-dipole-int], thus we are able to obtain dressing by the number of atoms $N$. Other virtue of atom correlations may provide the possibility of achieving Heisenberg limit of quantum noise $1/N$, rather than standard quantum noise limit of $1/\sqrt{N}$.
An experimental test for this theory should start with a cold atoms with anti-reflection coating optical cell inserted into a strong coupling regime cavity. A first attempt setup may involve cold Rb atoms to reduce thermal relaxations and extend $T_2$. We suggest to employ lasing without inversion in a cavity with one bi-chromatic mirror, such that a laser at $D_2(D_1)$ can pass and $D_1(D_2)$ is reflected in the cavity. Monitoring the $D_1$ line intensity vs the number density of the cold atoms by controlling the temperature would provide the test for atoms correlation effect. Successful result may pave the way for atoms correlation. We believe that atoms correlation may prove robust effect even at hot vapor case.
Acknowledgment {#acknowledgment .unnumbered}
==============
We acknowledge the support of Office of Naval Research Grant No. ONRG GRANT- NICOP-N62909-19-1-2030.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work, we construct and release a multi-domain and multi-modality event dataset (MMED), containing 25,165 textual news articles collected from hundreds of news media sites (e.g., Yahoo News, Google News, CNN News.) and 76,516 image posts shared on Flickr social media, which are annotated according to 412 real-world events. The dataset is collected to explore the problem of organizing heterogeneous data contributed by professionals and amateurs in different data domains, and the problem of transferring event knowledge obtained from one data domain to heterogeneous data domain, thus summarizing the data with different contributors. We hope that the release of the MMED dataset can stimulate innovate research on related challenging problems, such as event discovery, cross-modal (event) retrieval, and visual question answering, etc.'
author:
- Zhenguo Yang
- Zehang Lin
- Min Cheng
- Qing Li
- Wenyin Liu
bibliography:
- 'Multimedia.bib'
title: 'MMED: A Multi-domain and Multi-modality Event Dataset'
---
Introduction
============
Real-world events can be defined as something happened or happening, which attract people’s attention and impact our lives. Instances of events can include festivals, politics, natural disasters, etc. In reality, whenever an event happens, it can be witnessed by different groups of people (e.g., professionals and amateurs) and get attention of the public. Therefore, tremendous data related to the same events may be distributed in different data domains (e.g., online news domain and social media domain), due to the convenience of sharing data in the era of Web 2.0. The data distributed on the Internet platforms are rich in data modalities, such as text, image, and video, etc. Multimodal data in multiple domains have brought new opportunities for people and researchers to “sense the world”.
In terms of multi-modality, one of the most popular application is cross-modal retrieval, which aims to bridge the gap between the data modalities. For instance, an image can be a query to match the corresponding texts in a text database, and vice versa. There are quite a few existing datasets, e.g., Wikipedia dataset [@Rasi2010], PASCAL sentences [@Rash2010], NUS-WIDE [@Chua2009], Flickr-30K [@Young2014], MS COCO dataset [@Lin2014], etc. However, the existing datasets are strongly-aligned paired data [@Yang2017a; @Yang2017b; @Yang2019b], i.e., the textual contents are the exact descriptions of their corresponding images. In reality, the multimodal data may be associated with each other by sharing the same labels, yet they are not trying to describe each other exactly, e.g., news articles on a news media like BBC News reporting a real-world event and images shared by social media users on Flickr that are about the same event. In contrast, we denote such cases as weakly-aligned unpaired data [@Yang2019b]. The datasets used for cross-modal retrieval are collected from one data domain usually, which do not take into account the cross-domain relationships. In addition, the datasets usually have a fix number of object categories, which is not as complex as real-world events.
{width="480pt"}
In terms of events, there exist quite a few works and some datasets for real-world event detection, e.g., event detection from traditional news media [@Allan98], Twitter-like social media [@Sakaki2010], and Flickr-like photo-sharing social media [@Papado2011; @Yang2017a; @Yang2017b], etc. However, these datasets about real-world events involve one data domain merely. In reality, an influential event happens, the related data may be distributed on different Internet platforms by both amateur users and professional journalist. The rich data provide multiple viewpoints about the real-world event, benefiting to the comprehensive understanding of the events. The distributed data can be complementary or even sometimes contradictory. The single domain data losses the possibility of exploring real-world event related issues crossing data domains.
In this paper, we present a large-scale event dataset collected from multiple data domains in multimodal forms, denoted as MMED, including 25,165 textual news articles from hundreds of news media sites, and 76,516 images from Flickr-like photo-sharing social media. The two-domain data samples are related to 412 real-world events, covering various event categories, such as public security (e.g., shooting, attack, killing, conflict, explosion, epidemic, crash, etc.), natural disaster (e.g. earthquake, flood, fire, etc.), protest, sport, election, festival, etc. Compared to the existing datasets, MMED has several characteristics. Firstly, it is a weakly-aligned unpaired dataset, where the data samples are not trying to describe with each other, which is unlike the strongly-aligned datasets being used for cross-modal retrieval as mentioned previously [@Yang2019b]. Secondly, the multimodal data describing real-world event concepts crossing different data domains, and contributed by both amateur users and professional journalists, covering both folk and official viewpoints about events. The dataset has been released on GitHub (https://github.com/zhengyang5/ACM-MMSys19-MMED400). Hopefully, it can promote the research on related applications.
The rest of the paper is organized as follows. Section 2 reviews the principle of data collecting and data annotations. Section 3 introduces the statistics of the dataset. Section 4 sheds light on the possible applications based on the dataset. Section 5 shows the concluding remarks.
Collection and Annotation
=========================
The event labels in MMED cover a wide range of event categories (or event types) like emergency, natural disaster, sport, ceremony, election, protest, military intervention, economic crisis, etc. We collect the dataset considering the aspects of high relevance in supporting the application needs, wide range of event types, non-ambiguity of the event labels, imbalance of the event clusters, and difficulty of discriminating the event labels, etc.
{width="430pt"}
To collect textual news articles, there are two challenging problems. The first problem is the dispersibility of the data. For instance, news articles reporting the same events may be reported by any news media sites in the world, thus it is laborious and even impossible to collect the data by accessing each news media sites for annotation. In addition, most online news media sites do not provide effective and convenient interface for retrieval. The second problem is the quality of the data, such as authority, credibility, etc. We have to ensure the news articles are publicly-accepted, and avoid rumors, fake news, or inaccurate news stories as much as possible. Therefore, instead of collecting the data directly from the individual news media sites, we resort to Wikipedia and collect the data in a top-down manner. More specifically, we manually look up the Wikipedia entries about events recently, and check the crowd-sourced articles describing the events that are available online, which have cited quite a few news articles in the references as shown in Figure 1. These articles are contributed by different news media sites, and usually focus on different aspects or hold different views on the events, but are quite high in terms of quality and authority. Furthermore, we crawl these news articles further by accessing the links of the articles in different news media sites. As a result, the name of the Wikipedia entry about the event (i.e., event label) are well-accepted by the public as they are edited in a crowdsourcing manner, and there is no ambiguity in terms of event labels. In addition, Wikipedia provides quite a few portals to access similar events in terms of event types, regions, etc., which helps to collect similar events in certain aspects to ensure the difficulty of discriminating the event labels.
-------- --------- ---------- ---------- -------------- ----------- -------------- ---------- -------------------
Domain \#Items \#Events Time (%) Location (%) Title (%) Data source/ Tags (%) Textual content /
Username (%) Description (%)
$X$ 25,165 412 97.06 n.a. 100 100 n.a. 100
$Y$ 76,516 412 100 16.35 100 100 81.53 100
-------- --------- ---------- ---------- -------------- ----------- -------------- ---------- -------------------
For the image posts on social media, we take Flickr as an example for collecting data. Given the event names collected from Wikipedia, we retrieve the related data by using different queries (i.e., keywords) and some strategies like filtering by time to obtain a number of returned results and remove the replicated ones. On one hand, we annotate the image posts manually to verify whether the samples are related to the query events, and crawl the responding data samples. On the other hand, we further access the Flickr albums (as shown in Figure 2) of these image posts. Whenever the topics of the albums are in accord with the events, we will craw them further. The event labels of the Flickr image posts are consistent with the ones obtained from Wikipedia to ensure non-ambiguity and difficulty of differentiating.
The MMED Dataset
================
Statistics of the MMED Dataset
------------------------------
Specifically, 25,165 textual news articles are collected from hundreds of data sources in the online news media domain, including Yahoo News, Google News, Huffington Post, CNN News, New York Times, NBC News, Fox News, Washington Post, The Guardian, etc. A number of 76,516 Flickr image posts are included, which are shared by 4,473 Flickr social media users. All the data samples are related to 412 real-world events. The number of data samples corresponding to the event labels are summarized in Figure 3.
For each textual news article, in addition to its *textual context*, *time reporting*, *title*, *data source* indicating which news media it comes from, are also available.
For a Flickr image, in addition to the *photo*, *time-taken*, *title*, *GPS-tags*, *tags*, *textual description*, *username*, have been provided. In summary, the statistics of the feature modalities possessed by the data in the dataset are summarized in Table 1.
Event Labels of the MMED Dataset
--------------------------------
The real-world events in the dataset can be roughly summarized by several categories, i.e., public security, natural disaster, protest, sport, election, festival, etc. The following Tables summarize the names of event labels and the number of data samples in each domain for the six categories. In particular, there is a Wikipedia entry for the name of each event label, and more details about the events can be obtained by directly accessing these entries in Wikipedia.
------------------------------------- -- --
**[Event Label]{}&&\
2014 Isla Vista killings&81&143\
2015 Copenhagen shootings&131&100\
2015 San Bernardino attack&269&272\
2015 Bangkok bombing&103&68\
2015 Tianjin explosions&58&125\
...&...&...\
**
------------------------------------- -- --
: Examples of public security events in the dataset. Note that the numbers indicate the number of data samples in each domain.
------------------------------------- -- --
**[Event Label]{}&&\
2013 Bohol earthquake&64&189\
2015 Nepal blockade&51&58\
2015 Sabah earthquake&52&399\
April 2015 Nepal earthquake&91&850\
Cyclone Oswald&75&254\
...&...&...\
**
------------------------------------- -- --
: Examples of natural disaster events in the dataset.
----------------------------------- -- --
**[Event Label]{}&&\
2013 Stanley Cup playoffs&62&130\
2013 World Series&94&200\
2014 24 Hours of Le Mans&164&113\
2015 24 Hours of Le Mans&48&107\
2015 MotoGP season&58&97\
...&...&...\
**
----------------------------------- -- --
: Examples of sport events in the dataset.
---------------------------------------- -- --
**[Event Label]{}&&\
2013 Shahbag protests&106&83\
2014 Hong Kong protests&46&64\
Belfast City Hall flag protests&63&73\
Boko Haram insurgency&63&80\
Euromaidan&105&100\
...&...&...\
**
---------------------------------------- -- --
: Examples of protest events in the dataset.
---------------------------------------------- -- --
**[Event Label]{}&&\
Australian federal election, 2013&69&117\
Canadian federal election, 2015&66&369\
Egyptian presidential election, 2014&72&112\
European Parliament election, 2014&112&122\
Indian general election, 2014&190&125\
...&...&...\
**
---------------------------------------------- -- --
: Examples of election events in the dataset.
------------------------------------ -- --
**[Event Label]{}&&\
2014 Sundance Film Festival&75&52\
2015 Tour of California&89&399\
85th Academy Awards&342&1154\
86th Academy Awards&159&67\
Miss USA 2015&214&184\
...&...&...\
**
------------------------------------ -- --
: Examples of festival events in the dataset.
Application Scenarios and Evaluations
=====================================
{width="250pt"}
Event discovery
---------------
Event discovery [@Yang2017a; @Yang2017b; @Yang2019a] aims to detect and organize the data distributed on the Internet platforms based on the real-world events they depict. The collected dataset supports event detection from multiple data domains, such as social media sites and news media sites. The data are in multiple modalities, such as images and texts, and can be contributed by different uses, such as amateur users and professional journalists. In addition, it can be used for event discovery in a transfer learning or zero-shot learning manner, i.e., there are no event labels in the target domain that are available.
In terms of evaluations, F1-measure NMI can be adopted as the event labels are well-defined.
Cross-modal (event) retrieval
------------------------------
The scenario of cross-modal retrieval is that using a text query to retrieve the related images. Similarly, we extend the scenario to cross-modal event retrieval [@Yang2019b] with two unique characteristics. Firstly, the data are weakly-aligned and unpaired, where the multimodal data are not trying to describe with each other while both are describing high-level semantics like events. In contrast, the existing datasets as mentioned previously are all strongly-aligned and paired data.
In term of the evaluations, MAP, Precision, Recall and other widely-used metrics in cross-modal retrieval can be adopted similarly.
Visual Question Answering
-------------------------
A VQA system takes as input an image and a free-form, open-ended, natural-language question about the image and produces a natural-language answer as the output [@Antoal]. The collected dataset can support the VQA related applications, especially for real-world events. Figure 4 gives an example of questions on a Flickr image, whiles the answers can be found from news media articles. Actually, cross-modal event retrieval tries to answer the first question in the example.
In terms of evaluations, the performance depends on the distance or similarity between the answers of the methods and the groundtruth, which can be measured by BLEU, ROUGE, METEOR, etc. In addition, user studies on Amazon Mechanical Turk platform can be adopted for evaluations.
Conclusion
==========
In this paper, we have released an event dataset from social media and news media platforms, denoted as MMED. MMED can be used for three scenarios at least, i.e., event discovery, cross model (event) retrieval, and visual question answering, etc. The real-world dataset is well-labeled, can be used for other scenarios with an open mind.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the automorphism group of the disk complex is isomorphic to the handlebody group. Using this, we prove that the outer automorphism group of the handlebody group is trivial.'
address:
- |
- Department of Mathematics\
Middle East Technical University\
06531 Ankara, Turkey
- |
- Department of Mathematics\
University of Warwick\
Coventry, CV4 7AL, UK
author:
- Mustafa Korkmaz
- Saul Schleimer
bibliography:
- 'bibfile.bib'
title: Automorphisms of the disk complex
---
[^1]
Introduction
============
We show that the automorphism group of the disk complex is isomorphic to the handlebody group. Using this, we prove that the outer automorphism group of the handlebody group is trivial. These results and many of the details of the proof are inspired by Ivanov’s work [@Ivanov97] on the mapping class group and the curve complex.
Let $V = V_{g,n}$ be the genus $g$ handlebody with $n$ [*spots*]{}: a regular neighborhood of a finite, polygonal, connected graph in $\RR^3$ with $n$ disjoint disks chosen on the boundary. See for a picture of $V_{2,2}$. We write $V = V_g$ when $n = 0$. Let $\bdy_0 V$ denote the union of the spots. Let $\bdy_+ V$ be the closure of $\bdy V \setminus \bdy_0 V$. So $\bdy_+
V \homeo S = S_{g,n}$ is a compact connected orientable surface of genus $g$ with $n$ boundary components. We write $S = S_g$ when $n =
0$. Define $e(V) = -\chi(\bdy_+ V) = 2g - 2 + n$.
$$\begin{array}{c}
\epsfig{file=twoholedgenustwo.eps, height = 3.5 cm}
\end{array}$$
A simple closed curve $\alpha$ in $S = S_{g,n}$ is [*inessential*]{} if it cuts a disk off of $S$; otherwise $\alpha$ is [*essential*]{}. The curve $\alpha$ is [*peripheral*]{} if it cuts an annulus off of $S$; otherwise $\alpha$ is [*non-peripheral*]{}. A properly embedded disk $D$ in $V = V_{g,n}$, with $\bdy D \subset \bdy_+ V$, is [*essential*]{} or [*non-peripheral*]{} exactly as its boundary is in $\bdy_+ V$. We require any proper isotopy of $D \subset V$ to have track disjoint from the spots of $V$. This yields a proper isotopy of $\bdy D$ in $\bdy_+ V$.
\[Def:CurveComplex\] The [*curve complex*]{} $\calC(S)$ is the simplicial complex with vertex set being isotopy classes of essential, non-peripheral curves in $S$. The $k$–simplices are given by collections of $k+1$ vertices having pairwise disjoint representatives.
\[Def:DiskComplex\] The [*disk complex*]{} $\calD(V)$ is the simplicial complex with vertex set being proper isotopy classes of essential, non-peripherial disks in $V$. The $k$–simplices are given by collections of $k+1$ vertices having pairwise disjoint representatives.
Note that there is a natural inclusion $\calD(V) \to \calC(\bdy_+ V)$ taking a disk to its boundary. This map is simplicial and injective.
If $\calK$ is a simplicial complex then $\Aut(\calK)$ denotes the group of simplicial automorphisms of $\calK$. The elements of $\Aut(\calC(S))$ and $\Aut(\calD(V))$ are to be contrasted with mapping classes on the underlying spaces.
\[Def:MCG\] The [*mapping class group*]{} $\MCG(S)$ is the group of homeomorphisms of $S$, up to isotopy. The [*handlebody group*]{} $\calH(V)$ is the group of homeomorphisms of $V$, fixing the spots setwise, up to spot preserving isotopy.
Some authors refer to our $\MCG(S)$ as the [*extended*]{} mapping class group, as orientation reversing homeomorphisms are allowed. Note there is a natural map $\calH(V) \to \MCG(\bdy_+ V)$ which takes $f \in \calH(V)$ to $f|\bdy_+ V$. Again, this map is an injective homomorphism.
Note finally that there is a natural homomorphism $\MCG(S) \to
\Aut(\calC(S))$ (and similarly for $V$). We will call any element of the image of this map a [*geometric automorphism*]{}. Our main theorem is:
[Theorem]{}[Thm:AutOfDiskComplex]{} If a handlebody $V = V_{g,n}$ satisfies $e(V) \geq 3$ then the natural map $\calH(V) \to \Aut(\calD(V))$ is a surjection.
In the language above: every element of $\calH(V)$ is geometric. The plan of the proof of is given in and completed in . shows that is sharp; all handlebodies $V$ with $e(V) \leq 2$ exhibit some kind of exceptional behaviour. has a corollary:
[Theorem]{}[Thm:AutOfDiskComplexII]{} If a handlebody $V = V_{g,n}$ satisfies $e(V) \geq 3$ then the natural map $\calH(V) \to \Aut(\calD(V))$ is an isomorphism.
In we use to prove:
[Theorem]{}[Thm:AutOfHandlebodyGroup]{} If $e(V) \geq 3$ then the outer automorphism group of the handlebody group is trivial.
These results are inspired by work of Ivanov, Korkmaz and Luo [@Ivanov97; @Korkmaz99; @Luo00]:
\[Thm:Ivanov\] If $3g - 3 + n \geq 3$, or if $(g, n) = (0, 5)$, then all elements of $\Aut(\calC(S_{g,n}))$ are geometric. Also, the outer automorphism group of $\MCG(S)$ is trivial.
Background {#Sec:Background}
==========
The genus zero case of is contained in the thesis of the first author [@Korkmaz99 Theorem 1].
\[Thm:Korkmaz\] If $g = 0$ and $n \geq 5$ then all elements of $\Aut(\calC(S_{0,n}))$ are geometric.
Spotted balls are the simplest handlebodies. Accordingly:
\[Lem:SpottedBalls\] The natural maps $\calD(V_{0,n}) \to \calC(S_{0,n})$ and $\calH(V_{0,n}) \to \MCG(S_{0,n})$ are isomorphisms.
The three-manifold $V_{0,n}$ is an $n$–spotted ball. Every simple closed curve in $\bdy_+ V$ bounds a disk in $V$. This proves that $\calD(V_{0,n}) \to \calC(S_{0,n})$ is a surjection and thus, by the remark immediately after , an isomorphism.
It follows from the Alexander trick that the inclusion of mapping class groups is an isomorphism.
The genus zero case of is an immediate corollary. We now give basic definitions.
Suppose that $V$ is a handlebody. Two disks $D, E \in \calD(V)$ are [*topologically equivalent*]{} if there is a mapping class $f \in
\calH(V)$ so that $f(D) = E$. The [*topological type*]{} of $D$ is its equivalence class in $\calD(V)$.
For any simplicial complex, $\calK$, if $\sigma \in \calK$ is a simplex then recall that $$\link(\sigma) = \{ \tau \in \calK \st \sigma \cap \tau =
\emptyset,~\sigma \cup \tau \in \calK \}.$$ So if $\DD$ is a simplex of $\calD(V)$ then $\link(\DD)$ is the subcomplex of $\calD(V)$ spanned by disks $E$ disjoint from some $D
\in \DD$ and distinct from all $D \in \DD$.
If $X \subset Y$ is a properly embedded submanifold then we write $\neigh(X)$ and $\Neigh(X)$ to denote open and closed regular neighborhoods of $X$ in $Y$. If $X$ is codimension zero then the [*frontier*]{} of $X$ in $Y$ is the closure of $\bdy X \setminus \bdy Y$.
A simplex $\DD \in \calD(V)$ is a [*cut system*]{} if $V \setminus
\neigh(\DD)$ is a spotted ball. Note that every disk of $\DD$ yields two spots of $V \setminus \neigh(\DD)$.
Recall that for simple curves $\alpha, \beta$ properly embedded in $S$ the [*geometric intersection number*]{} $i(\alpha, \beta)$ is the minimum possible intersection number between proper isotopy representatives.
Two disks $D, E \in \calD(V)$ are [*dual*]{} if $i(\bdy D, \bdy E) =
2$; equivalently, after a suitable proper isotopy $D$ and $E$ intersect along a single arc; equivalently, after a suitable proper isotopy a regular neighborhood of $D \cup E$ is a four-spotted ball with all spots essential in $V$. See .
$$\begin{array}{c}
\epsfig{file=dual.eps, height = 3.5 cm}
\end{array}$$
If $\DD = \{ D_i \}$ is a cut system we define $\dual_i(\DD)$ to be the subcomplex spanned by the disks $E \in \calD(V)$ which are dual to $D_i$ and disjoint from $D_j$ for all $j \neq i$. We take $\dual(\DD)$ to be the complex spanned by $\cup_i \dual_i(\DD)$.
The proof of {#Sec:Sketch}
=============
Let $V = V_{g,n}$ be a genus $g$ handlebody with $n$ spots. We suppose that $g \geq 1$ and $e(V) \geq 3$. Let $\phi$ be any automorphism of $\calD(V)$. proves that $\phi$ preserves the topological types of disks. In addition, $\phi$ sends cut systems to cut systems (). Next shows that $\phi$ preserves duality. Also, for any cut system $\DD = \{ D_i
\}$, the complex $\dual_i(\DD)$ is connected ().
Pick any geometric automorphism $f_{{\operatorname{cut}}}$ so that $f_{{\operatorname{cut}}}(\DD) =
\phi(\DD)$, vertex-wise; $f_{{\operatorname{cut}}}$ exists by . Define $\phi_{{\operatorname{cut}}} = f^{-1}_{{\operatorname{cut}}} \circ \phi$. Thus $$\phi_{{\operatorname{cut}}} | \DD = \Id.$$
Let $V' \isom V_{0, 2g + n}$ be the spotted ball obtained by cutting $V$ along a regular neighborhood of $\DD$. Now, since $\phi_{{\operatorname{cut}}}$ preserves $\link(\DD) \isom \calD(V')$, by and there is a homeomorphism $f \from V' \to V'$ so that the induced automorphism $f \in \Aut(\calD(V'))$ satisfies $f =
\phi_{{\operatorname{cut}}} | \link(\DD)$. proves that $f$ preserves the $g$ pairs of spots of $V'$ coming from $\DD$. Thus $f$ can be glued to give a homeomorphism $f_{\link} \from V \to V$ as well as an induced geometric automorphism $f_{\link} \in \Aut(\calD(V))$. Define $\phi_{\link} = f^{-1}_{\link} \circ \phi_{{\operatorname{cut}}}$. Thus $$\phi_{\link} | \DD \cup \link(\DD) = \Id.$$
Recall that $\phi_{\link}$ preserves duals by . For every $D_i \in \DD$ pick some dual $E_i \in \dual_i(\DD)$. By there is an integer $m_i \in \ZZ$ so that $T_i^{m_i}(E_i) = \phi_{\link}(E_i)$, where $T_i$ is the Dehn twist about $D_i$. Define $f_{\dual} = \prod T_i^{m_i}$ and define $\phi_{\dual} = f_{\dual}^{-1} \circ \phi_{\link}$. Letting $\EE = \{
E_i \}$ we have $$\phi_{\dual} | \DD \cup \link(\DD) \cup \EE = \Id.$$
Recall that proves that $\dual_i(\DD)$ is connected. Therefore, a [*crawling*]{} argument, given in , proves that $$\phi_{\dual} | \DD \cup \link(\DD) \cup \dual(\DD) = \Id.$$ Wajnryb [@Wajnryb98] proves that the [*cut system complex*]{} is connected. Thus we may likewise crawl through $\calD(V)$ and prove () that $$\phi_{\dual} = \Id$$ and so prove that $$\phi = f_{{\operatorname{cut}}} \circ f_{\link} \circ f_{\dual}.$$ Thus $\phi$ is geometric.
Small handlebodies {#Sec:Small}
==================
In this section we deal with the small cases, where $e(V) =
2g - 2 + n \leq 2$. We start with genus zero. If $n \leq 3$ then $\calD(V_{0,n})$ is empty. By the mapping class groups of $V$ and $\bdy_+ V$ are equal. Thus $$\calH(V_{0}),~\calH(V_{0, 1}) \isom \ZZ/2\ZZ$$ while $$\calH(V_{0, 2}) \isom K_4 \quad \rm{and} \quad \calH(V_{0, 3})
\isom \ZZ/2\ZZ \cross \Sigma_3.$$ Here $K_4$ is the Klein four-group and $\Sigma_3$ is the symmetric group on three objects [@RafiSchleimer08 Appendix A].
If $n = 4$ then $\calD$ is a countable collection of vertices with no higher dimensional simplices. Thus $\Aut(\calD) = \Sigma_\infty$ is uncountable. However, there are only countably many geometric automorphisms. In fact, by , the mapping class group $\calH(V_{0,4})$ is isomorphic to $K_4 \rtimes \PGL(2,
\ZZ)$ [@RafiSchleimer08 Appendix A].
For genus one, if $n = 0$ or $1$ then $\calD$ is a single point and $\Aut(\calD)$ is trivial. On the other hand $$\calH(V_{1}),~\calH(V_{1, 1}) \isom \ZZ \rtimes K_4.$$ For $V = V_{1,2}$ matters are more subtle. The subcomplex ${\operatorname{NonSep}}(V)
\subset \calD(V)$, spanned by non-separating disks, is a copy of the Bass-Serre tree for the meridian curve in $S_{1,1} = \bdy_+
V_{1,1}$ [@KentEtAl06]. Thus ${\operatorname{NonSep}}(V)$ is a copy of $T_\infty$: the regular tree with countably infinite valance. Now, if $E \in \calD(V)$ is separating then there is a unique disk $D$ disjoint from $E$; also, $D$ is necessarily non-separating. It follows that $\calD(V)$ is a copy of ${\operatorname{NonSep}}(V)$ with countably many leaves attached to every vertex. Thus $\Aut(\calD)$ contains a copy of $\Aut(T_\infty)$ as well as countably many copies of $\Sigma_\infty$ and is therefore uncountable. As usual $\calH(V)$ is countable and so $\Aut(\calD)$ contains non-geometric elements. However, following Luo’s treatment of $\calC(S)$ [@Luo00] suggests the following problem:
\[Prob:Dual\] Suppose that $V = V_{1,2}$. Let $\calG$ be the subgroup of $\Aut(\calD(V))$ consisting of automorphisms preserving duality: if $\phi \in \calH$ and $D, E$ are dual then so are $\phi(D), \phi(E)$. Is every element of $\calG$ geometric?
Note that this approach of recording duality is precisely correct for the four-spotted ball; the complex where simplices record duality in $V_{0,4}$ is the [*Farey tessellation*]{}, $\calF$, and every element of $\Aut(\calF)$ is geometric. See [@Luo00 Section 3.2]. The last exceptional case is $V = V_2$. Let ${\operatorname{NonSep}}(V)$ be the subcomplex of $\calD(V)$ spanned by non-separating disks. Then ${\operatorname{NonSep}}(V)$ is an increasing union, as follows: $\calN_0$ is a single triangle, $\calN_{i+1}$ is obtained by attaching (to every free edge of $\calN_i$) a countable collection of triangles, and ${\operatorname{NonSep}}(V)$ is the increasing union of the $\calN_i$. A careful discussion of ${\operatorname{NonSep}}(V)$ is given by Cho and McCullough [@ChoMcCullough06 Section 4] We obtain $\calD(V)$ by attaching a countable collection of triangles to every edge of ${\operatorname{NonSep}}(V)$. To see this note that every separating disk $E$ divides $V$ into two copies of $V_{1,1}$. These copies of $V_{1,1}$ have meridian disks, say $D$ and $D'$. Thus $\link(E)$ is an edge and the triangle $\{ E, D, D' \}$ has two free edges in $\calD(V)$, as indicated. Finally, there is a countable collection of separating disks lying in $V \setminus (D \cup D')$, again as indicated.
It follows that $\Aut(\calD(V_2))$ is uncountable. Again, as in , we may ask: are all “duality-respecting” elements $f
\in \Aut(\calD(V_2))$ geometric? We end with another open problem:
\[Prob:NonSep\] Suppose that $V$ is a handlebody with $e(V)$ and genus both sufficiently large. Show that $\Aut({\operatorname{NonSep}}(V)) = \calH(V)$.
A solution to may lead to a simplified proof of .
Topological types {#Sec:Types}
=================
The goal of this section is:
\[Lem:Types\] Suppose that $\phi \in \Aut(\calD(V))$. Then $\phi$ preserves topological types of disks.
The [*complexity*]{} of $V_{g,n}$ is $\xi(V) = 3g - 3 + n$. If $\xi(V) \geq 1$ then $\xi(V)$ is the number of vertices of a maximal simplex of $\calD(V)$. Note that $V_1$, $V_{1,1}$ and $V_{0,4}$ are the only handlebodies where $\calD(V)$ has dimension zero. (When $\calD(V)$ is empty its dimension is $-1$.) Further $V_1$ and $V_{1,1}$ are the only handlebodies where $\calD(V)$ is a single point. We will call $V_{0,3}$, the three-spotted ball, a [*solid pair of pants*]{}. Thus $\xi(V)$ is the number of disks in a [*pants decomposition*]{} of $V$ while $e(V) = 2g - 2 + n$ is the number of solid pants in the decomposition. We will call $V_{1,1}$ a [*solid handle*]{}. Suppose now that $E$ is separating with $V
\setminus \neigh(D) = X \cup Y$. If $X$ or $Y$ is a solid pants then we call $E$ a [*pants disk*]{}. If $X$ or $Y$ is a solid handle then we call $E$ a [*handle disk*]{}.
Recall that if $\calK$ and $\calL$ are non-empty simplicial complexes with disjoint vertex sets then $\calK {\mathbin{\vee}}\calL$, their [*join*]{}, is the complex $$\calK \cup \{ \sigma \cup \tau \st \sigma \in \calK,~\tau \in \calL \}
\cup \calL.$$
\[Clm:NotJoin\] For any handlebody $V$ the complex $\calD(V)$ is not a join.
When $e(V) \leq 2$ this can be checked case-by-case, following . The remaining handlebodies all admit disks $D, E$ that [*fill*]{}: every disk $F$ meets at least one of $D$ or $E$. It follows that any edge-path in $\calD^{(1)}(V)$ connecting $D$ to $E$ has length at least three. However, the diameter of the one-skeleton of a join is either one or two.
The complex $\calD(V)$ is [*flag*]{}: minimal non-faces have dimension one.
Observe that $\phi$ preserves the combinatorics of $\calD(V)$. Thus any topological property of $V$ that has a combinatorial characterization will be preserved by $\phi$. We proceed with a sequence of claims.
\[Clm:Join\] The disk $E$ is a separating disk yet not a pants disk if and only if $\link(E)$ is a join. Furthermore, in this case $\link(E)$ is realized as a join in exactly one way, up to permuting the factors.
Suppose that $V \setminus \neigh(E) = X \cup Y$, where neither $X$ nor $Y$ is a solid pants. Since $E$ is essential and non-peripheral both $\calD(X)$ and $\calD(Y)$ are non-empty. It follows that $\link(E) =
\calD(X) {\mathbin{\vee}}\calD(Y)$, and neither factor is empty. Furthermore, this join is realized uniquely, because $\calD(X)$ is never itself a join (by ), $\calD(X)$ is flag and join is associative.
On the other hand, if $E$ is non-separating then $\link(E)$ is isomorphic to $\calD(V_{g - 1, n + 2})$. If $E$ is a pants disk then $\link(E) \isom \calD(V_{g, n - 1})$. Neither of these is a join by .
A [*cone*]{} is the join of a point with some non-empty simplicial complex.
\[Clm:Handle\] Suppose that $V \neq V_{1,2}$. Then $E \in \calD(V)$ is a handle disk if and only if $\link(E)$ is a cone.
Suppose that $E$ cuts off a solid handle $X$ with meridian $D$. Let $Y$ be the other component of $V \setminus \neigh(E)$. Since $V \neq
V_{1,2}$ we have that $\calD(Y)$ is non-empty; in particular $E$ is not a pants disk. By we have $\link(E) = \calD(X) {\mathbin{\vee}}\calD(Y)$. As $\calD(X) = \{ D \}$ we are done with the forward direction.
Now suppose that $\link(E)$ is a cone from $D$. Since a cone is the join of the apex with the base, by the disk $E$ is separating. Let $V \setminus \neigh(E) = X \cup Y$. Thus $\link(E) =
\calD(X) {\mathbin{\vee}}\calD(Y)$. However, by the decomposition of $\link(E)$ is unique; breaking symmetry we may assume that $\calD(X) = \{ D \}$. Thus $X$ is a solid handle and we are done.
It immediately follows that:
\[Clm:NonSep\] Suppose that $V \neq V_{1,2}$. Then $D \in \calD(V)$ is non-separating if and only if there is an $E \in \calD(V)$ so that $\link(E)$ is a cone with apex $D$.
\[Clm:Cut\] Suppose that $e(V) \geq 3$. A simplex $\DD \in \calD(V)$ is a cut system if and only if the following properties hold:
- for every pair of disks $D, E \in \link(\DD)$ the complex $\link(E)
\cap \link(\DD)$ is not a cone with apex $D$ and
- for every proper subset $\sigma \subsetneq \DD$ there is a pair of disks $D, E \in \link(\sigma)$ so that the complex $\link(E) \cap
\link(\sigma)$ is a cone with apex $D$.
The forward direction follows from and the definition of a cut system. (When $V$ is a spotted ball the only cut system is the empty set; the empty set has no proper subsets.)
Now for the backwards direction: From the first property and by deduce that $V' = V \setminus \neigh(\DD)$ is a collection of spotted balls. If $V'$ has at least two components then there is a proper subset $\sigma \subset \DD$ which is a cut system for $V$. Thus $V \setminus \neigh(\sigma)$ is a spotted ball and this contradicts the second property.
\[Lem:Characterize\] Suppose that $V, W$ are handlebodies with $\calD(V) \isom \calD(W)$. Then either:
- $V \homeo W$ or
- $V, W \in \{ V_1, V_{1,1} \}$ or
- $V, W \in \{ V_0, V_{0,1}, V_{0,2}, V_{0,3} \}$.
This is the handlebody version of [@Korkmaz99 Lemma 4.5] and [@Luo00 Lemma 2.1].
When $e(V) \leq 2$ this can be checked case-by-case, following . When $V$ has $e(V) \geq 3$ then $\xi(V) = \xi(W)$. By the handlebodies $V$ and $W$ have cut systems of the same size. It follows that $V, W$ have the same genus and thus the same number of spots.
We now have:
Let $V = V_{g,n}$ and fix $\phi \in \Aut(\calD(V))$. When $e(V) \leq
2$, can be checked case-by-case, following . So suppose that $e(V) \geq 3$.
The automorphism $\phi$ must preserve the set of non-separating disks by .
Suppose that $E \in \calD(V)$ is a separating disk yet not a pants disk. Writing $V \setminus \neigh(E) = X \cup Y$ we have $\link(E) =
\calD(X) {\mathbin{\vee}}\calD(Y)$. By this join is realized uniquely and so we can recover $\calD(X)$ and $\calD(Y)$. By we may deduce, combinatorially, the genus and number of spots of $X$ and $Y$. Thus $\phi$ preserves the topological type of $E$.
The only topological type remaining is the set of pants disks. Since all other types are preserved, so are the pants disks. We are done.
Regluing {#Sec:Reglue}
========
Suppose that $\phi_\DD \in \Aut(\calD(V))$ fixes $\DD$. By there is a homeomorphism $f$ of $V' = V
\setminus \neigh(\DD)$ so that the induced geometric automorphism equals $\phi_\DD|\link(\DD)$. We must show that $f$ gives a homeomorphism of $V$: that is, for every $i$ the spots $D_i^\pm$ are preserved by $f$.
Let ${\operatorname{handle}}_i(\DD) \subset \link(\DD)$ be the collection of handle disks $E \in \calD(V)$ such that
- one component of $V \setminus \neigh(E)$ is a solid handle containing $D_i$ and
- $E$ is disjoint from all of the $D_j$.
Let ${\operatorname{pants}}_i(\DD) \subset \calD(V')$ be the collection of pants disks $E$ such that one component of $V' \setminus \neigh(E)$ is a solid pants meeting the spots $D_i^\pm$.
By the claims in the previous section the set ${\operatorname{handle}}_i(\DD)$ is, for all $i$, combinatorially characterized and so preserved by $\phi_{\DD}$. It follows that the homeomorphism $f \in \Homeo(V')$ preserves the set ${\operatorname{pants}}_i(\DD)$, for all $i$. Now, suppose that $f(D_1^+), f(D_1^-) = A, B$ where $A, B$ are spots of $V'$. Let $E
\in {\operatorname{pants}}_1(\DD)$ be any pants disk. Then $f(E)$ is a pants disk cutting off $A$ and $B$. It follows that the spots $A, B$ (in some order) equal the spots $D_1^\pm$ as desired.
Duality {#Sec:Dual}
=======
Recall that two disks $D, E \in \calD(V)$ are [*dual*]{} if $i(\bdy D,
\bdy E) = 2$ (see ). A [*pentagon*]{} $P \subset
\calD(V_{0,5})$ is a collection of five disks $P = \{ E_i \}_{i =
0}^4$ so that $E_i$ and $E_{i+1}$ are disjoint, for all $i$ (modulo five). We say that the disks $E_i, E_{i+2}$ are [*non-adjacent*]{} in $P$, for all $i$ (modulo five).
\[Lem:Pentagon\] Suppose that $V = V_{0,5}$. Two disks $D, E \in \calD(V)$ are dual if and only if there is a pentagon $P$ so that $D, E \in P$ and $D, E$ are non-adjacent in $P$.
Recall that $\calD(V_{0,5}) \isom \calC(S_{0,5})$, by . The [*pentagon lemma*]{} for $S_{0,5}$ (see [@Korkmaz99 Theorem 3.2] or [@Luo00 Lemma 4.2]) implies that there is only one pentagon in $\calD(V_{0,5})$, up to the action of the handlebody group.
\[Lem:Dual\] Suppose that $V = V_{g,n}$ has $e(V) \geq 3$. Two disks $D, E \in \calD(V)$ are dual if and only if there is a simplex $\sigma \in \calD(V)$ with
- $\link(\sigma) \isom \calD(V_{0,5})$,
- $D, E$ are non-adjacent in some pentagon of $\link(\sigma)$.
It follows that every $\phi \in \Aut(\calD(V))$ preserves duality. We will say that a handlebody $W \subset V$ is [*cleanly embedded*]{} if:
- all spots of $W$ are essential in $V$ and
- if a spot of $W$ is peripheral in $V$ then it is also a spot of $V$.
Suppose that $D, E$ are dual. Let $X$ be the four-spotted ball containing them. Isotope $X$ to be cleanly embedded. Let $\EE$ be a pants decomposition of $V' = V \setminus \neigh(X)$. Now, there is at least one solid pants $P$ in $V' \setminus \neigh(\EE)$ which has a spot, say $F$, which is parallel to a spot of $X$. If not then $e(V)
\leq 2$, a contradiction.
Let $Y = X \cup \neigh(F) \cup P$ and notice that this is a five-spotted ball containing $D$ and $E$, our original disks. Isotope $Y$ to be cleanly embedded. Let $\EE'$ be any pants decomposition of $V \setminus \neigh(Y)$. Add to $\EE'$ any spots of $Y$ which are non-peripheral in $V$. This then is the desired simplex $\sigma \in
\calD(V)$. Since $D$ and $E$ are dual the pentagon lemma implies that there is a pentagon in $\calD(Y)$ making $D, E$ non-adjacent.
The backwards direction follows from , the combinatorial characterization of genus and number of spots, and from the pentagon lemma.
We now discuss the dual complex. Fix a cut system $\DD = \{ D_i \}$. Recall that $\dual_i(\DD)$ is the subcomplex of $\calD(V)$ spanned by the disks $E \in \calD(V)$ which are dual to $D_i$ and disjoint from $D_j$ for all $j \neq i$.
Define $V_i$ to be the spotted solid torus obtained by cutting $V$ along all disks of $\DD$ [*except*]{} $D_i$. Note that $V_i$ has exactly $e(V)$–many spots, and this is at least three. Also, $D_i$ is a meridian disk for $V_i$. Note that $\dual_i(\DD) \subset
\calD(V_i)$. A disk $E \in \dual_i(\DD)$ is a [*simple dual*]{} if $E$ is a pants disk in $V_i$.
Let $\calA_i(\DD)$ be the complex where vertices are isotopy classes of arcs $\alpha \subset \bdy_+ V_i$ so that
- $\alpha$ meets $\bdy D_i$ exactly once, transversely, and
- $\bdy \alpha$ meets distinct spots of $V_i$.
A collection of vertices spans a simplex if they can be realized disjointly.
If an arc $\alpha \in \calA_i(\DD)$ meets spots $A, B \in \bdy_0 V_i$ then the frontier of $\Neigh(A \cup \alpha \cup B)$ is a simple dual, $E_\alpha$.
\[Lem:DualComplexConnected\] If $e(V) \geq 3$ then the complex $\dual_i(\DD)$ is connected.
It suffices to check this for $i = 1$. To simplify notation we write $D = D_1$, $U = V_1$, $\dual(D) = \dual_i(\DD)$ and $\calA(D) =
\calA_1(\DD)$. We will prove via a sequence of claims.
For any pair of arcs $\alpha, \gamma \in \calA(D)$ there is a sequence $\{ \alpha_k \}_{k = 0}^N \subset \calA(D)$ so that:
- the arcs $\alpha_k, \alpha_{k+1}$ are disjoint, for all $k < N$,
- $\alpha_0 = \alpha$ and $\alpha_N = \gamma$, and
- there is at most one spot in common between the endpoints of $\alpha_k$ and $\alpha_{k+1}$, for all $k < N$.
Fix, for the remainder of the proof, an arc $\beta \in \calA(D)$ so that $\alpha$ and $\beta$ are disjoint and so that the endpoints of $\alpha$ and $\beta$ share at most one spot. This is possible as $U$ has at least three spots. Define the complexity of $\gamma$ to be $c(\gamma) = i(\alpha, \gamma) + i(\beta, \gamma)$. Notice if $c(\gamma) = 0$ then we are done: one of the sequences $$\{ \alpha, \gamma \} \quad \mbox{or} \quad \{ \alpha, \beta, \gamma \}$$ has the desired properties.
Now induct on $c(\gamma)$. Suppose, breaking symmetry, that $\alpha$ meets a spot, say $A \in \bdy_0 U$, so that $\gamma \cap A =
\emptyset$. If $i(\alpha, \gamma) = 0$ then the sequence $\{ \alpha,
\gamma \}$ has the desired properties. If not, then let $x$ be the point of $\alpha \cap \gamma$ that is closest, along $\alpha$, to the endpoint $\alpha \cap A$. Let $\alpha' \subset \alpha$ be the subarc connecting $x$ and $\alpha \cap A$. Let $N$ be a regular neighborhood, taken in $\bdy_+ U$, of $\gamma \cup \alpha'$. The frontier of $N$, in $\bdy_+ U$, is a union of three arcs: one arc properly isotopic to $\gamma$ and two more arcs $\gamma', \gamma''$.
The arcs $\gamma'$ and $\gamma''$ are disjoint from $\gamma$ and satisfy $c(\gamma') + c(\gamma'') \leq c(\gamma) - 1$. Also, since $\gamma'$ and $\gamma''$ each have one endpoint on the spot $A$ the arcs $\gamma'$ and $\gamma''$ have exactly one spot in common with $\gamma$. Now, if $\alpha' \cap \bdy D = \emptyset$ then one of $\gamma', \gamma''$ meets $\bdy D$ once and the other is disjoint. On the other hand, if $\alpha' \cap \bdy D \neq \emptyset$ then $\alpha'$ meets $\bdy D$ once. Thus one of $\gamma', \gamma''$ meets $\bdy D$ once and the other meets $\bdy D$ twice. In either case we are done.
Recall that if $\alpha \in \calA(D)$ is an arc then $E_\alpha$ is the associated simple dual.
If $\alpha, \beta \in \calA(D)$ are disjoint arcs, with at most one spot in common between their endpoints, then there is an edge-path in $\dual(D)$ of length at most four between $E_\alpha$ and $E_\beta$.
If $\alpha$ and $\beta$ share no spots then $\{ E_\alpha, E_\beta \}$ is a path of length one. Suppose that $\alpha$ and $\beta$ share a single spot. Let $A, B, C$ be the three spots that $\alpha$ and $\beta$ meet, with both meeting $C$. Let $\alpha', \beta'$ be the subarcs of $\alpha, \beta$ connecting $C$ to $\bdy D$. There are two cases: either $\alpha'$ and $\beta'$ are incident on the same side of $\bdy D$ or are incident on opposite sides.
Suppose that $\alpha'$ and $\beta'$ are incident on the same side of $\bdy D$. Then $\alpha'$ and $\beta'$, together with subarcs of $\bdy
C$ and $\bdy D$ bound a disk $\Delta \subset \bdy U$. Note that $\Delta$ may contain spots, but it meets $A \cup B \cup C$ only along the subarc in $\bdy C$. It follows that the disk $F$, defined to be the frontier of $$\Neigh\left( (A \cup B \cup C) \cup (\alpha \cup \beta) \cup \Delta
\right),$$ is dual to $D$. The disk $F$ is also essential as it separates at least three spots from a solid handle. So $\{ E_\alpha, F, E_\beta
\}$ is the desired path.
Suppose that $\alpha'$ and $\beta'$ are incident on opposite sides of $\bdy D$. Let $d \subset \bdy D$ be either component of $\bdy D
\setminus (\alpha \cup \beta)$. Let $\alpha'' = \closure{\alpha
\setminus \alpha'}$ and define $\beta''$ similarly. Define $\gamma
\in \calA(D)$ by forming the arc $\alpha'' \cup d \cup \beta''$ and using an proper isotopy of $\bdy_+ U$ to make $\gamma$ transverse to $\bdy D$. Now apply the previous paragraph to the pairs $\{ \alpha,
\gamma \}$ and $\{ \gamma, \beta \}$ to obtain the desired path of length four.
For every dual $E \in \dual(D)$ there is a simple dual connected to $E$ by an edge-path of length at most two.
The graph $\bdy E \cup \bdy D$ cuts $\bdy U$ into a pair of disks $B,
C$ and an annulus $A$. Each of $B, C$ contain at least one spot.
Suppose $E$ is separating. Then the disks $B, C$ are adjacent along an subarc of $\bdy D$. Connect a spot in $B$ to a spot in $C$ by an arc $\alpha$ that meets $\bdy D$ once and that is disjoint from $\bdy E$. Thus $E_\alpha$ is disjoint from $E$.
Suppose $E$ is non-separating. Then the two disks $B, C$ meet only at the points of $\bdy D \cap \bdy E$. Now, if the annulus $A$ contains a spot then we may connect a spot in $B$ to a spot in $A$ by an arc $\alpha$ meeting $\bdy D$ once and $\bdy E$ not at all. In this case we are done as in the previous paragraph.
If $A$ contains no spots then, breaking symmetry, we may assume that $B$ contains at least two spots while $C$ contains at least one. Let $\delta$ be an arc connecting some spot, say $B' \subset B$, to $E$. Let $N$ be a regular neighorhood of $E \cup \delta \cup B'$. Then the frontier of $N$ contains two disks. One of these is isotopic to $E$ while the other, say $E'$, is non-separating, dual to $D$, and divides the spots as described in the previous paragraph.
Equipped with these claims we have:
The first two claims imply that the set of simple duals in $\dual(D)$ is contained in a connected set. The third claim shows that every vertex in $\dual(D)$ is distance at most two from the set of simple duals. This completes the proof.
Crawling through the complex of duals {#Sec:CrawlDual}
=====================================
\[Lem:PreCrawl\] Suppose that $\phi_{\link}$ fixes $\DD$ and $\link(\DD)$. For any $E
\in \dual_i(\DD)$ the disks $E$ and $\phi(E)$ differ by some power of $T_i$, the Dehn twist about $D_i$.
As usual, it suffices to prove this for $D = D_1$. Let $U = V_1$.
Let $X \subset U$ be the four-spotted ball filled by $D$ and the dual disk $E$. Isotope $X$ to be cleanly embedded. Let $\FF$ be the components of $\bdy_0 X$ which are not spots of $U$. Note that $\phi_{\link}$ fixes $D$ as well as every disk of $\FF$. This, together with , implies that $\phi_{\link}$ preserves the set of disks that are contained in $X$ and dual to $D$.
Since $\calD(X)$ equipped with the duality relation is a copy of $\calF$, the Farey graph, it follows that $E$ and $F =
\phi_{\link}(E)$ differ by some number of half-twists about $D$. If $E$ and $F$ differ by an odd number of half-twists then $E$ and $F$ have differing topological types, contradicting applied to $\phi_{\link}|\calD(U)$. Thus $E$ and $F$ differ by an even number of half-twists, as desired.
\[Lem:CrawlDual\] Suppose that $\phi_{\dual}$ fixes $\DD$, $\link(\DD)$, and $\EE$, a collection of duals (that is, $E_i \in \dual_i(\DD)$). Then $\phi_{\dual}$ fixes every vertex of $\dual_i(\DD)$, for all $i$.
As usual, it suffices to prove this for $D = D_1$. Let $E = E_1$ and let $U = V_1$. We [*crawl*]{} through $\dual(D) = \dual_1(\DD)$, as follows.
Suppose that $F, G \in \dual(D)$ are adjacent vertices and suppose that $\phi_{\dual}(F) = F$. By , the disks $G$ and $G' = \phi_{\dual}(G)$ differ by some number of Dehn twists about $D$. Also, as $\phi_{\dual}$ is a simplical automorphism the disks $F$ and $G'$ are disjoint. Let $X$ be the four-spotted ball filled by $D$ and $F$. If $G$ and $G'$ are not equal then $G \cap X$ and $G' \cap X$ are also not equal and in fact differ by some non-zero number of twists; thus one of $G \cap X$ or $G' \cap X$ must cross $F$, a contradiction.
Recall that $\phi_{\dual}(E) = E$. Suppose that $G$ is any vertex of $\dual(D)$. Since $\dual(D)$ is connected () there is a path $\calP \subset
\dual(D)$ connecting $E$ to $G$. Induction along $\calP$ completes the proof.
Crawling through the disk complex {#Sec:CrawlDisk}
=================================
Before continuing we will need the following complex:
\[Def:CutGraph\] The [*cut system graph*]{} ${\mathcal{CG}}(V)$ is the graph with vertex set being isotopy classes of unordered cut systems in $V$. Edges are given by pairs of cut systems with $g - 1$ disks in common and the remaining pair of disks disjoint.
Wajnryb also gives a two-skeleton, but we will only require:
\[Thm:Wajnryb\] The cut system graph ${\mathcal{CG}}(V)$ is connected.
For the remainder of this section suppose that $\Phi = \phi_{\dual}$ is an automorphism of $\calD(V)$ and $\DD$ is a cut system so that $\Phi$ fixes $\DD$, $\link(\DD)$ and $\dual(\DD)$.
For the crawling step, suppose that $\EE, \FF$ are adjacent in ${\mathcal{CG}}(V)$ and that $\Phi$ fixes $\EE$, $\link(\EE)$ and $\dual(\EE)$. Let $\GG$ be a pants decomposition obtained by adding the new disk of $\FF$ to $\EE$ and then adding non-separating disks until we have $3g
- 3 + n$ disks. Let $\{ P_k \}$ enumerate the solid pants of $\GG$. Let $X_i = P_k \cup P_\ell$ be the four-spotted ball containing $G_i$ in its interior.
Let $\HH, \II = \{ H_i \}, \{ I_i \}$ be collections of disks so that $H_i, I_i$ are contained in $X_i$ and $G_i, H_i, I_i$ are pairwise dual in $X_i$. Now, all of these disks $\GG \cup \HH \cup \II$ lie in $\EE \cup \link(\EE) \cup \dual(\EE)$. Thus $\Phi$ fixes all of them. Thus $\Phi$ fixes $\FF$. Consider $\Phi|\link(\FF)$. By the automorphism $f = \Phi|\link(\FF)$ is geometric. Let $f$ also denote the given homeomorphism of $V' = V \setminus
\neigh(\FF)$. Let $\GG' = \GG \setminus \FF$ and $\HH', \II'$ be the disks of $\HH, \II$ contained in $V'$. Thus $f$ fixes all disks of $\GG', \HH', \II'$. It follows that $f$ permutes the solid pants $\{
P_k \}$.
If $f$ nontrivially permutes $\{ P_k \}$ then, since each $G_i$ is fixed, we find that adjacent solid pants are interchanged. This implies that $V' = P_1 \cup P_2$, a contradiction.
So $f$ fixes every $P_k$. Since all disks in $\GG'$ are fixed, $f$ is either orientation reversing, isotopic to the identity, or isotopic to a half-twist on each of the $P_k$. Let $G_i \in \GG'$ be any disk meeting $P_k$. Then $f|P_k$ cannot be orientation reversing because the triple $G_i, H_i, I_i$ determines an orientation on $X_i$ and hence on $P_k$. If $f|P_k$ is a half-twist then $P_k$ meets two spots of $V'$. Thus $G_i$ meets two solid pants $P_k, P_\ell$ so that $X_i
= P_k \cup P_\ell$. Now, as $e(V') \geq 3$, the solid pants $P_\ell$ meets at most one spot of $V'$. Thus $f|P_\ell$ is isotopic to the identity. So if $f|P_k$ is a half-twist then $f(H_i) \neq H_i$, a contradiction. Deduce that $f$, when restricted to any solid pants, is isotopic to the identity. Now, since $f$ fixes all of the $H_i$, $f$ is isotopic to the identity on $V'$, as desired.
Deduce that $\Phi|\link(\FF)$ is the identity. As $\Phi$ fixes duals to $\FF$ by the automorphism $\Phi$ fixes all of $\dual(\FF)$. This completes the crawling step and so completes the proof of:
\[Thm:AutOfDiskComplex\] If a handlebody $V = V_{g,n}$ satisfies $e(V) \geq 3$ then the natural map $\calH(V) \to \Aut(\calD(V))$ is a surjection.
As a corollary:
\[Thm:AutOfDiskComplexII\] If a handlebody $V = V_{g,n}$ satisfies $e(V) \geq 3$ then the natural map $\calH(V) \to \Aut(\calD(V))$ is an isomorphism.
Note that Theorems \[Thm:AutOfDiskComplex\] and \[Thm:AutOfDiskComplexII\] are sharp: when $e(V) \leq
2$ the conclusions are false. See .
shows that the natural map is surjective. Suppose that the mapping class $f$ lies in the kernel. As in the discussion of crawling through ${\mathcal{CG}}(V)$ given above, let $\GG = \{ G_i
\}$ be a pants decomposition of $V$ so that all of the $G_i$ are non-separating. Let $\{ P_k \}$ enumerate the solid pants of this decomposition. Let $X_i = P_j \cup P_k$ be the four-spotted ball containing $G_i$ in its interior. Let $\HH, \II = \{ H_i \}, \{ I_i
\}$ be collections of disks so that $H_i, I_i$ are contained in $X_i$ and $G_i, H_i, I_i$ are pairwise dual in $X_i$. All of these disks are fixed by $f$. It follows that $f$ is isotopic to the identity.
An application {#Sec:AutOfHandlebodyGroup}
==============
\[Thm:AutOfHandlebodyGroup\] If $e(V) \geq 3$ then the outer automorphism group of the handlebody group is trivial.
This may be restated as: $\Aut(\calH) \isom \calH$. When $g = 0$ then follows from and the first author’s thesis [@Korkmaz99 Theorem 3]. For the rest of this section we restrict to the case $g \geq 1$.
The idea of the proof is to turn an element $\phi \in \Aut(\calH)$ into an automorphism of the disk complex $\calD(V)$. We do this, following [@Ivanov88], by giving an algebraic characterization of first Dehn twists about non-separating disks and then Dehn twists generally. We then apply Theorem \[Thm:AutOfDiskComplex\] to $\phi$ to find the correspoding geometric automorphism. An algebraic trick then gives the desired result.
A finite index subgroup $\Gamma < \calH$ is [*pure*]{} if every reducible class in $\Gamma$ fixes every component of every reducing set. For example, the kernel of $\calH \to \Aut(H_1(\bdy_+ V,
\ZZ/3\ZZ))$ is pure.
\[Lem:NonSeparatingTwists\] Suppose $\Gamma < \calH$ is pure and finite index. Then $\{f_i\}
\subset \calH$ is a collection of Dehn twists along a pants decomposition of non-separating disks in $V$ if and only if
- the subgroup $A = \langle f_i \rangle$ is free Abelian of rank $\xi(V)$,
- $f_i$ and $f_j$ are conjugate in $\calH$, for all $i, j$,
- $f_i$ is [*primitive*]{} in $C_\calH(A)$: $f_i$ is not a proper power of any $h \in C_\calH(A)$, and
- the center of the centralizer of the class $f_i^n$ in $\Gamma$ is infinite cyclic (for all $i$ and for all $n$ so that $f_i^n \in
\Gamma$): $$C(C_\Gamma(f_i^n)) \isom \ZZ.$$
The forwards direction is identical to the forwards direction of [@Ivanov88 Theorem 2.1]. The backwards direction is similar in spirit to the backwards direction of [@Ivanov88 Theorem 2.1] but some details differ. Accordingly we sketch the backwards direction.
The mapping class $f_i$ can not be periodic or pseudo-Anosov as that would contradict the first property. Let $\Theta \subset S = \bdy_+
V$ be the [*canonical reduction system*]{} for the Abelian group $A$ [@BirmanEtAl83]. Let $\{ X_j \}$ be the components of $S
\setminus \neigh(\Theta)$ and let $\{ Y_k \}$ be the collection of annuli $\Neigh(\Theta)$. By [@BirmanEtAl83 Lemma 3.1(2)] the number of annuli in $\{ Y_k \}$ plus the number of non-pants in $\{
X_j \}$ equals $\xi(V)$. It follows that every non-pants $X_j$ has complexity one (so is homeomorphic to $S_{0,4}$ or $S_{1,1}$).
Fix a power $n$ (independent of $i$) to ensure that $f_i^n \in
\Gamma$. For each $X_j$ of complexity one there is some $f_i^n$ so that $f_i^n|X_j$ is pseudo-Anosov. Suppose that $f = f_1^n$, $X =
X_1$ has complexity one, and $f|X$ is pseudo-Anosov. Let $\lambda^\pm$ be the stable and unstable laminations of $f|X$. For every $i$, the mapping $f_i^n|X$ is either the identity or pseudo-Anosov. Note that in the latter case the stable and unstable laminations of $f_i^n|X$ agree with $\lambda^\pm$: otherwise a ping-pong argument gives a rank two free group in $A$, a contradiction. Thus, perhaps taking a larger power $n$, we may assume that for each $i$ either $f_i^n|X$ is the identity or identical to $f|X$. For each $i$ where $f_i^n|X = f|X$ we temporarily replace $f_i$ by $f_i^n f^{-1}$.
Continuing in this manner we find a free Abelian group $B < A \cap
\Gamma$ of rank at least $|\Theta|$ where all elements are supported inside of the union of annuli $\{ Y_k \}$. Since $B$ is pure, it follows that all elements of $B$ are compositions of powers of Dehn twists along disjoint curves. A theorem of McCullough [@McCullough02] implies that every curve in $\Theta$ either bounds a disk or cobounds an annulus with some other curve of $\Theta$. However, each annulus reduces the possible rank of $B$ by one; it follows that every curve in $\Theta$ bounds a disk.
Let $\gamma$ be any essential non-peripherial component of $\bdy X$. It follows that $f$ commutes with $T_\gamma$, that $T_\gamma$ lies in $\calH$ by the above paragraph, and that $T_\gamma$ to some power lies in $C(C_\Gamma(f))$. But this contradicts the fourth property. It follows that every component $X_j$ is a pants and that $|\Theta| =
\xi(V)$. Thus every $f_i$ is a compositions of powers of disjoint twists. Again, by the fourth property each $f_i$ is some power of a single twist. By the third property (following [@Ivanov88]) $f_i$ is in fact a twist. Finally, by the second property, each twist is supported on a disk of the same topological type. As every pants decomposition of $V$ must contain a non-separating disk all of the twists $f_i$ are supported by non-separating disks.
We now give the general characterization:
\[Lem:GeneralTwists\] Suppose $\Gamma < \calH$ is pure and finite index. Then $\{f_i\}
\subset \calH$ is a collection of Dehn twists along a pants decomposition of $V$ if and only if
- the subgroup $A = \langle f_i \rangle$ is free Abelian of rank $\xi(V)$,
- $f_i$ is primitive in $C_\calH(A)$,
- for all $i$ and for all $n$ so that $f_i^n \in \Gamma$ either $C(C_\Gamma(f_i^n)) \isom \ZZ$ or there is a $j$ so that $C(C_\Gamma(f_i^n)) \isom \ZZ^2$ with the latter given by $\langle
f_i, f_j \rangle$ and $f_j$ is a twist on a non-separating disk.
Suppose that $\{ D_i \}$ is a pants decomposition and $f_i$ is the positive twist on $D_i$. Then $A = \langle f_i \rangle$ is free Abelian of the correct rank. The second property follows as $A =
C_\calH(A)$. The third property follows from Ivanov’s discussion [@Ivanov88] except if $D_i$ is a handle disk. In this case the meridian of the handle, say $D_j$, gives a twist $f_j$ which lies in the center of the centralizer.
The backwards direction is similar to that of the proof of . The only change occurs when $f|X$ is pseudo-Anosov: when the center of the centralizer has rank two then the additional element is a twist about a separating disk and this contradicts the third property.
The following lemmas follow from the idential statements for the mapping class group of $S$ [@IvanovMcCarthy99]:
\[Lem:FarCommutivity\] Suppose $D$ and $E$ are essential disks. The twists $T_D, T_E$ commute if and only if $D$ and $E$ can be made disjoint via proper isotopy.
\[Lem:Conjugation\] For any twist $T_D$ and for any homeomorphism $h$ we have $h T_D
h^{-1} = T_{h(D)}$.
\[Lem:Powers\] For any pair of disks $D$ and $E$ and any pair of integers $n$ and $m$, if $T_D^n = T_E^m$ then $D = E$ and $n = m$.
The proof of now follows, essentially line-by-line, the proof of either [@Ivanov97 Theorem 2] or [@Korkmaz99 Theorem 3].
To extend our algebraic characterization of twists in $\calH(V)$ (Lemmas \[Lem:NonSeparatingTwists\] and \[Lem:GeneralTwists\]) to a characterization of powers of twists inside of finite index pure subgroups $\Gamma < \calH(V)$ appears to be a delicate matter. Solving this problem would, following Ivanov [@Ivanov97], solve:
\[Prob:Comm\] Show that the abstract commensurator of $\calH(V)$ is $\calH(V)$ itself. Show that $\calH(V)$ is not arithmetic.
[^1]: This work is in the public domain
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is shown whether Higgs inflation can be saved with high-scale supersymmetry critically depends on the magnitude of non-minimal coupling constant $\xi$. For small $\xi \leq 500$, the threshold correction at scale $M_{P}/\xi$ is constrained in high precision. Its magnitude is in the narrow range of $(-0.03, -0.02)$ and $(-0.05, -0.04)$ for the wino and higgsino/singlino dark matter, respectively. While in the large $\xi$-region with $\xi \geq 10^{4}$, such high-scale supersymmetry is excluded by too large threshold correction as required by Higgs inflation.'
author:
- Sibo Zheng
title: 'Can Higgs Inflation be Saved with High-scale Supersymmetry ?'
---
I. Introduction
================
The collider facilities such as the Large Hadron Collider (LHC) and astrophysical experiments are two main tools for exploring new physics beyond the standard model (SM). Recently, the Plank Collaboration, designed to detect the cosmic microwave background (CMB) temperature anisotropy and polarization, has reported the latest value of tensor-to-scalar ratio $r\leq 0.11$ at $95\%$ CL [@1502.01589; @1502.02114]. This data excludes a few well known inflation models such as quadratic inflation, but still allows some simple examples such as Starobinsky-like inflation [@Starobinsky], $\alpha$-attractor inflation [@1502.07733] and Higgs inflation [@0710.3755].
Among these survivors, Higgs inflation is rather special due to two considerations. Firstly, there is only one new parameter $\xi$ in this model in the Lagrangian in this model, which reads as in the Jordan frame, $$\begin{aligned}
\label{Lagrangian}
\mathcal{L}_{J}=\frac{M^{2}+\xi h^{2}}{2} R-\frac{1}{2} \left(\partial h\right)^{2}-\frac{\lambda_{H}}{4}\left(h^{2}-\upsilon^{2}\right)^{2}.\end{aligned}$$ Here, $\xi$ measures the non-minimal coupling between Higgs scalar $h$ and gravity. It was found that the Higgs scalar potential $V_{E}(h)$ in the Einstein frame rapidly approaches to a plateau potential in the large field region $h> M_{P}/\sqrt{\xi}$, and the model predicts that for e-foldings number $N=60$, $$\begin{aligned}
\label{nr}
n_{s}\simeq 0.970, ~~~~r\simeq 0.0033,\end{aligned}$$ which is in perfect agreement with the Plank 2015 data [@1502.01589; @1502.02114]. In terms of the present cosmological data $M$ in Eq.(\[Lagrangian\]) approximates to the reduced Plank mass $M_{P}=2.4\times 10^{18}$ GeV for our present universe. Moreover, unlike Starobinsky-like inflation or $\alpha$-attractor inflation, it is obvious that Higgs inflation is of special interest from the viewpoint of particle physics.
Unfortunately, there are two subjects against the Higgs inflation. The first subject shows that the SM is not valid above scale $M_{P}/\xi$ [@0902.4465; @0903.0355; @1002.2730]. It implies that the plateau potential $V(h)$ at large field value $h> M_{P}/\sqrt{\xi}$ may be significantly modified by the ultra-violet completion so that realistic inflation driven by the Higgs field can not occur at all. It is proposed in [@1008.5157] that a scale symmetry in the ultra-violet completion may be the prescription to this problem. The second subject arises since the discovery of the Higgs mass $m_{h}=125.5\pm 0.5$ GeV [@Higgsmass1; @Higgsmass2; @Higgsmass3], which implies that the SM Higgs quartic coupling $\lambda_{H}$ at the weak scale is not large enough so that $\lambda_{H}<0$ above the high-energy scale $10^{9}-10^{11}$ GeV. The uncertainty about this critical scale mainly arises from the the uncertainty of top quark mass, as the renormalization group equation (RGE) for $\lambda_{H}$ is very sensitive to the top Yukawa coupling. Obviously, positive $\lambda_{H}$ is required during inflationary epoch. This problem can be solved through introducing some new fields into the SM, which give rise to either correct threshold correction to $\lambda_{H}$ or slowing down the RG evaluation for $\lambda_{H}$ along high energy scale, see, .e.g., [@1203.0237; @1405.7331].
In this letter, we consider saving the Higgs inflation with high-scale supersymmetry (SUSY). High-scale SUSY can provide the observed Higgs mass at the LHC and stable DM candidate [@1409.2939]. However, the detection of such models at colliders is unpromising. In this sense the astrophysical probe is an important direction. Fortunately, they can be probed directly or indirectly in the light of the cosmological observations on the early universe [@1109.0292; @1409.7462].
We identify the cut-off scale $M_{P}/\xi $ as the typical mass scale $\tilde{m}$ in the SUSY mass spectrum. Above the scale $\tilde{m}=M_{P}/\xi $ supergravity is the natural ultra-violet (UV) completion, and for the matter content same as the next-to-minimal SUSY model (NMSSM) it may maintain the plateau potential [@1008.2942]. For earlier discussions, see [@1009.2276; @0912.2718; @1004.0712]. On the other hand, the SUSY dark matter (DM), whose mass is around the weak scale, is the natural choice on the new fields added to the SM below the cut-off scale $\tilde{m}$. In this paper, we consider two possibilities [@1409.2939], where wino $\tilde{w}$ and higgsino/singlino mixing state serves as the DM, respectively.
The paper is organized as follows. In the next section, we discuss the RGE for $\lambda_{H}$ in two classes of high-scale SUSY below scale $\tilde{m}$. In Section III, we discuss embedding the Higgs inflation to supergravity, where we uncover the constraint on parameters $\lambda_{H}$ and $\xi$ at the end of inflation $h_{\text{end}}=M_{P}/\sqrt{\xi}$ arising from the present cosmological data. In section IV, we discuss the constraint on the SUSY mass spectrum in terms of the constraint on the threshold correction to $\lambda_{H}$. Finally, we discuss our results in section V.
II. RGE for $\lambda_{H}$ Below Scale $M_{P}/\xi$
=================================================
Below the scale $\tilde{m}$, the effective theory is described by the SM together with either wino-like or higgsino/singlino-like DM. In this section we use the 2-loop RGEs for relevant couplings for our analysis. We refer the reader to appendix in our previous work [@1409.2939], where the 2-loop RGEs for SM gauge and Yukawa couplings, and Higgs quartic coupling in models of SM$+\tilde{w}$ and SM$+\tilde{h}/\tilde{s}$ are explicitly shown. We ignore other SM Yukawa couplings in the following discussion. For the calculation of beta functions, see the references therein.
![Two-loop RG running for $\lambda_{H}$ in the SM (black), SM$+\tilde{w}$ (blue) and SM$+\tilde{h}/\tilde{s}$ (green), respectively. The solid lines correspond to the central value $m_{t}=173.1$ GeV, $m_{h}=125.5$ GeV and $\alpha_{s}(m_{Z})=0.1184$. The dotted lines show the uncertainty due to the top quark mass with 1$\sigma$ deviation. In the model of SM$+\tilde{h}/\tilde{s}$, we have chosen small value $g_{\lambda}(m_{Z})=0.2$. See the text for the reason for this choice.[]{data-label="lambda1"}](1sigma.pdf){width="45.00000%"}
Fig.\[lambda1\] shows the RG running for $\lambda_{H}$ in high-scale SUSY discussed in this letter, which is modified by the SUSY DM (wino or higgisno-singlino mixing state) with mass of weak scale in compared with the SM. In this figure, black, blue and green curves corresponds to the SM, SM$+\tilde{w}$, and SM$+\tilde{h}/\tilde{s}$, respectively. The solid lines refer to the central value $m_{t}=173.1$ GeV, $m_{h}=125.5$ GeV and $\alpha_{s}(m_{Z})=0.1184$, while the dotted lines show the uncertainty due to the top quark mass with 1$\sigma$ deviation. From Fig.\[lambda1\] we observe that more positive $\lambda_{H}$ is obtained in high-scale SUSY with wino DM for smaller top quark pole mass $m_{t}=172.2$ GeV along RG running to scale $M_{P}/\xi$. This is mainly due to the correction to the SM beta function $\beta_{\lambda_{H}}$ induced by the wino DM.
Unlike the case with wino DM, the sign of the correction induced by higgsino/singlino DM is only positive when model parameter $g_{\lambda}$ is small. Otherwise, for large $g_{\lambda}\geq 0.4$ the sign of the correction will reverse and $\lambda_{H}$ will be negative more rapidly along RG running to scale $M_{P}/\xi$. This implies that the correction in this case is upper bounded.
![Same as Fig.\[lambda1\] but with $m_{t}=173.1$ GeV and $m_{h}=125.5\pm 0.5$ GeV.[]{data-label="lambda2"}](1sigmah.pdf){width="45.00000%"}
Fig.\[lambda2\] shows the sensitivity of the RG running for $\lambda_{H}$ to the Higgs mass. One finds that for about $\sim 0.5$ GeV deviation to the central value $m_{h}=125.5$ GeV the correction in each case is tiny. Among the three models, $\lambda_{H}$ approaches to zero mostly in the case with wino DM, similarly to what Fig.\[lambda1\] indicates.
In summary, for about $1\sigma$ deviation to the central value of top quark mass, $\lambda_{H}$ is tuned to be more positive in high-scale SUSY with the low-energy theory below scale $M_{P}/\xi$ described by either SM$+\tilde{w}$ or SM$+$higgsino/singlino DM in compared with the SM. As we will see in the next section, more positive $\lambda_{H}$ at scale $M_{P}/\xi$ is favored by the Higgs inflation. Because the threshold correction at scale $M_{P}/\xi$, may be unable to tune $\lambda_{H}$ into a positive parameter above this scale as required by Higgs inflation.
III. Embedding Higgs Inflation into Supergravity
================================================
We proceed to discuss the constraint arising from the condition of plateau potential for Higgs scalar in the context of supergravity, which is automatically the ultra-violet completion. For this purpose, we focus on the scalar-gravity part of supergravity Lagrangian, which is defined by the frame function $\Omega(z_{i},\bar{z}_{i})$, Kahler potential $K(z_{i},\bar{z}_{i})$ and superpotential $W(z_{i})$. We use $z_i$ to label the chiral superfields including the two Higgs doublets $H_{u,d}$.
We follow the notation and conventions in [@1008.2942], where the Kahler potential and frame function are given by in unit of Plank mass, $$\begin{aligned}
\label{frame}
K(z_{i},\bar{z}_{i})&=&-3\log(\Omega),\nonumber\\
\Omega(z_{i},\bar{z}_{i})&=&1-\frac{1}{3}\delta_{i\bar{j}}z^{i}z^{\bar{j}}+\cdots.\end{aligned}$$ Given the explicit form of these functions, the scalar potential in the Einstein frame is directly derived from the well-known formula [^1], $$\begin{aligned}
\label{potential}
V_{E}=e^{K}\left(D_{i}W K^{i\bar{j}}D_{\bar{j}}\bar{W}-3W\bar{W}\right)+V_{E}^{(D)},\end{aligned}$$ where $D_{i}W=\partial_{i}W+K_{i}W$, $V_{E}^{(D)}$ represents the D-term contribution.
The first attempt to embedding Higgs inflation into supergravity was shown in [@0912.2718]. In the light of [@0912.2718] there are two important observations. $1)$, The supergravity version of Eq.(\[Lagrangian\]) requires a holomorphic function $X=-\frac{1}{2}\chi H_{u}H_{d}+\text{h.c}$ in order to reproduce the $\xi$-term. $\chi$ is a dimensionless coupling constant. $2)$, The matter content of MSSM is not viable for the Higgs inflation. By following this line, the authors in [@1008.2942] proposed that the matter context of NMSSM is a realistic choice, in which the potential in the NMSSM depends on three complex superfields, $$\begin{aligned}
\label{z}
z_{i}&=&\{S, H^{0}_{u}, H^{0}_{d}\}\nonumber\\
&=&\{se^{i\alpha}/\sqrt{2}, h\cos\beta e^{i\alpha_{1}}/\sqrt{2},
h\sin\beta e^{i\alpha_{2}}/\sqrt{2} \}\nonumber\\\end{aligned}$$ , as long as the scalar $s$ in the singlet superfiled $S$ can be stabilized at $s=0$ and $D$-flat condition $\beta=\pi/4$ is satisfied. This idea was firstly considered in [@0912.2718], but a new $\zeta$-term should be added to the frame function [@1008.2942] for stabilizing the $s$ scalar.
In summary, the frame function and superpotential for the purpose of Higgs inflation are given by, respectively, $$\begin{aligned}
\label{frame2}
\Omega(z_{i},\bar{z}_{i})&=&1-\frac{1}{3}\left(\mid H^{0}_{u}\mid^{2}+\mid H^{0}_{d}\mid^{2}+\mid S\mid^{2}\right)\nonumber\\
&+&\frac{\zeta}{3} \mid S\mid^{4}-\left(\frac{1}{2}\chi H^{0}_{u}H^{0}_{d}+\text{h.c}\right),\nonumber\\
W(z_{i},\bar{z}_{i})&=&\lambda SH^{0}_{u}H^{0}_{d}+\frac{\rho}{3} S^{3},\end{aligned}$$ Substituting Eq.(\[frame2\]) into Eq.(\[potential\]) gives rise to the potential, $$\begin{aligned}
\label{potential2}
V_{E}=V^{(F)}_{E}=\frac{9\lambda^{2}h^{4}}{(3\chi h^{2}-2h^{2}+6)^{2}},\end{aligned}$$ for $s=0$ and $\beta=\pi/4$. Here we have used $W\mid_{s=0}=0$. For more details about the stabilization of $s$ and angles in Eq.(\[z\]), see [@1008.2942; @1004.0712]. In the region $\chi h^{2}>>1>>h^{2}$, Eq.(\[potential2\]) approaches to $(\lambda/\chi)^{2}$ (in unit of Plank mass), which verifies our statements above. Moreover, in terms of Eq.(\[potential2\]) one can also verify Eq.(\[nr\]).
Parameters $\xi$ and $\lambda_{H}$ in Eq.(\[Lagrangian\]) are related to parameters $\chi$ and $\lambda$ in Eq.(\[frame2\]) as, $$\begin{aligned}
\label{matching}
\xi=-\frac{1}{6}+\frac{1}{4}\chi,~~~~~~\lambda_{H}(\mu_{I})=\frac{\lambda^{2}}{4}.\end{aligned}$$ Here $\mu_{I}$ denotes the RG scale corresponding to inflation. They are constrained by the present cosmological data as follows. Recall that $V^{1/4}_{E}=\left(24\pi^{2}M^{4}_{P}\epsilon A_{s}\right)^{1/4}$, where $A_s$ is the amplitude of the power spectrum of the curvature perturbation and $\epsilon=r/16$ in the context of single field inflation. For $A^{1/2}_{S}\simeq 3.089\times 10^{-5}$ as reported by Plank Collaboration [@1303.5076] one obtains, $$\begin{aligned}
\label{constraint}
\xi \simeq 58789 \sqrt{\lambda_{H}(\mu_{I})}. \end{aligned}$$
IV. Constraints on SUSY mass spectrum
=====================================
Until now, we have obtained the value of $\lambda_{H}$ below the RG scale $\mu=M_{P}/\xi$ (as shown in Fig.\[lambda1\] and Fig. \[lambda2\] ) and above the RG scale $M_{P}/\sqrt{\xi}$ (as shown in Eq.(\[constraint\])) given a choice on $\xi$. The effective theory at the intermediate scale between scale $M_{P}/\xi$ and $M_{P}/\sqrt{\xi}$ is described by the SM together with gauginos and squarks. Other SUSY particles such as higgsinos and charged Higgs have masses of order $M_{P}/\sqrt{\xi}$ [@1008.2942; @1004.0712], so they should be integrated out at this intermediate energy scale [@0812.4946], especially for the discussion about RGE for $\lambda_{H}$.
The threshold correction $\delta\lambda_{H}$ at scale $\tilde{m}=M_{P}/\xi$ arising from integrating out gauginos and squarks can be determined in terms of the differences between the values of $\lambda_{H}$ below and above scale $\tilde{m}=M_{P}/\xi$, $$\begin{aligned}
\label{threshold1}
\lambda_{H}(\tilde{m}-\epsilon)=\lambda_{H}(\tilde{m}+\epsilon)+\delta\lambda_{H}\left(m_{g_{1}},m_{g_{3}}, m_{\tilde{q}_{i}}\right),\nonumber\\\end{aligned}$$ where $0<\epsilon<<1$, and $m_{g_{1}}$, $m_{g_{3}}$, $m_{\tilde{q}_{i}}$ refers to the gluino mass, bino mass and squark masses, respectively. As we will see below, the threshold correction $\delta\lambda_{H}\left(m_{g_{1}},m_{g_{3}}, m_{\tilde{q}_{i}}\right)$ can be measured in high precision, so it is a new and useful factor to constrain the GUT-scale SUSY mass spectrum.
The value of $\lambda_{H}(M_{P}/\xi+\epsilon)$ in Eq.(\[threshold1\]) is determined by the RGE for $\beta_{\lambda_{H}}$ in the effective theory at the intermediate scale, with the boundary value at the end of inflation
$$\begin{aligned}
{\label{bound}}
\lambda_{H}(M_{P}/\sqrt{\xi})=\lambda_{H}(\mu_{I})+\delta \lambda_{H}\left(m_{\tilde{h}_{u,d}}, m_{s}\right).\end{aligned}$$
Here, we have ignored the effects due to higgsino-induced operators with mass dimension higher than four, which are at least one order of magnitude smaller than the threshold correction in our case [^2].
The threshold corretion $\delta \lambda_{H}\left(m_{\tilde{h}_{u,d}}, m_{s}\right)$ in Eq.(\[bound\]) arises from the heavy higgsinos and singlet $s$, which are given by, respectively [@1108.6077], $$\begin{aligned}
{\label{values}}
\delta\lambda_{H}\mid_{\tilde{h}_{u,d}}&\simeq&-\frac{1}{6}\cos^{2}(2\beta)\left(\frac{9}{25}g^{4}_{1}+g^{4}_{2}\right)\text{ln}\left(\frac{\mu^{2}}{\tilde{m}^{2}}\right),\nonumber\\
\delta\lambda_{H}\mid_{s}&\simeq& 2\lambda_{H}(\mu_{I})\sin^{2}(2\beta).\end{aligned}$$ where we have assumed that the A-term $A_{s}$ is smaller than $m_{s}$. Substituting the stabilized value $\beta=\pi/4$ into Eq.(\[values\]) we find that $\lambda_{H}(M_{P}/\sqrt{\xi})\simeq \frac{5}{2}\lambda_{H}(\mu_{I})$. As long as the mass scale $m_{\beta}$, which is of order $M_{P}/\sqrt{\xi}$, is far larger than the Hubble parameter $H$ and the reheating temperature after inflation [@0812.4624], $\beta$ rapidly approaches to the stabilized value during and after inflation.
The matter content of the effective theory at the intermediate scale is composed of SM fermions, squarks, gauginos and electrically neutral scalars in the Higgs sector. The beta function coefficients $b_i$ for the SM gauge coupling in this effective theory are given by, respectively, $$\begin{aligned}
{\label{rges}}
b_{1}=\frac{n}{2}, ~~~~
b_{2}=-\frac{3}{2}+\frac{n}{2},~~~~
b_{3}=-\frac{9}{4}+\frac{n}{2},\end{aligned}$$ where $n=3$ is the number of SM fermion generations. In terms of Eq.(\[rges\]) and one-loop RGEs for $\lambda_{H}$ and SM top Yukawa coupling we show in Fig.\[threshold\] the constraint on the magnitude of $\delta\lambda_{H}\left(m_{g_{1}},m_{g_{3}}, m_{\tilde{q}_{i}}\right)$ for the range $300\leq \xi\leq 50000$. It is shown that the value for $\delta\lambda_{H}\left(m_{g_{1}},m_{g_{3}}, m_{\tilde{q}_{i}}\right)$ as required by Higgs inflation is very sensitive to parameter $\xi$.
In the light of Fig.\[threshold\] we conclude that for small non-minimal coupling $\xi \leq 500$, the threshold correction at scale $M_{P}/\xi$ induced by Higgs inflation is about $-0.03\sim-0.02$ and $-0.05\sim -0.04$ in wino and higgsino/singlino DM, respectively. The uncertainties mainly arise from the uncertainty of top quark pole mass. In the large $\xi$ region with $\xi\geq 1\times 10^{4}$, the required threshold correction is $\leq -0.1$, which is too large to exclude such models.
![Threshold correction $\delta\lambda_{H}\left(m_{g_{1}},m_{g_{3}}, m_{\tilde{q}_{i}}\right)$ in the wino DM model. We have chosen $m_{t}=172.2$ GeV, $m_{h}=125.5$ GeV and $\alpha_{s}(m_{Z})=0.1184$. Similar plot can be obtained for the case with higgsino/singlino DM.[]{data-label="threshold"}](threshold.pdf){width="45.00000%"}
V. Discussion and Summary
=========================
So far we have only discussed the constraint on the threshold correction $\delta\lambda$ at scale $M_{P}/\xi$ in the case for wino DM. In the case for higgsino/singlino DM the one-loop RGEs for relevant couplings at the intermediate scale are the same as in section IV. So the conclusion there is also applied to the model of higgsino/singlino DM. Actually, given the same $\xi$ it is expected that the threshold correction for the case with higgsino/singlino DM is larger than what is shown in Fig.\[threshold\].
One may wonder whether adopting larger $m_{t}$ than $172.2$ GeV can reduce the threshold correction in Fig.\[threshold\] $?$ Fig.\[lambda1\] and Fig.\[lambda2\] have shown that $\lambda_{H}(M_{P}-\epsilon)$ becomes more negative when one chooses larger $m_{t}$, and the value of $\lambda(M_{P}/\xi+\epsilon)$, which is positive, will be smaller simultaneously. Therefore the deviation to the threshold correction is not too obvious to violate our conclusions.
In summary, the answer to the question whether Higgs inflation can be saved by high-scale SUSY critically depends on the magnitude of $\xi$. For small non-minimal coupling $\xi \leq 500$, the threshold correction at scale $M_{P}/\xi$ required by Higgs inflation is constrained in high precision, the magnitude of which is about $-0.03\sim-0.02$ and $-0.05\sim -0.04$ for the wino and higgsino/singlino DM, respectively. The uncertainties mainly arise from the uncertainty of top quark pole mass. This amount of threshold correction can be explained, for example, for universal squark masses $m_{\tilde{q}}\simeq \frac{M_{P}
}{6\xi}$ with vanishing $A_t$ terms. For large $\xi\sim 10^{4}$, the required threshold correction is smaller than $\simeq -0.1$, which excludes such high-scale SUSY. Finally, we believe our analysis can be applied to Split SUSY, although this model is more complicated in compared with what have been discussed here.
acknowledgments
===============
We would like to thank the referee for valuable suggestions. The work is supported in part by National Natural Science Foundation of China under grant No. 11247031 and 11405015.
[99]{} P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], “Planck 2015 results. XIII. Cosmological parameters,” arXiv:1502.01589 \[astro-ph.CO\].
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], “Planck 2015 results. XX. Constraints on inflation,” arXiv:1502.02114 \[astro-ph.CO\].
A. A. Starobinsky, “A New Type of Isotropic Cosmological Models Without Singularity,” Phys. Lett. B [**91**]{}, 99 (1980).
R. Kallosh and A. Linde, “Planck, LHC, and $\alpha$-attractors,” arXiv:1502.07733 \[astro-ph.CO\].
F. L. Bezrukov and M. Shaposhnikov, “The Standard Model Higgs boson as the inflaton,” Phys. Lett. B [**659**]{}, 703 (2008), arXiv:0710.3755 \[hep-th\].
C. P. Burgess, H. M. Lee and M. Trott, “Power-counting and the Validity of the Classical Approximation During Inflation,” JHEP [**0909**]{}, 103 (2009), arXiv:0902.4465 \[hep-ph\].
J. L. F. Barbon and J. R. Espinosa, “On the Naturalness of Higgs Inflation,” Phys. Rev. D [**79**]{}, 081302 (2009), arXiv:0903.0355 \[hep-ph\].
C. P. Burgess, H. M. Lee and M. Trott, “Comment on Higgs Inflation and Naturalness,” JHEP [**1007**]{}, 007 (2010), arXiv:1002.2730 \[hep-ph\].
F. Bezrukov, A. Magnin, M. Shaposhnikov and S. Sibiryakov, “Higgs inflation: consistency and generalisations,” JHEP [**1101**]{}, 016 (2011), arXiv:1008.5157 \[hep-ph\].
S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], “Measurement of the properties of a Higgs boson in the four-lepton final state,” Phys. Rev. D [**89**]{}, no. 9, 092007 (2014). arXiv:1312.5353 \[hep-ex\].
V. Khachatryan [*et al.*]{} \[CMS Collaboration\], “Observation of the diphoton decay of the Higgs boson and measurement of its properties,” Eur. Phys. J. C [**74**]{}, no. 10, 3076 (2014), arXiv:1407.0558 \[hep-ex\].
G. Aad [*et al.*]{} \[ATLAS Collaboration\], “Measurement of the Higgs boson mass from the $H\rightarrow \gamma\gamma$ and $H \rightarrow ZZ^{*} \rightarrow 4\ell$ channels with the ATLAS detector using 25 fb$^{-1}$ of $pp$ collision data,” Phys. Rev. D [**90**]{}, no. 5, 052004 (2014), arXiv:1406.3827 \[hep-ex\].
J. Elias-Miro [*et al.*]{}, “Stabilization of the Electroweak Vacuum by a Scalar Threshold Effect,” JHEP [**1206**]{}, 031 (2012), arXiv:1203.0237 \[hep-ph\].
H. J. He and Z. Z. Xianyu, “Extending Higgs Inflation with TeV Scale New Physics,” JCAP [**1410**]{}, 019 (2014), arXiv:1405.7331 \[hep-ph\].
S. Zheng, “On Dark Matter Selected High-Scale Supersymmetry,” JHEP [**1503**]{}, 062 (2015), arXiv:1409.2939 \[hep-ph\].
D. Baumann and D. Green, “Signatures of Supersymmetry from the Early Universe,” Phys. Rev. D [**85**]{}, 103520 (2012), arXiv:1109.0292 \[hep-th\].
S. Zheng, “The Early Universe with High-Scale Supersymmetry,” arXiv:1409.7462 \[hep-ph\].
S. Ferrara [*et al.*]{}, “Superconformal Symmetry, NMSSM, and Inflation,” Phys. Rev. D [**83**]{}, 025008 (2011), arXiv:1008.2942 \[hep-th\].
I. Ben-Dayan and M. B. Einhorn, “Supergravity Higgs Inflation and Shift Symmetry in Electroweak Theory,” JCAP [**1012**]{}, 002 (2010), arXiv:1009.2276 \[hep-ph\].
M. B. Einhorn and D. R. T. Jones, “Inflation with Non-minimal Gravitational Couplings in Supergravity,” JHEP [**1003**]{}, 026 (2010), arXiv:0912.2718 \[hep-ph\].
S. Ferrara [*et al.*]{}, “Jordan Frame Supergravity and Inflation in NMSSM,” Phys. Rev. D [**82**]{}, 045003 (2010), arXiv:1004.0712 \[hep-th\].
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], “Planck 2013 results. XVI. Cosmological parameters,” Astron. Astrophys. [**571**]{}, A16 (2014), arXiv:1303.5076 \[astro-ph.CO\].
J. Garcia-Bellido, D. G. Figueroa and J. Rubio, “Preheating in the Standard Model with the Higgs-Inflaton coupled to gravity,” Phys. Rev. D [**79**]{}, 063531 (2009), arXiv:0812.4624 \[hep-ph\].
G. F. Giudice and A. Strumia, “Probing High-Scale and Split Supersymmetry with Higgs Mass Measurements,” Nucl. Phys. B [**858**]{}, 63 (2012), arXiv:1108.6077 \[hep-ph\].
A. De Simone, M. P. Hertzberg and F. Wilczek, “Running Inflation in the Standard Model,” Phys. Lett. B [**678**]{}, 1 (2009), arXiv:0812.4946 \[hep-ph\].
C. P. Burgess, S. P. Patil and M. Trott, “On the Predictiveness of Single-Field Inflationary Models,” JHEP [**1406**]{}, 010 (2014), arXiv:1402.1476 \[hep-ph\].
E. E. Jenkins, A. V. Manohar and M. Trott, “Renormalization Group Evolution of the Standard Model Dimension Six Operators I: Formalism and lambda Dependence,” JHEP [**1310**]{}, 087 (2013), arXiv:1308.2627 \[hep-ph\].
[^1]: Correspondingly, the scalar potential in the Jordan frame is given by $V_{J}=\Omega^{2}V_{E}$.
[^2]: Operators of type $c_{HG}\mid H\mid^{2} F_{\mu\nu}F^{\mu\nu}/M^{2}$ with mass dimension six, as induced by heavy SUSY particles with mass $M$, contribute to the leading corrections to the RGEs for SM EW gauge coupling [@1402.1476; @1308.2627]. As a result, they modify the RGE for $\lambda_H$ indirectly, the significance of which is determined by the ratio $(m_{h}^{2})_{\text{eff}}/M^{2}$, as seen from the modified beta function to SM gauge coupling, $16\pi^{2}\delta \beta_{g_{i}}\sim c_{HG}g_{i}(m_{h}^{2})_{\text{eff}}/M^{2}$. For the three stages we set in this paper, the modification to the RGE for $\lambda_H$ during inflation is the maximal, with heavy SUSY particle identified as $\tilde{h}_{\mu,d}$. In contrast, the modification after inflation is only mild. In our case, $(m_{h}^{2})_{\text{eff}}\leq H$ and $M\sim M_{P}$, which implies that this modification is smaller than the threshold correction $\delta\lambda_{H}(m_{\tilde{h}_{u,d}}, m_{s})$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Sign problems in path integrals arise when different field configurations contribute with different signs or phases. Phase unwrapping describes a family of signal processing techniques in which phase differences between elements of a time series are integrated to construct noncompact unwrapped phase differences. By combining phase unwrapping with a cumulant expansion, path integrals with sign problems arising from phase fluctuations can be systematically approximated as linear combinations of path integrals without sign problems. This work explores phase unwrapping in zero-plus-one-dimensional complex scalar field theory. Results with improved signal-to-noise ratios for the spectrum of scalar field theory can be obtained from unwrapped phases, but the size of cumulant expansion truncation errors is found to be undesirably sensitive to the parameters of the phase unwrapping algorithm employed. It is argued that this numerical sensitivity arises from discretization artifacts that become large when phases fluctuate close to singularities of a complex logarithm in the definition of the unwrapped phase.'
author:
- William Detmold
- Gurtej Kanwar
- 'Michael L. Wagman'
bibliography:
- 'refs.bib'
title: 'Phase Unwrapping and One-Dimensional Sign Problems'
---
[^1]
[^2]
[^3]
Introduction
============
If the properties of quantum states with large baryon number could be calculated directly from the Standard Model, open questions could be answered regarding the boundaries of the periodic table, the composition of neutron stars, and the interpretation of low-energy experimental searches for beyond-the-Standard-Model interactions in nuclear targets. Many electroweak effects can be calculated accurately in perturbation theory, but at low energies relevant to nuclear systems the effects of the strong nuclear force described by quantum chromodynamics (QCD) can only be accurately computed nonperturbatively. Lattice Quantum Field Theory (LQFT) provides a method for nonperturbatively calculating path integrals in many QFTs in which ultraviolet divergences have been regularized by replacing spacetime with a discrete lattice of points and infrared divergences have been regularized by restricting spacetime to a finite volume. Renormalized QFT observables are obtained from the continuum and infinite-volume limits of LQFT results. Monte Carlo (MC) methods can be used to calculate path integrals representing observables in lattice QCD (LQCD) and other LQFTs by stochastically sampling field configurations from appropriately chosen probability distributions and averaging observables over quantum fluctuations. If the probability distribution used for MC sampling is proportional to the contribution of each field configuration to the thermal partition function, then equilibrium thermodynamic observables can be computed from MC ensemble averages.
Sign problems arise in MC calculations when contributions of different field configurations to path integrals have different signs or more generally when path integral contributions are complex and have different phases. For example, the partition function for QCD at nonzero baryon chemical potential has a sign problem and cannot be used to define a probability distribution for MC simulations of nonzero baryon density systems. When a partition function has a sign problem, one can instead MC sample according to a different probability distribution and then attempt to reweight the contribution of each field configuration by the ratio of the desired complex weight to the positive weight used for MC importance sampling. In reweighting approaches to the baryon chemical potential sign problem, the signal-to-noise (StN) ratio of the reweighting factor, that is the average reweighting factor divided by the square root of its variance, vanishes exponentially quickly as the spacetime volume is taken to infinity [@Gibbs:1986ut; @Cohen:2003kd; @Cohen:2003ut; @Splittorff:2006fu; @Splittorff:2006vj; @Splittorff:2007ck; @deForcrand:2010ys; @Alexandru:2014hga]. Sign problems can arise for particular observables even when the partition function does not have a sign problem if different contributions to the path integrals representing these observables have different phases. Ensemble averages of the observable may then have small StN ratios analogous to the StN ratios of complex reweighting factors for nonzero baryon density partition functions.
Baryon and nuclear correlation functions have StN ratios that decrease exponentially at a rate predicted by the moment analysis of Parisi [@Parisi:1983ae] and Lepage [@Lepage:1989hd] when the total baryon number integrated over spacetime is increased. Many LQCD calculations of baryon and nuclear correlation functions rely on a golden window of intermediate source/sink separations where signals are consistent with single-state behavior but this StN problem is not too severe [@Beane:2009kya; @Beane:2009gs; @Beane:2009py; @Beane:2010em; @Beane:2011pc; @Beane:2014oea; @Detmold:2014hla]. The baryon correlation function StN problem arises from phase fluctuations that lead to sign problems in correlation functions and is absent when phase fluctuations are ignored [@Wagman:2016bam]. The average values of baryon and nuclear correlation functions decay exponentially faster in the limit of large source/sink separation than their average magnitudes, which demonstrates that exponentially precise cancellations must occur between different contributions to MC ensemble averages and signals an exponentially severe StN problem. LQCD simulations of mesons and baryons with the quark masses tuned to reproduce experimental observables are being used today to predict increasingly complex observables directly from the Standard Model. However, LQCD simulations of nuclei are still generally limited to simulations of small (baryon number $B=1-5$) nuclei with quark masses tuned to unphysically heavy values [@Beane:2006mx; @Yamazaki:2009ua; @Doi:2011gq; @Beane:2015yha; @Berkowitz:2015eaa; @Chang:2015qxa; @Doi:2015oha; @Detmold:2015daa; @Orginos:2015aya; @Yamazaki:2015asa; @Yamazaki:2015vjn; @Parreno:2016fwu; @Savage:2016kon; @Chang:2017eiq; @Doi:2017cfx; @Doi:2017zov; @Gongyo:2017fjb; @Nemura:2017bbw; @Shanahan:2017bgi; @Tiburzi:2017iux; @Wagman:2017tmp; @Winter:2017bfs; @Francis:2018qch; @Iritani:2018zbt]. Although physical quark mass simulations of small nuclei are becoming computationally feasible, current methods for calculating nuclear correlation functions require computational resources that grow exponentially with baryon number. These problems motivate the study of alternative approaches to calculating path integrals with sign problems arising from phase fluctuations.
A zero-plus-one-dimensional ($(0+1)D$) complex scalar field StN problem arising from phase fluctuations in correlation functions with nonzero $U(1)$ charge is studied in this work as a toy model of the baryon correlation function StN problem in LQCD.[^4] Scalar field phase fluctuations are found to qualitatively resemble the LQCD baryon correlation function phase fluctuations described in Ref. [@Wagman:2016bam] and in particular are shown to be wrapped normally distributed and have exponential StN problems in an analytically tractable approximation where magnitude fluctuations are neglected and phase fluctuations are assumed to be small. Analytically integrating over phase fluctuations using a change of variables similar to the dual lattice variables employed in Ref. [@Endres:2006xu] enables calculations of correlation functions that avoid sign and StN problems in $(0+1)D$ complex scalar field theory, but it remains challenging to extend similar analytic integration methods to more complicated theories such as LQCD [@Chandrasekharan:2008gp; @deForcrand:2010ys; @Gattringer:2012df; @Gattringer:2012ap; @Gattringer:2014nxa; @Gattringer:2016kco; @Giuliani:2017qeo]. Instead, a new method is explored in this work in which phase differences are “unwrapped,” or numerically integrated over a series of spacetime separations. The resulting unwrapped phases are noncompact random variables rather than circular random variables defined modulo $2\pi$. Moments of unwrapped phase differences can be calculated from positive-definite path integrals that do not have sign problems and do not generically require computational resources that increase exponentially with increasing $U(1)$ charge. Correlation functions can be calculated from moments of unwrapped phases using cumulant expansion techniques similar to those of Ref. [@Endres:2011jm], although beyond leading order in the cumulant expansion sign and StN problems can reemerge from differences of cumulants.[^5] Phase unwrapping in conjunction with this cumulant expansion allows generic complex correlation functions with nonzero $U(1)$ charge to be represented by series of path integrals without sign problems; however, it is shown below that finding a robust numerical implementation of phase unwrapping is challenging even in $(0+1)D$ complex scalar field theory.
The phase unwrapping techniques used here to map series of random phases to series of noncompact random variables are analogous to phase unwrapping techniques used in signal processing, radar interferometry, x-ray crystallography, magnetic resonance imaging, and other areas of science and engineering [@Judge:1994; @Ghiglia:98; @Ying:2006; @Kitahara:2015]. The idea of phase unwrapping correlation functions in LQFT was briefly mentioned in Ref. [@Wagman:2016bam] but its numerical implementation faces challenges arising from large phase fluctuations and was not pursued in detail. This work presents a detailed study of phase unwrapping in $(0+1)D$ complex scalar field theory in order to explore its potential applicability in physically relevant higher-dimensional LQFTs. The $1D$ phase unwrapping algorithms studied here can be immediately applied to three-momentum-projected correlation functions in LQCD and other higher-dimensional LQFTs; however, $1D$ phase unwrapping algorithms generically suffer from numerical instabilities and do not provide a robust solution to LQFT sign and StN problems. These numerical instabilities are argued to arise from an accumulation of phase unwrapping ambiguities related to large phase jumps that occur with non-negligible probability even on lattices with very fine levels of discretization in $(0+1)D$ complex scalar field theory. Multidimensional phase unwrapping algorithms are known to avoid analogous numerical instabilities, and more robust phase unwrapping algorithms might be achieved in future investigations of phase unwrapping in multidimensional LQFTs.
The remainder of this work is organized as follows. LQFT for $(0+1)D$ complex scalar fields is reviewed in Sec. \[sec:scalarstats\]. After discussing sign and StN problems in free complex scalar field theory in Sec. \[sec:freestats\], analytic integration over phase fluctuations and its effect on sign and StN problems is discussed in Sec. \[sec:dual\] and is used to derive wrapped phase statistics in Sec. \[sec:wrappedstats\] that confirm the phase fluctuation origin of the sign and StN problems. Phase unwrapping is introduced in Sec. \[sec:unwrappedstats\]. A cumulant expansion method for relating wrapped and unwrapped phase distributions is introduced in Sec. \[sec:cumulant\]. Numerical studies comparing 1D phase unwrapping and the cumulant expansion, analytic phase integration, and standard MC methods are presented in Sec. \[sec:unwrap1D\]. After brief remarks on applications to higher-dimensional LQFTs in Sec. \[subsec:cho-results\], conclusions are presented in Sec. \[sec:conclusions\].
Complex scalar field statistics {#sec:scalarstats}
===============================
Sign and signal-to-Noise problems {#sec:freestats}
---------------------------------
Consider a complex scalar field $\varphi(t)$ defined on a uniform lattice of points $t=0,\dots,L-1$ representing a discretized $(0+1)D$ flat Euclidean spacetime. Units where the lattice spacing is set to unity are used throughout. Periodic boundary conditions (PBCs) are imposed on $\varphi$, and $L$ is assumed to be even for simplicity. The free complex scalar field action is then given by $$\begin{split}
S(\varphi) &\equiv \sum_{t=0}^{L-1} \left( \varphi^*(t+1) - \varphi^*(t) \right)\left( \varphi(t+1) - \varphi(t) \right) + M^2 |\varphi(t)|^2 \\
&= \sum_{n=-L/2}^{L/2-1} |\varphi(n)|^2\left( 4 \sin^2\left( \frac{n \pi}{L} \right) + M^2 \right),
\end{split}\label{eq:Sdef}$$ where the $L$ dependence of the action and other quantities below is left implicit, $\varphi(L) \equiv \varphi(0)$ by PBCs, and the last expression introduces a discrete Fourier transform $$\begin{split}
\varphi(n) \equiv \frac{1}{\sqrt{L}}\sum_{t=0}^{L-1} \varphi(t)e^{-2\pi i n t /L}, \hspace{20pt} \varphi(t) = \frac{1}{\sqrt{L}}\sum_{n=-L/2}^{L/2-1} \varphi(n)e^{2\pi i n t /L}.
\end{split}\label{eq:phidft}$$ The partition function can be evaluated by Gaussian integration in polar coordinates $\varphi(n) = |\varphi(n)|e^{i\theta(n)}$, $$\begin{split}
Z &\equiv \int \mathcal{D}\varphi\; e^{-S(\varphi)} \equiv \int \prod_{t=0}^{L-1} \left[ \frac{1}{\pi}\; d\text{Re}\varphi(t)\; d\text{Im}\varphi(t) \right] e^{-S(\varphi)} \\
&= \prod_{n=-L/2}^{L/2-1} \int_{-\pi}^\pi \frac{1}{\pi}d\theta(n) \int_0^\infty |\varphi(n)|d|\varphi(n)| e^{-|\varphi(n)|^2\left( 4 \sin^2\left( \frac{n\pi}{L} \right) + M^2 \right)} \\
&= \exp\left\lbrace - \sum_{n=-L/2}^{L/2-1} \ln\left[ 4 \sin^2\left( \frac{n \pi}{L} \right) + M^2 \right] \right\rbrace.
\end{split}\label{eq:Zdef}$$ The scalar field propagator is given by $$\begin{split}
G(t) &\equiv \avg{\varphi(t)\varphi^*(0)} \equiv \frac{1}{Z}\int \mathcal{D}\varphi\; e^{-S} \varphi(t)\varphi^*(0) \\
&= \frac{1}{L}\sum_{n=-L/2}^{L/2-1} \frac{e^{2 \pi i n t/L}}{4 \sin^2\left( \frac{n\pi}{L} \right) + M^2}.
\end{split}\label{eq:Gprop}$$ In the limit $L\rightarrow\infty$ the sum can be replaced by an integral that can be evaluated by contour integral methods [@Smit:2002]. Corrections to this form are exponentially suppressed in $L$ and can be evaluated analytically if the LQFT action in Eq. is replaced by its continuum counterpart. This provides a spectral representation for the propagator, $$\begin{split}
G(t) &= Z_{1;0,1} e^{-E t}\left[ 1 + O\left(e^{-E (L-t)}\right) \right],
\end{split}\label{eq:Gspec}$$ where the $L\rightarrow \infty$ complex scalar propagator pole $E$ and residue $Z_{1;0,1}$ are given by $$\begin{split}
E = 2\; \text{arcsinh}\left( \frac{M}{2} \right), \hspace{20pt} Z_{1;0,1} = \avg{|\varphi|^2} = \frac{1}{M \sqrt{4 + M^2}}.
\end{split}\label{eq:lattpole}$$
The free complex scalar field action Eq. has a $U(1)$ symmetry $$\begin{split}
\varphi \rightarrow e^{-i\alpha}\varphi,\hspace{20pt} \varphi^* \rightarrow e^{i\alpha}\varphi^*.
\end{split}\label{eq:U1def}$$ This $U(1)$ symmetry can be used to classify sectors of states in the LQFT Hilbert space that do not mix under (Euclidean) time evolution. The charge of the vacuum is defined to be $Q=0$. Field products of the form $$\begin{split}
\mathcal{O}_{Q,2P}(t) \equiv \begin{cases} \varphi(t)^{|Q|} |\varphi(t)|^{2P}, & Q \geq 0 \\ {\varphi^*(t)}^{|Q|} |\varphi(t)|^{2P} , & Q < 0 \end{cases}
\end{split}\label{eq:Odef}$$ transform under $U(1)$ in the charge $Q$ representation, $$\begin{split}
\mathcal{O}_{Q,2P} \rightarrow e^{-iQ \alpha}\mathcal{O}_{Q,2P}.
\end{split}\label{eq:OU1}$$ In operator language, the $\mathcal{O}_{Q,2P}$ for all $Q, P \in \mathbb{Z}$ with $P \geq 0$ form a complete basis of interpolating operators for Hilbert space states connected to the Euclidean vacuum by polynomial functions of field operators. Two-point correlation functions involving these operators can be evaluated in terms of the scalar propagator by Wick’s theorem as $$\begin{split}
G_{Q,2P}(t) &\equiv \avg{ \mathcal{O}_{Q,2P}(t) \mathcal{O}_{Q,2P}^*(0) } \\
&= P!(|Q|+P)!\sum_{k=0}^P \binom{P}{k} \binom{|Q| + P}{k} G(t)^{|Q| + 2P - 2k} G(0)^{2k} \\
&= P!(|Q|+P)!\avg{|\varphi|^2}^{|Q| + 2P} \sum_{k=0}^P \binom{P}{k} \binom{|Q| + P}{k} e^{-(|Q|+2P-2k)Et}\left[ 1 + O\left(e^{-E (L-t)}\right) \right].
\end{split}\label{eq:GQP}$$ Comparing to the general spectral representation guaranteed by unitarity, $$\begin{split}
G_{Q,2P}(t) \equiv \sum_n Z_{n;Q,2P} e^{-E_n t}\left[ 1 + O\left(e^{-E (L-t)}\right) \right],
\end{split}\label{eq:GQPspec}$$ the energies of states produced by $\mathcal{O}_{Q,2P}$ are seen to be $E_n \in \{|Q|E,\ (|Q|+2)E,\ \dots,\ (|Q|+2P)E\}$, and the overlaps $Z_{n;Q,2P}$ of the $n$th energy eigenstate can be determined straightforwardly. The full spectrum of the theory includes energies $E_m = m E$ for all integers $m\geq 0$. Each eigenvalue $E_m$ appears in $(m+1)$ different charge sectors $Q = -m,\ -m+2,\ \dots,\ m-2,\ m$, signaling an $(m+1)$-fold degeneracy of energy eigenstates.
Since the action in Eq. is real, $e^{-S}$ is a positive-definite function that can be interpreted as a probability distribution $$\begin{split}
\mathcal{P}(\varphi) \equiv \frac{1}{Z} e^{-S(\varphi)}.
\end{split}\label{eq:CHOprob}$$ MC methods can be used to produce stochastic samples of complex scalar fields distributed by Eq. using for example the Metropolis algorithm.[^6] After generating a MC ensemble of field configurations $\varphi_i$ with $i = 1,\dots,N$ and $N\gg 1$, the scalar field propagator can be determined by approximating the path integral with the ensemble average $$\begin{split}
\overline{G}(t) = \frac{1}{N} \sum_{i=1}^N \varphi_i(t){\varphi_i}^*(0) = G(t)\left[ 1 + O\left( \sqrt{\frac{\text{Var}(G(t))}{N}} \right) \right].
\end{split}\label{eq:sampleG}$$ At fixed $t$ and asymptotically large $N$, convergence of the sample mean to the true average with $1/\sqrt{N}$ scaling is guaranteed (neglecting MC autocorrelations) by the Berry–Esseen theorem since all moments of $\mathcal{P}(\varphi)$ are finite. The $1/\sqrt{N}$ prefactor of $\text{Var}(G(t))$ describes the size of statistical errors at asymptotically large $N$ according to the central limit theorem but may only provide a rough guide to the error scaling for arbitrary $\mathcal{P}(\varphi)$ and finite $N$. It is noteworthy that $G(t)$ is real by Eq. , but individual samples $\varphi_i(t){\varphi_i}^*(0)$ and $\overline{G}(t)$ at finite $N$ are complex. At large $N$ the real part of $\overline{G}(t)$ converges to $G(t)$ as in Eq. and the imaginary part of $\overline{G}(t)$ converges to zero with similar $1/\sqrt{N}$ scaling. Ensemble average estimators can similarly be constructed for general correlation functions, $$\begin{split}
\overline{G}_{Q,2P}(t) &= \frac{1}{N}\sum_{i=1}^N \mathcal{O}^i_{Q,2P}(t)\mathcal{O}^i_{-Q,2P}(0) \\
&= G_{Q,2P}(t)\left[ 1 + O\left( \sqrt{\frac{\text{Var}(G_{Q,2P}(t))}{N}} \right) \right],
\end{split}\label{eq:sampleGQP}$$ where $\mathcal{O}^i_{Q,2P}$ is defined by Eq. with $\varphi$ replaced by $\varphi_i$ and $1/\sqrt{N}$ error scaling follows from the Berry-Esseen and central limit theorems. Ground-state energies can be determined from the large-$t$ behavior of the effective masses derived from ensemble average correlation functions, $$\begin{split}
E_{Q,2P}(t) \equiv - \partial_t \ln \left(\overline{G}_{Q,2P}(t)\right) \equiv - \ln\left( \overline{G}_{Q,2P}(t+1) \right) + \ln\left( \overline{G}_{Q,2P}(t) \right).
\end{split}\label{eq:EQ2P}$$
![ The left plot shows MC ensemble average ground-state energies $E_{Q,0}$ of $U(1)$ charge sectors $Q=1,\dots,4$ in $(0+1)D$ complex scalar field theory. Analytic results valid in the $L\rightarrow \infty$ limit from Eq. are shown as red lines. The right plot shows the variance in these ground-state energies. The red lines show the theoretically predicted $e^{2E_{Q,0} t}$ scaling. Error bars denote $68\%$ confidence intervals determined by bootstrap resampling correlation functions calculated from $L$ source points on $N = 5000$ MC field configurations of complex scalar fields with $M^2 = 0.00625$ and $L=512$ generated using the Metropolis algorithm. This ensemble is denoted $C_0$ below, see Sec. \[sec:unwrap1D\] and Appendix \[app:MC\] for more details.[]{data-label="fig:freeQ"}](free_Q.png){width="\textwidth"}
Following standard Parisi-Lepage arguments [@Parisi:1983ae; @Lepage:1989hd], the variance of $G_{Q,2P}$ can be described by a linear combination of correlation functions $$\begin{split}
\text{Var}(\text{Re}[G_{Q,2P}(t)]) &\equiv \frac{1}{4}\avg{ \Big( \mathcal{O}_{Q,2P}(t)\mathcal{O}_{Q,2P}(0)^* + \mathcal{O}_{Q,2P}(t)^*\mathcal{O}_{Q,2P}(0) \Big) ^2 } - G_{Q,2P}(t)^2 \\
&= \frac{1}{2} G_{0,2|Q| + 4P}(t) + \frac{1}{2}G_{2Q,4P}(t) - G_{Q,2P}(t)^2,
\end{split}\label{eq:GQPvar}$$ where $G_{-Q,2P} = G_{Q,2P}$ has been used following Eq. . The variance of $\overline{G}_{Q,2P}$ is related to the variance of $G_{Q,2P}$ by $1/\sqrt{N}$ in the large-$N$ limit by the central limit theorem, giving at large $N$ $$\begin{split}
\text{StN}(\text{Re}[\overline{G}_{Q,2P}(t)]) &\equiv \frac{G_{Q,2P}(t)}{\sqrt{\text{Var}(\text{Re}[\overline{G}_{Q,2P}(t)])}} \\
&= \sqrt{N} \frac{G_{Q,2P}(t)}{\sqrt{ \frac{1}{2} G_{0,2Q + 4P}(t) + \frac{1}{2}G_{2Q,4P}(t) - G_{Q,2P}(t)^2 }} + O(N^0) \\
&= \sqrt{2 N} e^{-Q E t} \left[ 1 + O\left( e^{-2 E t} \right) + O(N^{-1/2}) \right].
\end{split}\label{eq:GPQStN}$$ Correlation functions describing sectors with $U(1)$ charge $Q\neq 0$ face an exponentially severe StN problem where the exponent is proportional to the $U(1)$ charge of the system. This result is confirmed numerically in MC results shown in Fig. \[fig:freeQ\].
In LQCD, baryon correlation functions similarly face an exponentially severe StN problem whose exponent is proportional to $U(1)_B$ baryon number charge. MC studies indicate that this StN problem is related to the sign problem caused by correlation function phase fluctuations [@Wagman:2016bam]. Analogous features can be seen analytically in free complex scalar field theory correlation functions. A magnitude-phase decomposition of the complex scalar field $$\begin{split}
\varphi(t) = e^{R(t) + i\theta(t)}, \hspace{20pt} R(t) \equiv \ln|\varphi(t)|, \hspace{20pt} \theta(t) \equiv \text{arg}(\varphi(t)),
\end{split}\label{eq:phimagphase}$$ and an analogous decomposition of the scalar boson propagator and correlation functions $$\begin{split}
G(t) \equiv \avg{C(t)} &= \avg{e^{\mathcal{R}(t) + i \Theta(t)}}, \hspace{20pt} \mathcal{R}(t) \equiv R(t) + R(0), \hspace{20pt} \Theta(t) \equiv \theta(t) - \theta(0),
\end{split}\label{eq:Gmagphasedef}$$ can be inserted into the path integral representation of the propagator, Eq. , to give $$\begin{split}
G(t) = \frac{1}{Z}\int \mathcal{D} \varphi\; e^{-S + \mathcal{R}(t) + i\Theta(t)}.
\end{split}\label{eq:Gmagphase}$$ Fluctuations of the scalar field phase give scalar boson propagators a sign problem. The path integrand in Eq. is not positive-definite and cannot be interpreted as a probability distribution in MC simulations.[^7]
Applying a similar decomposition to generic correlation functions gives $$\begin{split}
G_{Q,2P}(t) = \frac{1}{Z}\int \mathcal{D}\varphi\; e^{-S + (2P+|Q|)\R(t) + iQ\Theta(t)}.
\end{split}\label{eq:GQPmagphase}$$ Free complex scalar field correlation functions have a sign problem if they describe states with $U(1)$ charge $Q \neq 0$. Identical considerations apply to the StN problem in Eq. , demonstrating that sign and StN problems for complex scalar field correlation functions both arise in the presence of nonzero $U(1)$ charge. The average of the correlation function magnitude, $$\begin{split}
\avg{ |\mathcal{O}_{Q,2P}(t)\mathcal{O}_{Q,2P}^*(0)| } &= \avg{ e^{(|Q|+2P)\R (t)} } = \avg{ |\varphi(t)|^{|Q|+2P}|\varphi(0)|^{|Q|+2P}} ,
\end{split}\label{eq:magprop}$$ depends nonanalytically on the Fourier modes $\varphi(n)$ but can be calculated simply in MC studies of $(0+1)D$ complex scalar field theory. As shown in Fig. \[fig:freeStN\], the scalar boson propagator magnitude is $O(1)$ both sample-by-sample and in expectation with no severe StN problem. Analogous behavior is seen for the magnitudes of generic correlation functions. The phase of the scalar boson propagator is $O(1)$ sample-by-sample by definition but $O(e^{-Et})$ in expectation with a severe $O(e^{-Et})$ StN problem as shown in Fig. \[fig:freeStN\]. The phase of a general correlation function depends only on the $U(1)$ charge of the correlation function $$\begin{split}
\Theta_Q(t) \equiv \text{arg}\left[ \mathcal{O}_{Q,2P}(t)\mathcal{O}_{Q,2P}^*(0) \right] = iQ \Theta(t),
\end{split}\label{eq:phaseprop}$$ and $\avg{e^{i\Theta_Q}}$ has both an expectation value and a StN problem of $O(e^{-Q E t})$ as shown in Fig. \[fig:freeStN\].
![The left plot shows a magnitude-phase decomposition of the scalar boson propagator using the same MC ensemble $C_0$ of free $(0+1)D$ complex scalar field configurations as in Fig. \[fig:freeQ\]. Ensemble average calculations of the propagator $\avg{\varphi(t)\varphi^*(0)}$ are shown in blue, calculations of the propagator magnitudes $\avg{|\varphi(t)\varphi(0)|}$ are shown in orange, and calculations of the propagator phase factor $\avg{e^{i\theta(t) - i\theta(0)}}$ are shown in green. The right plot shows the corresponding effective masses $E = -\partial_t \ln \avg{\varphi(t)\varphi^*(0)}$ in blue, $E_R = -\partial_t \ln \avg{|\varphi(t)\varphi(0)|}$ in orange, and $E_\theta = -\partial_t \ln \avg{e^{i\theta(t) - i \theta(0)}}$ in green.[]{data-label="fig:freeStN"}](free_polar_M0p1.png){width="\textwidth"}
Additional StN degradation is present in calculations of excited-state energies. Correlation functions $G_{0,2P}$, which include both vacuum and $Q=0$ excited-state contributions, have qualitatively similar behavior to $G_{0,1} = \avg{ |\varphi(t)\varphi(0)| }$ in Fig. \[fig:freeStN\] with $O(1)$ signal and root-mean-square variance independent of $t$. After subtracting vacuum contributions, connected $Q=0$ correlation functions $G^{conn}_{0,2P}$ are given by $$\begin{split}
G^{conn}_{0,2P} &\equiv \avg{|\varphi(t)\varphi(0)|^{2P}} - \avg{|\varphi(t)|^{2P}}\avg{|\varphi(0)|^{2P}} \\
&= (P!)^2\avg{|\varphi|^2}^{2P} \sum_{k=1}^P { P \choose P-k } e^{-2k E t} \\
&\sim e^{-2 E t},
\end{split}\label{eq:GPconn}$$ where $\sim$ denotes proportionality at large $t$. Connected $Q=0$ correlation functions are $O(e^{-2 E t})$ in expectation but $O(1)$ sample-by-sample and have a StN problem identical to $Q=2$ charged correlation functions. Eq. provides a simple example of a sign problem: expectation values of $O(1)$ random variables must cancel to exponentially increasing precision in order to achieve constant precision in calculations of $G^{conn}_{0,2P}$ at increasing $t$. Interpreting the vacuum-subtracted correlation functions $G^{conn}_{0,2P}$ as belonging to the $E_n \neq 0$ “charge sector,” excited-state spectroscopy in free $(0+1)D$ scalar field theory can be interpreted as possessing a sign problem and StN problem associated with exponentially small differences of averages that is associated with the presence of nonzero energy “charge.” This appearance of a sign problem and StN problem for interpolating operators overlapping on to excited-states of the $Q=0$ vacuum sector is analogous to the exponentially worse StN problem faced by excited-state correlation functions than ground-state correlation functions in LQCD when excited-state correlation functions are constructed from nonpositive linear combinations of correlation functions with the same quantum numbers.
Dual lattice variable sign problem solution {#sec:dual}
-------------------------------------------
Generalizing Eq. to include an arbitrary $U(1)$ invariant potential $V(|\varphi|)$ and inserting the magnitude-phase decomposition of Eq. gives $$\begin{split}
S(\varphi) = \sum_{t=0}^{L-1} \left\lbrace 2|\varphi(t)|^2 - \kappa(t) \cos(\Delta(t)) + V(|\varphi(t)|) \right\rbrace
\end{split}\label{eq:actiondecomp}$$ with $$\begin{split}
\kappa(t) \equiv 2 |\varphi(t)||\varphi(t-1)|
\qquad \text{and} \qquad
\Delta(t) \equiv \theta(t) - \theta(t-1).
\end{split}\label{eq:kappaDeltaDefn}$$ The partition function can be factored into magnitude and phase contributions as $$\begin{split}
Z &= \int_0^\infty \prod_{t= 0}^{L-1} \left[
d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \right]
\int_{-\pi}^\pi \prod_{t= 0}^{L-1} \left[
\frac{1}{\pi }d \theta(t) \; e^{\kappa(t) \cos(\Delta(t))} \right].
\end{split}\label{eq:Zdecomp}$$ It is shown in Appendix \[app:dual\] that the integral over phase fluctuations in Eq. can be performed analytically using a change of variables analogous to the dual lattice variables used in Ref. [@Endres:2006xu]. Analytically integrating over phase fluctuations provides a positive-definite representation for correlation functions, $$\begin{aligned}
G_{Q,2P}(t) &= \sum_{q \in \mathbb{Z}} \int \mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) \; |\varphi(t)|^{|Q|+2P} |\varphi(0)|^{|Q| + 2P} \prod_{t^\prime = 1}^t \frac{I_{|Q + q|}(\kappa(t))}{I_0(\kappa(t))} \prod_{t^\prime = t+1}^L \frac{I_{|q|}(\kappa(t))}{I_0(\kappa(t))},
\end{aligned}\label{eq:GQPdualrep}$$ where the $I_q$ are modified Bessel functions, $\mathcal{P}_0$ is a probability distribution suitable for MC sampling of scalar field magnitudes after phase fluctuations have been analytically integrated out, $$\begin{split}
\mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) = \frac{2}{Z_0} \prod_{t=0}^{L-1} \left[ |\varphi(t)| d|\varphi(t)| e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \; 2\; I_{0}\left( \kappa(t) \right) \right],
\end{split}\label{eq:Pprimedef}$$ and $Z_0$ is a normalization coefficient defined by $\int \mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) = 1$. The explicit sum over scalar field phase winding numbers $q$ is included in Eq. in order to avoid “topological freezing” arising with stochastic sampling over winding numbers; see Appendix \[app:dual\] for details. Significant contributions to Eq. arise for $q = -Q,\dots,+Q$ but topological charge sectors with $|q| > |Q|$ make subdominant contributions that rapidly converge to zero and allow the sum over topological charge sector to be truncated in practical calculations. Given a finite MC ensemble of scalar field magnitude $|\varphi_i|$, $i=1,\dots,N$ sampled from Eq. , correlation functions can be estimated from the corresponding ensemble averages $$\begin{aligned}
\overline{G}_{Q,2P}^{dual}(t) &= \frac{1}{N} \sum_{i=1}^{N}
\sum_{q \in \mathbb{Z}} \Bigg\{
|\varphi_i(t)|^{|Q|+2P} |\varphi_i(0)|^{|Q| + 2P} \prod_{t^\prime = 1}^t \frac{I_{|Q + q|}(\kappa(t))}{I_0(\kappa(t))} \prod_{t^\prime = t+1}^L \frac{I_{|q|}(\kappa(t))}{I_0(\kappa(t))} \Bigg\},
\end{aligned}\label{eq:GQ2Pdual}$$ where $\overline{G}_{Q,2P}^{dual}$ denotes ensemble average calculations of $G_{Q,2P}$ in this dual-variable approach.
![ The left plot shows the standard effective mass of the scalar boson propagator $E = -\partial_t \ln \overline{G}_{1,0}$ in gray as well as the dual effective mass $E_{dual} = -\partial_t \ln \overline{G}^{dual}_{1,0}$ in purple calculated with MC sampling of the dual representation defined in Eq. - Eq. where phase fluctuations have been integrated out analytically. The right plot shows bootstrap estimates of the variance of the effective mass for the same dual MC estimate of the propagator in purple as well as estimates of the variance of the standard MC propagator shown in Fig. \[fig:freeStN\] in gray. For both standard and dual representations, error bars denote $68\%$ confidence intervals determined by bootstrap resampling correlation functions calculated using $5000$ field configurations with $M^2 = 0.00625$ and $L=512$. The same MC ensemble $C_0$ as in Fig. \[fig:freeStN\] is used to determine $\overline{G}_{1,0}$.[]{data-label="fig:dualStN"}](dual_M0p1.png){width="\textwidth"}
The variance of correlation functions after integrating over dual lattice phase variables is given by $$\begin{split}
\frac{1}{\sqrt{N}}\text{Var}\left( \overline{G}_{Q,2P}^{dual} \right) &= -G_{Q,2P}^2 + \sum_{q \in \mathbb{Z}} \int \mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) \; |\varphi(t)|^{2|Q|+4P} |\varphi(0)|^{2|Q| + 4P} \prod_{t^\prime = 1}^t \frac{I_{|Q + q|}(\kappa(t))^2}{I_0(\kappa(t))^2} \prod_{t^\prime = t+1}^L \frac{I_{|q|}(\kappa(t))^2}{I_0(\kappa(t))^2} \\
&\sim e^{-2|Q|Et},
\end{split}\label{eq:Gdualvar}$$ where the scaling estimate arises from counting each positive-definite factor of $I_{|Q|}/I_0$ as $O(e^{-|Q|E})$ for $t \ll L$ and $1/\sqrt{N}$ corrections have been neglected. This suggests that charged scalar correlation functions with analytically integrated dual phase variables avoid both the sign problem for importance sampling correlation functions and the $O(e^{-QEt})$ StN degradation associated with correlation functions from standard MC methods where the phase is stochastically sampled. One still expects StN degradation arising from numerically estimating the average product of an increasingly large number of variables as $t$ is increased, but this residual StN problem is not associated with estimating a signal that vanishes in the large-$t$ limit and should therefore be much less severe.[^8] For numerical verification, the free scalar boson propagator for ensemble $C_0$ is compared to the propagator determined by MC sampling of the dual representation Eq. with identical parameters $M^2 =0.00625$ and $L=512$ in Fig. \[fig:dualStN\]. The dual representation provides calculations of the propagator and effective mass with slower StN degradation than the standard representation.
The integrand in Eq. includes products of $I_{|q|}$ functions reminiscent of transfer matrix products appearing in symmetry-projected path integral constructions of Ref. [@DellaMorte:2007zz; @DellaMorte:2008jd; @DellaMorte:2010yp]. Times between the scalar source and sink are associated with factors of $I_{|Q+q|}$ Bessel functions orthogonal to the $I_{|q|}$ Bessel functions associated with the partition function and with the propagator for times outside the source and sink. This suggests that integration over the phase projects the transfer matrix to sectors of definite scalar $U(1)$ charge. Formally, group-theoretical projectors are constructed for path integrals by integration over all elements of a symmetry group. Integration over decoupled $e^{i\Delta}$ factors associated with each link is equivalent to integration over the group of local $U(1)$ transformations $\varphi(t) \rightarrow e^{i\alpha(t)}\varphi(t)$, and so dual lattice phase integration acts as a projector to sectors of the Hilbert space with definite $U(1)$ charge. This projection provides the essential mechanism by which dual lattice phase integration avoids the $U(1)$ charged correlation function sign and StN problem.
Wrapped phase statistics {#sec:wrappedstats}
------------------------
The magnitude-phase decomposition of the partition function in Eq. shows that for a given scalar field magnitude the phase differences $\Delta(t)$ are independent in the $L\rightarrow \infty$ limit where the PBC constraint $\sum_{t=0}^{L-1} \Delta(t) = 2\pi w$ can be neglected. The $L\rightarrow \infty$ distribution for $\Delta(t)$ is given from Eq. in terms of $\kappa(t)$ by $$\begin{split}
\mathcal{P}(\Delta(t)) = \frac{1}{2\pi I_0(\kappa(t))}\; e^{\kappa(t)\cos(\Delta(t))}.
\end{split}\label{eq:VM}$$ This distribution is known as a von Mises distribution and is well studied in circular statistics [@Fisher:1995; @Mardia:2009]. The resulting probability distribution describing phase differences $\Theta(t) = \theta(t) - \theta(0)$ as sums of independent von Mises random variables can be expressed as $$\begin{split}
\mathcal{P}(\Theta) = \frac{1}{2\pi}\sum_{n\in\mathbb{Z}} e^{-in\Theta} \prod_{t^\prime = 1}^t \left[ \frac{I_{|n|}(\kappa(t^\prime))}{I_0(\kappa(t^\prime))} \right].
\end{split}\label{eq:PthetaVM}$$ It is difficult to calculate many properties of this probability distribution analytically. The remainder of Sec. \[sec:scalarstats\] studies a simpler approximation to Eq. where StN ratios can be calculated analytically for correlation function estimators using phase unwrapping introduced in Sec. \[sec:unwrappedstats\].
A simpler approximation to Eq. can be derived under the assumptions $$\begin{split}
\frac{|\varphi(t)||\varphi(t^\prime)| - \avg{|\varphi(t)||\varphi(t^\prime)|}}{\avg{|\varphi(t)||\varphi(t^\prime)|}} \ll 1
\qquad \text{and} \qquad
\Delta(t) \ll 1.
\end{split}\label{eq:small}$$ For fine discretizations with $M^2 \ll 1$, the gradient term provides the dominant contribution to the action and Eq. should approximately hold for generic neighborhoods of generic field configurations. Note however that Eq. is not exact in any limit of complex scalar field theory. The non-trivial consequences of relaxing Eq. are explored below by comparing numerical MC results to analytic expressions derived to leading order in Eq. , see in particular Figs. \[fig:phasehist\] and \[fig:phase-jumps\]. Throughout the remainder of this section $\approx$ will be used to denote equality to leading order in the small quantities indicated in Eq. and ignoring terms that vanish in the large $t$ limit below Eq . In this approximation, phase differences between adjacent lattice sites are identically distributed as well as independent since $$\begin{split}
\kappa(t) \approx \kappa \equiv \frac{1}{L} \sum_{t} \avg{\kappa(t)}.
\end{split}\label{eq:kappaapprox}$$ The assumption $\Delta \ll 1$ can be used to further simplify Eq. . Expanding the cosine to second order in $\Delta$, restoring invariance under $\Delta \rightarrow \Delta + 2\pi k$ shifts through explicit summation, and adjusting the overall normalization to enforce $\int d\Delta \;\mathcal{P}(\Delta) = 1$ exactly at this order gives $$\begin{split}
\mathcal{P}(\Delta) &\approx \sqrt{\frac{\kappa}{2\pi}}\sum_{k\in\mathbb{Z}} e^{-\kappa (\Delta + 2\pi k)^2 / 2} \\
&= \frac{1}{2\pi}\sum_{n\in\mathbb{Z}} e^{i n \Delta-n^2 /(2\kappa)}
\end{split}\label{eq:WN}$$ where the second line can be obtained using Poisson summation. Eq. defines the wrapped normal probability describing a normally distributed random variable defined modulo $2\pi$. The wrapped normal and von Mises distributions both approach normal distributions in the limit of small width $\kappa \rightarrow \infty$ and uniform distribution in the limit of large width $\kappa \rightarrow 0$, but the distributions differ at intermediate $\kappa$.
![ Histograms of differences of scalar field phases separated by 1, 10, and 25 lattice sites from left to right for ensemble $C_0$. These phase difference correspond to the phases of $G_{1,0}(1)$, $G_{1,0}(10)$, and $G_{1,0}(25)$ respectively. The 2,560,000 samples resulting from $L=512$ differences on each of $N=5000$ field configurations are grouped into 500 bins and normalized so that the histograms represent the empirical probability distribution functions $\mathcal{P}(\theta(t) - \theta(t-1))$, $\mathcal{P}(\theta(t) - \theta(t-10))$ and $\mathcal{P}(\theta(t) - \theta(t-25))$ respectively. The blue curves show the predictions for these distributions from Eq. -Eq. , which assume that magnitude fluctuations do not affect the phase distribution and are only exact to leading order in Eq. .[]{data-label="fig:phasehist"}](phase_diffs.png){width="\textwidth"}
The wrapped normal characteristic function is identical to the normal characteristic function, $$\begin{split}
\Phi_{\mathcal{P}(\Delta)}(n) \equiv \avg{e^{in\Delta}} = \int d\Delta \; e^{in\Delta} \mathcal{P}(\Delta) \approx e^{-n^2 /2\kappa}.
\end{split}\label{eq:Deltachar}$$ The characteristic function of $\Theta$ can be described as a product of characteristic functions of $\Delta$, $$\begin{split}
\Phi_{\mathcal{P}(\Theta)}(n) &\equiv \avg{ e^{i n \Theta} }
= \avg{\prod_{t^\prime=1}^t e^{i n \Delta(t^\prime)} }
= \prod_{t^\prime=1}^t \Phi_{\mathcal{P}(\Delta(t^\prime))}(n).
\end{split}\label{eq:thetachar}$$ The probability distribution of $\Theta$ is given by a Fourier transform of this characteristic function, $$\begin{split}
\mathcal{P}(\Theta) &= \frac{1}{2\pi}\sum_{n\in\mathbb{Z}} e^{-in\Theta} \Phi_{\mathcal{P}(\Theta)}(n)\\
&= \frac{1}{2\pi} \sum_{n\in\mathbb{Z}} e^{-in \Theta} \prod_{t^\prime =1}^t \Phi_{\mathcal{P}(\Delta(t^\prime))}(n) \\
&\approx \frac{1}{2\pi}\sum_{n\in\mathbb{Z}} e^{-in \Theta} e^{ - t n^2 /2\kappa}.
\end{split}\label{eq:WNtheta}$$ Under the assumption of Eq. , the scalar boson propagator is given by $$\begin{split}
G(t) &\approx \avg{|\varphi(t)\varphi(0)|} \avg{e^{i\Theta}} = \avg{|\varphi(t)\varphi(0)|} \Phi_{\mathcal{P}(\Theta)}(1) \\
&\approx \avg{|\varphi(t)\varphi(0)|} e^{-t/(2\kappa)}.
\end{split}\label{eq:propapprox}$$ Comparing this to the large-time spectral representation Eq. , the correct ground-state energy and overlap factor are reproduced if $$\begin{split}
\kappa \approx \frac{1}{2E}, \hspace{20pt} \avg{|\varphi(t)\varphi(0)|} \approx Z_{1;0,1},
\end{split}\label{eq:kappa}$$ where $t$ is assumed to be large. The expectation value of the ensemble average correlation function can be calculated in this approximation as $$\begin{split}
\avg{\overline{G}} &= \frac{1}{N}\sum_{i=1}^N \avg{|\varphi_i(t)||\varphi_i(0)|\cos(\Theta_i) }\\
&\approx Z_{1;0,1} \frac{1}{N}\sum_{i=1}^N \avg{\cos(\Theta_i) } \\
&\approx Z_{1;0,1} e^{-E t}.
\end{split}\label{eq:Gbardef}$$ Its variance is given by $$\begin{split}
\text{Var}(\overline{G}) &\approx Z_{1;0,1}^2 \left\lbrace \avg{\left( \frac{1}{N}\sum_i \cos(\Theta_i) \right)^2} - \avg{\left( \frac{1}{N}\sum_i \cos(\Theta_i) \right)}^2 \right\rbrace \\
&= \frac{Z_{1;0,1}^2}{2N}\left(1 + \avg{\cos(2 \Theta_i)} - \avg{\cos(\Theta_i)}^2 \right) \\
&\approx \frac{Z_{1;0,1}^2}{2N}\left(1 - e^{-2Et} \right),
\end{split}\label{eq:GbarVar}$$ and its StN ratio is $$\begin{split}
\text{StN}(\overline{G}) &= \frac{\avg{\overline{G}}}{\sqrt{\text{Var}(\overline{G})}} \approx \sqrt{2N} \frac{e^{-Et}}{\sqrt{1 - e^{-2Et}}}.
\end{split}\label{eq:GbarStN}$$ It is noteworthy that the full StN problem for the scalar propagator arises at leading order in Eq. where magnitude fluctuations are neglected and phase differences are wrapped normally distributed. Determination of the scalar propagator pole mass from MC sampling phases distributed according to Eq. is equivalent to parameter inference for a wrapped normal distribution with variance $1/\kappa \approx 2E$. Avoiding large finite sample size errors in wrapped normal parameter inference requires [@Fisher:1995] $$\begin{split}
\frac{1}{\sqrt{N}} \lesssim \avg{\cos(\Theta)} \approx e^{-E t}
\end{split}\label{eq:circstatsbound}$$ indicating the window of time in which reliable parameter inference is possible has size scaling only as $\log{N}$.
As shown in Fig. \[fig:phasehist\], Eq. roughly captures the $t$ dependence of phase difference distributions for MC ensemble $C_0$ but does not provide a precise fit to MC results. The empirical distribution of $\Delta$ is better described by a heavy-tailed wrapped stable distribution than by a wrapped normal distribution. Similar heavy-tailed phase derivatives were seen to arise for baryon correlation functions in Ref. [@Wagman:2016bam], where it was conjectured that these heavy tails arose from non-perturbative strong interaction physics.[^9] The appearance of heavy tails in free scalar field theory suggests that they have a generic origin. The von Mises distribution describing phase derivatives for a fixed field magnitude does not have heavy tails, and so the heavy tails present in MC distributions of phase derivatives must arise from integration over magnitude fluctuations. Large phase jumps leading to deviations from Eq. and their relation to magnitude fluctuations are discussed further below. It is also noteworthy that correlation functions for higher charge sector can be computed under the assumptions of Eq. by Eq. as $$\begin{split}
G_{Q,2P} \approx \avg{|\varphi(t)\varphi(0)|}^{|Q|+2P}\avg{e^{iQ\Theta}} \approx Z_{1;0,1}^{|Q|+2P} e^{-t Q^2 /(2\kappa)},
\end{split}\label{eq:GQ2Papprox}$$ which does not reproduce the linear spectrum of free-field theory in Eq. . These deficiencies are addressed in Sec. \[sec:unwrap1D\] with numerical MC studies not relying on the assumptions of Eq. .
Unwrapped phase statistics {#sec:unwrappedstats}
--------------------------
The results of the last section demonstrate that exponential StN degradation appears in calculations of the average cosines of wrapped normal phase differences. This suggests that to avoid sign and StN problems, one needs to avoid numerical sampling of circular random variables. For $(0+1)D$ complex scalar field theory phase fluctuations can be integrated analytically and the resulting path integrals describing charged scalar correlation functions have positive-definite weights and a less severe StN problem. Similar methods can be applied to more complicated scalar field theories in more dimensions [@Endres:2006xu] and also have a long history of application to lattice gauge theory [@Ukawa:1979yv]. The search for a more general dual representation of LQCD where properties of finite-density matter can be computed with path integrals with positive-definite weights is an active area of ongoing research, see for instance Refs. [@Chandrasekharan:2008gp; @deForcrand:2010ys; @Gattringer:2012df; @Gattringer:2012ap; @Gattringer:2014nxa; @Gattringer:2016kco; @Giuliani:2017qeo]. It is also possible to look for path integral deformations or changes of variables that reduce the severity of phase fluctuations. Methods based on Lefschetz thimbles and more general classes of complex path integral deformations have successfully transformed path integrals with phase fluctuations in a variety of LQFTs into (sums of) path integrals where the phase is exactly fixed or at least fluctuating less severely than in the original theory [@Witten:2010cx; @Cristoforetti:2012su; @Aarts:2013fpa; @Fujii:2013sra; @Cristoforetti:2013wha; @Tanizaki:2014tua; @Fujii:2015bua; @Alexandru:2015sua; @Alexandru:2015xva; @Tanizaki:2016xcu; @Alexandru:2016san; @Alexandru:2016gsd; @Mori:2017pne; @Alexandru:2018fqp]. Still, it is an open challenge to find an efficient representation for computing finite-density observables or multibaryon correlation functions in LQCD that avoids sign and StN problems.
The problems inherent to numerical sampling of circular random variables could be avoided if one could instead numerically sample a noncompact real random variable. Intuitively, one may imagine stochastically sampling a real random variable representing the angular displacement accumulated by the scalar field phase in the interval $[0,t]$ including any $2\pi$ revolutions around the unit circle. Other works have explored accessing distributions of an analogous “extensive phase” in the context of QCD and other theories at nonzero chemical potential [@Ejiri:2007ga; @Nakagawa:2011eu; @Ejiri:2012wp; @Greensite:2013gya; @Garron:2017fta; @Bloch:2018yhu]. This alternative has also been explored in other areas of science and engineering where circular random variables appear. A variety of “phase unwrapping” techniques have been developed to extract noncompact variables representing angular displacement from numerical samples of compact phases, see Refs. [@Judge:1994; @Ghiglia:98; @Ying:2006; @Kitahara:2015] for reviews. An unwrapped phase difference describes a difference between phases at opposite ends of a parametrized path plus $2\pi$ times the “winding number” counting the number of full revolutions of the unit circle accumulated along the path. A formal definition of the unwrapped phase dating back to the 1975 homomorphic signal processing of Oppenheim and Schafer [@Oppenheim:1975] can be used algebraically to compute the unwrapped phase of a complex polynomial [@McGowan:1982; @Steiglitz:1983; @AlNashi:1989; @Yamada:1998; @Yamada:2002; @Yamada:2011; @Kitahara:2015], while numerical techniques can be used to approximately compute the unwrapped phase from sufficiently finely sampled time series of phases [@Oppenheim:1975; @Tribolet:1977; @Itoh:1982] and higher-dimensional arrays of wrapped phases [@Goldstein:1988; @Huntley:1989; @Bone:91; @Huntley:01; @Jenkinson:2003; @Hooper:07; @Abdul-Rahman:09].
A continuous function denoted Arg that describes accumulated phase differences along a $1D$ path is defined in Appendix \[app:unwrapping\] and when applied to $1D$ correlation functions gives $$\begin{split}
\text{Arg}(C(t)) = \text{Arg}(|C(t)|e^{i\theta(t) - i\theta(0)}) = \int_{\theta(0)}^{\theta(t)} d\theta^\prime = \int_0^t \frac{d\theta}{dt^\prime} dt^\prime.
\end{split}\label{eq:ArgGdef}$$ Since $\Theta(t) = \text{arg}(C(t))$ is not continuous at branch cut crossings, the fundamental theorem of calculus cannot be directly applied to Eq. if there is a branch cut crossing in the interval $[0,t]$. By deforming the integration contour to replace branch cut crossings by integrals encircling a neighborhood of the origin (see Appendix \[app:unwrapping\]), Eq. can be transformed into an integral over a domain where $\theta(t)$ is analytic plus $2\pi$ times the total number of oriented branch cut crossings to give $$\begin{split}
\text{Arg}(C(t)) = \Theta(t) + 2\pi \nu(t)
= \text{arg}(C(t)) + 2\pi \nu(t)
\equiv \widetilde{\theta}(t) - \widetilde{\theta}(0)
\equiv \widetilde{\Theta}(t).
\end{split}\label{eq:ArgGnu}$$ The unwrapped phase difference $\widetilde{\Theta}(t)$ associated with a LQFT propagator therefore differs from the principal-valued or “wrapped” phase difference by $2\pi$ times an integer winding number $\nu(t)$ that counts the number of oriented branch cut crossings of the propagator in $[0,t]$.
![The left plot shows real and imaginary parts of a particular $(0+1)D$ free complex scalar field propagator $C(t)$ generated with $M^2 = 0.01$ exhibiting near-zeros of the magnitude indicated by the arrows. The right plot shows the wrapped correlator phase $\Theta(t) = \text{arg}C(t)$ defined with $-\pi < \Theta \leq \pi$ and three calculations of the unwrapped phase $\text{Arg}C(t) \in \mathbb{R}$ obtained by numerical integration of Eq. for $\widetilde{\Theta}_1$, determination of winding numbers by assuming Eq. for $\widetilde{\Theta}_2$, and algebraic phase unwrapping of a linear polynomial interpolation according to Ref. [@Kitahara:2015] for $\widetilde{\Theta}_3$. The numerical and algebraic winding number methods agree exactly at all lattice sites and only differ in their interpolation between lattice sites. Numerical integration of Eq. leads to $O(\pi)$ deviations from both winding number methods for all $t > 30$.[]{data-label="fig:hardunwrapping"}](hard_unwrapping.png){width="\textwidth"}
In LQFT as well as in other applications of complex time series, wrapped phases $\theta(t)$ are determined directly from “data” by applying $\text{arg}$ to complex random variables. The phase unwrapping problem is to determine winding numbers $\nu(t)$ that make the unwrapped phase $\widetilde{\theta}(t)$ a continuous function of $t$ across the branch cuts of $\theta(t)$. For a complex time series that samples a smooth function with sufficiently fine resolution, one expects that branch cut discontinuities of $\theta(t)$ can be identified and winding numbers can be assigned to keep $\widetilde{\theta}(t)$ continuous across these branch cuts of $\theta(t)$. It was explicitly demonstrated by Itoh in Ref. [@Itoh:1982] that the assumption $$\begin{split}
|\widetilde{\theta}(t) - \widetilde{\theta}(t-1)| < \pi,
\end{split}\label{eq:1Dbound}$$ is sufficient to uniquely define winding numbers for a time series of wrapped phases $$\begin{split}
\nu(0) = 0,\hspace{20pt} \nu(t) = \nu(t-1) + \begin{cases}
1, & -2\pi < \theta(t) - \theta(t-1) \leq -\pi \\
0, & -\pi < \theta(t) - \theta(t-1) \leq \pi \\
-1, & \pi < \theta(t) - \theta(t-1) \leq 2\pi \end{cases}.
\end{split}\label{eq:nu1Ddef}$$ In LQFT, one might hope that Eq. is valid in generic field configurations when the lattice spacing is much smaller than all physical correlation lengths. However, Eq. shows that points with $|\varphi(t)|=0$ have infinite $d\widetilde{\theta}/dt$ even in the continuum. As demonstrated by example in Fig. \[fig:hardunwrapping\], near-zeros of $|C(t)|$ can occur for $(0+1)D$ free complex scalar field theory with $M^2 = 0.01$. The wrapped phase of the same correlation function is also shown in Fig. \[fig:hardunwrapping\] along with results for three different $(0+1)D$ phase unwrapping schemes:
1. Numerical integration of $d\widetilde{\theta}/dt$ according to a linear discretization of Eq. .
2. Numerical integration of $\theta(t) - \theta(t-1)$ with winding numbers $\nu(t)$ assigned by Eq. to satisfy $|\partial_t \widetilde{\theta}| < \pi$.
3. Algebraic phase unwrapping of a linear polynomial interpolation of $\varphi$ using the numerically stabilized Strum sequence method of Kitahara and Yamada [@Kitahara:2015].
The unwrapped phase defined by numerically integrating Eq. does not satisfy Eq. at $t=30$, which coincides with a near-zero of the magnitude as indicated in Fig. \[fig:hardunwrapping\]. This violation of Eq. leads to $O(\pi)$ discrepancies between the results of unwrapping based on numerical integration of Eq. and both the numerical and algebraic winding number determination methods. Notably, $O(\pi)$ discrepancies occur at all lattices points with $t$ larger than the point where Eq. is violated. Near-zeros therefore produce an accumulating $O(\pi)$ sensitivity in the unwrapped phase which increases with increasing $t$ unless Eq. holds at all points. The accumulation-of-errors problem will present numerical difficulties for the simple phase unwrapping schemes explored in Sec. \[sec:unwrap1D\]. These difficulties become more tractable in higher-dimensional phase unwrapping problems [@Goldstein:1988; @Huntley:1989; @Ghiglia:98], essentially because redundancies in the multidimensional gradient of a smooth function provide additional information that can be used to guide unwrapping across regions where Eq. is violated. The theory of multidimensional phase unwrapping is much richer than the $1D$ theory explored below and deferred to future work. The remainder of this section addresses the StN behavior of the unwrapped phase and how properties of its distribution can be used to usefully estimate the average phase cosine. The small fluctuation assumptions of Eq. are used to leading order and phase differences are therefore wrapped normally distributed. It is also assumed that the unwrapped phase $\widetilde{\theta}$ differs from $\theta$ by $2\pi$ times an integer winding number $\nu$ and therefore that $$\begin{split}
W[\widetilde{\theta}] \equiv \widetilde{\theta} \text{ mod } 2\pi = \theta ,
\end{split}\label{eq:Wdef}$$ where the wrapping operator $W$ restricts the unwrapped phase to the interval $(-\pi,\pi]$. Wrapped normal phase differences, Eq. , can be generated by applying $W$ to a normally distributed unwrapped phase difference, $$\begin{split}
\mathcal{P}(\widetilde{\Theta}) = \sqrt{\frac{\kappa }{2\pi t}} e^{-\kappa \widetilde{\Theta}^2/(2t)}.
\end{split}\label{eq:N}$$ By construction the average cosine of the wrapped and unwrapped phases are identical $$\begin{split}
\avg{\cos(\widetilde{\Theta})} &= \avg{\cos(\Theta + 2\pi\nu )} \\
&= \avg{\cos(\Theta)} \\
&= e^{-t/(2\kappa)} \approx e^{-E t}.
\end{split}\label{eq:cosUW}$$ The sample mean cosine of an ensemble of unwrapped phases could be used to estimate the ground-state energy with identical results and identical StN degradation as calculations based on the sample mean of the wrapped phase cosine. However, the boson mass can be estimated more efficiently from a MC ensemble of normally distributed unwrapped phase differences by $$\begin{split}
\widetilde{E}(t) \equiv \frac{1}{2N}\sum_{i=1}^N \left[ \widetilde{\Theta}_i(t)^2 - \widetilde{\Theta}_i(t+1)^2 \right],
\end{split}\label{eq:Mbartildedef}$$ where $\widetilde{\Theta} \rightarrow - \widetilde{\Theta}$ symmetry has been assumed on the basis of $\Theta \rightarrow - \Theta$ symmetry (which follows in the infinite-statistics $N\rightarrow \infty$ limit from the reality of correlation functions guaranteed by unitarity). The corresponding estimate of the correlation function is $$\begin{split}
\widetilde{G}(t) \equiv \left(\frac{1}{N}\sum_{i=1}^N |\varphi_i(t)\varphi_i(0)| \right) \exp\left( -\frac{1}{2N} \sum_{i=1}^N \widetilde{\Theta}_i(t)^2 \right).
\end{split}\label{eq:Gtildedef}$$ Under the present assumptions this provides an accurate estimate of the correlation function as $N\rightarrow \infty$, $$\begin{split}
\avg{ \widetilde{G}(t) } &\approx Z_{1;0,1} \prod_{i=1}^{N}\left[
\frac{1}{\sqrt{4 \pi E t}}
\int d \UWTheta_i(t)\ e^{-\UWTheta_i(t)^2 \kappa/(2 t) - \frac{1}{2N}\UWTheta_i(t)^2} \right] \\
&= Z_{1;0,1} \paren{1 + \frac{2 E t}{N}}^{-N/2} \\
&= Z_{1;0,1} e^{-E t}\left[ 1 + \frac{(Et)^2}{N} + O\left(N^{-2}\right) \right] .
\end{split}\label{eq:Gtildeav}$$ The variance of $\widetilde{G}$ in the $N\rightarrow\infty$ limit can be computed similarly, $$\begin{split}
\text{Var}\left( \widetilde{G}(t) \right) &\approx
-\avg{ \widetilde{G}(t) }^2 + Z_{1;0,1}^2 \prod_{t=0}^{L-1}\left[
\frac{1}{\sqrt{4 \pi E t}}
\int d \UWTheta_i(t)\ e^{-\UWTheta_i(t)^2 \kappa/(2 t) - \frac{1}{N}\UWTheta_i(t)^2} \right]\\
&= Z_{1;0,1}^2 \paren{1 + \frac{4 E t}{N}}^{-N/2} - Z_{1;0,1}^2 \paren{1 + \frac{2 E t}{N}}^{-N} \\
&= Z_{1;0,1}^2 e^{-2 E t}\left[ \frac{2 (E t)^2}{N} + O\left(N^{-2}\right) \right].
\end{split}\label{eq:Gtildevar}$$ The correlation function computed from normally distributed unwrapped phases therefore has a StN ratio $$\begin{split}
\text{StN}\left( \widetilde{G}(t) \right) \approx \frac{\sqrt{N}}{\sqrt{2} Et}\left[ 1 + O(N^{-1/2})\right].
\end{split}\label{eq:StNMbartilde}$$ Eq. demonstrates that normally distributed unwrapped phases provide correlation function estimates whose StN ratios decrease polynomially as $t^{-1}$ rather than exponentially as $e^{-E t}$ as the spacetime volume $t$ containing nonzero $U(1)$ charge is increased.
Unwrapped characteristic function and cumulant expansion {#sec:cumulant}
--------------------------------------------------------
For field configurations violating the small fluctuation assumptions of Eq. , as demonstrated to occur even in free-field theory in Figs. \[fig:phasehist\]-\[fig:hardunwrapping\], it is necessary to construct an estimator for $\avg{\cos\Theta}$ from the unwrapped phase that does not depend on assumptions about the distribution of $\Theta$. The constraint that the winding numbers of the unwrapped phase are integer values, $W[\widetilde{\theta}] = \theta$, can be interpreted as a statement that the characteristic functions of the wrapped and unwrapped phase differences agree at every integer, $$\begin{split}
\Phi_\Theta(n) = \avg{e^{i\Theta n}} = \avg{e^{i\widetilde{\Theta}n}} = \Phi_{\widetilde{\Theta}}(n),\hspace{20pt} n\in\mathbb{Z}.
\end{split}\label{eq:charUW}$$ For noninteger $n$ the wrapped and unwrapped characteristic functions can differ. By constraining the unwrapped characteristic function with results not limited to integer $n$, winding number information present in the unwrapped phase can be incorporated. Once the unwrapped phase characteristic function is fit to numerical results by some method, the mean cosine of the (wrapped or unwrapped) phase is given by evaluating the resultant fit function at $n=1$, $$\begin{split}
\Phi_{\widetilde{\Theta}}(1) = \avg{\cos(\Theta)} \approx e^{-E t}.
\end{split}\label{eq:avecos}$$
Cumulant expansion methods similar to those explored in Refs. [@Endres:2011jm; @Endres:2011er; @Endres:2011mm; @Lee:2011sm; @Nicholson:2012xt; @Grabowska:2012ik; @Wagman:2016bam] can be used to estimate $\Phi_{\widetilde{\Theta}}(1)$ with systematic uncertainties whose size can be assessed by varying the truncation order of the expansion. For a generic complex random variable $z$ with characteristic function $\Phi_z(k) = \avg{e^{ikz}}$, cumulants can be defined as the coefficients of a Taylor series for $\ln(\Phi_z)$, $$\begin{split}
\Phi_z(k) = \avg{e^{ikz}} \equiv \exp\left[ \sum_{n=1}^\infty \frac{(ik)^n}{n!}\kappa_n(z) \right].
\end{split}\label{eq:charfundef}$$ Equivalent expansions can be constructed that perform a dual expansion in cumulants of the real and imaginary parts of $z$. The cumulants appearing in Eq. can be related to the moments of $z$ by comparing Taylor series expansions for the exponentials in Eq. , $$\begin{split}
\kappa_n(z) = \avg{z^n} - \sum_{m=1}^{n-1} {n-1 \choose m-1} \kappa_m(z)\avg{z^{n-m}}.
\end{split}\label{eq:cumulantmom}$$ Noting that scalar field propagators are given in terms of the field’s log-magnitude and unwrapped phase by $$\begin{split}
C(t) = \varphi(t)\varphi^*(0) = e^{\R + i \widetilde{\Theta}},
\end{split}\label{eq:CUW}$$ an estimator for the scalar boson mass can be defined by $$\begin{split}
\widetilde{E}^{(n_{max})} &= -\sum_{n=1}^{n_{max}} \frac{1}{n!} \; \partial_t \kappa_n(\R+i\widetilde{\Theta}) \\
&= -\sum_{n=1}^{n_{max}} \frac{1}{n!} \; \partial_t \kappa_n(\text{ln}|C| + i \text{Arg}(C)).
\end{split}\label{eq:cumulantEMdef}$$ In the limits $n_{max}\rightarrow \infty$ and $N\rightarrow \infty$, Eq. should approach the scalar boson mass or the ground-state energy for two-point correlation functions in general LQFTs. Consistency between unwrapped and wrapped phase cumulant expansions requires that $W[\widetilde{\theta}] = \theta$ but is otherwise independent of the particular choice of phase unwrapping algorithm used to define $\widetilde{\Theta}$.
The leading contributions to Eq. are $$\begin{split}
\kappa_1(\R) = \avg{\R},\hspace{20pt}
\kappa_2(\R) = \avg{\R^2}- \avg{\R}^2,\hspace{20pt} \kappa_2(\widetilde{\Theta}) = \avg{\widetilde{\Theta}^2},
\end{split}\label{eq:leadingcumulants}$$ since $\kappa_1(\UWTheta)$ and the covariance of $\R$ and $\UWTheta$ are guaranteed to vanish by $\UWTheta \rightarrow -\UWTheta$ symmetry. In general LQFTs the magnitude and phase might make very different contributions to the effective mass, and so an arbitrary hierarchy is possible between odd cumulant contributions only involving the magnitude and even cumulant contributions that also involve the phase. In particular, $\kappa_2(\widetilde{\Theta})$ dominates $\kappa_1(\R)$ for free complex scalar field theory and the leading contribution to the cumulant effective mass above is $$\begin{split}
\widetilde{E}^{(2)} &= -\partial_t \kappa_1(\R) - \frac{1}{2}\partial_t \kappa_2(\R) + \frac{1}{2}\partial_t \kappa_2(\widetilde{\Theta}) + \dots,
\end{split}\label{eq:leadingcumuEM}$$ where the ellipsis denotes contributions from $\kappa_n(\R+i\widetilde{\Theta})$ with $n\geq 3$. All omitted contributions with $n\geq 3$ would vanish in the infinite statistics $N\rightarrow \infty$ limit if $\R$ and $\widetilde{\Theta}$ were exactly normally distributed and independent. For distributions with finite moments, contributions from $n\geq 3$ provide subdominant corrections that will be small for approximately normally distributed $\R$ and $\widetilde{\Theta}$. The size of these contributions can be assessed in practice by comparing results for $\widetilde{E}^{(n_{max})}$ with multiple truncation points $n_{max}$ and systematic uncertainties can be assigned based off sensitivity of $\widetilde{E}^{(n_{max})}$ to the truncation point. Since terms with odd $n$ have vanishing phase contributions by $\Theta \rightarrow -\Theta$ symmetry, the convergence pattern of $E^{(n_{max})}$ should be expected to strongly depend on whether $n_{max}$ is even or odd and be comparatively smooth as a function of even (odd) cumulant number $n_{max} = 2,4,6,\dots$ ($n_{max} = 3,5,7,\dots$).
An analogous expansion to Eq. could be defined for the wrapped phase; however, the $\mathcal{P}(\Theta)$ approaches a uniform distribution at large times and large cumulants will make sizable contributions to the wrapped analog of Eq. . Much faster convergence is expected for Eq. if $\widetilde{\theta}(t)$ differs from $\theta(t)$ by nonzero winding numbers and $\mathcal{P}(\widetilde{\Theta})$ is approximately normal. Any unwrapping algorithm with $W[\widetilde{\theta}]=\theta$ will define a $\widetilde{\Theta}$ such that Eq. is a consistent estimator in the $n_{max}\rightarrow \infty$ limit, but different algorithms may have different convergence rates.
Estimators for correlation functions including cumulant expansions of unwrapped phases are constructed by generalizing Eq. as $$\begin{split}
\widetilde{G}^{(n_{max})}(t) \equiv \exp\left[ \sum_{n=1}^{n_{max}} \frac{1}{n!}\kappa_n\left( \mathcal{R}(t) + i \widetilde{\Theta}(t) \right) \right].
\end{split}\label{eq:Gtildedefn}$$ Despite the vanishing of all cumulants with $n\geq 3$ under the assumption of uncorrelated $\R$ and $\UWTheta$ and in the $N\rightarrow \infty$ limit of an exactly normal unwrapped phase distribution, the statistical uncertainties of these higher cumulants increase with increasing $n$. For large $n$, the variance of the $n$th cumulant will be dominated by the variance of the $n$th moment. The large moment $n$ and large statistical ensemble size $N$ behavior of cumulant expansion contributions $\kappa_n / n!$ is therefore determined by the statistical behavior of $$\begin{split}
\frac{1}{(2n)!}\avg{\widetilde{\Theta}^{2n}} = \frac{1}{(2n)!}\avg{ \frac{1}{N}\sum_{i=1}^N \widetilde{\Theta}_i^{2n}}.
\end{split}\label{eq:momdef}$$ For normally distributed unwrapped phases, these sample moments have expectation values $$\begin{split}
\frac{1}{(2n)!}\avg{\widetilde{\Theta}^{2n}} \approx \frac{(2n-1)!!}{(2n)!} \left( \frac{t}{\kappa} \right)^n.
\end{split}\label{eq:normalscaling}$$ The variance of $\widetilde{\Theta}$ can be calculated straightforwardly from Eq. and leads to StN behavior for large moments given by $$\begin{split}
\text{StN}\left( \frac{1}{(2n!)}\widetilde{\Theta}^{2n} \right) \approx \sqrt{N} 2^{-n + 1/4} \left[ 1 + O(n^{-1}) + O(N^{-1}) \right].
\end{split}\label{eq:momStN}$$ Cancellations between contributions from different moments could lead to a smaller variance for $\widetilde{E}^{(n_{max})}$ and $\widetilde{G}^{(n_{max})}$ than that of $\widetilde{\Theta}^{2n}/(2n)!$, but exponentially precise cancellations that would be required to avoid $2^{-n}$ StN degradation are not expected to occur at finite $N$. This suggests that $\widetilde{E}^{(n_{max})}$ and $\widetilde{G}^{(n_{max})}$ have StN ratios proportional to $\sqrt{N}2^{-n}$ as in Eq. . This suggests that even under the assumptions of Eq. , the construction of a complete solution to the sign problem using phase unwrapping and the cumulant expansion still requires an extrapolation $n_{max}\rightarrow\infty$ where $N$ must be taken exponentially large in $n_{max}$ to remove all truncation errors at fixed statistical precision.
The appearance of such an exponentially hard extrapolation should be expected: phase unwrapping is applicable to generic LQFT correlation functions and the sign problem has been demonstrated to be NP-hard for some quantum systems by Troyer and Wiese [@Troyer:2004ge]. For LQFTs including LQCD, observations of the ubiquity of (complex-)log-normally distributed correlation functions [@Hamber:1983vu; @Guagnelli:1990jb; @Endres:2011jm; @DeGrand:2012ik; @Drut:2015uua; @Porter:2016vry; @Wagman:2016bam; @Rammelmuller:2017vqn] suggest that useful results might be obtained using modest $n_{max}$ despite the exponential difficulty of extrapolating to $n_{max}\rightarrow \infty$. Understanding the size of truncation errors in practical calculations and systematic limitations of this method will likely require specific studies for particular LQFTs of interest.
Generic correlation functions in $(0+1)D$ complex scalar field theory can be analyzed similarly to the scalar boson propagators above. The wrapped phase for a general correlation function $G_{Q,2P}$ only depends on its $U(1)$ charge $Q$ and not on $P$ and is denoted by $$\begin{split}
\Theta_Q \equiv \Theta_{Q,2P} = W[Q \Theta]
\end{split}\label{eq:ThQdef}$$ The unwrapped phase can similarly be chosen to be independent of $P$ and defined by[^10] $$\begin{split}
\widetilde{\Theta}_Q &\equiv \text{Arg}\left( \mathcal{O}_{Q,2P}(t) \mathcal{O}^*_{Q,2P}(0) \right) \\
&= \text{Arg} \left( e^{i Q \Theta } \right).
\end{split}\label{eq:uwQ}$$ Note that because Arg is a nonlinear function, $\widetilde{\Theta}_Q$ is not simply related to $\widetilde{\Theta}$ despite the identity $W[Q\widetilde{\Theta}] = \Theta_Q$. In particular $Q \widetilde{\Theta}$ is sensitive to branch cut crossings of the variable $Q\Theta$ with principle domain $(-\pi Q, \pi Q]$ rather than branch cut crossings of $\Theta_{Q} = W[Q\Theta]$ with principle domain $(-\pi,\pi]$. With $\widetilde{\Theta}$ defined by Eq. , $Q\widetilde{\Theta}$ will include jumps of $2\pi Q$ at branch cut crossings of $\Theta$ rather than jumps of $2\pi$ at branch cut crossings of $\Theta_Q$ and is therefore not equal to $\widetilde{\Theta}_Q$. With a consistent phase unwrapping of $\Theta_{Q,2P} \in (-\pi,\pi]$, correlation functions and ground-state energies can be estimated with the cumulant expansions $$\begin{split}
\widetilde{G}_{Q,2P}^{(n_{max})} = \exp\left[ \sum_{n=1}^{n_{max}} \frac{1}{n!}\kappa_n \left( (|Q| + 2P)\R + i \widetilde{\Theta}_Q \right) \right], \\
\widetilde{E}_{Q,2P}^{(n_{max})} = -\sum_{n=1}^{n_{max}} \frac{1}{n!}\partial_t \kappa_n \left( (|Q| + 2P)\R + i \widetilde{\Theta}_Q \right).
\end{split}\label{eq:EQPdef}$$ For any phase unwrapping algorithm satisfying $W[\widetilde{\Theta}_Q] = \Theta_{Q,2P}$, these provide unbiased estimators for correlation functions and effective masses in the dual limit $N\rightarrow \infty$ and $n_{max} \rightarrow \infty$. In general charge sectors, $\Theta_Q$ is wrapped normally distributed under the assumptions of Eq. and $\widetilde{\Theta}_Q$ can be consistently defined to be normally distributed with variance $1/\kappa_Q$ chosen to reproduce the ground-state energy charge $Q$ sector. In analogy to $\widetilde{\Theta}$, the correct ground-state energy $E_Q \equiv E_{Q,0}$ is reproduced if the variance of $\widetilde{\Theta}$ is taken to be $1/\kappa_Q \approx 2 E_Q$. The StN results of Eq. can therefore be applied to $\widetilde{G}_{Q,2P}^{(n_{max})}$ if $\kappa$ is replaced by $\kappa_Q$ to give $$\begin{split}
\text{StN}\left( \widetilde{G}_{Q,2P}^{(2)} \right) &\approx \frac{\sqrt{N}}{2E_Q t}\left[ 1 + O(N^{-1}) \right].
\end{split}\label{eq:EQPStN}$$ The moment analysis of Eq. can be applied to $\widetilde{\Theta}_Q$, and suggests that StN ratios for $\widetilde{G}_{Q,2P}^{(n_{max})}$ decrease polynomially with increasing $E_Q t$ but exponentially with increasing $n_{max}$.
The avoidance of exponential StN degradation with increasing $E_Q t$ at fixed order in the cumulant expansion can also be understood in the language of sign problems. Integration over phase fluctuations making nonpositive-definite contributions to path integrals is replaced by calculation of the moments of the unwrapped phase, $$\begin{split}
\avg{\widetilde{\Theta}_Q^n} = \frac{1}{Z}\int_\varphi e^{-S} \widetilde{\Theta}_Q^{n} ,
\end{split}\label{eq:GQPmagphaseUW}$$ that vanish for odd $n$ by unitarity and are path integrals of positive-definite quantities without sign problems or phase fluctuations for even $n$. A sign problem can reemerge beyond leading order in the cumulant expansion from linear combinations of positive-definite moments that enter the cumulant expansion with opposite signs. In particular if $\widetilde{E}^{(2)}$ approximates $E$ poorly, then the sum of cumulants could be $O(e^{-E_Q t})$ while the individual cumulant contributions are $O(1)$ and the full sign problem could reemerge at full strength.
![ The left plot shows the scalar boson mass $\widetilde{E}^{(n_{max})}$ on ensemble $C_0$ obtained using cumulant expansions of the propagator log-magnitude and unwrapped phase truncated at order $n_{max} = 2,\ 4,\ 6$. The $L\rightarrow\infty$ analytic result is shown as a red line and the standard effective mass is shown in gray. The right plot shows the variances of these effective masses. Gaussian-weighted integration with $\sigma = 1.41$ is used to calculate the unwrapped phase.[]{data-label="fig:unwrapStN"}](free_uw_stn.png){width="\textwidth"}
![ The left plot shows the ground-state energies $\widetilde{E}_{Q,0}^{(2)}$ of charge sectors $Q=1,\dots,8$ for ensemble $C_0$ that involve second-order truncations of cumulant expansions of the log-magnitudes and unwrapped phases of the correlation functions $G_{Q,0}$. Results for $E^{(2)}_{1,0}$ are identical to those shown in Fig. \[fig:unwrapStN\]. The right plot shows the average inverse StN of these ground-state energy measurements for a time region $t = 10\rightarrow 20$ as a function of $Q$ for various cumulant expansion truncation orders. Gaussian-weighted integration with $\sigma = 1.41$ is used to calculate the unwrapped phase.[]{data-label="fig:unwrapStNQ"}](free_uw_stn_Q.png){width="\textwidth"}
One-dimensional phase unwrapping {#sec:unwrap1D}
================================
Phase unwrapping was shown analytically above to remove exponential StN degradation at fixed order in a cumulant expansion under the assumptions of Eq. . Relaxing these assumptions, the probability distribution of phase fluctuations becomes more complicated and an unwrapped phase distribution satisfying $W[\widetilde{\theta}] = \theta$ cannot be easily found analytically. Numerical MC simulations are used in this section to analyze the accuracy and precision of cumulant expansions involving the log-magnitude and unwrapped phase in $(0+1)D$ scalar field theory without the small fluctuation assumptions of Eq. .
Numerical phase unwrapping schemes {#sec:unwrapnum}
----------------------------------
For a field defined on a discrete lattice of points, the unwrapped phase is not uniquely defined without further assumptions that for instance could be based on a discretized definition of smoothness. Precisely defining this smoothness assumption is essential for constructing a numerical phase unwrapping algorithm. The assumptions $|\partial_t \widetilde{\theta}| < \pi$ and $W[\widetilde{\theta}]=\theta$ lead uniquely to the path unwrapping algorithm of Eq. . This section employs this phase unwrapping algorithm and two variations with alternative smoothness criteria that enforce smoothness on larger distances than a single lattice spacing.
1. Single-point integration: $\widetilde{\theta}(t)$ is determined by demanding $$\abs{\widetilde{\theta}(t) - \widetilde{\theta}(t-1)} < \pi$$ as in Eq. . This technique assumes a finely sampled lattice, such that the probability density of phase jumps near $\pi$ is vanishing.
2. Windowed integration, with window $w$: $\widetilde{\theta}(t)$ is determined by demanding $$\abs{\widetilde{\theta}(t) - \frac{1}{\text{min}(w,t)} \sum_{t' = \text{max}(t- w, 0)}^{t-1} \widetilde{\theta}(t')} < \pi.$$ This technique more robustly handles large phase jumps by considering the average of previously unwrapped phases. This locally may allow the unwrapped phase jump magnitude to exceed $\pi$, but in such a way that global fluctuation is reduced. When $w = 1$, this reduces to single-point integration.
3. Gaussian-weighted integration, with width $\sigma$: $\widetilde{\theta}(t)$ is determined by demanding $$\abs{\widetilde{\theta}(t) - \sum_{t' = 0}^{t-1} \mathcal{N} e^{-(t'-t)^2/(2 \sigma^2)} \widetilde{\theta}(t')} < \pi.$$ The normalization $\mathcal{N}$ is fixed by $\sum_{t' = 0}^{t-1} \mathcal{N} e^{-(t'-t)^2/(2 \sigma^2)} = 1$. This technique allows one to smoothly interpolate between integer window sizes by providing a non-integer tunable parameter. When $\sigma \sim w$, we expect the two techniques to perform similarly.
Only phase unwrapping algorithms satisfying $W[\widetilde{\theta}] = \theta$ are considered, since this condition guarantees that the unwrapped cumulant expansion reproduces the exact correlation function in the dual limit of infinite truncation order and infinite statistics.
A time-reversal-symmetric integration path is used in which one of the forward integration techniques above is applied to determine the unwrapped phase in $[0,L/2]$ and the corresponding reverse integration technique is applied to determine the unwrapped phase in $[L/2+1,L]$. This symmetric integration path has the advantage of beginning with regions closest to the source where phase gradients are smallest and the probability of unwrapping ambiguities due to physical fluctuations violating $|\partial_t \widetilde{\theta}| < \pi$ is correspondingly smallest. The $t=0$ phase $\widetilde{\theta}(0) = \theta(0)$ is used as an initial condition for unwrapping, although this is irrelevant for correlation functions since they only involve phase differences $\widetilde{\theta}(t) - \widetilde{\theta}(0)$. With this scheme the unwrapped phase is discontinuous at $L/2$ if the wrapped phase is associated with a $q\neq 0$ field configuration with nonzero $U(1)$ winding number. Since the same point $t=0$ is used for the correlation function source and the initial unwrapping point in this scheme, the unwrapped phase should be separately calculated for each correlation function in a MC ensemble when multiple source points are used with each field configuration. As discussed above, the phase differences $Q\theta(t) - Q\theta(0)$ for each charge sector $Q$ also need to be unwrapped individually since the unwrapped phase is a nonlinear function of the wrapped phase.
Complex scalar field MC ensembles {#subsec:cho-ensembles}
---------------------------------
Nineteen different choices of the $(0+1)D$ complex scalar field theory parameters $M^2$ and $\lambda$ indicated in Table \[tbl:cho-ensembles\] are employed to generate a variety of free and interacting MC ensembles. Free-field ensembles $A_0$, $B_0$, and $C_0$ are generated with parameters $M^2= 0.1,\ 0.025,$ and $0.00625$ and serve as toy models for LQFTs with coarse, moderate, and fine lattice spacings, respectively. In lattice units, the free-field correlation lengths defined by $\xi = 1/E$ are $$\begin{split}
\xi_{A_0} = 3.175, \hspace{20pt} \xi_{B_0} = 6.331, \hspace{20pt} \xi_{C_0} = 12.652,
\end{split}\label{eq:corrdef}$$ where the $L\rightarrow \infty$ approximation Eq. to the free scalar boson mass $E$ has been used. For each choice of $M^2$, the length of the lattice has been rescaled to $L = 128,\ 256,$ and $512$ respectively to enforce $M L = 128\sqrt{0.1} \approx 40.48$ and maintain a roughly constant temporal extent in units of the free-field correlation length. Two additional free-field ensembles $D_0$ and $E_0$ are generated with finer lattice spacing to explore lattice spacing dependence. In addition to the free-field ensembles, a variety of interacting scalar field theory ensembles are generated. Quartic self-interactions are used that correspond to potentials $$\begin{split}
V(|\varphi|) = \lambda |\varphi|^4,
\end{split}\label{eq:phi4}$$ with a variety of couplings $\lambda$ indicated in Table \[tbl:cho-ensembles\]. Repulsive couplings $\lambda > 0$ are necessary for the action to be bounded from below and for the path integral representing the thermal partition function to converge. With $\lambda > 0$, the partition function is well defined for $M^2 < 0$. In higher dimensions, this corresponds to a phase of complex scalar field theory where the $U(1)$ global symmetry is spontaneously broken. In $(0+1)D$ at finite $L$ the correlation length is finite in the $M^2 < 0$ phase but much larger than in the $M^2 > 0$ phase with the same $\lambda$ and $|M^2|$. Ensembles $A_n^\pm$, $B_n^\pm$ and $C_n^\pm$ describe interacting scalar field theories with the same $|M^2|$ as $A_0$, $B_0$, and $C_0$ with the sign of $M^2$ indicated by a superscript and two different values of $\lambda$ shown in Table \[tbl:cho-ensembles\] denoted by subscripts $n=1,2$. Ensembles with different $|M^2|$ but the same $\lambda$ subscript correspond to choices of $\lambda$ that keep $\lambda L / |M^2|$ fixed. Two additional negative $M^2$ ensembles $D_1^-$ and $E_1^-$, corresponding to $D_0$ and $E_0$, are also generated for a detailed study of lattice spacing dependence in the interacting case.
To interpret calculations at different $M^2$ and $L$ as having a fixed physical correlation length and varying lattice spacing $a$ for a $(0+1)D$ field of mass dimension $[\varphi] = -1/2$, the dimensionless parameters used in the MC calculations should be interpreted as $(Ma)^2$, $L/a$, and $a^3 \lambda$. This scaling is obtained if the dimensionless parameter $a^3 \lambda$ is chosen for calculations at different $(aM)^2$ and $L/a$ such that $\lambda L/M^2$ is held fixed. In appropriately rescaled units, the spectrum $Ea$ obtained at different $(Ma)^2$, $L/a$, and $a^3\lambda$ but fixed $\lambda L/M^2$ will differ by $O(\lambda)$ in small-$\lambda$ perturbation theory. It is tempting to interpret this as a renormalization condition that permits quantitative comparison of results at different $Ma$ and $L/a$; however, $(0+1)D$ complex scalar field theory is nonrenormalizable and an infinite number of renormalization conditions need to be imposed to consistently match results for the spectrum of LQFTs with different parameters to all orders in $\lambda$. Spectral results from LQFT calculations with different parameters but fixed $\lambda L/M^2$ should approximately agree for $\lambda \ll 1$ and $L/M^2 \gg 1$ but will differ in general by nonuniversal corrections. This nonuniversality is an artifact of working in $(0+1)D$ and arises even at $\lambda = 0$ where it can be understood as arising from higher-derivative operators in a Symanzik-improved lattice action [@Symanzik:1983dc]. Since there is no universal continuum limit for $(0+1)D$ scalar field theory, no attempt is made to employ non-perturbative renormalization conditions to match observables between ensembles with different parameters. In the $ML \approx 40.48$ MC ensembles considered here, the two different values of $(\lambda L / M^2)_n$ held fixed among $A_n^\pm$, $B_n^\pm$ and $C_n^\pm$ correspond to $(\lambda L / |M^2|)_1 = 16$ and $(\lambda L / |M^2|)_2 = 32$ respectively.
$A_0$ $A_1^\pm$ $A_2^\pm$
----------- -------- ----------- -----------
$L$ 128
$M^2$ $+0.1$ $\pm$ $\pm$
$\lambda$ 0 0.0125 0.025
: Ensembles used for complex scalar investigation, segmented by size. For the free-field cases $A_0$, $B_0$, $C_0$, $D_0$, and $E_0$, a consistent positive $M^2$ is chosen to match physical lattice extent. For the interacting cases of series $A$, $B$, and $C$, both $\pm M^2$ are used, while for series $D$ and $E$ only $-M^2$ is used. Each ensemble is updated via a Metropolis sweep over the odd/even lattice sites $N_{\text{skip}} = 10000$ times between each measurement. Before saving lattice measurements, $N_{\text{therm}} = 50$ iterations of the complete measurement cycle are performed for thermalization. Following thermalization, $N_{\text{meas}} = 5000$ measurement cycles are performed with the values of $\varphi(t)$ saved each time.[]{data-label="tbl:cho-ensembles"}
$B_0$ $B_1^\pm$ $B_2^\pm$
---------------- ---------- ----------- -----------
$L$ 256
$M^2$ $+0.025$ $\pm$ $\pm$
$10^3 \lambda$ 0 1.5625 3.125
: Ensembles used for complex scalar investigation, segmented by size. For the free-field cases $A_0$, $B_0$, $C_0$, $D_0$, and $E_0$, a consistent positive $M^2$ is chosen to match physical lattice extent. For the interacting cases of series $A$, $B$, and $C$, both $\pm M^2$ are used, while for series $D$ and $E$ only $-M^2$ is used. Each ensemble is updated via a Metropolis sweep over the odd/even lattice sites $N_{\text{skip}} = 10000$ times between each measurement. Before saving lattice measurements, $N_{\text{therm}} = 50$ iterations of the complete measurement cycle are performed for thermalization. Following thermalization, $N_{\text{meas}} = 5000$ measurement cycles are performed with the values of $\varphi(t)$ saved each time.[]{data-label="tbl:cho-ensembles"}
$C_0$ $C_1^\pm$ $C_2^\pm$
---------------- ------------ ----------- -----------
$L$ 512
$M^2$ $+0.00625$ $\pm$ $\pm$
$10^3 \lambda$ 0 0.1953125 0.390625
: Ensembles used for complex scalar investigation, segmented by size. For the free-field cases $A_0$, $B_0$, $C_0$, $D_0$, and $E_0$, a consistent positive $M^2$ is chosen to match physical lattice extent. For the interacting cases of series $A$, $B$, and $C$, both $\pm M^2$ are used, while for series $D$ and $E$ only $-M^2$ is used. Each ensemble is updated via a Metropolis sweep over the odd/even lattice sites $N_{\text{skip}} = 10000$ times between each measurement. Before saving lattice measurements, $N_{\text{therm}} = 50$ iterations of the complete measurement cycle are performed for thermalization. Following thermalization, $N_{\text{meas}} = 5000$ measurement cycles are performed with the values of $\varphi(t)$ saved each time.[]{data-label="tbl:cho-ensembles"}
$D_0$ $D_1^-$
---------------- -------------- ------------
$L$ 1024
$M^2$ $+0.0015625$ $-$
$10^6 \lambda$ 0 24.4140625
: Ensembles used for complex scalar investigation, segmented by size. For the free-field cases $A_0$, $B_0$, $C_0$, $D_0$, and $E_0$, a consistent positive $M^2$ is chosen to match physical lattice extent. For the interacting cases of series $A$, $B$, and $C$, both $\pm M^2$ are used, while for series $D$ and $E$ only $-M^2$ is used. Each ensemble is updated via a Metropolis sweep over the odd/even lattice sites $N_{\text{skip}} = 10000$ times between each measurement. Before saving lattice measurements, $N_{\text{therm}} = 50$ iterations of the complete measurement cycle are performed for thermalization. Following thermalization, $N_{\text{meas}} = 5000$ measurement cycles are performed with the values of $\varphi(t)$ saved each time.[]{data-label="tbl:cho-ensembles"}
$E_0$ $E_1^-$
---------------- ---------------- --------------
$L$ 2048
$M^2$ $+0.000390625$ $-$
$10^6 \lambda$ 0 3.0517578125
: Ensembles used for complex scalar investigation, segmented by size. For the free-field cases $A_0$, $B_0$, $C_0$, $D_0$, and $E_0$, a consistent positive $M^2$ is chosen to match physical lattice extent. For the interacting cases of series $A$, $B$, and $C$, both $\pm M^2$ are used, while for series $D$ and $E$ only $-M^2$ is used. Each ensemble is updated via a Metropolis sweep over the odd/even lattice sites $N_{\text{skip}} = 10000$ times between each measurement. Before saving lattice measurements, $N_{\text{therm}} = 50$ iterations of the complete measurement cycle are performed for thermalization. Following thermalization, $N_{\text{meas}} = 5000$ measurement cycles are performed with the values of $\varphi(t)$ saved each time.[]{data-label="tbl:cho-ensembles"}
In addition to MC ensembles generated using the standard action in Eq. with the parameter choices described above, ensembles are also generated using the analytically phase-integrated dual form of the theory given in Eq. -. These dual ensembles only involve MC sampling over the magnitude of the scalar field, and as demonstrated analytically above they have no sign problem and a much milder (though still exponentially severe) StN problem with increasing $t$. It is verified below that these dual ensembles lead to more precise calculations of ground-state energies $E_{Q,0}$ in charge sectors $Q=1,\dots,4$. These results are used as a precise check on the accuracy of results obtained using the standard ensembles with and without phase unwrapping. Results show small but statistically significant differences between the low-lying energy levels of ensembles with different parameters but equal $\lambda L / M^2$. More details on free-field consistency checks and autocorrelation times can be found in Appendix \[app:MC\].
1D phase unwrapping results {#subsec:cho-results}
---------------------------
Where the phase varies smoothly, a nearest-neighbor unwrapping scheme accurately captures the variation in phase across the ensemble. Close to large phase jumps with $|\partial_t \theta| > \pi/2$, winding number assignment can vary depending on the phase unwrapping algorithm. Different algorithms will lead to distributions that broaden more or less quickly with $t$ and therefore different low-order cumulant expansion truncation errors. Finding an unwrapping scheme with fast convergence in the cumulant expansion amounts to choosing an algorithm for winding number assignment in neighborhoods of large phase jumps that appropriately tunes the unwrapped phase variance growth controlling $\widetilde{E}^{(2)}_{Q,0}$.
![Effective mass measurements comparing estimation using cumulants up to order $n_{max} = 2,4,6$ with several window sizes for ensembles $C_2^+$ and $C_2^-$. The red band indicates the dual variables estimate of the correct mass. The choice $w=16$ provides an accurate estimate for $C_2^-$ with $n_{max} = 2$ and little variation with increasing $n_{max}$, while other choices have much larger truncation errors at $n_{max} = 2$ that are reduced by increasing the truncation order. The choice $w=10$ provides the most accurate estimate for $C_2^+$ with $n_{max} = 2$, although statistically significant truncation errors are still visible. For all window choices shown, results with $n_{max} = 6$ are statistically consistent with dual variable calculations at the $1\sigma -2 \sigma$ level. []{data-label="fig:meff-vs-cumu-pos-neg-mass"}](m0p00625_l0p000390625_meff_vs_ncumu.png "fig:"){width="80.00000%"} ![Effective mass measurements comparing estimation using cumulants up to order $n_{max} = 2,4,6$ with several window sizes for ensembles $C_2^+$ and $C_2^-$. The red band indicates the dual variables estimate of the correct mass. The choice $w=16$ provides an accurate estimate for $C_2^-$ with $n_{max} = 2$ and little variation with increasing $n_{max}$, while other choices have much larger truncation errors at $n_{max} = 2$ that are reduced by increasing the truncation order. The choice $w=10$ provides the most accurate estimate for $C_2^+$ with $n_{max} = 2$, although statistically significant truncation errors are still visible. For all window choices shown, results with $n_{max} = 6$ are statistically consistent with dual variable calculations at the $1\sigma -2 \sigma$ level. []{data-label="fig:meff-vs-cumu-pos-neg-mass"}](m0p00625n_l0p000390625_meff_vs_ncumu.png "fig:"){width="80.00000%"}
It is empirically found that the single-point integration phase unwrapping method described in Sec. \[sec:unwrapnum\] that enforces $|\widetilde{\theta}(t) - \widetilde{\theta}(t-1)| < \pi$ gives poor results for the scalar boson mass across all ensembles. Results do not markedly improve at finer lattice spacing. Statistical precision is generally good for correlation functions and ground-state energies estimated from cumulant expansions truncated at low orders with qualitatively similar $t$ scalings to those shown for Gaussian-weighted integration of $C_0$ in Figs. \[fig:unwrapStN\]-\[fig:unwrapStNQ\]. However, the truncation errors of second- and third-order results for $\widetilde{E}^{(n)}$ are large, sometimes an order of magnitude larger than both the statistical uncertainties and central values of standard ensemble average estimates of $E$. Truncation errors decrease at higher orders in the expansion, generally with a pattern of visible decreases at even orders that are sensitive to the shape of the phase distribution, but statistical errors increase dramatically.
![The scalar boson mass determined with a variety of phase unwrapping algorithm parameters and truncated cumulant expansion for ensembles $C_2^\pm$. The green points show $\widetilde{E}^{(2)}$ obtained using windowed integration with window sizes $w$ shown on the horizontal axis. Dark green error bars on these points show $68\%$ confidence intervals including statistical uncertainties. The dark purple points and error bars show $\widetilde{E}^{(2)}$ and its statistical uncertainties obtained using Gaussian integration. The Gaussian widths $\sigma$ used are proportional to the window size $w$ with $w = 1.65\sigma$ for $C_2^-$ (left) and $w = 1.62\sigma$ for $C_2^+$ (right) empirically found to provide agreement between windowed and Gaussian integration. The lighter purple error bars on both Gaussian and windowed unwrapping points show the extent of the variation in central values of $\widetilde{E}^{(2)}$, $\widetilde{E}^{(4)}$, and $\widetilde{E}^{(6)}$ and demonstrate that results tend to converge towards dual ensemble results as $n_{max}$ is increased and also that the size of truncation errors is sensitive to the unwrapping algorithm parameters used. The red bands show dual ensemble results and statistical uncertainties for comparison. []{data-label="fig:gauss-window-sizes"}](window_vs_gaussian.png){width="\textwidth"}
![Effective mass plots for unwrapped phase cumulant expansion mass calculations for ensembles $C_2^\pm$. The upper plots use the optimal window sizes for $C_2^+$ and $C_2^-$ indicated in Fig. \[fig:meff-vs-cumu-pos-neg-mass\] and show little variation with truncation order at all $t$. The lower plots use suboptimal window sizes that include significant truncation errors with $n_{max} = 2$ and smaller truncation errors with larger $n_{max}$. The red band indicates the dual variables estimate of the scalar mass and its uncertainties.[]{data-label="fig:meff-vs-time-pos-neg-mass"}](m0p00625_m0p00625n_l0p000390625_cumulant_meff_by_time.pdf){width=".8\textwidth"}
The windowed and Gaussian-weighted integration methods that enforce smoothness on scales larger than the lattice scale provide estimates for correlation functions and energies with much smaller truncation errors than single-point integration. The statistical precision of results at various orders in the cumulant expansion is roughly independent of the choice of smearing scale (the window size $w$ or Gaussian width $\sigma$ used when calculating winding numbers) but the central values of low-order results depend sensitively on the smearing scale. A representative demonstration of this tendency is shown for $\widetilde{E}^{(n)}$ for ensembles $C_2^+$ and $C_2^-$ in Figs. \[fig:meff-vs-cumu-pos-neg-mass\]-\[fig:gauss-window-sizes\]. Results with Gaussian-weighted integration tend to match results with windowed integration up to a single $O(1)$ constant of proportionality between $w$ and $\sigma$. Gaussian-weighted integration can be tuned to interpolate between integer-valued window sizes in this way. An empirical condition relating the phase unwrapping smearing scale to the correlation length or another cost function penalizing large truncation errors could be used to self-consistently define an optimal smearing scale, but it is difficult to assess the systematic errors of optimized estimates without sacrificing precision by going to higher orders in the cumulant expansion.
![Statistical variance in ground-state energy estimate versus correlator time separation for ensembles $C_2^\pm$. The gray overlay plots the variance for the standard effective mass estimator, demonstrating the exponentially decaying StN problem where the variance estimate remains reliable. The red line indicates the theoretical Parisi-Lepage StN decay $\mathcal{N} e^{-Et}$ with $E$ given by the precise dual variables estimate and normalization $\mathcal{N}$ determined by a fit to the first $L/8$ values. The purple points show the variance of the effective mass in the dual lattice variable ensemble and demonstrate exponential variance growth that is significantly less severe than the standard effective mass. The $n_{max} = 2$ estimate with phase unwrapping has even less severe variance growth and becomes more precise than the dual variable estimate at large $t$. Phase unwrapped cumulant effective masses with $n_{max} = 2,\ 4,\ 6$ show variance growth with downward curvature on the logarithmic scale shown that is consistent with polynomial variance growth, though it is difficult to robustly distinguish high-order polynomial from exponential variance growth numerically.[]{data-label="fig:cho-stn"}](c2_uw_stn.png){width="\textwidth"}
![Ground-state energies $E_Q \equiv E_{Q,0}$ for charge sectors $Q=1,\dots,4$ in interacting complex scalar field theory with positive $M^2$ (top) and negative $M^2$ (bottom) as well as two choices of $|\lambda L / M^2| = 16$ (left) and $|\lambda L / M^2| = 32$ (right). Colored error bars indicate 68% confidence intervals including statistical errors only, while thin gray error bars include systematic uncertainties associated with variation of fitting window range and cumulant expansion truncation errors added to statistical errors in quadrature. Systematic uncertainties associated with fitting window range variation are estimated as one half the difference between maximum and minimum central values for fit windows shifted by one and two time slices. Systematic uncertainties associated with truncation errors are estimated by the maximum difference between the central value of $\widetilde{E}^{(2)}_{Q,0}$ and the central values of $\widetilde{E}^{(4)}_{Q,0}$ and $\widetilde{E}^{(6)}_{Q,0}$. In several cases the $Q = 4, M^2 = 0.1$ standard estimator does not reliably plateau and systematic errors cover the plot range with no estimate for the central value shown. The vertical axis shows ratios of interacting energies $E_{Q,0}(\lambda,M,L)$ to the noninteracting scalar boson mass $E$ to facilitate comparison with noninteracting energies $E_{Q,0}(\lambda=0,M,L) = Q E$. []{data-label="fig:cho-spectrum-phi4"}](spectrum.png){width="\textwidth"}
Empirically, window sizes tuned to reproduce the correlation length in each charge sector $w_Q \sim \xi_Q = 1/E_{Q,0}$ tend to give accurate results for $E_{Q,0}$ at low orders in the cumulant expansion. Results for the ground-state energies in $Q=1,\dots,4$ charge sectors obtained with optimally tuned Gaussian integration phase unwrapping for the free-field ensemble $C_0$ are summarized in Figs. \[fig:unwrapStN\]-\[fig:unwrapStNQ\]. It is noteworthy that second-order truncated cumulant expansion energy estimates $\widetilde{E}^{(2)}_{Q,0}$ have negligible StN loss with increasing $t$ at fixed $Q$ and with increasing $Q$ at fixed $t$. Close agreement between $\widetilde{E}^{(2)}_{Q,0}$ and the exact results for $E_{Q,0}$ is obtained but requires tuning a smearing parameter in the phase unwrapping algorithm, see Sec. \[sec:unwrap1D\] for further discussion. Higher-order truncations $\widetilde{E}^{(n_{max})}_{Q,0}$ have StN ratios that noticeably decrease with increasing $t$ and with increasing $Q$. This StN decrease shows less curvature on the log-log scale in Fig. \[fig:unwrapStNQ\] than the exponential StN decrease of $E_{Q,0}$, and numerical results are consistent with constant StN at second-order and increasingly high-order polynomial StN degradation at increasingly high cumulant expansion truncation order.
The cumulant expansion tends to converge from above or below depending on whether the smearing scale is tuned to be larger or smaller than the physical correlation length. A heuristic explanation for these observations is that unwrapping with an overly small smearing scale is overly sensitive to short-distance fluctuations and erroneously adds winding numbers while unwrapping with an overly large smearing scale penalizes diffusive motion away from physically uncorrelated points and leads to underbroadening. Fig. \[fig:meff-vs-time-pos-neg-mass\] demonstrates that this overbroadening or underbroadening is a time-independent feature when estimating the effective mass. In either case, truncation errors are reduced by going to higher order in the cumulant expansion at the cost of decreased statistical precision.
The StN behavior of phase unwrapped ensembles using optimally tuned smearing parameters is shown in Fig. \[fig:cho-stn\]. There is very little StN degradation in $\widetilde{E}^{(2)}$. Standard ensemble average correlation functions show exponential StN degradation with the expected $O(e^{-E_{Q,0}t})$ scaling, while dual variable correlation functions show much more mild but likely still exponential StN scaling. For the largest source/sink separations, the dual estimate precision grows to become worse than the precision of $\widetilde{E}^{(2)}$. The limiting factors on the accuracy of low-order results in the cumulant expansion extracted with optimally tuned phase unwrapping are the systematic uncertainties associated with truncation errors and phase unwrapping parameter tuning, not statistical precision. The precision of $\widetilde{E}^{(2)}_{Q,0}$ and systematic truncation errors are both clearly visible in results for the ground-state energies of charge sectors $Q=1,\dots,4$ in interacting complex scalar field MC ensembles in Fig. \[fig:cho-spectrum-phi4\]. Truncation errors are estimated from the maximum difference between the central value of $\widetilde{E}^{(2)}_{Q,0}$ and the central values of $\widetilde{E}^{(4)}_{Q,0}$ and $\widetilde{E}^{(6)}_{Q,0}$. After including this difference as a systematic uncertainty added in quadrature with the statistical errors, $\widetilde{E}^{(2)}_{Q,0}$ results are consistent with precise results from the dual ensembles. Systematic truncation uncertainties determined in this way are significantly larger than statistical uncertainties. The combination of large truncation errors in $\widetilde{E}^{(2)}_{Q,0}$ and significant variance growth with increasing $n_{max}$ prevents phase unwrapped results from providing precise and accurate results for the spectrum of $(0+1)D$ complex scalar field theory.
![The left plot shows the probability of large phase jumps defined by $|\theta(t) - \theta(t-1)| > \pi/2$ for a variety of ensembles. The blue curve shows results for free-field ensembles $A_0$, $B_0$, $C_0$, $D_0$, and $E_0$ as a function of $1/|M|$ and therefore approximately as a function of the correlation length. The green curve shows analogous results for interacting scalar field ensembles $A_1^-$, $B_1^-$, $C_1^-$, $D_1^-$ and $E_1^-$ with $M^2 < 0$ and fixed $-\lambda L/M^2 = 16$ corresponding to fixed coupling strength in units of the $O(\lambda^0)$ tree-level correlation length. Results for phase differences between all nearest-neighbor sites on the lattice are averaged, and error bars on each point indicate $68\%$ confidence intervals calculated using bootstrap techniques. The dotted red (dashed pink) curve shows the predictions of Eq. corresponding to von Mises distributed phase differences with $\kappa \approx 1/(2E)$ calculated for the free (interacting) ensembles. The right plot shows the same probabilities multiplied by the lattice size $L$ to represent the probability that a field configuration will have a large phase jump. []{data-label="fig:phase-jumps"}](lattice_spacing_jumps.png){width=".9\textwidth"}
As the lattice spacing is taken much smaller than the physical correlation length, phase differences between neighboring lattices become smaller on average. Under the fixed scalar field magnitude assumption, this probability can be calculated using the von Mises distribution derived exactly for $\partial_t \theta$ in Eq. to be $$\begin{split}
\mathcal{P}(|\partial_t \theta| > \pi - \varepsilon) = \frac{2}{I_0(\kappa)} \int_{\pi - \varepsilon}^\pi \frac{d \Delta}{2\pi} e^{\kappa \cos(\Delta)}.
\end{split}\label{eq:Panalytic}$$ Under small fluctuation assumptions and neglecting excited-states, $\kappa \approx 1/(2E) $ as in Eq. and $\kappa$ therefore becomes large as the correlation length becomes large in lattice units. The probability in Eq. vanishes rapidly as $\kappa \rightarrow\infty$ with $\varepsilon >0$, and so one may expect the probability of large phase jumps in a MC ensemble to vanish as $M^2 \rightarrow 0$. However, there is some non-negligible probability that $|\varphi|$ fluctuates to become arbitrarily small even at very small lattice spacing; for example, this occurs due to nearly coincident zero crossings of the real and imaginary parts of $\varphi$ as they fluctuate from one sign to the other as shown in Fig. \[fig:hardunwrapping\]. The distribution of $\partial_t \theta$ in LQFT MC ensembles is given by marginalizing over $\kappa$, and nontrivial correlations between the magnitude and phase could lead to significant departures from the expectations of Eq. . Such departures are seen in Fig. \[fig:phase-jumps\], where the probability of jumps larger than $\varepsilon = \pi/2$ appears to vanish as $M^2 \rightarrow 0 $ much more slowly than predicted by Eq. . Similar scaling is found in free-field theory, interacting field theory with $M^2>0$, and somewhat surprisingly also in the $M^2 < 0$ regime where the magnitude typically fluctuates about local minima where $\kappa$ in Eq. is nonzero. The expected number of large phase jumps per field configuration $L \times \mathcal{P}(|\partial_t \theta| > \pi/2)$ is empirically observed to grow as $|M^2|$ is decreased and $L$ is increased to hold $|M|L$ fixed, suggesting that there is never a physically relevant regime that is likely to be free of large phase jumps and the phase unwrapping ambiguities associated with them.
The result that the number of large phase jumps per configuration grows faster than the number of sites grows as the lattice spacing is decreased is particularly troubling because of an accumulation-of-errors (or differences) problem arising in $1D$ phase unwrapping. If there is a link connecting $t_{jump}$ and $t_{jump}+1$ with a large phase difference $|\theta(t_{jump}+1) - \theta(t_{jump})| \gtrapprox \pi$, then different phase unwrapping algorithms tend to assign different winding numbers following $t_{jump}$. With the forward integration schemes described above, this differing winding number at $t_{jump}$ will lead to differences of $2\pi$ in the phase unwrapped by different unwrapping schemes at all $t \geq t_{jump}$. This accumulation-of-errors problem means that large phase differences between nearest neighbor lattice sites, which might be considered lattice artifacts, lead to scheme-dependent variation of size $(2\pi)^n$ in contributions of a MC correlation function to the $n$th moment of the unwrapped phase that do not disappear as the lattice spacing is reduced. Further studies are necessary to understand whether this scaling is an artifact of the nonuniversality of $(0+1)D$ complex scalar field theory or a feature that persists in $1D$ unwrapping of momentum-projected correlation functions of renormalizable LQFTs.
While these results on the prevalence of large phase jumps and accumulation of errors in phase unwrapping suggest a pessimistic outlook for $(0+1)D$ complex scalar field theory, other applications of phase unwrapping provide encouraging results demonstrating that the $1D$ accumulation of errors problem becomes more tractable in higher dimensions. It was realized in the 1980s that accumulation of phase unwrapping errors along a $1D$ integration path is a generic problem in the presence of undersampling [@Itoh:1982] but can be avoided in alternative algorithms for $2D$ phase unwrapping [@Goldstein:1988; @Huntley:1989]. The basic source of greater robustness in higher dimensions is that in $1D$ only one integration path[^11] can be used to connect two points $t$ and $t^\prime$, while in two- and higher-dimensions multiple paths can be used to connect the same points. Assuming that $\Theta(x,y) = \text{arg} [G(x,y)]$ where $G(x,y)$ is an analytic function, $1D$ unwrapping will provide identical results for phase integration from $y$ to $x$ that do not depend on the choice of $1D$ integration contour, see e.g. Ref [@Ghiglia:98]. Under this analyticity assumption, path dependence that arises in numerical data must be the result of numerical noise and sampling several $1D$ unwrapping paths adds error correction through redundancy. Applications in 3D have been found to be even more robust to noise than applications in $2D$, suggesting that phase unwrapping generically becomes more robust as the number of dimensions is increased [@Huntley:01; @Hooper:07; @Abdul-Rahman:09]. A simple argument supporting this idea is that phase unwrapping makes smoothness assumptions informed by nearest-neighbor phases, and as the number of dimensions increases the number of nearest neighbors that can be used to inform a phase unwrapping algorithm also increases.
Successful applications of numerically robust phase unwrapping algorithms in higher dimensions crucially rest on the assumption that wrapped phases are discrete samples of the complex logarithm of an underlying analytic function. In this case noise may locally produce regions in which phase unwrapping along differing paths produces results that differ by multiples of $2\pi$, but unwrapping along paths avoiding these regions is guaranteed by the underlying analyticity to produce identical results. If the wrapped phase is sampled with sufficiently high resolution and low noise, then the density of points with large phase jumps leading to unwrapping ambiguities is guaranteed to be vanishingly small. Field configurations in LQFTs are not analytic and are not expected to approach smooth or even continuous functions as the continuum limit is approached. Instead, field configurations may approach distributions including isolated singularities that will lead to nonlocal unwrapping ambiguities. It is not clear without further studies of multidimensional LQFTs whether appropriately smeared configurations calculated on finely discretized lattices will be smooth enough for phase unwrapping algorithms to determine unwrapped phases without ambiguities arising from large phase jumps. Even if large phase jumps are unavoidable in multidimensional LQFTs, more robust multidimensional phase unwrapping algorithms can still be used to avoid the $1D$ unwrapping accumulation of errors problem encountered here. By applying multidimensional phase unwrapping algorithms to correlation functions in coordinate space, $2\pi$ ambiguities from large phase jumps leading to the $1D$ accumulation of errors problem could be localized to isolated neighborhoods of spacetime. This might improve the convergence of the cumulant expansion and reduce StN degradation at higher orders. Precise but approximate results at low orders in the cumulant expansion could also be used as starting points for subsequent calculations of differences between exact and approximate correlation functions that might be more efficient than calculations of exact correlation functions alone.
Conclusions {#sec:conclusions}
===========
In $(0+1)D$ complex scalar field theory, phase fluctuations distinguish correlation functions in $Q\neq 0$ charge sectors from vacuum sector correlation functions. These phase fluctuations result in sign problems for the path integrals representing correlation functions, even though the vacuum sector partition function is positive-definite. A method for avoiding $(0+1)D$ scalar field sign problems is introduced that relies on numerically integrating time series of phase differences at a range of source/sink separation using phase unwrapping techniques developed for signal processing and a variety of engineering applications. The nonzero moments of the unwrapped phase distribution can be computed with positive-definite path integrals without sign problems. A cumulant expansion involving moments of correlation function log-magnitudes and unwrapped phases can be used to reproduce the spectrum of $(0+1)D$ complex scalar field theory. The numerical results presented here include large systematic truncation errors at low orders in the cumulant expansion and decreased precision as well as the reemergence of a mild StN problem at higher orders. It is argued that the large truncation errors arise from isolated large phase jumps that lead to errors in the $n$th moment proportional to $(2\pi)^n$ at all subsequent times. This accumulation-of-errors problem makes results using a cumulant expansion of the unwrapped phase numerically sensitive to the presence of large phase jumps. Numerical MC studies suggest that in $(0+1)D$ scalar field theory the probability of having one or more large phase jumps per lattice extent grows as the $M^2\rightarrow 0$ limit is taken to increase the correlation length, and this accumulation-of-errors problem leads to large systematic errors even at very fine lattice spacing. This may be due to the non-renormalizability of $(0+1)D$ scalar field theory and these investigations should be extended to renormalizable field theories to better understand this issue. The appearance of heavy-tailed phase derivative distributions in free-field theory as well as in LQCD baryon correlation functions [@Wagman:2016bam] suggests that these problematic large phase jumps are present in physically relevant LQFTs and are possibly generic features of correlation functions with phase fluctuations. If heavy-tailed phase differences are generic features of LQFT, then high moments of the unwrapped phase sensitive to the tails of the distribution may be noisy and convergence of the cumulant expansion may be slow. Leading-order cumulant expansion results using appropriately tuned phase unwrapping algorithms provide precise approximations to correlation functions that avoid sign or StN problems, but robust applications of phase unwrapping in multidimensional LQFTs will require a solution to the $1D$ accumulation of errors problem and perhaps alternative methods of including corrections from noisy higher-order terms.
[**Acknowledgments:**]{} We would like to thank Adam Bene Watts, Dorota Grabowska, David Kaplan, Christopher Monahan, Andrew Pochinsky, Martin Savage, Phiala Shanahan, and Daniel Trewartha for helpful discussions. This work was partially supported by the U. S. Department of Energy through Early Career Research Award No. de-sc0010495 and Grant No. de-sc0011090 and by the SciDAC4 Grant No. de-sc0018121. MLW was supported by a MIT Pappalardo Fellowship.
Dual Lattice Variable Phase Integration {#app:dual}
=======================================
The action for a $(0+1)D$ complex scalar field with an arbitrary $U(1)$ invariant, spacetime translation invariant potential energy function $V(|\varphi|)$ can be decomposed into magnitude and phase contributions as $$\begin{split}
S(\varphi) &= \sum_{t=0}^{L - 1} \left\lbrace |\varphi(t)|\left[ -|\varphi(t-1)|e^{i\theta(t) - i\theta(t-1)} + 2|\varphi(t)| - |\varphi(t+1)|e^{i\theta(t) - i\theta(t+1)} \right] + V(|\varphi(t)|) \right\rbrace \\
&= \sum_{t=0}^{L - 1} \left\lbrace 2|\varphi(t)|^2 + V(|\varphi(t)|) - \kappa(t)\cos(\theta(t) - \theta(t-1)) \right\rbrace
\end{split}\label{eqA:actiondecomp}$$ where translation invariance has been used to shift field arguments. The partition function for the interacting theory can be similarly decomposed as, $$\begin{split}
Z &= \int_0^\infty \prod_{t= 0}^{L-1} \left[ d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \right] \int_{-\pi}^\pi \prod_{t= 0}^{L-1} \left[ \frac{1}{\pi }d \theta(t) \; e^{\kappa(t)\cos(\theta(t) - \theta(t-1))} \right]
\end{split}\label{eqA:Zdecomp}$$ The phase integral can be evaluated analytically be introducing dual lattice variables representing the differences between phases at adjacent lattice sites, $$\begin{split}
\Delta(t) &\equiv \theta(t) - \theta(t-1).
\end{split}\label{eq:Detladef}$$ This transformation is related to dual lattice variable methods that have a long history in lattice gauge theory [@Ukawa:1979yv] and can be viewed as a $1D$ analog of the $O(N)$ model dual lattice variable transformation introduced in Ref. [@Endres:2006xu]. To simplify the change of variables, we first use $2\pi$-periodicity to rotate the integration domains for the sequence of $\theta(t)$ integrals, $$\int_{-\pi}^\pi d\theta(0) \prod_{t=1}^{L-1} \sq{\int_{-\pi}^\pi d\theta(t)}
\qquad \rightarrow \qquad
\int_{-\pi}^\pi d\theta(0) \prod_{t=1}^{L-1} \sq{\int_{\theta(t-1)-\pi}^{\theta(t-1)+\pi} d\theta(t)}.$$ We then change variables from $\theta(t)$ to $\Delta(t)$ for all $t \geq 1$, with trivial Jacobian, $$\begin{split}
\det \left(\frac{\partial(\theta(0),\Delta(1),\dots,\Delta(L - 1))}{\partial(\theta(0),\theta(1),\dots,\theta(L-1))}\right) = \det \begin{pmatrix} 1&0&0&0&\cdots \\ -1&1&0&0&\cdots \\ 0&-1&1&0& \cdots \\ \vdots&\vdots&\vdots&\vdots&\ddots \end{pmatrix} = 1
\end{split}.
\label{eq:Jacobian}$$ Having first rotated the integration ranges, we are left with simple integration bounds for the newly-introduced dual variables, $$\int_{-\pi}^\pi d\theta(0) \prod_{t=1}^{L-1} \sq{\int_{\theta(t-1)-\pi}^{\theta(t-1)+\pi} d\theta(t)}
= \int_{-\pi}^\pi d\theta(0) \prod_{t=1}^{L-1}
\sq{\int_{-\pi}^{\pi} d\Delta(t)}.$$ To completely decouple the integrals, we would like to introduce the final variable $\Delta(0) = \theta(0) - \theta(L-1)$. The presence of PBCs slightly complicates this transformation by introducing the constraint $$\Delta(0) \equiv \theta(0) - \theta(L-1) - \sum_{t=1}^{L-1} \Delta(t)$$ which can be implemented by a $\delta$-function $$1 = \int_{-\infty}^{\infty} d\Delta(0) \delta \paren{\sum_{t=0}^{L-1} \Delta(t)}.$$ The compact nature of the phase variables requires this final integral to run over all reals to satisfy PBCs in all winding number sectors. To treat all dual variables on equal footing, we instead handle this sum over winding number sectors directly, $$\int_{-\infty}^{\infty} d\Delta(0) \delta \paren{\sum_{t=0}^{L-1} \Delta(t)}
= \sum_{w \in \mathbb{Z}} \int_{-\pi}^{\pi} d\Delta(0) \delta \paren{\sum_{t=0}^{L-1} \Delta(t) + 2 \pi w}.$$
This path integral change of variables from $\theta$ to $\Delta$ turns the phase integrals turns into a product of decoupled integrals. This representation allows the integral over phases to be explicitly evaluated as $$\begin{split}
Z &= \int_0^\infty \prod_{t= 0}^{L-1} \left[ d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \right] \\
&\hspace{20pt} \times \int_{-\pi}^\pi \frac{1}{\pi}d\theta(0) \int_{-\pi}^{\pi} \prod_{t= 0}^{L-1} \left[ \frac{1}{\pi} d \Delta(t) \; e^{\kappa(t)\cos(\Delta(t))} \right] \sum_{w \in \mathbb{Z}} \delta\left( \sum_{t=0}^{L-1} \Delta(t) + 2\pi w \right)\\
&= 2 \int_0^\infty \prod_{t= 0}^{L-1} \left[ d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \right] \\
&\hspace{20pt} \times \int_{-\infty}^\infty dq \; \sum_{w \in \mathbb{Z}} e^{2\pi i q w} \int_{-\pi}^{\pi} \prod_{t= 0}^{L-1} \left[ \frac{1}{\pi} d \Delta(t) \; e^{\kappa(t)\cos(\Delta(t)) + i q \Delta(t) } \right] \\
&= 2 \sum_{q\in\mathbb{Z}} \int_0^\infty \prod_{t= 0}^{L-1} \left[ d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \; 2\; I_{\abs{q}}\left( \kappa(t) \right) \right]
\end{split}\label{eq:CHOZdecomp}$$ where we have used an integral representation to factorize the $\delta$-function and explicitly integrated the $\Delta(t)$ to produce modified Bessel functions of the first kind, $I_{|q|}(z)$. The remaining integrals over $|\varphi(t)|$ cannot be evaluated in closed form for arbitrary $V(|\varphi|)$; however, since $I_{|q|}(z) \geq 0$ for $z\geq 0$, the form of the partition function given in the final line of Eq. defines a positive-definite probability density for $q, \varphi$: $$\begin{split}
1 &= \sum_{q\in\mathbb{Z}} \int \mathcal{D}|\varphi|\; \mathcal{P}(q,|\varphi|) \\
&\equiv \frac{2}{Z} \sum_{q\in\mathbb{Z}}
\int
\prod_{t=0}^{L-1} d|\varphi(t)| \curly{ |\varphi(t)| e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \;
2\; I_{\abs{q}}\left( \kappa(t) \right)
}.
\end{split}
\label{eq:CHOP1}$$ The probability distribution $\mathcal{P}(q,|\varphi|)$ is a positive-definite, normalizable function that can be used for MC sampling of $|\varphi|$ and $q$ as described below.
The integrals over phase variables can similarly be performed analytically for scalar field correlation functions. The general correlation function can first be written in terms of the new dual variables, $$\mathcal{O}_{Q,2P}(t)\mathcal{O}_{Q,2P}^*(0) = |\varphi(t)|^{|Q|+2P}|\varphi(0)|^{|Q|+2P} e^{i Q \sum_{t^\prime = 1}^t \Delta(t^\prime)}.$$ Inserting this observable into the path integration and explicitly evaluating gives $$\begin{split}
G_{Q,2P}(t) &= \frac{1}{Z} \int_0^\infty \prod_{t= 0}^{L-1} \left[ d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \right] |\varphi(t)|^{|Q|+2P}|\varphi(0)|^{|Q|+2P} \\
&\hspace{20pt} \times \int_{-\pi}^\pi \frac{1}{\pi } d\theta(0) \int_{-\pi}^{\pi} \prod_{t= 0}^{L-1} \left[ \frac{1}{\pi} d \Delta(t) \; e^{\kappa(t)\cos(\Delta(t))} \right] \sum_{w \in \mathbb{Z}} \delta\left( \sum_{t=0}^{L-1} \Delta(t) + 2\pi w \right) e^{i Q \sum_{t^\prime = 1}^t \Delta(t^\prime)} \\
&= \frac{2}{Z} \int_0^\infty \prod_{t= 0}^{L-1} \left[ d|\varphi(t)|\; |\varphi(t)| \; e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \right] |\varphi(t)|^{|Q|+2P}|\varphi(0)|^{|Q|+2P} \\
&\hspace{20pt} \times \int_{-\infty}^\infty dq \; \sum_{w \in \mathbb{Z}} e^{2\pi i q w} \int_{-\pi}^{\pi} \prod_{t= 0}^{L-1} \left[ \frac{1}{\pi} d \Delta(t) \; e^{\kappa(t)\cos(\Delta(t)) + i q \Delta(t) } \right]e^{i Q \sum_{t^\prime = 1}^t \Delta(t^\prime)} \\
&= \sum_{q\in\mathbb{Z}} \int \mathcal{D}|\varphi|\; \mathcal{P}(q,|\varphi|)
\; |\varphi(t)|^{|Q|+2P}|\varphi(0)|^{|Q|+2P}
\prod_{t^\prime=1}^{t}\left[ \frac{I_{\abs{Q+q}}\left( 2|\varphi(t^\prime)||\varphi(t^\prime-1)| \right)}{I_{\abs{q}}\left( 2|\varphi(t^\prime)||\varphi(t^\prime-1)| \right)} \right].
\end{split}\label{eq:CHOpropdecomp}$$ The integrand is again positive-definite and can be interpreted as an integration measure without a sign problem, in contrast to Eq. . It is also possible to calculate correlation functions by MC sampling field configurations according to $\mathcal{P}(q,|\varphi|)$ and then including the ensemble average of the product of Bessel functions in Eq. as a reweighting factor.
Care must be taken in defining MC updates of $q$. For instance, a Metropolis scheme in which updates $q\rightarrow q^\prime$ are proposed and then accepted with probability $\text{min}\left( 1,\; e^{-S_{\eff} (q,|\varphi|) + S_{\eff}(q^\prime,|\varphi|)} \right)$ with $$\begin{split}
S_{\eff} (q,|\varphi|) = \sum_{t=0}^{L-1} 2|\varphi(t)|^2 + V(|\varphi(t)|) - \ln|\varphi(t)| - \ln\left[ I_{|q|}(\kappa(t))\right],
\end{split}\label{eq:Sqdef}$$ will experience “topological freezing”. The minimum action $q=0$ sector is sampled effectively but $q \neq 0$ sectors make $O(e^{-L})$ suppressed contributions to the partition function, because they involve products of $L$ small factors $\prod_{t=0}^{L-1} \frac{I_{|q|}}{I_0}$, so are scarcely or never present in a finite-$N$ MC ensemble. This is problematic for MC calculations of correlation functions, because $q \neq 0$ contributions can provide significant contributions to correlation functions with nonzero $U(1)$ charge. Considering Eq. for the case of the scalar field propagator $G = G_{1,0}$, the $q=-1$ sector makes a contribution at large $t \sim L$ involving the exponentially large product $\prod_{t^\prime = 1}^L \frac{I_0}{I_1} \sim e^{L}$. This situation of exponentially rare MC configurations making exponentially large contributions to observables suggests this MC scheme has an overlap problem where the distribution being importance sampled has poor overlap with the region of configuration space making dominant contributions to observables of interest.
Figure \[fig:cho-standard-autocorrs\] plots the integrated autocorrelation time for $\ob = \braket{|\varphi|^2}$ for all standard ensembles used in this work. Figure \[fig:cho-pi-autocorrs\] analogously plots the integrated autocorrelation time for the dual variable ensembles. Autocorrelation times between corresponding standard and dual ensembles are similar. There is significant autocorrelation only on the finest lattice ($M^2 = 0.00625$), and as such all analyses of methods we introduce applied binning with bin size $\geq 10$ on the finest lattice to more accurately estimate errors in the presence of this autocorrelation. Larger bin sizes were used in some cases to further improve the $\chi^2$/DoF estimates. Tables \[tbl:free-fit-data\]-\[tbl:l16-neg-fit-data\] indicate the bin and window sizes used for all fits to energy levels required to produce the various spectrum plots in this work.
Instead, MC sampling over $q$ can be replaced by explicit summation over a finite subset of winding numbers that make dominant contributions to particular observables. MC sampling can be performed using the modified probability distribution $$\begin{split}
\mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) = \frac{2}{Z_0} \prod_{t=0}^{L-1} \left[ |\varphi(t)| d|\varphi(t)| e^{-2|\varphi(t)|^2 - V(|\varphi(t)|)} \; 2\; I_{0}\left( \kappa(t) \right) \right],
\end{split}\label{eqA:Pprimedef}$$ where $Z_0$ represents the $q=0$ contribution to the partition function and is defined to ensure that $\int \mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) = 1$. Importance sampling with respect to $\mathcal{P}_0(|\varphi|)$ can be performed with local Metropolis update steps of $|\varphi(t)|$ and an accept-reject probability determined by changes in the action $S_{\eff}(q=0,|\varphi|)$. Correlation functions $G_{Q,2P}$ can be calculated from field configurations importance sampled according to $\mathcal{P}_0(|\varphi|)$ by explicit summation over all $q$, $$\begin{aligned}
G_{Q,2P}(t) &= \sum_{q \in \mathbb{Z}} \int \mathcal{D}|\varphi| \mathcal{P}_0(|\varphi|) \; |\varphi(t)|^{|Q|+2P} |\varphi(0)|^{|Q| + 2P} \\
&\hspace{20pt} \times \prod_{t^\prime = 1}^t \frac{I_{|Q + q|}(\kappa(t))}{I_0(\kappa(t))} \prod_{t^\prime = t+1}^L \frac{I_{|q|}(\kappa(t))}{I_0(\kappa(t))}.
\end{aligned}$$ Given a finite MC ensemble of scalar field magnitude $|\varphi_i|$, $i=1,\dots,N$ sampled from Eq. , correlation functions can be estimated from the corresponding ensemble averages $$\begin{aligned}
\overline{G}_{Q,2P}^{dual}(t) &= \frac{1}{N} \sum_{i=1}^{N}
\sum_{q \in \mathbb{Z}} \Bigg\{
|\varphi_i(t)|^{|Q|+2P} |\varphi_i(0)|^{|Q| + 2P} \\
&\hspace{20pt} \times \prod_{t^\prime = 1}^t \frac{I_{|Q + q|}(\kappa_i(t))}{I_0(\kappa_i(t))} \prod_{t^\prime = t+1}^L \frac{I_{|q|}(\kappa_i(t))}{I_0(\kappa_i(t))} \Bigg\},
\end{aligned}\label{eqA:GQ2Pdual}$$ where $\overline{G}_{Q,2P}^{dual}$ denotes ensemble estimates of $G_{Q,2P}$ in this dual variables approach. Significant contributions to Eq. arise for $q = -Q,\dots,+Q$ but topological charge sectors with $|q| > |Q|$ make subdominant contributions that rapidly converge to zero and allow the sum over topological charge sector to be truncated in practical calculations.
Unwrapped Phase Definition {#app:unwrapping}
==========================
The complex exponential function is not injective because $e^{z} = e^{z + 2\pi i}$. The complex logarithm function, intuitively describing the inverse of the complex exponential function, therefore requires care to define. The principal value of the complex logarithm of an analytic function $f$ of a complex variable $z$ defined such that $-\pi < \text{Im}\ln f(z) \leq \pi$ is given by $$\begin{split}
\ln f(z) &\equiv \ln(|f(z)|e^{i\theta(z)}) \equiv \ln |f(z)| + i \text{arg}(e^{i\theta(z)}) \equiv \ln |f(z)| + i \theta(t),
\end{split}\label{eq:logdef}$$ where $-\pi < \theta \leq \pi$. The principal-valued logarithm is not a continuous function on the punctured complex plane $\mathbb{C} \setminus \{0\}$ because $$\begin{split}
\lim_{\theta\rightarrow \pi^-}\text{arg}(e^{i\theta})= \pi \neq -\pi = \lim_{\theta\rightarrow \pi^+}\text{arg}(e^{i\theta}) .
\end{split}\label{eq:logdisc}$$ The standard method of defining a continuous logarithm function involves analytic continuation and construction of a Riemann surface including infinitely many copies of the punctured complex plane glued together at branch cuts $\theta = \pi + 2\pi \nu$. Instead, one can assume the existence of a single-valued, analytic function $\text{Ln}: \mathbb{C}\setminus \{0\} \rightarrow \mathbb{C}$ satisfying $$\begin{split}
\frac{d}{dz}\text{Ln}(f(z)) \equiv \frac{1}{f(z)}\frac{df}{dz}, \hspace{20pt} \text{Ln}(1) \equiv 0.
\end{split}\label{eq:Lnderivdef}$$ Applying the fundamental theorem of calculus for integration along a curve $\gamma : [a,b] \rightarrow \mathbb{C}\setminus \{0\}$ gives $$\begin{split}
\text{Ln}(f(b)) - \text{Ln}(f(a)) = \int_{\gamma(z)} \frac{d}{dz}\text{Ln}(f(z)) dz = \int_{\gamma(z)} \frac{\frac{df}{dz}}{f(z)}.
\end{split}\label{eq:Lndef}$$ Note that this construction implicitly depends on the function $f$ under consideration. To gain intuition for Eq. first consider the case $f(z) = z$. By analyticity of $\text{Ln}(z)$ in the domain $\mathbb{C}\setminus \{0\}$, any line integral along a curve $\gamma(z)$ that is assumed to be in the trivial homotopy class of $\mathbb{C}\setminus \{0\}$ can be deformed into an integral along a piecewise continuous path $\gamma = \gamma_R \cup \gamma_\theta$ composed of a purely radial path $\gamma_R : |a|e^{i\text{arg}(a)} \rightarrow |b|e^{i\text{arg}(a)}$ and a purely angular path $\gamma_\theta : |b|e^{i\text{arg}(a)} \rightarrow |b|e^{i\text{arg}(b)}$. Since a radial path $\gamma_R$ does not cross any branch cuts of $\ln(z)$, the integral along $\gamma_R$ can be evaluated as $$\begin{split}
\int_{\gamma_R(z)} \frac{dz}{z} = \int_{|a|e^{i\text{arg}(a)}}^{|b|e^{i\text{arg}(a)}} \frac{dz}{z} = \ln|a| - \ln|b|.
\end{split}\label{eq:Lnreal}$$ If the angular segment of the path crosses the branch cut of $\ln(z)$ placed at $z = \pi$, then a further path deformation should be made so that the angular path stops a distance $\varepsilon$ before the branch cut, runs radially along the branch cut from radius $|b|$ to the origin, encircles the origin in a circle of radius $\varepsilon$, runs radially back to radius $|b|$ along the opposite side of the branch cut, and then continues along the remainder of the angular path. In the $\varepsilon\rightarrow 0$ limit the contributions from the radial paths vanish and contribution from the path encircling the origin can be evaluated by Cauchy’s theorem, $$\begin{split}
\int_{\gamma_\theta(z)} \frac{dz}{z} = \fint_{|b|e^{i\text{arg}(a)}}^{|b|e^{i\text{arg}(b)}} \frac{dz}{z} + \nu \oint \frac{dz}{z} = 2\pi i \nu + i \fint_{\text{arg}(a)}^{\text{arg}(b)} d\theta = 2\pi i \nu + i (\text{arg}(b) - \text{arg}(a)),
\end{split}\label{eq:Lnim}$$ where $\nu = +1$ if $\gamma$ includes a positive branch cut crossing (since $\gamma$ is simply connected this occurs if and only if $a$ is in quadrant II and $b$ is in quadrant III), $\nu= -1$ if $\gamma$ includes a negative branch cut crossing (if $b$ is in quadrant II and $a$ is in quadrant III), and zero otherwise. Choosing instead $a = 1$ such that $\text{Ln}(a) = 0$, $b = z = |z|e^{i\theta}$ arbitrary, and $\gamma$ a counter-clockwise path that does not encircle the origin gives $$\begin{split}
\text{Ln}(z) &= \int_{\gamma_R(z)} \frac{dz}{z} + \int_{\gamma_\theta(z)} \frac{dz}{z} = \int_1^{|z|} \frac{dz}{z} + \fint_{|z|}^{|z|e^{i\text{arg}(z)}} \frac{dz}{z} + \nu \oint \frac{dz}{z} \\
&= \ln |z| + \text{arg}(z) + 2\pi i \nu,
\end{split}\label{eq:Lnln}$$ where $\nu = +1$ if $\text{Im}z < 0$ and $\nu=0$ otherwise. The imaginary part of this expression defines a function that agrees with $\text{arg}(z)$ modulo $2\pi$, $$\begin{split}
\text{Arg}(z) \equiv \text{Im}\text{Ln}(z) = \text{arg}(z) + 2\pi i \nu.
\end{split}\label{eqA:Argzdef}$$ For an arbitrary analytic function $f(z)$, $\text{Arg}(f(z))$ can be computed by integrating $\text{Im}\frac{d}{dz}\text{Ln}(f(z))$ along the angular piece of $\gamma: [0, z] \rightarrow \mathbb{C} \setminus \{0\}$, $$\begin{split}
\text{Arg}(f(|z|e^{i\theta})) = \int_0^\theta \text{Im}\left[ \frac{\frac{df}{d\theta^\prime} }{f(|z|e^{i\theta^\prime})} \right]d\theta^\prime.
\end{split}\label{eqA:Argdef}$$ Assuming $f(|z|e^{i\theta^\prime}) \neq 0$, the integral in Eq. is well-defined and $\text{Arg}(f(|z|e^{i\theta}))$ is continuous and analytic provided that $f$ is continuous and analytic [@Kitahara:2015].
Monte Carlo Ensembles {#app:MC}
=====================
![In the left plots, we compare free-field $Z_{0;1,0} = \braket{|\varphi|^2}$ estimated using both the standard and phase-integrated ensembles versus the analytical value given in Eq. . The three main ensembles $A_0$, $B_0$, and $C_0$, agree with the analytical prediction to the percent level. The auxiliary $D_0$ and $E_0$ ensembles agree to the few percent level. In the right plot, GEVP methods are used to determine the lowest six energy levels in the spectrum of the free theory, rescaled into physical units. The coarsest ensemble $A_0$ exhibits large statistical and systematic uncertainties in fitting, and for $Q=3,4$ no plateau could be reliably fit (indicated by vertical gray lines). Where reliable estimates are possible, the data agree with analytical predictions.[]{data-label="fig:cho-phisq"}](ensemble_phisq_grid.pdf){width="\linewidth"}
![In the left plots, we compare free-field $Z_{0;1,0} = \braket{|\varphi|^2}$ estimated using both the standard and phase-integrated ensembles versus the analytical value given in Eq. . The three main ensembles $A_0$, $B_0$, and $C_0$, agree with the analytical prediction to the percent level. The auxiliary $D_0$ and $E_0$ ensembles agree to the few percent level. In the right plot, GEVP methods are used to determine the lowest six energy levels in the spectrum of the free theory, rescaled into physical units. The coarsest ensemble $A_0$ exhibits large statistical and systematic uncertainties in fitting, and for $Q=3,4$ no plateau could be reliably fit (indicated by vertical gray lines). Where reliable estimates are possible, the data agree with analytical predictions.[]{data-label="fig:cho-phisq"}](free_spectrum.png){width="\linewidth"}
Ensembles are generated via Metropolis sweeps over the sites in a red-black alternating pattern for efficient execution, with $N_{\text{skip}}/2$ odd and $N_{\text{skip}}/2$ even updates between each measurement. Both the standard $1D$ complex scalar field action defined in Eq. with the potential Eq. and the analytically phase-integrated dual form of the theory given in Eq. - are used to perform MC calculations using identical values of the parameters $M^2$, $L$, and $\lambda$ given in Table. \[tbl:cho-ensembles\]. The phase unwrapping techniques based on smoothed numerical integration of the wrapped phase described above are applied to all correlation functions generated using the standard complex scalar action. The cumulant expansion is then used to estimate correlation functions from sample moments of the corresponding unwrapped phases and log-magnitudes, and a generalized eigenvalue problem (GEVP) is solved to numerically extract the low-lying spectrum of the theory from the resulting correlation function estimates [@Luscher:1990ck].
We perform some checks for ensemble consistency. Eq. gives the noninteracting expectation for $Z_{1;0,1} = \braket{|\varphi|^2}$. We can reliably estimate this overlap on our noninteracting lattices and compare against the theoretically predicted value. The left plot of Figure \[fig:cho-phisq\] compares the standard ensemble and dual variable ensemble estimates versus the theoretical prediction, finding agreement to the percent level for the three main ensembles $A_0$, $B_0$, and $C_0$, while the auxiliary ensembles (used only for investigation of lattice spacing effects) match the prediction at the few percent level. Eqs. and describe the noninteracting spectrum in terms of $Q=0$ ground state energy $E = E_0 = 2 \text{arcsinh} \paren{M/2}$. There are six low-lying states (energy $E_i \leq 4E_0$), with two states in each of the $Q=0$ and $Q=1$ channels, and one state in each of the $Q=2$ and $Q=3$ channels. Figure \[fig:cho-phisq\] further demonstrates that our free-field ensembles correctly reproduce this low-lying spectrum to within statistical and systematic fitting errors.
![Integrated autocorrelation times $\rho(\tau)/\rho(0)$ versus MCMC stream separation $\tau$ for the $\braket{|\phi|^2}$ observables on all dual variable ensembles. Similarly to the standard ensembles, there is significant autocorrelation up to roughly $\author \approx 10$ for the finest lattice ($M^2 = 0.00625$), and significantly less autocorrelation on the two coarser lattices.[]{data-label="fig:cho-pi-autocorrs"}](m0p1_autocorrs.pdf "fig:"){height="5cm"} ![Integrated autocorrelation times $\rho(\tau)/\rho(0)$ versus MCMC stream separation $\tau$ for the $\braket{|\phi|^2}$ observables on all dual variable ensembles. Similarly to the standard ensembles, there is significant autocorrelation up to roughly $\author \approx 10$ for the finest lattice ($M^2 = 0.00625$), and significantly less autocorrelation on the two coarser lattices.[]{data-label="fig:cho-pi-autocorrs"}](m0p025_autocorrs.pdf "fig:"){height="5cm"} ![Integrated autocorrelation times $\rho(\tau)/\rho(0)$ versus MCMC stream separation $\tau$ for the $\braket{|\phi|^2}$ observables on all dual variable ensembles. Similarly to the standard ensembles, there is significant autocorrelation up to roughly $\author \approx 10$ for the finest lattice ($M^2 = 0.00625$), and significantly less autocorrelation on the two coarser lattices.[]{data-label="fig:cho-pi-autocorrs"}](m0p00625_autocorrs.pdf "fig:"){height="5cm"}
![Integrated autocorrelation times $\rho(\tau)/\rho(0)$ versus MCMC stream separation $\tau$ for the $\braket{|\phi|^2}$ observables on all dual variable ensembles. Similarly to the standard ensembles, there is significant autocorrelation up to roughly $\author \approx 10$ for the finest lattice ($M^2 = 0.00625$), and significantly less autocorrelation on the two coarser lattices.[]{data-label="fig:cho-pi-autocorrs"}](m0p1_pi_autocorrs.pdf "fig:"){height="5cm"} ![Integrated autocorrelation times $\rho(\tau)/\rho(0)$ versus MCMC stream separation $\tau$ for the $\braket{|\phi|^2}$ observables on all dual variable ensembles. Similarly to the standard ensembles, there is significant autocorrelation up to roughly $\author \approx 10$ for the finest lattice ($M^2 = 0.00625$), and significantly less autocorrelation on the two coarser lattices.[]{data-label="fig:cho-pi-autocorrs"}](m0p025_pi_autocorrs.pdf "fig:"){height="5cm"} ![Integrated autocorrelation times $\rho(\tau)/\rho(0)$ versus MCMC stream separation $\tau$ for the $\braket{|\phi|^2}$ observables on all dual variable ensembles. Similarly to the standard ensembles, there is significant autocorrelation up to roughly $\author \approx 10$ for the finest lattice ($M^2 = 0.00625$), and significantly less autocorrelation on the two coarser lattices.[]{data-label="fig:cho-pi-autocorrs"}](m0p00625_pi_autocorrs.pdf "fig:"){height="5cm"}
Figure \[fig:cho-standard-autocorrs\] plots the integrated autocorrelation time for $\ob = \braket{|\varphi|^2}$ for all standard ensembles used in this work. Figure \[fig:cho-pi-autocorrs\] analogously plots the integrated autocorrelation time for the dual variable ensembles. Autocorrelation times between corresponding standard and dual ensembles are similar. There is significant autocorrelation only on the finest lattice ($M^2 = 0.00625$), and as such all analyses applied binning with bin size $\geq 10$ on the finest lattice to more accurately estimate errors in the presence of this autocorrelation. Larger bin sizes were used in some cases to further improve the $\chi^2$/DoF estimates. Tables \[tbl:free-fit-data\]-\[tbl:l16-neg-fit-data\] indicate the bin and window sizes used for all fits to energy levels required to produce the various spectrum plots in this work.
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
---------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_0$ $Q=1$ 0.317 0.005 0.001 0.466 \[1,8\] 1
$Q=1^*$ 0.972 0.159 0.329 1.289 \[1,2\]
$Q=2$ 0.603 0.026 0.066 1.605 \[1,3\]
$Q=2^*$ 0.854 0.494 0.812 0.696 \[1,2\]
$Q=3$ – – – – –
$Q=4$ – – – – –
$B_0$ $Q=1$ 0.157 0.001 0.000 0.333 \[1,16\] 1
$Q=1^*$ 0.444 0.013 0.024 0.396 \[1,8\]
$Q=2$ 0.316 0.004 0.003 0.459 \[1,8\]
$Q=2^*$ 0.662 0.037 0.060 0.301 \[1,5\]
$Q=3$ 0.476 0.011 0.022 1.571 \[1,8\]
$Q=4$ 0.664 0.036 0.012 1.003 \[1,3\]
$C_0$ $Q=1$ 0.080 0.000 0.000 2.117 \[1,15\] 1
$Q=1^*$ 0.237 0.003 0.003 1.989 \[1,6\]
$Q=2$ 0.160 0.001 0.000 0.491 \[1,15\]
$Q=2^*$ 0.315 0.011 0.006 1.376 \[1,6\]
$Q=3$ 0.241 0.002 0.002 1.783 \[1,4\]
$Q=4$ 0.326 0.007 0.006 2.093 \[1,2\]
: Free spectrum fitting information. Charge eigenvalues decorated with an asterisk indicate first excitations. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, while those without indicate the standard method.[]{data-label="tbl:free-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
---------------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_0^\bigstar$ $Q=1$ 0.316 0.001 0.000 0.515 \[1,6\] 5
$Q=1^*$ 1.051 0.124 0.048 0.946 \[1,2\]
$Q=2$ 0.636 0.005 0.04 0.795 \[1,6\]
$Q=2^*$ 1.505 0.360 0.900 1.056 \[1,2\]
$Q=3$ 0.974 0.016 0.020 0.716 \[1,4\]
$Q=4$ 1.375 0.037 0.094 0.941 \[1,4\]
$B_0^\bigstar$ $Q=1$ 0.159 0.001 0.000 0.736 \[1,10\] 10
$Q=1^*$ 0.495 0.016 0.010 0.343 \[1,4\]
$Q=2$ 0.316 0.002 0.001 0.480 \[1,10\]
$Q=2^*$ 0.711 0.043 0.052 0.213 \[1,3\]
$Q=3$ 0.481 0.005 0.003 0.948 \[1,5\]
$Q=4$ 0.656 0.010 0.008 0.711 \[1,5\]
$C_0^\bigstar$ $Q=1$ 0.080 0.000 0.000 1.493 \[1,15\] 20
$Q=1^*$ 0.246 0.006 0.004 0.503 \[1,6\]
$Q=2$ 0.162 0.001 0.000 0.963 \[1,15\]
$Q=2^*$ 0.330 0.013 0.005 0.290 \[1,6\]
$Q=3$ 0.242 0.003 0.001 0.406 \[1,4\]
$Q=4$ 0.324 0.006 0.003 0.212 \[1,4\]
: Free spectrum fitting information. Charge eigenvalues decorated with an asterisk indicate first excitations. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, while those without indicate the standard method.[]{data-label="tbl:free-fit-data"}
Systematic errors presented in the fit tables indicate the variation in central value as the fit window is offset by up to two lattice points. Systematic errors for the phase unwrapping technique additionally include variation in the central value as higher-order cumulants are included. In this work, these errors include variation up to cumulant order 6 (the second subleading order in phase variations), as higher-order cumulants include too much noise to be reliably estimated.
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
---------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_2^+$ $Q=1$ 0.430 0.009 0.016 0.480 \[1,8\] 1
$Q=2$ 0.848 0.102 0.412 0.469 \[1,3\]
$Q=3$ – – – – –
$Q=4$ – – – – –
$B_2^+$ $Q=1$ 0.222 0.002 0.001 0.689 \[1,8\] 1
$Q=2$ 0.457 0.007 0.017 1.420 \[1,5\]
$Q=3$ 0.719 0.031 0.045 0.550 \[1,3\]
$Q=4$ 0.912 0.127 0.467 0.699 \[1,3\]
$C_2^+$ $Q=1$ 0.113 0.000 0.000 1.354 \[1,16\] 1
$Q=2$ 0.238 0.002 0.001 0.878 \[1,10\]
$Q=3$ 0.373 0.005 0.006 0.536 \[1,5\]
$Q=4$ 0.504 0.013 0.015 0.664 \[1,5\]
: Interacting spectrum fitting information for $\lambda L / M^2 = +32$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l32-pos-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
-------------------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_2^{+\bigstar} $ $Q=1$ 0.442 0.002 0.001 0.523 \[1,8\] 5
$Q=2$ 0.916 0.008 0.003 1.464 \[1,4\]
$Q=3$ 1.373 0.027 0.028 1.081 \[1,2\]
$Q=4$ 1.803 0.070 0.092 0.716 \[1,2\]
$B_2^{+\bigstar} $ $Q=1$ 0.225 0.001 0.000 1.804 \[1,15\] 10
$Q=2$ 0.470 0.002 0.001 1.893 \[1,10\]
$Q=3$ 0.735 0.006 0.001 1.336 \[1,8\]
$Q=4$ 1.033 0.013 0.007 2.003 \[1,8\]
$C_2^{+\bigstar} $ $Q=1$ 0.113 0.000 0.000 1.306 \[1,24\] 20
$Q=2$ 0.237 0.001 0.000 1.148 \[1,16\]
$Q=3$ 0.370 0.002 0.001 1.613 \[1,10\]
$Q=4$ 0.511 0.004 0.002 1.091 \[1,6\]
: Interacting spectrum fitting information for $\lambda L / M^2 = +32$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l32-pos-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
------------------------- ------------ --------------- --------------- --------------- -------------- ----------- ---------- --
$A_2^{+\blacklozenge} $ $Q=1$ 0.493 0.003 0.016 1.226 \[8,32\] 10
$Q=2$ 1.172 0.007 0.531 0.737 \[8,24\]
$Q=3$ 1.438 0.011 0.514 1.447 \[8,24\]
$Q=4$ 1.536 0.011 1.514 1.574 \[8,24\]
$B_2^{+\blacklozenge} $ $Q=1$ 0.249 0.001 0.027 1.922 \[16,64\] 10
$Q=2$ 0.571 0.003 0.032 1.794 \[16,48\]
$Q=3$ 0.671 0.004 0.753 1.482 \[16,48\]
$Q=4$ 0.721 0.004 0.401 0.979 \[16,48\]
$C_2^{+\blacklozenge} $ $Q=1$ 0.108 0.001 0.011 1.185 \[32,64\] 10
$Q=2$ 0.274 0.001 0.046 1.179 \[32,64\]
$Q=3$ 0.414 0.002 0.282 1.245 \[32,64\]
$Q=4$ 0.434 0.002 0.103 1.362 \[32,64\]
: Interacting spectrum fitting information for $\lambda L / M^2 = +32$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l32-pos-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
---------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_2^-$ $Q=1$ 0.248 0.003 0.002 0.702 \[1,8\] 1
$Q=2$ 0.571 0.016 0.009 0.433 \[1,4\]
$Q=3$ 0.696 0.089 0.429 2.164 \[1,4\]
$Q=4$ 0.801 0.443 1.007 0.305 \[1,2\]
$B_2^-$ $Q=1$ 0.124 0.001 0.001 1.458 \[1,8\] 1
$Q=2$ 0.285 0.003 0.005 0.937 \[1,8\]
$Q=3$ 0.469 0.009 0.006 1.066 \[1,8\]
$Q=4$ 0.682 0.028 0.050 1.340 \[1,4\]
$C_2^-$ $Q=1$ 0.062 0.000 0.000 1.194 \[1,16\] 1
$Q=2$ 0.142 0.001 0.001 2.057 \[1,16\]
$Q=3$ 0.235 0.002 0.001 1.294 \[1,16\]
$Q=4$ 0.339 0.005 0.003 0.796 \[1,8\]
: Interacting spectrum fitting information for $\lambda L / M^2 = -32$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l32-neg-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
-------------------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_2^{-\bigstar} $ $Q=1$ 0.242 0.001 0.000 1.412 \[1,8\] 5
$Q=2$ 0.558 0.003 0.003 1.981 \[1,8\]
$Q=3$ 0.930 0.006 0.008 1.395 \[1,8\]
$Q=4$ 1.378 0.014 0.032 1.007 \[1,8\]
$B_2^{-\bigstar} $ $Q=1$ 0.123 0.000 0.000 0.527 \[1,10\] 10
$Q=2$ 0.283 0.001 0.001 0.886 \[1,8\]
$Q=3$ 0.473 0.002 0.000 0.469 \[1,6\]
$Q=4$ 0.683 0.005 0.004 1.737 \[1,4\]
$C_2^{-\bigstar} $ $Q=1$ 0.062 0.000 0.000 1.417 \[1,24\] 20
$Q=2$ 0.143 0.000 0.001 2.386 \[1,12\]
$Q=3$ 0.238 0.001 0.001 2.638 \[1,8\]
$Q=4$ 0.341 0.002 0.002 1.436 \[1,6\]
: Interacting spectrum fitting information for $\lambda L / M^2 = -32$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l32-neg-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
------------------------- ------------ --------------- --------------- --------------- -------------- ----------- ---------- --
$A_2^{-\blacklozenge} $ $Q=1$ 0.250 0.002 0.010 1.096 \[8,24\] 10
$Q=2$ 0.588 0.003 0.047 1.832 \[8,24\]
$Q=3$ 1.211 0.009 0.033 1.352 \[8,24\]
$Q=4$ 1.397 0.009 2.369 0.818 \[8,24\]
$B_2^{-\blacklozenge} $ $Q=1$ 0.127 0.001 0.004 1.947 \[16,48\] 10
$Q=2$ 0.289 0.002 0.036 1.930 \[16,48\]
$Q=3$ 0.579 0.003 0.058 1.303 \[16,48\]
$Q=4$ 0.665 0.004 0.122 1.171 \[16,48\]
$C_2^{-\blacklozenge} $ $Q=1$ 0.061 0.000 0.003 2.091 \[32,48\] 10
$Q=2$ 0.156 0.001 0.010 0.713 \[32,48\]
$Q=3$ 0.239 0.002 0.079 0.806 \[32,48\]
$Q=4$ 0.407 0.003 0.274 1.097 \[32,48\]
: Interacting spectrum fitting information for $\lambda L / M^2 = -32$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l32-neg-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
---------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_1^+$ $Q=1$ 0.394 0.006 0.011 0.511 \[1,8\] 1
$Q=2$ 0.779 0.056 0.247 0.942 \[1,4\]
$Q=3$ 0.885 0.451 1.289 0.497 \[1,2\]
$Q=4$ – – – – –
$B_1^+$ $Q=1$ 0.200 0.001 0.001 0.449 \[1,8\] 1
$Q=2$ 0.418 0.005 0.006 0.692 \[1,8\]
$Q=3$ 0.634 0.021 0.033 0.612 \[1,4\]
$Q=4$ 0.731 0.065 0.367 1.944 \[1,4\]
$C_1^+$ $Q=1$ 0.099 0.000 0.000 1.083 \[1,16\] 1
$Q=2$ 0.207 0.001 0.001 0.498 \[1,16\]
$Q=3$ 0.324 0.003 0.006 1.807 \[1,16\]
$Q=4$ 0.451 0.010 0.013 1.223 \[1,8\]
: Interacting spectrum fitting information for $\lambda L / M^2 = +16$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l16-pos-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
-------------------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_1^{+\bigstar} $ $Q=1$ 0.398 0.002 0.001 0.883 \[1,8\] 5
$Q=2$ 0.824 0.006 0.004 2.060 \[1,8\]
$Q=3$ 1.272 0.020 0.015 0.377 \[1,4\]
$Q=4$ 1.768 0.055 0.068 1.465 \[1,4\]
$B_1^{+\bigstar} $ $Q=1$ 0.198 0.001 0.000 1.177 \[1,10\] 10
$Q=2$ 0.410 0.002 0.000 1.679 \[1,8\]
$Q=3$ 0.632 0.005 0.002 0.326 \[1,6\]
$Q=4$ 0.876 0.011 0.005 0.540 \[1,4\]
$C_1^{+\bigstar} $ $Q=1$ 0.100 0.000 0.000 1.069 \[1,24\] 20
$Q=2$ 0.209 0.001 0.001 1.906 \[1,12\]
$Q=3$ 0.322 0.002 0.001 0.717 \[1,8\]
$Q=4$ 0.443 0.003 0.003 0.746 \[1,6\]
: Interacting spectrum fitting information for $\lambda L / M^2 = +16$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l16-pos-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
------------------------- ------------ --------------- --------------- --------------- -------------- ----------- ---------- --
$A_1^{+\blacklozenge} $ $Q=1$ 0.460 0.003 0.077 1.890 \[8,32\] 10
$Q=2$ 1.111 0.007 1.083 1.825 \[8,32\]
$Q=3$ 1.408 0.009 1.448 1.125 \[8,32\]
$Q=4$ 1.549 0.009 0.483 0.753 \[8,32\]
$B_1^{+\blacklozenge} $ $Q=1$ 0.206 0.001 0.010 2.013 \[16,48\] 10
$Q=2$ 0.403 0.003 0.034 1.306 \[16,48\]
$Q=3$ 0.662 0.004 0.071 1.666 \[16,40\]
$Q=4$ 1.345 0.007 4.357 1.047 \[16,40\]
$C_1^{+\blacklozenge} $ $Q=1$ 0.103 0.001 0.008 1.602 \[32,64\] 10
$Q=2$ 0.227 0.001 0.029 1.055 \[32,64\]
$Q=3$ 0.403 0.003 0.326 2.010 \[32,64\]
$Q=4$ 0.434 0.003 0.801 1.274 \[32,64\]
: Interacting spectrum fitting information for $\lambda L / M^2 = +16$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l16-pos-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
---------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_1^-$ $Q=1$ 0.141 0.001 0.002 1.919 \[1,8\] 1
$Q=2$ 0.345 0.006 0.011 1.136 \[1,4\]
$Q=3$ 0.586 0.023 0.064 1.423 \[1,4\]
$Q=4$ 0.739 0.090 0.388 1.607 \[1,4\]
$B_1^-$ $Q=1$ 0.072 0.000 0.001 1.664 \[1,8\] 1
$Q=2$ 0.176 0.001 0.004 0.618 \[1,8\]
$Q=3$ 0.301 0.004 0.006 0.468 \[1,8\]
$Q=4$ 0.430 0.009 0.018 1.301 \[1,4\]
$C_1^-$ $Q=1$ 0.037 0.000 0.000 2.926 \[1,10\] 1
$Q=2$ 0.091 0.000 0.001 1.855 \[1,10\]
$Q=3$ 0.157 0.001 0.002 1.284 \[1,8\]
$Q=4$ 0.231 0.002 0.003 0.248 \[1,6\]
: Interacting spectrum fitting information for $\lambda L / M^2 = -16$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l16-neg-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
-------------------- ------------ --------------- --------------- --------------- -------------- ---------- ---------- --
$A_1^{-\bigstar} $ $Q=1$ 0.144 0.000 0.000 2.531 \[1,8\] 5
$Q=2$ 0.354 0.001 0.002 1.506 \[1,8\]
$Q=3$ 0.612 0.003 0.005 1.366 \[1,8\]
$Q=4$ 0.907 0.007 0.014 1.733 \[1,8\]
$B_1^{-\bigstar} $ $Q=1$ 0.073 0.000 0.000 1.561 \[1,10\] 10
$Q=2$ 0.179 0.001 0.001 1.636 \[1,8\]
$Q=3$ 0.307 0.001 0.002 1.610 \[1,6\]
$Q=4$ 0.453 0.003 0.004 0.848 \[1,4\]
$C_1^{-\bigstar} $ $Q=1$ 0.037 0.000 0.000 1.676 \[4,24\] 20
$Q=2$ 0.091 0.000 0.000 2.067 \[4,12\]
$Q=3$ 0.156 0.001 0.001 1.211 \[2,8\]
$Q=4$ 0.231 0.001 0.001 0.810 \[2,6\]
: Interacting spectrum fitting information for $\lambda L / M^2 = -16$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l16-neg-fit-data"}
Ensemble Eigenvalue Central value Error (stat.) Error (syst.) $\chi^2$/DoF Window Bin Size
------------------------- ------------ --------------- --------------- --------------- -------------- ----------- ---------- --
$A_1^{-\blacklozenge} $ $Q=1$ 0.144 0.001 0.006 1.614 \[12,24\] 10
$Q=2$ 0.363 0.003 0.027 1.967 \[12,24\]
$Q=3$ 0.619 0.006 0.111 0.874 \[12,24\]
$Q=4$ 1.195 0.008 0.355 1.857 \[12,24\]
$B_1^{-\blacklozenge} $ $Q=1$ 0.075 0.001 0.001 1.412 \[24,48\] 10
$Q=2$ 0.185 0.002 0.016 0.795 \[24,48\]
$Q=3$ 0.303 0.003 0.111 1.254 \[24,48\]
$Q=4$ 0.441 0.003 0.061 1.585 \[24,48\]
$C_1^{-\blacklozenge} $ $Q=1$ 0.036 0.000 0.005 2.258 \[40,70\] 10
$Q=2$ 0.098 0.001 0.021 1.632 \[40,70\]
$Q=3$ 0.168 0.001 0.059 2.488 \[40,70\]
$Q=4$ 0.243 0.002 0.039 1.993 \[40,70\]
: Interacting spectrum fitting information for $\lambda L / M^2 = -16$. Ensembles reported with a $\bigstar$ symbol indicate the dual-variable method, ensembles reported with a $\blacklozenge$ symbol indicate the phase unwrapping method, and those without any symbol indicate the standard method.[]{data-label="tbl:l16-neg-fit-data"}
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
[^4]: Other interesting LQCD observables face distinct StN problems. For instance, isoscalar meson correlation functions are uncharged under $U(1)$ symmetries and possess exponential StN problems but not $U(1)$ phase fluctuations. Excited-state energies are extracted from differences of correlation functions with the same quantum numbers and face StN problems arising from the exponentially precise cancellations needed to project out ground-state contributions and leave exponentially faster decaying excited-state contributions. In large nuclei, StN problems associated with MeV excitation energies are negligible compared to the phase fluctuation StN problem associated with the multi-GeV rest mass of the nucleus. LQCD calculations of isoscalar mesons and exotic hadrons conversely face StN problems primarily from sources besides $U(1)$ phase fluctuations, and phase unwrapping is not immediately applicable to these systems.
[^5]: Cumulant expansions of noncompact “extensive phases” have also been applied to sign problems in QCD and other theories at nonzero chemical potential [@Ejiri:2007ga; @Nakagawa:2011eu; @Ejiri:2012wp; @Greensite:2013gya; @Garron:2017fta; @Bloch:2018yhu].
[^6]: See Ref. [@Lepage:1998dt] for a pedagogical review of MC methods for the simple harmonic oscillator that can be readily applied to $(0+1)D$ complex scalar field theory.
[^7]: Note that the magnitude of the integrand of Eq. is positive-definite but not properly normalized as a probability distribution. A suitably normalized positive-definite probability distribution can be found by replacing $\frac{1}{Z}$ by $\frac{1}{\widetilde{Z}}$ where $\widetilde{Z} = \int \mathcal{D} \varphi\; e^{-S + \R}$. The scalar field propagator can be computed from field configurations sampled from the integral $\frac{1}{\widetilde{Z}}\int \mathcal{D} \varphi\; e^{-S + \R}$ by taking the ensemble average of $e^{i\Theta}$ and including an additional factor of $\frac{\widetilde{Z}}{Z}$ that can be computed from propagator magnitudes generated with standard MC sampling of Eq. .
[^8]: Multilevel hierarchical integration [@Luscher:2001up] can be used to exponentially reduce the StN problem associated with sampling products of increasingly many factors in Eqs. - as in Refs. [@DellaMorte:2007zz; @DellaMorte:2008jd; @DellaMorte:2010yp; @Ce:2016idq; @Ce:2016ajy]. This approach has been explored, and for instance a two-level hierarchical integration scheme for calculating correlation functions from Eqs. - achieves the expected $N_1^{-1}N_0^{-1/2}$ error scaling at moderately large $N_1 \ll N_0$.
[^9]: The real parts of baryon correlation functions are also heavy-tailed, as pointed out in Ref. [@davidkaplanLuschertalk].
[^10]: Note that this is not the most general choice. Since large phase jumps are associated with regions of small magnitude by Eq. , a phase unwrapping scheme that depends on the magnitude and therefore on $P$ may have advantages.
[^11]: It is not expected that additional information from phase unwrapping integration paths that wind around the circle of a finite volume can be used to resolve the accumulation of errors issue.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The question of existence is treated for near-minimizers for the distance functional (or $E$-functional in the interpolation terminology) that are stable under the action of certain operators. In particular, stable near-minimizers for the couple $(L^1, L^p)$ are shown to exist when the operator is the projection on wavelets and these wavelets possess only some weak conditions of decay at infinity.'
address:
- 'Chebyshev Laboratory, St. Petersburg State University, 14th Line V.O., 29B, Saint Petersburg 199178 Russia'
- 'St. Petersburg Department of Steklov Mathematical Institute, Fontanka 27, St. Petersburg 191023, Russia'
author:
- Anton Tselishchev
title: Stability of nearly optimal decompositions in Fourier Analysis
---
Introduction
============
Let $(X, Y)$ be a couple of Banach spaces and $f\in X$. Consider the distance functional from $f$ to the ball of radius $s$ in $Y$: $$E(s, f; X, Y)=\operatorname{dist}_X (f, B_Y(s))=\inf\{\|f-g\|_X: \|g\|_Y\leq s\}.$$ In the book [@KK] the near-minimizers for this functional (and some other functionals) are studied. By this we mean functions $g$ such that $$\|g\|_Y\leq Cs\ \ \hbox{and}\ \ \|f-g\|_X\leq C\operatorname{dist}_X\big(f, B_Y\big(\frac{s}{C}\big)\big).$$ We are interested in the behaviour of near-minimizers under the action of certain operators $T$. It is clear that if $T$ is bounded on $X$ and $Y$ then $Tg$ will also belong to the ball of radius nearly $s$ in $Y$ (which means that $\|Tg\|_Y\leq Cs$) and $\|Tf-Tg\|_X\leq C\operatorname{dist}(f, B_Y(\frac{s}{C}))$ (here $C$ stands for some other constant). In particular if $\operatorname{dist}_X(f, B_Y(t))\leq C\operatorname{dist}_X(Tf, B_Y(t))$ then $Tg$ will be a near-minimizer for $Tf$.
In this regard, we will be interested in operators which are unbounded on $X$ — can we say something about their action on near-minimizers? The corresponding stability theorems are helpful in reducing the problems of evaluation of various functionals in interpolation theory (and thus the interpolation spaces) for complicated pairs of Banach spaces to the case of more simple embracing pairs. Stable near-minimizers for $K$-functionals are the most effective tools for these problems, cf., for example the “shift of smoothness” theorem in §10.2.2. in the book [@KK]. In this article, however, we study the more “demonstrative” distance functional (or $E$-functional in the interpolation terminology). However, the problems about near-minimizers for $E$- and $K$- functional can in a sense be reduced to one another — cf. §5.4. in [@KK].
In the book [@KK] $T$ usually stands for a Calder' on–Zygmund operator and $X$ — for the space $L^1$. As for the space $Y$, the $L^p$ spaces with $1<p<\infty$, $L^\infty$ or (homogeneous) Campanato spaces $\dot{C}_p^{s,k}$ are considered. The minimizers which are in a sense “stable” under the action of $T$ are constructed there. The essential ingredients of these constructions are the Calder' on–Zygmund decompositions or its smooth analogues.
One of the statements proved in that book is the following.
Let $T$ be a Calder' on –Zygmund operator and $f\in L^1$ is a function for which $Tf\in L^1$. Then for any $s>0$ there exists such function $u^{(s)}\in L^1$ that the following conditions hold: $$\begin{aligned}
\|u^{(s)}\|_{L^p}\lesssim& s,\\
\|f-u^{(s)}\|_{L^1}\lesssim& \operatorname{dist}_{L^1}(f,B_{L^p}(s)),\\
\|Tf-Tu^{(s)}\|_{L^1}\lesssim& \operatorname{dist}_{L^1}(f,B_{L^p}(s))+\operatorname{dist}_{L^1}(Tf,B_{L^p}(s)).
\end{aligned}$$
Here we say that $A\lesssim B$ if $A\leq CB$ for some constant $C$. It will always be clear from the context from which parameters $C$ can depend and from which it can not (or it will be stated explicitly). Here these constants do not depend on $s$ and $f$.
The first two conditions in this theorem mean that $u^{(s)}$ is a near-minimizer for the distance functional for $f$ at $s$ and the third one says that $Tu^{(s)}$ behaves much like the near-minimizer for the distance functional for $Tf$ at $s$ (in particular, it will be the near-minimizer if the second term majorizes the first one).
One of the proofs presented in the book reproduces Bourgain’s arguement from paper [@Bo] — the arbitrary near-minimizer is turned to the stable one by adding the summand which is a “good” part of Calder' on–Zygmund decomposition of a certain function.
We will be intersted in stability of near-minimizers in some cases that are not treated in the book [@KK] — more precisely, when operator $T$ is a projection on wavelets which possess only some weak conditions of decay at infinity (in this case $T$ might not be the singular integral operator in the classic sense) or when $T$ is a usual singular intagral operator but $X$ and $Y$ are weighted $L^1$ and $L^p$ spaces. The proofs will also use the Bourgain’s arguments but instead of the standard Calder' on–Zygmund decomposition of a function into the “bad” and “good” parts some other suitable decompositions will be useful.
The author is kindly greatful to his scientific advisor, S. V. Kislyakov, for posing these problems and for the continuous support during the process of their solutions.
The stability theorem for projections on wavelets
=================================================
Some helpful information about wavelets
---------------------------------------
In this section we are using the notation $L^p$ for $L^p({\mathbb{R}})$. Let $\Psi$ be a wavelet. By this we mean that $\Psi\in L^2({\mathbb{R}})$ and functions $\{2^{j/2}\Psi(2^jx-k)\}_{(j, k)\in{\mathbb{Z}}^2}$ form orthonormal basis in $L^2({\mathbb{R}})$. We denote $2^{j/2}\Psi(2^jx-k)$ by $\Psi_{jk}(x)$.
Paper [@Wojt] contains a condition on $\Psi$ which guarantees that $\{\Psi_{jk}\}$ is unconditional basis not only for $L^2$ but for all $L^p$, $1<p<\infty$. Specifically, it says that there exists a function $\phi$ on ${\mathbb{R}}$ such that the following conditions hold for it:\
1) $\phi(x)=\phi(-x)$ for all $x\in{\mathbb{R}}$;\
2) $\phi$ is a decaying function on $[0, \infty]$;\
3) $\phi$ is a bounded function on ${\mathbb{R}}$;\
4) $\int_0^\infty \phi(x) \log(1+x)<\infty$;\
5) $|\Psi(x)|\leq\phi(x)$ for all $x\in{\mathbb{R}}$.
We are assuming that this condition holds. It implies that $\{\Psi_{jk}\}$ is an unconditional basis for $L^p$. The proof of this fact is also presented in a book [@NPS]. Its main ingredient is a decomposition of a function into a sum of two other functions which we are going to need. In order to present it we will use some convenient notations.
Let ${\varepsilon}=\{{\varepsilon}_{jk}\}_{j,k\in{\mathbb{Z}}}$ be a collection of numbers each of which equals to $\pm 1$. We introduce the following operator $U_{\varepsilon}$: $$U_{\varepsilon}f:=\sum_{j, k\in{\mathbb{Z}}} {\varepsilon}_{jk} \langle f, \Psi_{jk} \rangle \Psi_{jk}.$$ In paper [@Wojt] it is proved that these operators are continuous in $L^p$ for all $1<p<\infty$ and their norms are uniformly bounded (in ${\varepsilon}$). In fact, they are operators of weak type $(1, 1)$ with a constant which does not depend on ${\varepsilon}$. We note that all of the subsequent facts are also true for operators $T$ of the form $(\operatorname{Id}+U_{\varepsilon})/2$ which are simply orthogonal projections in $L^2$ on $\operatorname{span}\{\Psi_{jk}:(j, k)\in A\}$ where $A$ can be any subset of ${\mathbb{Z}}^2$ (by $\operatorname{span}$ we mean the closed linear span).
For integer numbers $r$ and $l$ we denote the dyadic inerval $[2^{-r}l, 2^{-r}(l+1)]$ by $I_{rl}$.
For a function $f\in L^1$ and number $\lambda>0$ using Calder' on–Zygmund decomposition we get the collection of intervals $\{I_{rl}\}_{(r,l)\in S}$ wih nonintersecting interiors such that for all of these intervals the following inequalities hold: $$\lambda< \frac{1}{|I_{rl}|}\int_{I_{rl}}|f|\leq 2\lambda$$ and if $x\not\in \cup_{(r,l)\in S} I_{rl}$ the inequality $|f(x)|\leq\lambda$ holds a.e. We set $f_{rl}:=f\chi_{I_{rl}}$, $F:={\mathbb{R}}\setminus \cup_{(r,l)\in S} I_{rl}$. Finally, we denote by $P_j$ the following orthogonal projection in $L^2$: $$P_jh:=\sum_{i<j}\sum_{k\in{\mathbb{Z}}} \langle h, \Psi_{ik}\rangle \Psi_{ik}$$ and by $Q_j$ — projection $\operatorname{Id}-P_j$: $$Q_jh:=\sum_{i\geq j}\sum_{k\in{\mathbb{Z}}} \langle h, \Psi_{ik}\rangle \Psi_{ik}$$
The “good” part of the decomposition from paper [@Wojt] is then the function $$f_\lambda:=f\cdot\chi_f + \sum_{(r,l)\in S} P_r(f_{rl}).$$ The remaining “bad” part is $$f-f_\lambda=\sum_{(r,l)\in S} Q_r(f_{rl}).$$
We are going to need the following statements about this decomposition which are proved in the book [@NPS].
Let $f$ be a function whith $\operatorname{supp}f\subset I_{rl}$. Then there exists a bounded even integrable function $\beta$ decaying on $[0, \infty)$ (and not depending on $f$) such that $\beta(2^j x)\leq 2^{4-j} \beta(x)$ if $|x|\geq 1$ and $j\in{\mathbb{Z}}_+$ and such that the following inequality holds: $$|P_rf(x)|\leq 2^r\|f\|_{L^1}\beta(2^rx-l).$$
Let $f$ be a function whith $\operatorname{supp}f\subset I_{rl}$. Then there exists an even integrable function $\eta$ that decays on the interval $[10, \infty)$ and such that if $|2^rx-l|>10$ the following inequality holds: $$|U_{\varepsilon}Q_r f(x)|\leq \|f\|_{L^1} 2^r \eta(2^rx-l).$$ Here $\eta$ does not depend on $f$ and ${\varepsilon}$.
We are going to need the following lemma which says that we can control $L^p$ norm of function $f_\lambda$.
For any $1\leq p<\infty$ and any function $f\in L^1$ the following inequality holds: $$\Big\|\sum_{(r,l)\in S}P_r(f_{rl})\Big\|_{L^p}\lesssim \lambda^{1-1/p}\|f\|_{L^1}^{1/p}.$$
We note that in [@NPS] and [@Wojt] this statement is proved only for $p=2$. However, our proof will be much like the proof in book [@NPS].
At first we note that it is enough to prove the statement of lemma for integer values of $p$ — in this case we can derive the required bound using interpolation (or simply H" older’s inequality). Thus we need to prove the inequality $$\int_{{\mathbb{R}}}\Big| \sum_{(r,l)\in S}P_r(f_{rl}) \Big|^{p}\lesssim \lambda^{p-1}\|f\|_{L^1}$$ where $p$ is an integer not less than 1.
According to fact 1, the left hand side of this inequality does not exceed $$\begin{aligned}
&\int_{{\mathbb{R}}}\Big| \sum_{(r,l)\in S} 2^r \|f_{rl}\|_{L^1}\beta(2^rx-l) \Big|^p dx\\
&\lesssim \sum_{(r_1,l_1)\in S}2^{r_1}\|f_{r_1l_1}\|_{L^1}\int_{{\mathbb{R}}}\beta(2^{r_1}x-l_1)\Big| \sum_{(r,l)\in S,r\geq r_1}2^r
\|f_{rl}\|_{L^1}\beta(2^rx-l) \Big|^{p-1}dx.\end{aligned}$$ Using the fact that $2^r\|f_{rl}\|_{L^1}\leq 2\lambda$ we see that this expression is bounded by the following: $$\lambda^{p-1}\sum_{(r_1,l_1)\in S}\|f_{r_1l_1}\|_{L^1}2^{r_1}\int_{{\mathbb{R}}}\beta(2^{r_1}x-l_1)
\Big| \sum_{(r,l)\in S,r\geq r_1}\beta(2^rx-l) \Big|^{p-1} dx.$$ Changing the variable in the integral, we can write this expression in the following way: $$\lambda^{p-1}\sum_{(r_1,l_1)\in S}\|f_{r_1l_1}\|_{L^1} \int_{{\mathbb{R}}}\beta(t)
\Big| \sum_{(r,l)\in S, r\geq r_1} \beta(2^{r-r_1}t-(l-2^{r-r_1}l_1)) \Big|^{p-1}dt.$$ For any fixed pair $(r_1, l_1)\in S$ we denote by $S'$ the set of pairs $\{(r-r_1, l-2^{r-r_1}l_1) : (r, l)\in S\}$. It is easy to see that $\{I_{rl}\}_{(r,l)\in S'}$ are also dyadic intervals with nonintersecting interiors. So we need to estimate the following expression: $$\lambda^{p-1}\sum_{(r_1,l_1)\in S}\|f_{r_1l_1}\|_{L^1} \int_{{\mathbb{R}}}\beta(t)
\Big| \sum_{(r,l)\in S', r\geq 0} \beta(2^rt-l)\Big|^{p-1} dt.$$ Now we prove that the integral in this expression is bounded by constant which does not depend on $S'$. Clearly the statement of lemma will follow immediately. So it is left to prove that for every $k\in{\mathbb{Z}}_+$ the following inuequality holds with constant $C$ depending on $k$ but not on $S'$: $$\int_{{\mathbb{R}}} \beta(t)\Big( \sum_{(r,l)\in S', r\geq 0}\beta(2^rt-l) \Big)^k dt\leq C.$$ We prove this by induction in $k$. The inequality is obvious for $k=0$ since $\beta$ is an integrable function. Now assume this inequality holds for $k-1$ and we prove that it holds also for $k$. Note that $$\begin{aligned}
&\int_{{\mathbb{R}}} \beta(t)\Big( \sum_{(r,l)\in S', r\geq 0}\beta(2^rt-l) \Big)^k dt\\
&\lesssim\sum_{(r,l)\in S', r\geq 0} \int_{{\mathbb{R}}}\beta(t)\beta(2^rt-l)\Big(\sum_{(r_1,l_1)\in S', r_1\geq r}
\beta(2^{r_1}t-l_1)\Big)^{k-1}dt.\end{aligned}$$ Let us denote by $S_{nr}$ the set $\{l: (r,l)\in S', I_{rl}\subset [n,n+1]\}$ and by $\varkappa_{nr}$ — the cardinality of $S_{nr}$. Since $r\geq 0$, every interval $I_{rl}$ is contained in the interval of the form $[n, n+1]$ with integer $n$, so we can rewrite our expression in the following way: $$\sum_{n\in{\mathbb{Z}}}\sum_{r=0}^{\infty}\sum_{l\in S_{nr}}\int_{{\mathbb{R}}}
\beta(t)\beta(2^rt-l)\Big( \sum_{(r_1,l_1)\in S', r_1\geq r} \beta(2^{r_1}t-l_1) \Big)^{k-1}dt.$$ Now for any integer $n$ we can split our integral into three parts: $$\begin{aligned}
J_{n1}&:=\sum_{r=0}^{\infty}\sum_{l\in S_{nr}}\int_{n-10}^{n+10}
\beta(t)\beta(2^rt-l)\Big( \sum_{(r_1,l_1)\in S', r_1\geq r} \beta(2^{r_1}t-l_1) \Big)^{k-1}dt,\\
J_{n2}&:=\sum_{r=0}^{\infty}\sum_{l\in S_{nr}}\int_{-\infty}^{n-10}
\beta(t)\beta(2^rt-l)\Big( \sum_{(r_1,l_1)\in S', r_1\geq r} \beta(2^{r_1}t-l_1) \Big)^{k-1}dt,\\
J_{n3}&:=\sum_{r=0}^{\infty}\sum_{l\in S_{nr}}\int_{n+10}^{+\infty}
\beta(t)\beta(2^rt-l)\Big( \sum_{(r_1,l_1)\in S', r_1\geq r} \beta(2^{r_1}t-l_1) \Big)^{k-1}dt.\end{aligned}$$ Now we estimate each of these terms separately. We start with $J_{n1}$: $$\begin{aligned}
J_{n1}&\leq(\max_{[n-10,n+10]}\beta) \sum_{r=0}^{\infty}\sum_{l\in S_{nr}}\int_{{\mathbb{R}}}
\beta(2^rt-l)\Big( \sum_{(r_1,l_1)\in S', r_1\geq r} \beta(2^{r_1}t-l_1) \Big)^{k-1} dt \\
&=(\max_{[n-10,n+10]}\beta) \sum_{r=0}^{\infty}\sum_{l\in S_{nr}}2^{-r}\int_{{\mathbb{R}}}
\beta(t)\Big( \sum_{(r_2,l_2)\in S'', r_2\geq 0} \beta(2^{r_2}t-l_2) \Big)^{k-1}dt.\end{aligned}$$ In order to pass to the last line we used the change of variable which we have already done before. Here $S''$ is the set of pairs of integers depending on $(r, l)$ but it is true for it that $\{I_{r_2l_2}\}_{(r_2, l_2)\in S''}$ are non-intersecting intervals. Using induction hypothesis we conclude that the integral in the expression does not exceed some constant which does not depend on $(r,l)$ and thus our expression is less than or equal to $$C\max_{[n-10,n+10]}\beta\sum_{r=0}^{\infty}2^{-r}\varkappa_{nr}.$$ Here $\sum_{r=0}^{\infty}2^{-r}\varkappa_{nr}$ is the sum of lengths of nonintersecting intervals contained in $[n, n+1]$ and so it does not exceed 1. We conclude that $$J_{n1}\lesssim\max_{[n-10,n+10]}\beta.$$ Using the fact that $\beta$ is a decaying on $[0, +\infty]$ even integrable function we conclude: $$\sum_{n\in{\mathbb{Z}}}J_{n1}\lesssim\sum_{n\in{\mathbb{Z}}}\max_{[n-10,n+10]}\beta\leq C.$$
Now we estimate $J_{n2}$. If $l\in S_{nr}$, then $I_{rl}\subset [n, n+1]$ and so $2^{-r}l \geq n$. So if $t< n-10$ then $2^rt-l=2^r(t-2^{-r}l)\leq 2^r(t-n)<0$. Using the properties of $\beta$ from fact 1 we can conclude that the following inequality holds: $$\begin{aligned}
J_{n2}&\leq \sum_{r=0}^{\infty}\sum_{l\in S_{nr}} \int_{-\infty}^{n-10}
\beta(t)\beta(2^{r}(t-n))\Big( \sum_{(r_1,l_1)\in S', r_1\geq 0} \beta(2^{r_1}t-l_1) \Big)^{k-1}dt\\
&\lesssim\sum_{r=0}^{\infty}2^{-r}\varkappa_{nr}\int_{-\infty}^{n-10}
\beta(t)\beta(t-n)\Big( \sum_{(r_1,l_1)\in S', r_1\geq 0} \beta(2^{r_1}t-l_1) \Big)^{k-1}dt.\end{aligned}$$ As we already mentioned, $\sum_{r=0}^{\infty}2^{-r}\varkappa_{nr}\leq 1$. Then, using monotonicity and integrability of function $\beta$ we see that $\sum_{n\in{\mathbb{Z}}}\beta(t-n)$ is a uniformly bounded function and we get the following estimate: $$\sum_{n\in{\mathbb{Z}}}J_{n2}\lesssim \int_{{\mathbb{R}}}\beta(t) \Big( \sum_{(r_1,l_1)\in S', r_1\geq 0}\beta(2^{r_1}t-l_1) \Big)^{k-1}dt.$$ Using the induction hypothesis we see that the right hand side is bounded by some constant. The term $\sum_{n\in{\mathbb{Z}}}J_{n3}$ is estimated in exactly the same way — if $t\geq n+10$ and $I_{rl}\subset [n, n+1]$, then $2^{-r}l\leq n+1$ and $2^rt-l=2^r(t-2^{-r}l)\geq 2^r(t-n-1)>0$ and the estimates similar to that we have done above show that $\sum_{n\in{\mathbb{Z}}}J_{n3}\leq C$ and the lemma is proved.
Clearly, since on the set $F$ the inequality $|f|\leq\lambda$ holds, the lemma we just proved implies the inequality $$\|f_\lambda\|_{L^p}\lesssim \lambda^{1-1/p} \|f\|_{L^1}^{1/p}.$$
Stability theorem for couple $(L^1, L^p)$
-----------------------------------------
Now we pass to the proof of the stability theorem. Here $T$ will denote the projection on $\operatorname{span}\{\Psi_{jk}: (j,k)\in A\}$ described previously although any operator bounded on $L^p$ and for which fact 2 holds would suit us (every such operator is of weak type $(1,1)$). In this situation the analogue of theorem 1 is true.
Let $T$ be as above, $1<p<\infty$ and $f\in L^1$ is a function for which $Tf\in L^1$. Then for any $s>0$ there exists such function $u^{(s)}\in L^1$ that the following conditions hold: $$\begin{aligned}
\|u^{(s)}\|_{L^p}\lesssim& s, \label{1st}\\
\|f-u^{(s)}\|_{L^1}\lesssim& \operatorname{dist}_{L^1}(f,B_{L^p}(s)), \label{2nd}\\
\|Tf-Tu^{(s)}\|_{L^1}\lesssim& \operatorname{dist}_{L^1}(f,B_{L^p}(s))+\operatorname{dist}_{L^1}(Tf,B_{L^p}(s)). \label{3rd}
\end{aligned}$$
Let $h$ be any near-minimizer such that $\|h\|_{L^p}\leq s$ and $\|f-h\|_{L^1}\leq 2\operatorname{dist}_{L^1}(f,B_{L^p}(s))$. Then we set $u^{(s)}:=h+(f-h)_t$ where $t$ satisfies the condition $t^{p-1}\|f-h\|_{L^1}=s^p$. We remind the reader that here by $(f-h)_t$ we understand the “good” part of the decomposition which is described previously applied to the function $f-h$ and the number $t$. Now we check that $u^{(s)}$ is also a near-minimizer which means that conditions and hold for it. The inequality follows immediately from the fact that according to the lemma we proved $\|(f-h)_t\|_{L^1}\lesssim \|f-h\|_{L^1}$. In order to prove the condition , it is enough to check that $\|(f-h)_t\|_{L^p}\lesssim s$. But using our choise of $t$ and lemma 1 once again we can write: $\|(f-h)_t\|_{L^p}\lesssim t^{1-1/p}\|f-h\|_{L^1}^{1/p}=s$.
It is left to check the condition . In order to do it we choose a function $v\in L^1$ which is a near-minimizer for $Tf$: $\|v\|_{L^p}\leq s$, $\|Tf-v\|_{L^1}\leq 2\operatorname{dist}_{L^1}(Tf, B_{L^p}(s))$. Let $\{I_{rl}\}_{(r,l)\in S}$ be the set of dyadic intervals arising in the construction of function $(f-h)_t$. Then the following estimate holds: $$\sum_{(r,l)\in S}|I_{rl}|\leq t^{-1}\|f-h\|_{L^1}=\Big(\frac{\|f-h\|_{L^1}}{s}\Big)^{p'}
\lesssim \Big(\frac{\operatorname{dist}_{L^1}(f, B_{L^p}(s))}{s}\Big)^{p'}.$$ Here $p'=\frac{p}{p-1}$. Now we write: $$\|Tf-Tu^{(s)}\|_{L^1}\leq \int_{{\mathbb{R}}\setminus\cup 30 I_{rl}} |Tf-Tu^{(s)}| + \int_{\cup 30 I_{rl}}|Tf-v| +
\int_{\cup 30 I_{rl}}|Tu^{(s)}-v|.
\label{eq1}$$ Let us estimate the first summand. We note that it can be written in the following way: $$\begin{aligned}
&\int_{{\mathbb{R}}\setminus\cup 30 I_{rl}} |Tf-Tu^{(s)}|=\int_{{\mathbb{R}}\setminus\cup 30 I_{rl}}|T((f-h)-(f-h)_t)|\\
=&\int_{{\mathbb{R}}\setminus\cup 30 I_{rl}} \Big|T \Big(\sum_{(r,l)\in S}Q_r((f-h)_{rl})\Big)\Big|\leq
\sum_{(r,l)\in S}\int_{{\mathbb{R}}\setminus 30 I_{rl}} |T(Q_r((f-h)_{rl}))|dx.
\end{aligned}$$ According to fact 2 this expression can be bounded by the following: $$\begin{aligned}
&\sum_{(r,l)\in S}\int_{{\mathbb{R}}\setminus 30 I_{rl}} \|(f-h)_{rl}\|_{L^1} 2^r\eta(2^rx-l)dx\\
\leq &\sum_{(r,l)\in S} \|(f-h)_{rl}\|_{L^1} \int_{\mathbb{R}}\eta(x) dx\lesssim\|f-h\|_{L^1}.
\end{aligned}$$ Due to our choice of $h$ this expsession is less than or equal to $2\operatorname{dist}_{L^1}(f,B_{L^p}(s))$.
The second summand in is obviously less than or equal to $\|Tf-v\|_{L^1}\leq 2\operatorname{dist}_{L^1}(Tf,B_{L^p}(s))$. In order to estimate the third one we use the H" older’s inequality and conclude that it does not exceed the following expresion: $$\|Tu^{(s)}-v\|_{L^p}\Big(\sum_{(r,l)\in S}|30 I_{rl}|\Big)^{1/p'}\lesssim
(\|Tu^{(s)}\|_{L^p}+\|v\|_{L^p})\Big(\frac{\|f-h\|_{L^1}}{t}\Big)^{1/p'}.$$ Using the boundedness of $T$ on $L^p$ we get that the third summand in the right hand side of the inequality is estimated by $$s\Big(\frac{\|f-h\|_{L^1}}{t}\Big)^{1/p'}=\|f-h\|_{L^1}\leq 2\operatorname{dist}_{L^1}(f,B_{L^p}(s)).$$ So we checked that the property holds and the theorem is proved.
Now we turn to some corollaries of the theorem we just proved.
Suppose $1<p<\infty$, $T$ is an operator from the theorem and $f\in L^1$ is a function for which $Tf\in L^1$. Then there exists a sequence of functions $f_k\in L^1\cap L^p$ tending to $f$ in $L^1$ for which $Tf_k\in L^1$ and $\|Tf_k-Tf\|_{L^1}\rightarrow 0$.
This statement immediately follows from the theorem if we tend $s$ to infinity (in this case since $L^1\cap L^p$ is dense in $L^1$ the right hand sides of inequalities and tend to zero).
We note that if $T$ is a projection described previously and $E$ is a measurable subset of ${\mathbb{R}}$ then $\chi_E T$ is of course a bounded operator on $L^p$ and fact 2 holds for it. So we have the following generalization of the previous corollary.
Suppose $1<p<\infty$, $T$ is an operator from the theorem, $f$ is a function from $L^1$ and set $E\subset{\mathbb{R}}$ is such that $\chi_E Tf\in L^1$. Then there exist functions $f_k\in L^1\cap L^p$ tending to $f$ in $L^1$ for which $\chi_E Tf_k\in L^1$ and $\chi_E Tf_k\rightarrow \chi_E Tf$ in $L^1$.
It is enough to use the theorem for operator $\chi_E T$ and then repeat the proof of the previous corollary.
Using the first corollary it is easy to see that if a function from $L^1$ has some of the wavelet coefficients equal to zero then it can be approximated by functions from $L^1\cap L^p$ for which the same coefficients are also zero. Here is the precise statement of this fact.
Suppose $1<p<\infty$ and $f\in L^1$. Then there exist functions $g_k\in L^1\cap L^p$ tending to $f$ in $L^1$ such that if $\langle f, \Psi_{rl}\rangle = 0$ then $\langle g_k, \Psi_{rl}\rangle = 0$.
Denote by $A$ the set $\{(r,l):\langle f,\Psi_{rl} \rangle \neq 0\}$ and let $T$ be the orthogonal projection on $\operatorname{span}\{\Psi_{rl}:(r,l)\in A\}$. Then $Tf=f$ and we can set $g_k=Tf_k$ where $f_k$ are the functions from the first corollary.
Weighted stability for singular integrals
=========================================
In this section we will be interested in weighted spaces $L^p({\mathbb{R}}^d; w)$ and action of singular integral operators on them. The standart information about these things can be found for example in the book [@RdF]. By singular integral operator (or Calder' on–Zygmund operator) we mean the operator $T$ bounded on $L^2({\mathbb{R}}^d)$ and which possesses the kernel — the function $K(x, y)$ such that $$(Tf)(x)=\int_{{\mathbb{R}}^d}K(x,y)f(y) dy$$ for all $f$ with compact support and all $x$ outside this support. We assume that for the kernel $K$ and $x$, $y_1$, $y_2$ such that $y_1$ and $y_2$ are inside some cube $Q$ and $x\not \in 5Q$ the following inequality holds: $$|K(x,y_1)-K(x,y_2)|\leq C \frac{|y_1-y_2|^\alpha}{|x-y_1|^{d+\alpha}},$$ where $\alpha$ is a positive number (not depending on $x$, $y_1$ and $y_2$). Besides that, we will need weights from Muckenhoupt classes $A_p$ — all necessary information about them (in particular, the boundedness of Calderon – Zygmund operators on the spaces $L^p(w)$ with $w\in A_p$) can be found in the books [@RdF] and [@Graf]. In the book [@RdF], among other things, the following fact is proved, which is a weight analogue of the property of Calder' on – Zygmund operators, called in [@KK] long-range $L^1$ -regularity:
Let $T$ be a Calder' on–Zygmund operator and $f$ be a function with support in cube $Q$ such that $\int f=0$, $w\in A_1$. Then $$\int_{{\mathbb{R}}^d\setminus 2\sqrt{d}Q}|Tf(x)|w(x)\lesssim \int_{{\mathbb{R}}^d}|f(x)|w(x) dx.$$
We note that, strictly speaking, in [@RdF] only singular integrals of convolution type are considered, that is, for which $K(x, y)$ depends only on $ x-y $. However this plays no role in the proofs of the statements we need (in particular, fact 3).
So, our goal is to prove an analogue of Theorem 1 for spaces with weights. To do this, we use the analogue of Calder' on–Zygmund decomposition which can be found in the article [@AK]. For an arbitrary weight $w$ and a measurable set $E$, we will use the standart notation $w(E)$ for $\int_E w$. Suppose $a\in A_\infty$, $w\in A_1$, $ G\in L^1(w)$. We set $b = \frac{a}{w}$, $g = Gb^{-1}$. Then $g\in L^1 (a)$. The weight $a$ lying in $ A_\infty$ possesses the doubling condition (that is, $ a(2Q) \lesssim a(Q) $ for any cube $Q$), and therefore a Calder' on–Zygmund partition can be applied to $g$ with weight $a$ and the parameter $\lambda$ and we get a set of non-intersecting dyadic cubes $ \{Q_i \} $, such that $$\lambda\leq\frac{1}{a(Q_i)}\int_{Q_i} |Gb^{-1}|a\leq C\lambda,$$ and $|Gb^{-1}|\leq\lambda$ almost everywhere outside $\cup Q_i$. Then the “good part” of the decomposition is the function $G_t$, defined as follows: $$G_\lambda(x)=
\begin{cases}
G(x), x\not \in\cup Q_i,\\
\frac{b(x)}{b(Q_i)}\int_{Q_i} G, x\in Q_i.
\end{cases}$$ We present the properties of this decomposition; their prooves can be found in the article [@AK]. We denote by $ \tilde{Q}$ the cube $ 2\sqrt{d}Q $.
The cubes $Q_i$ we presented and the function $G_\lambda$ possess the following properties:\
1) $|G_\lambda|\lesssim \lambda b$;\
2) $\|G_\lambda\|_{L^1(w)}\lesssim \|G\|_{L^1(w)}$ and thus $\|G-G_\lambda\|_{L^1(w)}\lesssim\|G\|_{L^1(w)}$;\
3) $\int_{Q_i}(G-G_\lambda)=0$;\
4) $a(Q_i)\leq\frac{1}{\lambda}\int_{Q_i}|G|w$, and so $a(\cup \tilde{Q_i})\lesssim \frac{1}{\lambda}\|G\|_{L^1(w)}$.
We now pass to the stability theorem.
Suppose $ 1 <p <\infty $ and let the weights $ w $ and $ v $ be such that $ w \in A_1 $, $ v \in A_p $ and $ a:= (\frac{w^p}{v})^{\frac{1}{p-1}} \in A_\infty $. Suppose $ T $ is the Calder' on–Zygmund operator and the function $ f \in L^1(w) $ is such that $ Tf \in L^1(w)$. Then for any $s>0$ there exists a function $ u^{(s)} \in L^1(w) $ such that $$\begin{aligned}
\|u^{(s)}\|_{L^p(v)}\lesssim& s,\\
\|f-u^{(s)}\|_{L^1(w)}\lesssim& \operatorname{dist}_{L^1(w)}(f, B_{L^p(v)}(s)),\\
\|Tf-Tu^{(s)}\|_{L^1(w)}\lesssim& \operatorname{dist}_{L^1(w)}(f,B_{L^p(v)}(s))+\operatorname{dist}_{L^1(w)}(Tf, B_{L^p(v)}(s)).
\end{aligned}$$
Once the decomposition we need is described, for the proof of the theorem it remains only to repeat the argument from the book [@KK]. Let $ h $ be a function for which the inequalities $\|h\|_{L^p(v)}\leq s$, $\|f-h\|_{L^1(w)}\leq 2\operatorname{dist}_{L^1(w)}(f,B_{L^p(v)}(s))$ hold. We set $ u^{(s)}:= h + (f-h)_t$ where $ t $ is a number such that $ t^{p-1} \| f-h \|_{L^1(w)} = s^p $. Here $ (f-h)_t $ is the function described above (and it was constructed with respect to the weights $ w \in A_1 $ and $ a \in A_\infty $). We check that $ u^{(s)} $ is a near-minimizer. Indeed, $$\|f-u^{(s)}\|_{L^1(w)}\leq \|f-h\|_{L^1(w)}+\|(f-h)_t\|_{L^1(w)}$$ which by the fact 4 does not exceed $$C\|f-h\|_{L^1(w)}\lesssim \operatorname{dist}_{L^1(w)}(f,B_{L^p(v)}(s))$$ The norm of $u^{(s)}$ in $ L^p(v) $ is also easily estimated: $$\|u^{(s)}\|_{L^p(v)}\leq \|h\|_{L^p(v)}+\|(f-h)_t\|_{L^p(v)}\leq s+\Big(\int|(f-h)_t|v\Big)^{1/p}.$$ Taking into consideration that, according to fact 4, $|(f-h)_t|\lesssim tb$, where $b=aw^{-1}=(wv^{-1})^{\frac{1}{p-1}}$, the second term, up to a constant multiplication, is less than or equal to $$\Big(\int t^{p-1} b^{p-1} |(f-h)_t| v\Big)^{1/p}=t^{\frac{p-1}{p}}\|(f-h)_t\|_{L^1(w)}\lesssim
t^{\frac{p-1}{p}}\|(f-h)\|_{L^1(w)}=s.$$ Thus $\|u^{(s)}\|_{L^p(v)}\lesssim s$. It remains to check that the last propert holds, that is, the stability of $u^{(s)}$ under the action of $T$. In order to do this, we consider the near-minimizer $g$ for $Tf$ such that $\|g\|_{L^p(v)}\leq s$ и $\|Tf-g\|_{L^1(w)}\leq 2\operatorname{dist}_{L^1(w)}(Tf, B_{L^p(v)}(s))$ and write: $$\|T(f-u^{(s)})\|_{L^1(w)}\leq \int_{{\mathbb{R}}^d\setminus\cup\tilde{Q_i}} |Tf-Tu^{(s)}|w +
\int_{\cup\tilde{Q_i}}|Tf-g|w+\int_{\cup\tilde{Q_i}}|Tu^{(s)}-g|w.$$
We estimate the first term. Note that $f-u^{(s)}=(f-h)-(f-h)_t$ is a function with support in $\cup Q_i$, moreover, according to fact 4, its integral over each of the cubes $ Q_i $ is equal to zero. Therefore, using fact 3, the first summand can be estimated by $$\|(f-h)-(f-h)_t\|_{L^1(w)}\lesssim \|f-h\|_{L^1(w)}\lesssim \operatorname{dist}_{L^1(w)}(f, B_{L^p(v)}(s)).$$ The second summand is less than or equal to $$\|Tf-g\|_{L^1(w)}\leq 2\operatorname{dist}_{L^1(w)}(Tf, B_{L^p(v)}(s)).$$ In order to estimate the third one we use the H" older’s inequality: $$\int_{\cup\tilde{Q_i}}|Tu^{(s)}-g|w=\int_{\cup\tilde{Q_i}}|Tu^{(s)}-g|v^{1/p}a^{1/p'}\leq
\Big(\int_{\cup\tilde{Q_i}}|Tu^{(s)}-g|^p v\Big)^{1/p}a(\cup\tilde{Q_i})^{1/p'}.$$ Finally, using the last statement of fact 4 (as well as the facts that the operator $ T $ is bounded on $L^p(v)$ and $ \| g \|_{L^p(v)} \lesssim s $, $ \| u^{(s)} \|_{L^p(v)} \lesssim s $), we conclude that our expression is estimated by the following: $$(\|Tu^{(s)}\|_{L^p(v)}+\|g\|_{L^p(v)})\frac{1}{t^{1/p'}}\|f-h\|_{L^1(w)}^{1/p'}\lesssim
s\Big(\frac{\|f-h\|_{L^1(w)}}{t}\Big)^{\frac{p-1}{p}}=\|f-h\|_{L^1(w)}.$$ According to our choise of the function $h$, this expression is less than or equal to $$2\operatorname{dist}_{L^1(w)}(f,B_{L^p(v)}(s)).$$ It remains to collect the estimates and the theorem is proved.
[99]{} D. S. Anisimov, S. V. Kislyakov, *Double singular integrals: interpolation and correction*, Algebra i Analiz, **16**:5 (2004), 1–33. J. Bourgain, *Some consequences of Pisier’s approach to interpolation*, Isr. Math. J., 77 (1992), 165–185. J. Garcia-Cuerva, J. L. Rubio De Francia, *Weighted norm inequalities and related topics*, North-Holland Math. Stud., vol. 116. Notas. Math., vol. 104, North-Holland, Amsterdam, 1985. L. Grafakos, *Classical Fourier Analysis*, 3rd edition, Springer, 2014. S. Kislyakov and N. Kruglyak, *Extremal Problems in Interpolation Theory, Whitney–Besicovitch Coverings, and Singular Integrals*, Birkh" auser, 2013. I. Ya. Novikov, V. Yu. Protasov, M. A. Skopina, *Wavelet Theory*, American Mathematical Society, 2011. P. Wojtaszczyk, *Wavelets as unconditional bases in $L_p({\mathbb{R}})$*, The Journal of Fourier Analysis and Applications, **5**:1 (1999), 73–85.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the popular internal shock model for the prompt emission of gamma-ray bursts (GRBs), collisions between a series of relativistic shells generate lots of paired forward and reverse shocks. We show that the synchrotron emission produced by the forward and reverse shocks respectively could peak at two quite different energy bands if the Lorentz factors of these two types of shocks are significantly different with each other (e.g., one shock is relativistic and the other is Newtonian). We then investigate whether this scenario is applicable to the case of GRB 080319B and find that a bimodal distribution of the shell Lorentz factors, peaking at $\sim400$ and $\sim10^5$, is required. In addition, this scenario predicts an accompanying inverse-Compton (IC) GeV emission with a luminosity comparable to (not much higher than) that of the synchrotron MeV emission, which can be tested with future *Fermi* observations.'
author:
- 'Y. W. Yu$^{1,2}$, X. Y. Wang$^1$, and Z. G. Dai$^1$'
title: 'Optical and $\gamma$-ray emissions from internal forward-reverse shocks: application to GRB 080319B?'
---
Introduction
============
Since the pioneering works by Rees & Mészáros (1994) and Paczyński & Xu (1994), it has been widely argued that the prompt emission of GRBs arises from internal shocks in a relativistic fireball that consists of a series of shells with different Lorentz factors and that the observed $\gamma$-ray emission is usually attributed to synchrotron or inverse-Compton (IC) emission from power-law electrons in the shocks (Mészáros & Rees 1997). Especially, in the case of IC $\gamma$-ray emission, we can also expect a bright synchrotron low-energy (e.g., optical) emission as seed photons for IC scattering. Based on this synchrotron and synchrotron self-Compton (SSC) emission scenario, Mészáros & Rees (1999) accounted for the bright optical flash of GRB 990123.
In this paper we put forward an alternative model, in which prompt optical and $\gamma$-ray emissions may be produced only by synchrotron radiation in internal shocks. It is natural to expect that internal shocks should consist of paired forward and reverse shocks, which are produced simultaneously by collisions between the shells in the fireball. In the case where the Lorentz factors of these two types of shocks are quite different, two groups of electrons accelerated by a forward and a reverse shock respectively during one collision are expected to reach different characteristic energies. Therefore, by combining the two synchrotron components contributed by these two shocks, we can get a bimodal photon spectrum, which might be able to account for some GRBs that have two spectral components. In particular, under some peculiar conditions, an optical peak can be found to fit the prompt optical emission of some GRBs such as the naked-eye GRB 080319B.
For GRB 080319B, with its prompt $\gamma$-ray fluence $(6.13\pm0.13)\times10^{-4}~\rm erg~cm^{-2}$ (20 keV $-$ 7 MeV; Racusin et al. 2008b) and redshift $z=0.937$ (Vreeswijk et al. 2008), its isotropic equivalent $\gamma$-ray energy release is estimated to be $E_{\gamma,iso}=1.4\times10^{54}\rm ergs$ ($20\,{\rm keV}-7\,{\rm MeV}$), which is among the highest ever measured. More surprisingly, this GRB was found to be associated by an extraordinarily bright optical flash peaked at a visual magnitude of 5.3 (Racusin et al. 2008b), which is visible even for the unaided eye. Compared with the extrapolation of the $\gamma$-ray spectrum to the optical band, the observed flux density of the optical flash ($\sim20$ Jy) is about ten thousands times higher. In addition, thanks to the high time resolution, a fluctuating structure can be seen clearly in the optical light curve.
For this rare multi-wavelength prompt emission, any scenario invoking a single synchrotron component from only one emitting region is obviously not viable. Moreover, different from some other bright optical flashes such as those of GRBs 990123 and 061007, an external shock origin for the optical flash of GRB 080319B is disfavored due to the short duration of the optical pulse and the lack of an increasing optical pulse duration throughout the whole prompt phase, as argued by Kumar & Panaitescu (2008). Instead, an internal dissipation origin may be plausible. As argued by Mészáros & Rees (1999) for GRB 990123, Kumar & Panaitescu (2008) and Racusin et al. (2008b) explained the prompt emission of GRB 080319B in the synchrotron and SSC emission scenario. However, as a direct consequence, they predicted a remarkably strong GeV emission, whose luminosity is about 10 times higher than the observed MeV one. In contrast, for several bright GRBs detected by EGRET on [*CGRO*]{}, the GeV fluence is not higher than that in the MeV BATSE energy band (e.g. Sommer et al. 1994; Hurley et al. 1994).
However, in our internal shock-produced two-component synchrotron emission scenario, the problem of GeV emission excess can be avoided, whereas the prompt optical and MeV $\gamma$-ray fluxes of GRB 080319B can be still explained if an unusually high variability of Lorentz factors exists in the fireball. Moreover, since internal shocks can take place many times, we can naturally understand the fluctuating structure of the optical and $\gamma$-ray light curves. However, the emissions in the two energy bands are not strictly correlated with each other, since the durations of the shocks determined by the widths of the shells could be different and the specific shape of the light curves is dependent on the specific structure of the shells. For simplicity, these complications will be not considered in our simple model.
Compared with the internal shock model suggested by Mészáros & Rees (1999) for GRB 990123, our work differs from theirs in two aspects. First, as discussed above, we suggest that both optical and $\gamma$-ray emissions arise from synchrotron emission, differing from the synchrotron plus SSC scenario proposed by them. Second, in their works, only nonrelativistic to semirelativistic internal shocks are considered, which are produced by collisions between shells with similar Lorentz factors, while in our paper some relativistic internal (reverse) shocks would be involved in order to fit the observations of GRB 080319B.
This paper is organized as follows: in §2, we describe the dynamics and synchrotron emission of internal forward-reverse shocks and the model parameters are expressed as functions of some observational quantities. In §3, we constrain the model parameters by the observations of GRB 080319B and then some implications from these results are discussed. In addition, the contribution to the prompt emission by IC scattering of the electrons is also considered with the inferred model parameters. Finally, conclusions and discussion are given in §4.
Two-component synchrotron emission from internal shocks
=======================================================
The dynamics and electron energy distributions
----------------------------------------------
In the internal shock model, the central engine of GRBs is assumed to eject a fireball consisting of a series of shells with different Lorentz factors $\gamma_{\rm shell}$ during the prompt phase. Considering two shells that are ejected subsequently, for example, if the posterior shell (denoted by 4) moves more rapidly than the prior one (denoted by 1), a collision takes place at radius $R_{is}\approx2\gamma_{\rm 1}^2c\delta t\phi_z^{-1}$ (Yu & Dai 2008) to produce an emission pulse, where $\delta t$ is the observed variability time and $\phi_z=1+z$ is introduced due to the cosmological dilution of time. Because of the collision, a pair of shocks (i.e., internal shocks) could arise: a forward shock propagating into shell 1 and a reverse shock propagating into shell 4. The shocked regions in shells 1 and 4 are denoted by 2 and 3, respectively.
According to the jump conditions between the two sides of a shock (Blandford & McKee 1976), we can calculate the comoving internal energy densities of the two shocked regions by $e_2=(\gamma_{21}-1)(4\gamma_{21}+3)n_1m_pc^2$ and $
e_3=(\gamma_{34}-1)(4\gamma_{34}+3)n_4m_pc^2$, where $\gamma_{21}$ or $\gamma_{34}$ is the Lorentz factor of region 2 or 3 relative to the unshocked region 1 or 4. The comoving proton number density $n_i$ of unshocked region $i$ can be calculated by $n_{i}={L_{k,i}/(
4\pi R_{is}^2\gamma_{i}^2m_pc^3})$ for an isotropic kinetic-energy luminosity $L_{k,i}$ and a bulk Lorentz factor $\gamma_i~(\gg1)$ of the unshocked shell. The mechanical equilibrium between the two shocked regions requires $e_2=e_3$, which yields $${(\gamma_{21}-1)(4\gamma_{21}+3)\over(\gamma_{34}-1)(4\gamma_{34}+3)}={n_4\over
n_1}=\left({\gamma_1\over \gamma_4}\right)^2,\label{dyne}$$ where $L_{k,1}=L_{k,4}=L_k$ is supposed. By assuming $\gamma_4\gg\gamma_1$, the above equation leads to (Yu & Dai 2008) $$\gamma_{21}-1={1\over2}\left({\gamma_1\over\gamma}+{\gamma\over\gamma_1}\right)-1\approx\xi\ll1,
~~\gamma_{34}={1\over2}\left({\gamma\over\gamma_4}+{\gamma_4\over\gamma}\right)\approx{\gamma_4\over2\gamma},$$ where $\gamma=\gamma_1(1+\sqrt{2\xi})$ is the Lorentz factor of the shocked regions. This indicates that the reverse shock is relativistic and the forward shock is Newtonian. Moreover, following Dai & Lu (2002), the total number of the electrons swept-up by the forward and reverse shocks during a period of $\delta t$ can be expressed by $N_2={2\sqrt{2\xi} L_k\delta
t/\left(\phi_z\gamma_1m_pc^2\right)}$ and $N_3={L_k\delta
t/\left(\phi_z\gamma_4m_pc^2\right)}$, respectively[^1].
Both forward and reverse shocks can accelerate particles to high energies and amplify magnetic fields. As usual, we assume that the energies of the hot electrons and magnetic fields are fractions $\epsilon_e$ and $\epsilon_B$ of the total internal energy, respectively. Thus, the strength of the magnetic fields is given $$B_i=\left(8\pi
\epsilon_{B}e_i\right)^{1/2}={1\over\sqrt{2}c^{3/2}}\left({\phi_z^2\epsilon_{B}L_k\over\delta
t^2\gamma^{6}}\right)^{1/2},~~i=2,3.$$ For the shock-accelerated electrons, a power-law energy distribution, $dn_e/d\gamma_e\propto\gamma_e^{-p}$ for $\gamma_e\geq\gamma_{e,m}$, is assumed. The characteristic random Lorentz factors of these hot electrons in regions 2 and 3 are determined respectively by $$\begin{aligned}
\gamma_{e,m,2}=\hat{\epsilon}_{e}{m_p\over m_e}(\gamma_{21}-1)\approx{\hat{\epsilon}_{e}}{m_p\over
m_e}\xi,\\
\gamma_{e,m,3}=\hat{\epsilon}_{e}{m_p\over
m_e}(\gamma_{34}-1)\approx{\hat{\epsilon}_{e}\over2}{m_p\over
m_e}{\gamma_{4}\over\gamma},\label{gem}\end{aligned}$$ where $\hat{\epsilon}_{e}=\epsilon_{e}(p-2)/(p-1)$. That $\gamma_{e,m,3}\gg\gamma_{e,m,2}$, due to $\gamma_4/\gamma\gg2\xi$, indicates that the characteristic energy of the reverse-shocked electrons is much higher than that of the forward-shocked electrons. Therefore, the resulting synchrotron photons emitted by these two types of electrons are expected to peak at two different energy bands and thus two distinct spectral components would be observed as in GRB 080319B. To be specific, the reverse shock is responsible for emission in a relatively high-energy band, while the forward shock contributes to a relatively low-energy component.
In both shocked regions, the hot electrons with energies above $\gamma_{e,c,i}m_ec^2$ lose most of their energies during a cooling time $t_{c,i}$, where the cooling Lorentz factor is determined by ${\gamma}_{e,c,i}={6\pi m_ec\phi_z/\left( y_i\sigma_T{B}_i^2\gamma
t_{c,i}\right)}$. The parameter $y_i$, defined as the ratio of the total luminosity to the synchrotron one, is introduced by considering the cooling effect due to the IC emission besides the synchrotron cooling. As pointed out by Ghisellini et al. (2000), the theoretical synchrotron spectrum arising from these electrons, calculated by using the standard assumption that the magnetic field maintains a steady value throughout the shocked region, leads to a spectral slope $F_\nu \propto \nu^{-1/2}$ below $\sim$100 keV, which is in contradiction to the much harder spectra observed. In order to overcome this problem, Pe’er & Zhang (2006) suggested that the magnetic field created by a shock could decay on a length scale ($\lambda_{B,i}$) much shorter than the comoving width ($\Delta_i$) of the shocked region, i.e., $\lambda_{B,i}=\Delta_i/f_{B,i}$ ($f_{B,i}\gg1$). In other words, the shocked region can be roughly divided into a magnetized part immediately after the shock front and a further unmagnetized part. Under this assumption, the cooling time of the electrons should be determined by the time during which the electrons traverse the magnetized region, i.e., $t_{c,i}=\delta
t/f_{B,i}$. Although the size of the magnetized region is reduced significantly by the field-decay effect (as found in §3.3), this region could be still wide enough for electrons to lose a great part of their energy when they traverse it. In this case, the cooling Lorentz factor $\gamma_{e,c,i}$ of electrons is not much higher than $\gamma_{e,m,i}$, so that the radiation efficiency of the electrons is not reduced drastically compared to the case without any magnetic field decay.
Two-component synchrotron emission
----------------------------------
With the electron distributions and the magnetic fields described above, we can give the resultant synchrotron spectra using the method developed by Sari et al. (1998). The reference peak energies of the synchrotron spectra of the forward and reverse shocks are taken to be in optical and $\gamma$-ray bands, respectively. Then, the model parameters can be expressed as functions of some observational quantities.
For the electrons in the magnetized reverse-shocked region, two break frequencies of the synchrotron spectrum are given by $$\begin{aligned}
\nu_{m,3} & = & { q_e\over2\pi
m_ec\phi_z}{\gamma}_{e,m,3}^2{B}_3\gamma \nonumber \\
& = & {\sqrt{2}m_p^2q_e\over16\pi
m_e^3c^{5/2}}\left({\epsilon_{B}\hat{\epsilon}_e^4L_k\gamma_4^4}\over\delta
t^2\gamma^8\right)^{1/2},\label{num,3}\end{aligned}$$ $$\begin{aligned}
\nu_{c,3} & = & {q_e\over2\pi
m_ec\phi_z}{\gamma}_{e,c,3}^2{B}_3\gamma \nonumber \\
& = & {36\sqrt{2}\pi m_eq_ec^{11/2}\over \sigma_{\rm T}^2}\left({\delta t^2f_{B,3}^4\gamma^{16}}\over
\phi_z^4y_3^4\epsilon_{B}^3L_k^3\right)^{1/2}.\label{nuc,3}\end{aligned}$$ The peak flux density of the spectrum at $\nu_{\rm
p}=\min[\nu_m,\nu_c]$ reads $$\begin{aligned}
F_{\nu,{\rm p},3} & = & {\phi_z\over4\pi d_{L}^2}{
m_{e}c^2\sigma_{T}\over 3q_e}{N_{3}\over
f_{B,3}}B_{3}\gamma\nonumber \\& = & {\sqrt{2}m_e\sigma_{\rm
T}\over24\pi m_pq_ec^{3/2}}\left({\phi_z^2\epsilon_{B}L_k^3}\over
d_L^4f_{B,3}^2\gamma^4\gamma_4^2\right)^{1/2},\label{fmax,3}\end{aligned}$$ where the parameter $f_{B,3}$ is introduced because, at any moment, only a fraction $1/f_{B,3}$ of the total reverse-shocked electrons locate at the magnetized region and other electrons in the unmagnetized region do not contribute to the synchrotron emission. The quantities in the left sides of equations (\[num,3\])-(\[fmax,3\]) can be inferred from an observed prompt $\gamma$-ray spectrum, while the right sides are functions of the model parameters. We can therefore solve these equations to find the values of some model parameters, $$\begin{aligned}
L_k&=&2.5\times10^{53}{\rm
erg~s^{-1}}~{y_{3,0}\hat{\epsilon}_{e,-1}^{-1}}\nonumber\\
&&\times\left(d_{L,28}^{2}{F_{\nu,{\rm
p},3,-25}}^{}\nu_{c,3,20}^{1/2}\nu_{m,3,20}^{1/2}\right)\nonumber\\
& \equiv & {{y_3\over\hat{\epsilon}_{e}}}L_k^*\label{lk},
\\
\gamma_{4}&=&4\times10^4~\gamma_{2.5}^{2}{y_{3,0}^{-1/4}\epsilon_{B,-1}^{-1/4}\hat{\epsilon}_{e,-1}^{-3/4}}\nonumber\\
&&\times\left( {\delta t_{0}}^{1/2}d_{L,28}^{-1/2}F_{\nu,{\rm
p},3,-25}^{-1/4}\nu_{c,3,20}^{-1/8}\nu_{m,3,20}^{3/8}\right),
\\
f_{B,3}&=&6\times10^{3}~
\gamma_{2.5}^{-4}{y_{3,0}^{7/4}\epsilon_{B,-1}^{3/4}\hat{\epsilon}_{e,-1}^{-3/4}}\nonumber\\
&&\times\left({\delta
t_0}^{-1/2}\phi_{z,0.3}d_{L,28}^{3/2}F_{\nu,{\rm
p},3,-25}^{3/4}\nu_{c,3,20}^{7/8}\nu_{m,3,20}^{3/8}\right),\end{aligned}$$ where and hereafter the convention $Q=10^xQ_x$ is adopted in cgs units. The quantities in the brackets are basically determined by the observational data and the values of $L_k$, $\gamma_4$, and $f_{B,3}$ are modulated by the remaining free parameters. In particular, $\gamma_4$ and $f_{B,3}$ are strongly dependent on $\gamma$ that can be constrained by optical observations. The fact that $L_{k}$ is independent of the parameter $f_{B,3}$ indicates that the hypothesis of magnetic field decay does not increase the energy requirement of the model. This is because all the reverse-shocked electrons, when they traverse the tiny magnetized region at different times, have released a great part of their energy to $\gamma$-rays (as indicated by $\nu_{m,3}\sim\nu_{c,3}$) via synchrotron emission. The observed $\gamma$-ray emission is mainly contributed by the emission from this tiny region behind the shock front, no matter whether the other part of the shocked region is magnetized or not.
In order to study the properties of the forward shock, we now focus on a possible low-energy emission, whose spectral information is however not as rich as the $\gamma$-ray component. We calculate the peak frequency of the synchrotron spectrum produced by the forward shock as $$\begin{aligned}
\nu_{m,2} & = & {q_e\over2\pi
m_ec\phi_z}{\gamma}_{e,m,2}^2{B}_2\gamma \nonumber
\\ & = & {\sqrt{14}m_p^2q_e\over4\pi
m_e^3c^{5/2}}\left({\epsilon_{B}\hat{\epsilon}_e^4L_k\xi^{5}}\over \delta t^2\gamma^4\right)^{1/2}.\label{num,2}\end{aligned}$$ In order to explain the optical spectrum measured by Raptor after 80 s for GRB 080319B (Kumar & Panaitascu 2008; Woźniak et al. 2008) in the next section, we assume that this peak frequency is below the optical band (i.e., $\nu_{m,2}<\nu_o\equiv5\times10^{14}\rm ~Hz$), which yields $\gamma>\tilde{\gamma}\equiv288~y_3^{1/4}\epsilon_{B,-1}^{1/4}\hat{\epsilon}_{e,-1}^{3/4}\delta
t_0^{-1/2}L_{52.4}^{*1/4}$. This requirement is however not necessary for common GRBs. For $\nu_{m,2}<\nu_o<\nu_{c,2}$, we can calculate the optical flux density and the synchrotron self-absorption thickness at $\nu_o$ by $$\begin{aligned}
F_{\nu,o}=F_{\nu,\rm
p,2}\left({\nu_o\over\nu_{m,2}}\right)^{-(p-1)/2},\label{fo2}\end{aligned}$$ $$\begin{aligned}
\tau_{sa,o}=\tau_{sa,m}\left({\nu_o\over\nu_{m,2}}\right)^{-(p+4)/2},\label{tao2}\end{aligned}$$ where $$\begin{aligned}
F_{\nu,{\rm p},2} & = & {\phi_z\over4\pi d_{L}^2}{ m_{e}c^2\sigma_{T}\over 3q_e}{N_{2}\over f_{B,2}}B_{2}\gamma
\nonumber \\&=& {\sqrt{7}m_e\sigma_{\rm T}\over6\pi m_pq_ec^{3/2}}\left({\phi_z^2\epsilon_{B}L_k^3\xi^2}\over
d_L^4f_{B,2}^2\gamma^6\right)^{1/2},\end{aligned}$$ $$\begin{aligned}
\tau_{sa,m} & = & {5q_e\over B_2\gamma_{e,m,2}^5}{N_2\over 4\pi
R^2f_{B,2}}\nonumber
\\ & = & {5\sqrt{7}m_e^5q_e\over28\pi
m_p^6c^{5/2}}\left({}L_k\over\epsilon_{B}\hat{\epsilon}_e^{10}f_{B,2}^2\gamma^4\xi^{10}\right)^{1/2}.\end{aligned}$$ Combining equations (\[num,2\]), (\[fo2\]) and (\[tao2\]), we write the parameters $\gamma$ and $f_{B,2}$ as functions of $\epsilon_B$ and $\hat{\epsilon}_e$, $$\begin{aligned}
\gamma&=&420~y_{3,0}^{1/16}\epsilon_{B,-1}^{1/16}\hat{\epsilon}_{e,-1}^{-1/16}\nonumber\\
&&\times\left(\delta
t_0^{-5/8}\phi_{z,0.3}^{-1/4}d_{L,28}^{1/2}F_{\nu,o,-22}^{1/4}L_{k,52.4}^{*1/16}\tau_{sa,o,-1}^{-1/4}\right),\\
f_{B,2}&=&180~y_{3,0}^{3(6+p)/16}\epsilon_{B,-1}^{(2+3p)/16}\hat{\epsilon}_{e,-1}^{-(34-13p)/16}\nonumber
\\& & \times\left[\delta
t_0^{(14+p)/8}\phi_{z,0.3}^{(6+p)/4}d_{L,28}^{-(6+p)/2}\right.\nonumber\\
&&\times\left.F_{\nu,o,-22}^{-(6+p)/4}L_{k,52.4}^{*{3(6+p)/16}}\tau_{sa,o,-1}^{(2+p)/4}\right],\end{aligned}$$ where the kinetic-energy luminosity $L_k=2.5\times10^{52}{\rm
erg~s^{-1}}(y_3/\hat{\epsilon}_{e})L^*_{k,52.4}$ obtained in equation (\[lk\]) has been substituted. The value of $\gamma$ is mainly determined by the observational data and insensitive to the remaining free parameters $\epsilon_B$ and $\hat{\epsilon}_e$. For common GRBs, the value of $\tau_{sa,o}$ is thought to be very high (e.g., $\sim10^3$) and thus the value of $\gamma$ could be lower than 100, which leads to hundreds for $\gamma_4$. However, for GRB 080319B, $\tau_{sa,o}$ is deemed to be not larger than unity to ensure a bright optical flash and we suggest $\tau_{sa,o}=0.1$ as a reference value hereafter.
Application to GRB 080319B
==========================
GRB 080319B triggered the Swift-Burst Alert Telescope (15-350 keV) at $T_0$=06:12:49 UT on March 19, 2008 (Racusin et al. 2008a) and was simultaneously detected by the Konus $\gamma$-ray detector onboard the Wind satellite (20 keV-15 MeV; Golenetskii et al. 2008). The time-averaged Konus-Wind $\gamma$-ray spectrum can be fitted well by the Band function (Band et al. 1993) with a low energy slope of $0.855^{+0.014}_{-0.013}$ below the peak of $E_p=675\pm22$ keV and a high-energy slope of $-3.59^{+0.32}_{-0.62}$ above the peak (Racusin et al. 2008b). The burst had a peak flux of $F_p=(2.26\pm0.21)\times10^{-5}\rm erg~cm^{-2}~s^{-1}$ and thus the peak isotropic equivalent luminosity was $L_{p,iso}=(1.01\pm0.09)\times10^{53}\rm erg~s^{-1}$. Using the values of $F_p$ and $E_p$, we roughly estimate the peak flux density of the $\gamma$-ray spectrum, $F_p/E_p\approx14$ mJy. Compared with the extrapolation of the $\gamma$-ray spectrum to the optical band, the observed flux density of the optical flash ($\sim20$ Jy) is about ten thousands times higher.
The model parameters
--------------------
Adopting $z=0.937$ ($d_{L}=1.9\times10^{28}\rm cm$), $\delta t\sim
3$ s, $F_{\nu,o}\sim 20 \rm Jy$ (at $\nu_o=5\times10^{14}\rm Hz$), $F_{\nu,\rm p,3}\sim14\rm mJy$, $h\nu_{m,3}\sim 675\rm keV$ and denoting $\nu_{c,3}\equiv x\times \nu_{m,3}\equiv 10^0x_0\times
\nu_{m,3} $ for GRB 080319B, we derive the model parameters, $$\begin{aligned}
L_k&\simeq &2\times10^{54}~{\rm
erg~s^{-1}}~x_0^{1/2}y_{3,0}\hat{\epsilon}_{e,-1}^{-1},\label{para1}\\
\gamma&\simeq&400~x_0^{1/32}y_{3,0}^{1/16}\epsilon_{B,-1}^{1/16}\hat{\epsilon}_{e,-1}^{-1/16},\label{para2}\\
\gamma_4&\simeq&9\times10^4~x_0^{-1/16}y_{3,0}^{-1/8}\epsilon_{B,-1}^{-1/8}\hat{\epsilon}_{e,-1}^{-7/8},\label{para3}\\
f_{B,2}&\simeq&700~x_0^{3(6+p)/32}y_{3,0}^{3(6+p)/16}\epsilon_{B,-1}^{(2+3p)/16}\hat{\epsilon}_{e,-1}^{-(34-13p)/16},\label{para4}\\
f_{B,3}&\simeq&7\times10^3~x_0^{3/4}y_{3,0}^{3/2}\epsilon_{B,-1}^{1/2}\hat{\epsilon}_{e,-1}^{-1/2},\label{para5}\end{aligned}$$ where a fiducial value of unity is assumed for $y_3$, which will be proved in §3.2. Besides a constraint by the maximum allowed equipartition value ($\epsilon_e\lesssim 0.3$ and $\epsilon_B\lesssim 0.3$), the remaining free parameters $\epsilon_B$ and $\hat{\epsilon}_e$ satisfy $\hat{\epsilon}_{e}<0.09(xy_3^2\epsilon_{B}^2)^{-3/26}$ given $\gamma>\tilde{\gamma}$. The upper limit of $\hat{\epsilon}_e$ is insensitive to the value of $\epsilon_B$ (strictly, with a decrease of $\epsilon_B$, the upper limit of $\hat{\epsilon}_e$ increases slightly). Taking $\hat{\epsilon}_e<0.09$ ($\epsilon_e\lesssim 0.3$) as a conservative estimate, we find:
\(i) $L_k\gtrsim2\times10^{54}\rm erg~s^{-1}$. This is a natural result due to the high observed $\gamma$-ray luminosity ($\sim10^{53}\rm erg~s^{-1}$) of GRB 080319B. The MeV $\gamma$-ray radiation efficiency of the reverse shock can be estimated by $\eta\sim0.05x_0^{-1/2}y_{3,0}^{-1}\hat{\epsilon}_{e,-1}$. We next calculate the total isotropic-equivalent energy release of GRB 080319B, $$\begin{aligned}
E_{k,iso}=2E_{\gamma,iso}/\eta\sim5\times10^{55}x_0^{-1/2}y_{3,0\hat{\epsilon}_{e,-1}^{-1}}~\rm ergs,\end{aligned}$$ where a factor 2 is introduced by considering a similar amount of energy carried by the forward shocks. Using a very small jet angle $\theta_j=0.2^o$ that is found by Racusin et al. (2008b), we get the beaming-corrected energy release of GRB 080319B, $$\begin{aligned}
E_{k,jet}=E_{k,iso}\theta_j^2/2\sim10^{51}x_0^{-1/2}y_{3,0}\hat{\epsilon}_{e,-1}^{-1}~\rm ergs,\end{aligned}$$ which is a typical value for common GRBs.
\(ii) $\gamma\sim400$. This is a typical value for the Lorentz factor of merged GRB ejecta after internal shocks. Thus, we can obtain the internal shock radius for GRB 080319B, $$\begin{aligned}
R_{is}\approx2\gamma_{\rm 1}^2c\delta t\phi_z^{-1}\approx10^{16}{\rm
cm}~\delta t_{0.5}\phi_{z,0.3}^{-1}\gamma_{1,2.6}^{2} ,\end{aligned}$$ which is larger than those of common GRBs ($\sim10^{14}$cm). So, the dynamic influence of the circum-stellar medium on the leading shell ($\gamma_1\sim \gamma\sim400$) needs to be assessed. We calculate the deceleration radius of the leading shell via $$\begin{aligned}
R_{dec}={L_k\delta t\phi_z^{-1}\over2\pi \gamma_1^2A c^2}\approx5\times10^{16}{\rm cm}~A_{*,-1}^{-1}\delta
t_{0.5}\phi_{z,0.3}^{-1}\gamma_{1,2.6}^{-2}L_{k,54.3},\end{aligned}$$ where a wind-like circumburst medium ($\rho=Ar^{-2}$ with $A=5\times10^{11}{\rm g~cm^{-1}}A_*$) is assumed. Moreover, as claimed by Racusin et al. (2008b), a tenuous wind with an upper limit of $A_*\sim 0.03$ is required by the afterglow data of GRB 080319B. In this case, the deceleration radius is larger. Therefore, we conclude that the deceleration of the shells can be ignored at the times where internal shocks among the shells occur.
\(iii) $\gamma_4\gtrsim 10^5$. This high Lorentz factor is allowable for acceleration of an initial fireball with very low baryon contamination (Piran 1999). (iv) $f_{B,3}\gtrsim 10f_{B,2}$. Our constraint on $f_{B,3}$, which is much larger than unity and much less than $\Delta/\lambda_s$ ($\lambda_s$ is the plasma skin depth), is consistent with that found by Pe’er & Zhang (2006) for other GRBs. However, the physical underpinning of these values of $f_{B}$ is unknown. So the difference in $f_{B}$ for different shocked regions lacks a reasonable physical explanation and it might be related to different shock strengths of the forward and reverse shocks.
From the above results, we can see that most of the inferred model parameters are reasonable and acceptable even in common GRBs, except for an unusually high variability of Lorentz factors denoted by $\gamma_4/\gamma_1\sim300$. Although the $\gamma$-ray emission of GRB 080319B seems to be not unusual, the relatively high value of $\delta t\sim3$s (versus $\sim$10 ms for common GRBs) implies an unusually large internal shock radius, which ensures the synchrotron self-absorption frequency below the optical band and reduces the magnetic field strength. Therefore, in order to produce sufficiently energetic $\gamma$-ray photons, it is necessary to invoke some highly relativistic internal shocks that require high variability of Lorentz factors in the fireball. Since the GRB central engine is far from being thoroughly understood, it is difficult to demonstrate whether the central engine can produce such a drastically varying fireball or not, but some possible origins can be still imagined. In the collapsar model, for example, a shell passing through the envelope of a progenitor star should sweep up and clear away the envelope material, leaving a channel behind the shell. A following shell will pass through this clear channel and thus have a very low baryon contamination. This might lead to a highly relativistic ($\gamma_{\rm shell}\sim10^5$) shell. However, due to lateral diffusion of the channel wall, this channel will be possibly contaminated by baryons again some time ($\sim$ order of seconds) later. Subsequently, such a switching-on-and-off process of the channel repeats again and again. As a result, relatively slow and rapid shells are generated alternately. In reality, this process is unlikely to be so regular because a slow/rapid shell could be followed by another slow/rapid shell. Therefore, the temporally-correlated optical and $\gamma$-ray emissions from this process could be polluted by the emission due to the collisions between slow-slow or rapid-rapid shells. This along with other effects (e.g., different shock-crossing times of shells and so on, see §4) may lead to the fact that the prompt optical and $\gamma$-ray emissions are not correlated finely, as observed in GRB 080319B.
To summarize, if some unknown physical processes of the central engine can give rise to a fireball in which the Lorentz factors and densities vary drastically in a variability timescale of few seconds, our internal shock-produced two-component synchrotron emission scenario may account for some fundamental features of the optical and MeV $\gamma$-ray emissions of GRB 080319B.
Inverse Compton emission
------------------------
We denote $\tau_{i}={\sigma_TN_i/\left(4\pi R^2f_{B,i}\right)}$ and $\tau^{\dag}_{i}={\sigma_TN_i/\left(4\pi R^2\right)}$ as the Thomson optical depth of the magnetized and unmagnetized regions, respectively. The SSC emission from the two magnetized regions is considered first. Following Sari & Esin (2001) for $p=2.5$, we estimate the two break energies of the SSC spectrum contributed by the forward shock by $$\begin{aligned}
h\nu_{m,2}^{\rm SSC}=2\gamma_{e,m,2}^2h\nu_{m,2}=1.5~{\rm
keV}~x_0^{3/16}y_{3,0}^{3/8}\epsilon_{B,-1}^{3/8}\hat{\epsilon}_{e,-1}^{29/8},\end{aligned}$$ $$\begin{aligned}
h\nu_{c,2}^{\rm SSC}=2\gamma_{e,c,2}^2h\nu_{c,2}=70~{\rm
GeV}~x_0^{2}y_{2,0}^{-4}y_{3,0}^{4}\hat{\epsilon}_{e,-1}^{2}.\end{aligned}$$ The peak flux at $\nu_{c,2}^{\rm SSC}$ can be estimated by $$\begin{aligned}
[\nu F_{\nu}]_{{\rm p},2}^{\rm SSC}&\sim&\nu_{c,2}^{\rm SSC}\tau_2F_{\nu,{\rm p},2}\left({\nu_{c,2}^{\rm
SSC}/\nu_{m,2}^{\rm SSC}}\right)^{-(p-1)/2}\nonumber\\
& \sim &2\times10^{-8} ~{\rm erg~cm^{-2}s^{-1}}
%~y_{2,0}^{-1}\epsilon_{B,-1}^{-1}\hat{\epsilon}_{e,-1}^{}.
\label{sscflux1}\end{aligned}$$ For the reverse shock, as the SSC peak enters the Klein-Nishina regime, the real peak energy can be determined by (Gupta & Zhang 2007; Fragile et al. 2004) $$\begin{aligned}
h\nu_{KN,3}^{\rm SSC}={\gamma^2m_e^2c^4\over h\nu_{m,3}}=60~{\rm
GeV}~x_0^{1/16}y_{3,0}^{1/8}\epsilon_{B,-1}^{1/8}\hat{\epsilon}_{e,-1}^{-1/8},\end{aligned}$$ at which a negligible flux is found from $$\begin{aligned}
[\nu F_{\nu}]_{{\rm p},3}^{\rm SSC}&\sim&\nu_{KN,3}^{\rm SSC}\tau_3F_{\nu,{\rm p},3}\left({\nu_{KN,3}^{\rm
SSC}\over2\gamma_{e,c,3}^2\nu_{c,3}}\right)^{1/3}\nonumber\\
&\sim &8\times10^{-11} ~{\rm erg~cm^{-2}s^{-1}}. \label{sscflux2}\end{aligned}$$ According to the definition of parameter $y_i$, we obtain $$\begin{aligned}
y_2=1+{L^{\rm SSC}_{2}\over L^{\rm syn}_{2}}\lesssim1+{[\nu F_{\nu}]_{\rm p,2}^{\rm SSC}\over[\nu F_{\nu}]_{\rm
opt}}\approx1.2,\\
y_3=1+{L^{\rm SSC}_{3}\over L^{\rm syn}_{3}}\approx1+{[\nu F_{\nu}]_{\rm p,3}^{\rm SSC}\over[\nu F_{\nu}]_{\rm
MeV}}\approx1,\end{aligned}$$ where $[\nu F_{\nu}]_{\rm opt}=10^{-7}{\rm erg~cm^{-2}s^{-1}}$ and $[\nu F_{\nu}]_{\rm MeV}=2.3\times10^{-5}~\rm erg~cm^{-2}s^{-1}$. This indicates that the SSC emission is relatively unimportant and the estimations in equations (\[para1\])-(\[sscflux2\]) under the assumption, $y_2\sim y_3\sim1$, are self-consistent.
Although a part of the electron energy can be released via synchrotron and SSC emissions when the electrons traverse the magnetized regions, about half energy of the reverse-shocked electrons and almost all energy of the forward-shocked electrons are still held by the electrons when they enter into the unmagnetized regions. This remaining energy can be no longer released via synchrotron radiation because of the lack of magnetic fields, but can via external inverse Compton (EIC) scattering due to the existence of the radiation fields. Determined by this EIC cooling, the cooling Lorentz factor of the electrons in both of the unmagnetized regions reads $$\begin{aligned}
{\gamma}^{\dag}_{e,c}={m_ec\phi_z\over {4\over3}\sigma_Tu_{\gamma}\gamma \delta t},\end{aligned}$$ where the radiation energy density contributed by the synchrotron radiation from the two magnetized regions can be calculated by $u_{\gamma}=(y_2+y_3-2)(B^2/8\pi)\equiv Y^{\dag}(B^2/8\pi)$ by considering $B_2=B_3\equiv B$. Then we have $$\begin{aligned}
{\gamma}^{\dag}_{e,c}={6\pi m_ec\phi_z\over Y^{\dag}\sigma_TB^2\gamma \delta
t}=15~x_0^{-11/32}{Y_{-0.7}^{\dag-1}}y_{3,0}^{-11/16}\epsilon_{B,-1}^{-11/16}\hat{\epsilon}_{e,-1}^{11/16}.\end{aligned}$$ and ${\gamma}^{\dag}_{e,c}<(\gamma_{e,m,2},\gamma_{e,m,3})$. For an EIC spectrum produced by upscattering the seed photons from magnetized region $j$ by the electrons in unmagnetized region $i$, we calculate its two characteristic break frequencies by $$\begin{aligned}
\nu_{L,i}^{(j)}=2{\gamma}_{e,L,i}^2\nu_{L,j},\end{aligned}$$ $$\begin{aligned}
\nu_{H,i}^{(j)}=2{\gamma}_{e,H,i}^2\nu_{H,j},\end{aligned}$$ and estimate its peak flux at the peak frequency $\nu_{H,i}^{(j)}$ roughly by $$\begin{aligned}
\left[\nu F_{\nu}\right]_{{\rm p},i}^{(j)}\sim\nu_{H,i}^{(j)}\tau_i^{\dag}F_{\nu,{\rm
p},(j)}\left({\nu_{H,i}^{(j)}/\nu_{L,i}^{(j)}}\right)^{-(p-1)/2},\label{eicflux}\end{aligned}$$ where the subscription $L$ represents the low break frequency of the seed photons and the low break Lorentz factor of the target electrons, while $H$ represents the high ones. Considering the two synchrotron components for seed photons and the two population unmagnetized electrons, four EIC components are expected:
\(i) For $i=3$ and $j=2$, we have $$\begin{aligned}
h\nu_{L,3}^{(2)}=2{{\gamma}^{\dag2}_{e,c}}h\nu_{m,2}=0.5~{\rm
keV}~x_0^{-1/2}Y_{-0.7}^{-2}y_{3,0}^{-1}\epsilon_{B,-1}^{-1}\hat{\epsilon}_{e,-1}^{3},\end{aligned}$$ $$\begin{aligned}
h\nu_{H,3}^{(2)}=2{{\gamma}^{2}_{e,m,3}}h\nu_{c,2}=6~{\rm
TeV}~x_0^{29/32}y_{2,0}^{-2}y_{3,0}^{29/16}\epsilon_{B,-1}^{-3/16}\hat{\epsilon}_{e,-1}^{19/16}.\end{aligned}$$ Although the peak frequency $\nu_{H,3}^{(2)}$ is just around the Klein-Nishina break determined by $h\nu_{KN,3}^{(2)}={\gamma^2m_e^2c^4/( h\nu_{c,2})}\approx6~{\rm
TeV}$, we may still derive a relatively low flux from equation (\[eicflux\]), $$\begin{aligned}
\left[\nu F_{\nu}\right]_{{\rm p},3}^{(2)}\sim8\times10^{-8}\rm ~erg~cm^{-2}s^{-1}.\label{eicflux3}\end{aligned}$$
\(ii) For $i=3$ and $j=3$, we have $$\begin{aligned}
h\nu_{L,3}^{(3)}=2{{\gamma}^{\dag2}_{e,c}}h\nu_{c,3}=0.3~{\rm
GeV}~x_0^{5/16}Y_{-0.7}^{-2}y_{3,0}^{-11/8}\epsilon_{B,-1}^{-11/8}\hat{\epsilon}_{e,-1}^{11/8},\end{aligned}$$ $$\begin{aligned}
h\nu_{H,3}^{(3)}=2{{\gamma}^{2}_{e,m,3}}h\nu_{m,3}=0.6~{\rm
EeV}~x_0^{-3/16}y_{3,0}^{-3/8}\epsilon_{B,-1}^{-3/8}\hat{\epsilon}_{e,-1}^{3/8}.\end{aligned}$$ The peak frequency $\nu_{H,3}^{(3)}$ is however higher than the Klein-Nishina break frequency, $h\nu_{KN,3}^{(3)}={\gamma^2m_e^2c^4/( h\nu_{m,3})}\approx60~{\rm
GeV}$, and thus the real peak flux can be calculated at $\nu_{KN,3}^{(3)}$ to be $$\begin{aligned}
\left[\nu F_{\nu}\right]_{{\rm p},3}^{(3)}\sim2\times10^{-7}\rm ~erg~cm^{-2}s^{-1}.\label{eicflux4}\end{aligned}$$ Summarizing the above two cases, the EIC process of the reverse-shocked electrons produces a TeV and a GeV emission component, both of which are much weaker than the observed synchrotron MeV emission with a flux of $\sim2.3\times10^{-5}\rm
~erg~cm^{-2}s^{-1}$.
\(iii) For $i=2$ and $j=2$, we have $$\begin{aligned}
h\nu_{L,2}^{(2)}=2{{\gamma}^{\dag2}_{e,c}}h\nu_{m,2}=0.5~{\rm
keV}~x_0^{-1/2}Y_{-0.7}^{-2}y_{3,0}^{-1}\epsilon_{B,-1}^{-1}\hat{\epsilon}_{e,-1}^{3},\end{aligned}$$ $$\begin{aligned}
h\nu_{H,2}^{(2)}=2{{\gamma}^{2}_{e,m,2}}h\nu_{c,2}=10~{\rm
MeV}~x_0^{35/32}y_{2,0}^{-2}y_{3,0}^{35/16}\epsilon_{B,-1}^{3/16}\hat{\epsilon}_{e,-1}^{45/16},\end{aligned}$$ and the peak flux at $\nu_{H,2}^{(2)}$ $$\begin{aligned}
\left[\nu F_{\nu}\right]_{{\rm p},2}^{(2)}\sim7\times10^{-7}\rm ~erg~cm^{-2}s^{-1}.\label{eicflux1}\end{aligned}$$ Due to the stronger synchrotron MeV emission, this relatively weaker MeV emission component is likely to be covered up.
\(iv) For $i=2$ and $j=3$, we have $$\begin{aligned}
h\nu_{L,2}^{(3)}=2{{\gamma}^{\dag2}_{e,c}}h\nu_{m,2}=0.3~{\rm
GeV}~x_0^{5/16}Y_{-0.7}^{-2}y_{3,0}^{-11/8}\epsilon_{B,-1}^{-11/8}\hat{\epsilon}_{e,-1}^{11/8},\end{aligned}$$ $$\begin{aligned}
h\nu_{H,2}^{(3)}=2{{\gamma}^{2}_{e,m,2}}h\nu_{m,3}=0.9~{\rm GeV}~\hat{\epsilon}_{e,-1}^{2}.\end{aligned}$$ Distinct from the emission components discussed in the previous three cases, the flux at $\nu_{H,2}^{(3)}$, $$\begin{aligned}
\left[\nu F_{\nu}\right]_{{\rm p},2}^{(3)}\sim2\times10^{-5}\rm ~erg~cm^{-2}s^{-1},\label{eicflux2}\end{aligned}$$ indicates a strong GeV emission that is as strong as the synchrotron MeV emission. This means that the energy of the forward-shocked electrons would be mainly released by upscattering the synchrotron MeV $\gamma$-ray photons from the reverse shock, while the reverse-shocked electrons lose a great part of their energy via synchrotron cooling directly. Due to the strong emission within the energy regime from sun-GeV to GeV, electron-positron pairs might be produced by collisions between the sub-GeV and GeV photons. According to the results obtained above, we can give approximately an upper limit luminosity of $L_{\rm lim}\sim10^{53}~\rm erg~s^{-1}$ for the sub-GeV ($\varepsilon_{\gamma}\sim0.1\rm GeV$) emission. Then, the optical depth due to pair production interactions can be roughly estimated as $$\begin{aligned}
\tau_{\gamma\gamma}\lesssim{3\over8}\sigma_T{L_{\rm lim}\over4\pi R^2
c\gamma\varepsilon_{\gamma}}{R\over\gamma}\sim0.3~L_{\rm lim,53}\gamma_{2.6}^{-2}R_{16}^{-1},\end{aligned}$$ which indicates that the pair production effect is not significant. For simplicity, the further contribution from the secondary electrons is ignored here.
To summarize, the contributions by the SSC and EIC emission to the observed optical and MeV $\gamma$-ray emissions are insignificant and the two synchrotron components are dominant in the observed bands for GRB 080319B. In contrast, some higher-energy emission components would be produced by the EIC process. Although most of these components are weak, the flux of the strong GeV emission can reach as high as $\sim10^{-5}\rm erg~cm^{-2}~s^{-1}$, which is comparable to the observed MeV one.
Summary and conclusions
=======================
In the popular internal shock model for the prompt emission of GRBs, paired forward and reverse shocks are produced by collisions between some relativistic shells with different Lorentz factors. In this paper, we have considered this model in an unusual situation where a bimodal distribution of the shell Lorentz factors exists. As a result, the Lorentz factors of the forward and reverse shocks are quite different (i.e., the forward shock is Newtonian and the reverse shock is relativistic) and the resulting two-component synchrotron emission is expected to provide a new scenario for some seldom GRBs that have two spectral peaks in the prompt emission.
As an example, we compare our scenario with the recently-observed naked-eye GRB 080319B and constrain the model parameters by fitting the observations. We find that, on one hand, the optical and MeV $\gamma$-ray fluxes of this unique GRB could be explained in our two-component synchrotron emission scenario, if some unknown physical processes of the central engine (e.g., the picture described in §2.1) can give rise to a fireball where the Lorentz factors of $\sim400$ and $\sim10^5$ appears alternately in a variability timescale of few seconds. On the other hand, although the internal shock-produced emission can roughly account for the fluctuating structure of the light curves and the mild temporal correlation between the optical and $\gamma$-ray emissions, it is still difficult to explain clearly why the observed optical emission varies relatively slower than the $\gamma$-ray emission. We speculate that this difference between the light curve variabilities may be due to a complicated distribution of Lorentz factors in the fireball, an inhomogeneous structure of each fireball shell, different shock-crossing times, and other more realistic properties of the system. So, a more detailed simulation is required to improve our present model.
Finally, for high energy emission, the synchrotron plus SSC scenario suggested by Kumar & Panaitescu (2008) and Racusin et al. (2008b) predicts significant GeV $\gamma$-ray emission by considering the second order IC-scattering, the flux of which is about 10 times higher than the observed MeV one. In contrast, our model predicts a relatively weaker GeV component, whose flux is lower than or at most comparable to that of the synchrotron MeV emission. Therefore, future observations for high energy counterparts of GRBs by the *Fermi* Space Telescope are expected to be able to discriminate these two models.
We would like to thank the referees for their comments that have allowed us to improve this paper. This work is supported by the National Natural Science Foundation of China (grants no. 10221001, 10640420144, 10403002 and 10873009) and the National Basic Research Program of China (973 program) No. 2007CB815404 and 2009CB824800. YWY is also supported by the Scientific Innovation Foundation of Huazhong Normal University and the Visiting PhD Candidate Foundation of Nanjing University.
Band, D. et al. 1993, ApJ, 413, 281 Blandford, R. D. & McKee, C. F. 1976, Phys. Fluids, 19, 1130 Bloom, J. S. et al. 2008, ApJ, submitted (arXiv:0803.3215) Dai, Z. G., & Lu, T. 2002, ApJ, 565, L87 Fragile, P. C., Mathews, G. J., Poirier, J., & Totani, T. 2004, Astropart. Phys., 20, 591 Ghisellini, G., Celotti, A., & Lazzati, D. 2000, MNRAS, 313, L1 Golenetskii, S. et al. 2008, GCN Circ. 7482 Gupta, N., & Zhang, B. 2007, MNRAS, 380, 78 Hurley, K. et al. 1994, Nature, 372, 652 Kumar, P., & Panaitescu, A. 2008, MNRAS, accepted (arXiv:0805.0144) Mészáros, P., & Rees, M. J. 1997, ApJ, 476, 232 Mészáros, P., & Rees, M. J. 1999, MNRAS, 306, L39 Paczyński, B., & Xu, G. 1994, ApJ, 427, 708 Pe’er, A., & Zhang, B. 2006, ApJ, 653, 454 Piran, T. 1999, Phys. Rep., 314, 575 Racusin, J. L. et al. 2008a, GCN Circ. 7427 Racusin, J. L. et al. 2008b, Nature, 455, 183 Rees, M. J., & Mészáros, P. 1994, ApJ, 430, L93 Sari, R., & Esin, A. A. 2001, ApJ, 548, 787 Sari, R., Piran, T., & Narayan, R. 1998, ApJ, 497, L17 Sommer, M. et al. 1994, ApJ, 422, L63 Vreeswijk, P. M. et al. 2008, GCN Circ. 7444 Woźniak, W. et al. 2008, ApJ, accepted (arXiv:0810.2481) Yu, Y. W., & Dai, Z. G. 2008, ApJ, accepted (arXiv:0811.1068)
[^1]: In these expressions, the possible spreading of the shells (i.e. decreasing of $L_k$) is not taken into account, since a detailed description for the dynamic evolution is not necessary for our calculations. In this paper, the constant $L_k$ can be seen as an effective kinetic-energy luminosity.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A semiclassical model is presented for characterizing the linear response of elementary quantum optical systems involving cavities, optical fibers, and atoms. Formulating the transmission and reflection spectra using a scattering-wave (transfer matrix) approach, the calculations become easily scalable. To demonstrate how useful this method is, we consider the example of a simple quantum network, i.e., two cavity-QED systems connected via an optical fiber. Differences between our quasi-exact transfer matrix approach and a single-mode, linearized quantum-optical model are demonstrated for parameters relevant to recent experiments with coupled nanofiber-cavity-QED systems.'
author:
- Nikolett Német
- Donald White
- Shinya Kato
- Scott Parkins
- Takao Aoki
bibliography:
- 'Citations.bib'
title: 'Transfer matrix approach to determining the linear response of all-fiber networks of cavity-QED systems'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Document-level context has received lots of attention for compensating neural machine translation (NMT) of isolated sentences. However, recent advances in document-level NMT focus on sophisticated integration of the context, explaining its improvement with only a few selected examples or targeted test sets. We extensively quantify the causes of improvements by a document-level model in general test sets, clarifying the limit of the usefulness of document-level context in NMT. We show that most of the improvements are not interpretable as utilizing the context. We also show that a minimal encoding is sufficient for the context modeling and very long context is not helpful for NMT.\
author:
- |
Yunsu Kim Duc Thanh Tran Hermann Ney\
Human Language Technology and Pattern Recognition Group\
RWTH Aachen University, Aachen, Germany\
[{surname}@cs.rwth-aachen.de]{}\
bibliography:
- 'references.bib'
title: |
When and Why is Document-level Context Useful\
in Neural Machine Translation?
---
Introduction {#sec:introduction}
============
Neural machine translation (NMT) [@bahdanau2015neural; @vaswani2017attention] has been originally developed to work sentence by sentence. Recently, it has been claimed that sentence-level NMT generates document-level errors, e.g. wrong coreference of pronouns/articles or inconsistent translations throughout a document [@guillou2018pronoun; @laubli2018has].
A lot of research addresses these problems by feeding surrounding context sentences as additional inputs to an NMT model. Modeling of the context is usually done with fully-fledged NMT encoders with extensions to consider complex relations between sentences [@bawden2018evaluating; @voita2018context; @zhang2018improving; @miculicich2018document; @maruf2019selective]. Despite the high overhead in modeling, translation metric scores (e.g. [[<span style="font-variant:small-caps;">Bleu</span>]{}]{}) are often only marginally improved, leaving the evaluation to artificial tests targeted for pronoun resolution [@jean2017does; @tiedemann2017neural; @bawden2018evaluating; @voita2018context; @voita2019when]. Even if the metric score gets significantly better, the improvement is limited to specific datasets or explained with only a few examples [@tu2018learning; @maruf2018document; @kuang2018fusing; @cao2018encoding; @zhang2018improving; @maruf2019selective].
This paper systematically investigates when and why document-level context improves NMT, asking the following research questions:
- In general, how often is the context utilized in an interpretable way, e.g. coreference?
- Is there any other (non-linguistic) cause of improvements by document-level models?
- Which part of a context sentence is actually meaningful for the improvement?
- Is a long-range context, e.g. in ten consecutive sentences, still useful?
- How much modeling power is necessary for the improvements?
To answer these questions, we conduct an extensive qualitative analysis on non-targeted test sets. According to the analysis, we use only the important parts of the surrounding sentences to facilitate the integration of long-range contexts. We also compare different architectures for the context modeling and check sufficient model complexity for a significant improvement.
Our results show that the improvement in [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} is mostly from a non-linguistic factor: regularization by reserving parameters for context inputs. We also verify that very long context is indeed not helpful for NMT, and a full encoder stack is not necessary for the improved performance.
Document-level NMT {#sec:doc-level}
==================
In this section, we review the existing document-level approaches for NMT and describe our strategies to filter out uninteresting words in the context input. We illustrate with an example of including one previous source sentence as the document-level context, which can be easily generalized also to other context inputs such as target hypotheses [@agrawal2018contextual; @bawden2018evaluating; @voita2019when] or decoder states [@tu2018learning; @maruf2018document; @miculicich2018document].
For the notations, we denote a source sentence by $\mathbf{f}$ and its encoded representations by $H$. A subscript distinguishes the previous (pre) and current (cur) sentences. $e_i$ indicates a target token to be predicted at position $i$, and $e_1^{i-1}$ are already predicted tokens in previous positions. $Z$ denotes encoded representations of a partial target sequence.
Single-Encoder Approach {#ssec:single-enc}
-----------------------
The simplest method to include context in NMT is to just modify the input, i.e. concatenate surrounding sentences to the current one and put the extended sentence in a normal sentence-to-sentence model [@tiedemann2017neural; @agrawal2018contextual]. A special token is inserted between context and current sentences to mark sentence boundaries (e.g. `_BREAK_`).
Figure \[fig:docnmt:concat\] depicts this approach. Here, a single encoder processes the context and current sentences together as one long input. This requires no change in the model architecture but worsens a fundamental problem of NMT: translating long inputs [@koehn2017six]. Apart from the data scarcity of a higher-dimensional input space, it is difficult to optimize the attention component to the long spans [@sukhbaatar2019adaptive].
Multi-Encoder Approach {#ssec:multi-enc}
----------------------
Alternatively, multi-encoder approaches encode each additional sentence separately. The model learns representations solely of the context sentences which are then integrated into the baseline model architecture. This tackles the integration of additional sentences on the architecture level, in contrast to the single-encoder approach. In the following, we describe two methods of integrating the encoded context sentences. The descriptions below do not depend on specific types of context encoding; one can use recurrent or self-attentive encoders with a variable number of layers, or just word embeddings without any hidden layers on top of them (Section \[ssec:architecture\]).
### Integration Outside the Decoder {#ssec:multi-enc:outside}
The first method combines encoder representation of all input sentences before being fed to the decoder [@maruf2018document; @voita2018context; @miculicich2018document; @zhang2018improving; @maruf2019selective]. It attends from the representations of the current sentence ($H_\textrm{cur}$) to those of the previous sentence ($H_\textrm{pre}$), yielding $\bar{H}$. Afterwards, a linear interpolation with gating is applied: $$\label{eq:gating}
g\bar{H} + (1 - g)H_\textrm{cur}$$ where $g=\sigma\left(W_g\left[\bar{H};H_\textrm{cur}\right]+b_g\right)$ is gating activation and $W_g,b_g$ are learnable parameters. This type of integration is depicted in Figure \[fig:docnmt:outside\]. By using such a gating mechanism, the model is capable of learning how much additional context information shall be included.
### Integration Inside the Decoder {#ssec:multi-enc:inside}
Another method integrates the context inside the decoder; the partial target history $e_1^{i-1}$ is available during the integration. Here, using the (encoded) target history as a query, the decoder attends directly to the context representations. It also has the original attention to the current sentence. Depending on the order of these two attention components, this type of integration has two variants.
**Sequential Attentions** The first variant is stacking the two attention components, with the output of one component being the query of another [@tu2018learning; @zhang2018improving].
Figure \[fig:decoder\_attention\_parallel:seq\] shows the case when the current sentence is attended by the decoder first, which is then used to attend to the context sentence. This refines the regular attention to the current source sentence with additional context information. The order of the attention components may be switched. To block signals of potentially unimportant context information, a gating mechanism can be employed between the regular and context attention outputs like Section \[ssec:multi-enc:outside\].
**Parallel Attentions** Figure \[fig:decoder\_attention\_parallel\] shows the case when performing the two attention operations in parallel and combining them with a gating afterwards [@jean2017does; @cao2018encoding; @kuang2018fusing; @bawden2018evaluating; @stojanovski2018coreference]. This method relates document-level context to the target history independently of the current source sentence, and lets the decoding computation faster.
For each category above, we have described a common architecture shared by previous works in that category. There are slight variations but they do not diverge much from our descriptions.
Filtering of Words in the Context {#ssec:summary-theory}
---------------------------------
Original source in recent years, I correctly foresaw that, in the absence of stronger fiscal stimulus (which was not forthcoming in either Europe or the United States), recovery from the Great Recession of 2008 would be slow.
---------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Remove stopwords recent years, I correctly foresaw absence stronger fiscal stimulus (forthcoming Europe United States), recovery Great Recession 2008 slow.
Remove most frequent words recent correctly foresaw absence stronger fiscal stimulus forthcoming either States recovery Great Recession 2008 slow
Retain named entities recent years Europe the United States the Great Recession 2008
Retain specific POS years I foresaw the absence stimulus was forthcoming either Europe or the United States recovery the Great Recession 2008 would be
Document-level NMT inherently has heavy computations due to longer inputs and additional processing of context. However, intuitively, not all of the words in the context are actually useful in translating the current sentence. For instance, in most literature, the improvements from using document-level context are explained with coreference, which can be resolved with just nouns, articles, and the conjugated words affected by them.
Under the assumption that we do not need the whole context sentence in document-level NMT, we suggest to retain only the context words that are likely to be useful. This makes the training easier with a smaller input space and less memory requirement. Concretely, we filter out words in the context sentences according to pre-defined word lists or predicted linguistic tags:
- Remove stopwords using a pre-defined list[^1]
- Remove $n\in\mathbb{N}$ most frequent words
- Retain only named entities
- Retain only the words with specific parts-of-speech (POS) tags
The first method has the same motivation as to ignore function words. The second method aims to keep infrequent words that are domain-specific or containing gender information. We empirically found that $n=150$ works reasonably well. For the last two methods, we use the <span style="font-variant:small-caps;">Flair</span>[^2] [@akbik2018flair] toolkit. We exclude the tags that are irrelevant to syntax/semantics of the current sentence. The detailed lists of retained tags can be found in the appendix.
The filtering is performed on word level in the preprocessing. When a sentence is completely pruned, we use a special token to denote an empty sentence (e.g. `_EMPTY_`). Table \[tab:summary-ex\] gives examples of the filtering. We can observe that the original sentence is shortened greatly by removing redundant tokens, but the topic information and the important subjects still remain.
Experiments
===========
We evaluate the document-level approaches in IWSLT 2017 English$\to$Italian[^3] and WMT 2018 English$\to$German[^4] translation tasks. We used TED talk or News Commentary v14 dataset as the training data respectively, preprocessed with theMoses tokenizer[^5] and byte pair encoding [@sennrich2016neural] trained with 32k merge operations jointly for source and target languages. In all our experiments, one previous source sentence was given as the document-level context. A special token was inserted at each document boundary, which was also fed as context input when translating sentences around the boundaries. Detailed corpus statistics are given in Table \[tab:corpus\].
All experiments were carried out with <span style="font-variant:small-caps;">Sockeye</span> [@hieber2017sockeye]. We used Adam optimizer [@kingma2014adam] with the default parameters. The learning rate was reduced by 30% when the perplexity on a validation set was not improving for four checkpoints. When it did not improve for ten checkpoints, we stopped the training. Batch size was 3k tokens, where the bucketing was done for a tuple of current/context sentence lengths. All other settings follow a 6-layer base Transformer model [@vaswani2017attention].
In all our experiments, a sentence-level model was pre-trained and used to initialize document-level models, which was crucial for the performance. We also shared the source word embeddings over the original and context encoders.
**en-it** **en-de**
------------------------------- ----------- -----------
Running Words 4.3M 8.1M
Sentences 227k 329k
Documents 2,045 8,891
Document Length (avg. \#sent) 111 37
: Training data statistics.[]{data-label="tab:corpus"}
---------------------- ---------------- ---------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- --
Approach Architecture \#layers [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} \[%\] [[<span style="font-variant:small-caps;">Ter</span>]{}]{} \[%\] [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} \[%\] [[<span style="font-variant:small-caps;">Ter</span>]{}]{} \[%\]
Baseline $\cdot$ $\cdot$ 31.4 56.1 28.9 61.8
Single-Encoder Transformer 6 31.5 57.2 28.9 61.4
Multi-Encoder (Out.) Transformer 6 31.3 56.1 29.1 61.4
Multi-Encoder (Seq.) Transformer 6 32.6 55.2 29.9 60.7
6 **32.7** **54.7** 30.1 60.3
2 32.6 55.2 30.2 60.5
1 32.2 55.8 30.0 60.4
Word Embedding $\cdot$ 32.5 54.8 **30.3** **59.9**
---------------------- ---------------- ---------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- --
Model Comparison {#ssec:architecture}
----------------
**Model Architecture** Firstly, we compare the performance of existing single-encoder and multi-encoder approaches (Table \[tab:architecture\]). For each category of document-level methods (Section \[sec:doc-level\]), we test one representative architecture (Figures \[fig:docnmt:outside\], \[fig:decoder\_attention\_parallel:seq\], \[fig:decoder\_attention\_parallel\]) which encompasses all existing work in that category except slight variations. The tested methods are equal or closest to:
- Single-Encoder:
- Integration outside the decoder: without sharing the encoder hidden layers over current/context sentences
- Integration inside the decoder
- Sequential attention: Decoder integration of with the order of attentions (current/context) switched
- Parallel attention: Gating version of
The training of the single-encoder method was quite unstable. It took about twice as long as other document-level models, yet yielding no improvements, which is consistent with . Longer inputs make the encoder-decoder attention widely scattered and harder to optimize. We might need larger training data, massive pre-training, and much larger batches to train the single-encoder approach effectively [@junczys2019microsoft]; however, these conditions are often not realistic.
For the multi-encoder models, if the context is integrated outside the decoder (“Out.”), it barely improves upon the baseline. By letting the decoder directly access context sentences with a separate attention component, they all outperform the single-encoder method, improving the sentence-level baseline up to +1.4% [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} and -1.9% [[<span style="font-variant:small-caps;">Ter</span>]{}]{}. Particularly, when attending to current and context sentences in parallel (“Para.”), it provides more flexible and selective information flow from multiple source sentences to the decoder, thus producing better results than the sequential attentions (“Seq.”).
**Model Complexity** In the linguistic sense, surrounding sentences are useful in translating the current sentence mostly by providing case distinctions of nouns or topic information (Section \[sec:analysis\]). The sequential relation of tokens in the surrounding sentences is important for neither of them. Therefore we investigate how many levels of sequential encoding is actually needed for the improvement by the context. From a 6-layer Transformer encoder, we gradually reduce the model complexity of the context encoder: 2-layer, 1-layer, and only using word embeddings without any sequential encoding. We remove positional encoding [@vaswani2017attention] when we encode only with word embeddings.
The results are shown in the lower part of Table \[tab:architecture\]. Context encoding without any sequential modeling (the last row) shows indeed comparable performance to using a full 6-layer encoder. This simplified encoding eases the memory-intensive document-level training by having 22% fewer model parameters, which allows us to adopt a larger batch size without accumulating gradients. For the remainder of this paper, we stick to using the multi-encoder approach with parallel attention components in the decoder and restricting the context encoding to only word embeddings.
---------------------------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- -- ----------
Context sentence [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} \[%\] [[<span style="font-variant:small-caps;">Ter</span>]{}]{} \[%\] [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} \[%\] [[<span style="font-variant:small-caps;">Ter</span>]{}]{} \[%\] \#tokens
None 31.4 56.1 28.9 61.8 -
Full sentence 32.5 54.8 30.3 59.9 100%
Remove stopwords 32.2 55.2 30.3 59.9 63%
Remove most frequent words 32.1 55.6 30.2 60.2 51%
Retain only named entities 32.3 55.4 30.3 60.3
Retain specific POS [**32.5**]{} [**55.2**]{} [**30.4**]{} [**60.0**]{} 59%
---------------------------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- -- ------------------------------------------------------------------ ----------------------------------------------------------------- -- ----------
[0.45]{} {width="\textwidth"}
[0.45]{} {width="\textwidth"}
Filtering Words in the Context {#ssec:summary-exp}
------------------------------
To make the context modeling even lighter, we analyze the effectiveness of the filtered context (Section \[ssec:summary-theory\]) in Table \[tab:summarization\]. All filtering methods shrink the context input drastically without a significant loss of performance. Each method has its own motivation to retain only useful tokens in the context; the results show that they are all reasonable in practice. In particular, using only named entities as context input, we achieve the same level of improvement with only 13% of tokens in the full context sentences. By filtering words in the context sentences, we can use more examples in each batch for a robust training.
Context Length {#ssec:length}
--------------
Filtered context inputs (Section \[ssec:summary-exp\]) with a minimal encoding (Section \[ssec:architecture\]) make it also feasible to include much longer context without much difficulty. Most of previous works on document-level NMT have not examined context inputs longer than three sentences.
Figure \[fig:length\] shows the translation performance with an increasing number of context sentences. If we concatenate full context sentences (plain curves), the performance deteriorates severely. We found that it is hard to fit such long sequences in memory as the training becomes very erratic.
The training is much more stable with filtered context; the dashed/dotted curves do not drop significantly even when using 20 context sentences. In the English$\to$Italian task, the performance slightly improves up to 15 context sentences. In the English$\to$German task, there is no improvement by extending the context length over 5 sentences. This discrepancy can be explained with document lengths in each dataset (Table \[tab:corpus\]). The TED talk corpus for English$\to$Italian has much longer documents, thus it is probable to benefit from larger context windows. However, in general we observe only marginal improvements by enlarging the context length to more than one sentence, as seen also in , , or .
Analysis {#sec:analysis}
========
Simplifying the context encoder (Section \[ssec:architecture\]) and filtering the context input (Section \[ssec:summary-exp\]) are both inspired by the intuition that only a small part of the context is useful for NMT. In order to verify this intuition rigorously, we conduct an extensive analysis on how document-level context helps the translation process, manually checking every output of sentence-level/document-level NMT models; automatic metrics are inherently not suitable for distinguishing document-level behavior. Our analysis is not constrained to certain discourse phenomena which are favored in evaluating document-level models. We quantify various causes of the improvements 1) regardless of its linguistic interpretability and 2) in a realistic scenario where not all the test examples require document-level context. Here are the steps we take:
1. Translate a test set with a sentence-level baseline and a document-level model.
2. Compute per-sentence [[<span style="font-variant:small-caps;">Ter</span>]{}]{} scores of outputs from both models.
3. Select those cases where the document-level model improves the per-sentence [[<span style="font-variant:small-caps;">Ter</span>]{}]{} over the sentence-level baseline.
4. Examine each case of 3 by looking at:
- Source, context, and translation outputs
- Attention distribution over the context tokens for each target token: averaged over all decoder layers/heads
- Gating activation (Equation \[eq:gating\])
5. Classify each case into “coreference”, “topic-aware lexical choice”, or “not interpretable”.
Statistics of each category on the test sets are reported in Table \[tab:analysis\]. The manual inspection of translation outputs is done by a native-level speaker of Italian or German, respectively.
Only a couple of cases belong to coreference, which is ironically the most advocated improvement in the literature on document-level NMT. One of them is shown in Table \[tab:analysis-co\]. In the document-level NMT, the English word “said” is translated to a correct conjugation of “sagen” (= say) for the third person noun “der Präsident” (= the President). This can be explained by the high attention energy on “Trump” (Figure \[fig:analysis-co\]) in the context sentence.
Another interpretable cause is topic-aware lexical choice (Table \[tab:analysis-to\]). The document-level model actively attends to “seized” and “cocaine” in the context sentence (Figure \[fig:analysis-to\]), and does not miss the source word “raids” in the translation (“Razzien”). When it corrects the translation of polysemous words, it is related to word sense disambiguation [@gonzales2017improving; @marvin2018exploring; @pu2018integrating]. This category includes also a coherence of text style in the translation outputs, depending on the context topic.
-------------------------------------------------------------------------- -- ------- -------
Category
Coreference 21 2
Topic-aware lexical choice 66 33
Not interpretable 292 1,211
Total [[<span style="font-variant:small-caps;">Ter</span>]{}]{} improved 379 1,246
Total 1,147 2,998
-------------------------------------------------------------------------- -- ------- -------
: Causes of improvements by document-level context.[]{data-label="tab:analysis"}
We found that only 7.5% of the [[<span style="font-variant:small-caps;">Ter</span>]{}]{}-improved cases can be interpreted as utilizing document-level context. The other cases are mostly general improvements in adequacy or fluency which are not related to the given context. Table \[tab:analysis-un\] shows such an example. It improves the translation by a long-range reordering and rephrasing some nouns, whose clues do not exist in the previous source sentence. Its attention distribution over the context words is totally random and blurry (Figure \[fig:analysis-un\]).
A possible reason for the non-interpretable improvements is regularization of the model, since the training data of our experiments are relatively small. Figure \[fig:gating\] shows that, for most of the improved cases, the model has non-negligible gating activation towards document-level context, even if the output seems not to benefit from the context. It means that, when combining the encoded representations of context/current sentences, the model can reserve some of its capacity to the information from context inputs. This might effectively mitigate overfitting to the given training data.
![Gating activation for all [[<span style="font-variant:small-caps;">Ter</span>]{}]{}-improved cases of the English$\to$German task, averaged over all layers and target positions.[]{data-label="fig:gating"}](hist.pdf){width="1.05\linewidth"}
-------------------- ----------------------------------------------------------------------------------------------------------------------------
**Previous src** inside the White House, addressed Sikorsky representatives, joking with the media about his own fleet of company products.
**Current src** “I know Sikorsky very well,” the President said, “I have three of them.”
**Reference** ich kenne Sikorsky sehr gut, sagte der Präsident, ich habe drei davon.
**Sent-level hyp** ich kenne Sikorsky sehr gut, **so** der Präsident, habe drei davon.
**Doc-level hyp** ich kenne Sikorsky sehr gut, **sagte** der Präsident, ich habe drei davon.
-------------------- ----------------------------------------------------------------------------------------------------------------------------
-------------------- ---------------------------------------------------------------------------------------------------------
**Previous src** in addition, officials large quantities of marijuana and , firearms and several hundred thousand euros.
**Current src** at simultaneous raids in Italy, two people were detained.
**Reference** bei zeitgleichen Razzien in Italien wurden zwei Personen festgenommen.
**Sent-level hyp** gleichzeitig wurden in Italien zwei Personen verhaftet.
**Doc-level hyp** bei gleichzeitigen **Razzien** in Italien wurden zwei Menschen inhaftiert.
-------------------- ---------------------------------------------------------------------------------------------------------
-------------------- -----------------------------------------------------------------------------------------------------------
**Previous src** other cities poach good officials and staff members and offer attractive conditions.
**Current src** the talk is of a downright “contest between public employers”.
**Reference** die Rede ist von einem regelrechten Wettbewerb der öffentlichen Arbeitgeber.
**Sent-level hyp** das Gerede über einen Wettkampf zwischen öffentlichen Arbeitgebern$\:$ ist von einem Gerechtigkeitstreit.
**Doc-level hyp** die Rede ist von einem herben Wettbewerb zwischen öffentlichen Arbeitgebern.
-------------------- -----------------------------------------------------------------------------------------------------------
[0.32]{} {width="\linewidth"}
[0.33]{} {width="\linewidth"}
[0.33]{} {width="\linewidth"}
----------- ---------------- -- ----------- -----------
Condition System **en-it** **en-de**
Sentence-level 31.4 28.9
Document-level **32.5** **30.3**
Sentence-level **33.7** **32.3**
Document-level 33.5 32.0
Sentence-level - **40.2**
Document-level - 39.9
----------- ---------------- -- ----------- -----------
: Sentence-level vs. document-level translation performance in different data/training conditions.[]{data-label="tab:regularized"}
We argue that the linguistic improvements with document-level NMT have been sometimes oversold, and the document-level components should be tested on top of a well-regularized NMT system. In our experiments, we obtain a much stronger sentence-level baseline by applying a simple regularization (dropout), which the document-level model cannot outperform (Table \[tab:regularized\]).
On a larger scale, we also built a sentence-level model with all parallel training data available for the WMT 2019 task and fine-tuned only with document-level data (Europarl, News Commentary, newstest2008-2014/2016). The document-level training does not give any improvement in [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} (last two rows of Table \[tab:regularized\]). There may exist document-level improvements which are not highlighted by the automatic metrics, but the amount of such improvements must be very small without a clear gain in [[<span style="font-variant:small-caps;">Bleu</span>]{}]{} or [[<span style="font-variant:small-caps;">Ter</span>]{}]{}.
Conclusion {#sec:conclusion}
==========
In this work, we critically investigate the advantages of document-level NMT with a thorough qualitative analysis and expose the limit of its improvements in terms of context length and model complexity. Regarding the questions asked in Section \[sec:introduction\], our answers are:
- In general, document-level context is utilized rarely in an interpretable way.
- We conjecture that a dominant cause of the improvements by document-level NMT is actually the regularization of the model.
- Not all of the words in the context are used in the model; we leave out redundant tokens without loss of performance.
- A long-range context gives only marginal additional improvements.
- Word embeddings are sufficient to model document-level context.
For a fair evaluation of document-level NMT methods, we argue that one should make a sentence-level NMT baseline as strong as possible first, i.e. by using more data or applying proper regularization. This will get rid of by-product improvements from additional information flows and help to focus only on document-level errors in translation. In this condition, we show that document-level NMT can barely improve translation metric scores against such strong baselines. Targeted test sets [@bawden2018evaluating; @voita2019when] might be helpful here to emphasize the document-level improvements. However, one should bear in mind that a big improvement in such test sets may not carry over to practical scenarios with general test sets, where the number of document-level errors in translation is inherently small.
Given these conclusions, a future research direction would be building a lightweight post-editing model to correct only document-level errors, not complicating the sentence-level model too much for a very limited amount of document-level improvements. To strengthen our arguments, we also plan to conduct the same qualitative analysis on other types of context inputs (e.g. translation history) and different domains.
Our implementation of document-level NMT methods is publicly available on the web.[^6]
Acknowledgments {#acknowledgments .unnumbered}
===============
[l]{}[0.2]{} {width="22.00000%"}
This work has received funding from the European Research Council (ERC) (under the European Union’s Horizon 2020 research and innovation programme, grant agreement No 694537, project “SEQCLAS”). The work reflects only the authors’ views and none of the funding agencies is responsible for any use that may be made of the information it contains.
The authors thank Tina Raissi for analyzing English$\rightarrow$Italian translations.
[l]{} **AContext Word Filtering Details**\
\
The tables below follow the tagging conventions of <span style="font-variant:small-caps;">Flair</span> ().
Tag Examples Tag Examples
-------------- ---------------------------------- ---------- --------------------------
Location mountain ranges, bodies of water Law named laws
Organization companies, agencies Language any named language
Person People, fictional characters Dates important holidays
NORP nationalities, political groups Time times smaller than a day
Facility buildings, airports Percent percentage values
GPE countries, cities Money monetary values
Product vehicles, food Ordinal first, second, $\ldots$
Event named hurricanes, battles Cardinal other numeric values
Tag Description Tag Description
------ -------------------------- ------- ---------------------------------------
CC Coordinating conjunction PDT Predeterminer
CD Cardinal number PRP Personal pronoun
CODE Code IDs PRP\$ Possessive pronoun
DT Determiner UH Interjection
FW Foreign word VBG Verb, gerund/present particle
MD Modal VBN Verb, past participle
NN Noun, singular VBP Verb, non-3rd person singular present
NNP Proper noun, singular VBZ Verb, 3rd person singular present
NNS Noun, plural VB Verb, base form
NNPS Proper noun, plural VBD Verb, past tense
[^1]: https://github.com/explosion/spaCy
[^2]: https://github.com/zalandoresearch/flair
[^3]: https://sites.google.com/site/iwsltevaluation2017
[^4]: https://www.statmt.org/wmt18/translation-task.html
[^5]: http://www.statmt.org/moses
[^6]: <https://github.com/ducthanhtran/sockeye_document_context>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Ahmad Shariati[^1]'
date: 7 Feb 2008
title: 'Comment on “Learning to build the bomb”'
---
According to Alisa Carrigan’s opinion in *Physics Today* Dec 2007 [@Carrigan], to prevent proliferation of nuclear weapons, certain rules should be set to prevent the spread of a particular kind of knowledge. Her argument goes as follows: To build nuclear weapons, scientists and engineers of potentially rogue countries need to know some technics which could be learnt in nonmilitary peaceful activities, for example in nuclear power plants. Therefore, to prevent some countries access to nuclear weapon knowledge, one should prevent their scientists and engineers being trained in such facilities. As Carrigan says, *knowledge proliferation* is as important as nuclear proliferation. To show this, Carrigan mentions the case of South Africa’s nuclear program—some scientists and engineers having trained in USA and Europe in non-military, peaceful, academic activities, obtained enough knowledge and expertise to make their own nuclear weapons. Carrigan says that the cases of North Korea, India, and Pakistan obtaining nuclear weapons, and Iran’s progress in uranium enrichment are alike.
I would like to comment on this line of reasoning.
Logical consequences
====================
First, nuclear weapons are not the only threats. Chemical and biological weapons are as dangerous as nuclear weapons. So if we accept this logic, the restriction should not be limited to nuclear physics and the related fields— by the same reasons, various fields in chemistry, chemical engineering, pharmaceutical and biological sciences, physics, and mechanics must be off-limits. After that comes various fields of mathematics, for example number theory, and software engineering; as they have applications in cryptography. Just think of a terrorist attack by some hacker to a computer that is controlling an airlines corridor traffic. Even quantum computation is also dangerous, because it has applications in deciphering. Where should one stop?
Carrigan distinguishes between explicit and tacit knowledge. But there is no permanent sharp line between explicit and tacit knowledge. For example, the need to use fabric gloves to assemble centrifuges, the problem mentioned in Carrigan’s article, now that it is being published, has been transformed from tacit to explicit. Since people do have access to explicit knowledge, through books and journals, it is not sufficient to monitor the sources of tacit knowledge—to prevent proliferation of the required knowledge, it is necessary to control the flow of explicit knowledge as well. This requires establishing a system of censorship.
I think the logical consequence of accepting Carrigan’s idea is a kind of “Knowledge Nonproliferation Treaty”—a system to monitor and control the flow of information through books, journals, internet, participation in conferences, sabbaticals, etc. Such a system, if implemented, simply means this: Humans are divided into two categories, those having the knowledge of making nuclear, chemical, and biological weapons, and those that have not yet this knowledge. The first category has the right and must do its best to prevent the second category obtaining the required knowledge and technology. At this limit, I think, it is nothing but a variant of apartheid.
An inevitable conclusion in line with Carrigan’s arguments would be that good people should control other people in the sense that if other people were approaching dangerous knowledge (even by themselves), good people should prevent them even if necessary by force, even if necessary by getting rid of the scientists of other people and destroying their scientific facilities, including their libraries and equivalent digital resources. This, simply would force people in the second category—those who are forbidden to have the sacred knowledge—to invoke dirty tricks.
Scientific apartheid doesn’t work
=================================
I am not saying that scientific apartheid is bad, for valuing something as good or bad could not be judged scientifically. What *is* important, I think, is that this scientific apartheid does not work, and is not a suitable means to establish a sustainable peace.
It does not work, because it is now almost impossible to impose it. Today, contrary to say 100 years ago, even people in developing countries do have access to the basics of the scientific method and the fundamentals of science. Once one knows these, it is in principle possible to produce the forbidden knowledge. After all, this is what scientists in the developed countries have done, and assuming that there is no meaningful distinction between the intelligence of people in different countries, if people in say USA have been able to learn or construct things by themselves, people in other countries can do that as well, though with some delay. So a Knowledge Nonproliferation Treaty does not help, since knowledge is not only transported, but also produced—the example of fabric gloves mentioned in Carrigan’s article is a very good example of this.
Reducing tensions
=================
Now let us consider this problem from another point of view. The case of South Africa’s nuclear program is worthy of discussing. Why South Africa made weapons, and why finally destroyed its weapons? I think the answer is that, 4 decades ago South Africa was a country, having trouble with its neighbors—and its own people as well. After the Apartheid era, the troubles being solved, and now South Africa does not need any nuclear weapons. Which other countries have made nuclear weapons? North Korea, having trouble with South Korea; Israel, having trouble with all its neighbors; Pakistan, having trouble with India; India, trouble with Pakistan. What Carrigan points, is that all these nations were able to obtain the required knowledge, and all of them from non-military activities. What I conclude from this, is that if some nation has enough motivation to build a dangerous weapon, it probably can obtain the required knowledge—and Carrigan says that this has always been achieved by native scientists. *Now, if we want to make a sustainable peace, why not try to reduce the motivation of nations to have weapons?*
In mathematical terms
=====================
Let me formulate my view more mathematically. Let $K(T)$ be the probability of nation $X$ to have the knowledge and technology required to produce a nuclear weapon before time $T$. Let $H(T)$ be the probability of nation $X$ to have nuclear weapons before time $T$. And let $U(T)$ be the probability of nation $X$ using nuclear weapons before time $T$. For time $T$ let’s consider 2020 for the moment.
One can argue that $K$ is an increasing function of the level of ease physicists from $X$ can visit foreign universities having nuclear physics departments. Denote this level of ease with $x$. One can also argue that $H$, and especially $U$ depend critically on the regional tensions—by region I mean the Middle East, Kashmir, Korean Peninsula, etc. Let $y$ denote the level of this tensions.
The most important task is to try to reduce $U$, and after that $H$. Carrigan is saying that $K$ is an increasing function of $x$, even though so far all those nations who had enough motivation, have succeeded in obtaining nuclear weapons. What I am saying is that we know that decreasing $y$ has quite profound effects on reducing $H$ and $U$, and we know that in the only case for which the regional tensions vanished, the country (South Africa) destroyed its weapons. So why not trying to reduce the regional tensions?
Besides, $K(T)$ is obviously an increasing function of time $T$, because it is an increasing function of the overall level of knowledge and technology of the world. Day by day it will become more and more difficult to make $K$ not approaching 1. However, for $H(T)$ and especially $U(T)$ it is not obvious that they are increasing functions of time, for they depend on the political conditions at times $t \leq T$. So again, it is quite wiser to try to reduce the regional tensions.
Finally, trying to reduce the level of knowledge of nation $X$, or preventing it from increasing its knowledge, by establishing a type of Knowledge Nonproliferation Treaty, will cause $X$ to become more aggressive and less developed. I think both of these would increase $H(T)$ and $U(T)$.
[9]{} Alisa L. Carrigan, “Learning to Build the Bomb”, [*Physics Today*]{}, Dec 2007, pp. 54-55.
[^1]: Dept of Physics, Alzahra University, Tehran 19935, Iran; e-mail: `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nuclear electromagnetic currents derived in a chiral-effective-field-theory framework including explicit nucleons, $\Delta$ isobars, and pions up to N$^2$LO, [*i.e.*]{} ignoring loop corrections, are used in a study of neutron radiative captures on protons and deuterons at thermal energies, and of $A$=2 and 3 nuclei magnetic moments. With the strengths of the $\Delta$-excitation currents determined to reproduce the $n$-$p$ cross section and isovector combination of the trinucleon magnetic moments, we find that the cross section and photon circular polarization parameter, measured respectively in $n$-$d$ and $\vec{n}$-$d$ processes, are significantly underpredicted by theory.'
address: |
$^{\rm a}$Physics Department, Old Dominion University, Norfolk, Virginia 23529, USA\
$^{\rm b}$Theory Center, Jefferson Laboratory, Newport News, Virginia 23606, USA\
$^{\rm c}$Physics Department, Hampton University, Hampton, Virginia 23668, USA\
[email protected], [email protected], [email protected]
author:
- 'S. Pastore$^{\rm a}$, R. Schiavilla$^{\rm a,b}$, and J.L. Goity$^{\rm b,c}$'
title: 'ELECTROMAGNETIC PROCESSES IN $\chi$EFT'
---
Introduction {#sec:sec1}
============
Nuclear electromagnetic currents have recently been derived in a chiral-effective-field-theory ($\chi$EFT) framework including nucleons, $\Delta$ isobars, and pions [@Pastore08]. Formal expressions up to one loop have been obtained in time-ordered perturbation theory with non-relativistic Hamiltonians constructed from the chiral Lagrangian formulation of Refs. [@Weinberg90; @vanKolck94; @Epelbaum98]. Thus, the study in Ref. [@Pastore08] is similar to the work of Park [*et al.*]{} [@Park96], albeit it uses a different formalism.
The present talk is a much abridged summary of Ref. [@Pastore08]. The currents up to next-to-next-to-leading order (N$^2$LO), that is ignoring loop corrections which enter at N$^3$LO, are used to calculate the magnetic moments of $A$=2 and 3 nuclei, and the thermal neutron radiative captures on protons and deuterons. Realistic two- and three-nucleon (for $A$=3) potentials are used to generate the bound and continuum wave functions. To have an estimate of the model dependence arising from short-range phenomena, the variation of the predictions is studied as function of the cutoff parameter, which is used to regularize the two-body operators, as well as function of the input potentials—either the Argonne $v_{18}$ (AV18) [@Wiringa95] or CD-Bonn (CDB) [@Machleidt01] in combination with respectively the Urbana IX [@Pudliner97] or Urbana IX$^*$ [@Viviani07]—used to generate the wave functions (the AV18 and CDB have rather different short-range behaviors).
We find that the N$^2$LO calculations do not provide a satisfactory description of the experimental data, particularly for the $^2$H($n,\gamma$)$^3$H process. It remains an interesting question whether N$^3$LO corrections will resolve the present discrepancies between theory and experiment.
Currents up to N$^2$LO {#sec:tree}
======================
The currents up to N$^2$LO are illustrated by the diagrams in Fig. \[fig:fig1\], where we show only one of the possible time orderings. The LO and NLO currents, panels a) and b)-c), are well known [@Park96] and will not be given here. At N$^2$LO there is a contribution originating from $(Q/M)^2$ corrections to the LO (one-body) current ($Q$ is the low momentum scale and $M \simeq 1$ GeV is the typical hadronic mass scale); it reads [@Pastore08]: $$\begin{aligned}
{\bf j}^{\rm N^2LO}_{\rm RC}=&-&\frac{e}{8 \, m_N^3}
e_{N,1}\, \Bigg[
2\, \left( K_1^2 +q^2/4 \right)
\left( 2\, {\bf K}_1+i\, {\bm \sigma}_1\times {\bf q } \right)
+ {\bf K}_1\cdot {\bf q}\,
\left({\bf q} +2i\, {\bm \sigma}_1\times {\bf K }_1 \right)\Bigg]
\nonumber \\
&-& \frac{i\,e}{8 \, m_N^3}
\kappa_{N,1}\, \Bigg[ {\bf K}_1\cdot {\bf q}\,
\left( 4\, {\bm \sigma}_1\times {\bf K}_1 -i\, {\bf q}\right)
- \left( 2\, i\, {\bf K}_1 -{\bm \sigma}_1\times {\bf q} \right)\, q^2/2 \nonumber \\
&& \qquad\qquad \qquad +2\, \left({\bf K}_1\times {\bf q}\right) \, {\bm \sigma}_1\cdot {\bf K}_1
\Bigg] + 1 \rightleftharpoons 2 \ ,
\label{eq:j1rc}\end{aligned}$$ where the momenta ${\bf k}_i$ and ${\bf K}_i$ are defined as ${\bf k}_i={\bf p}_i^\prime-{\bf p}_i$ and ${\bf K}_i=({\bf p}_i^\prime+{\bf p}_i)/2$, ${\bf q}$ is the photon momentum, and $$e_{N,i}=(1+\tau_{i,z})/2 \ ,\qquad \kappa_{N,i}=(\kappa_S+\kappa_V \tau_{i,z})/2 \ ,
\qquad \mu_{N,i}=e_{N,i}+\mu_{N,i} \ ,$$ with $\kappa_S$=–0.12 n.m and $\kappa_V$=3.706 n.m..
The N$^2$LO contributions, represented by diagrams e)-f) and involving $\Delta$-isobar excitation, read in the static limit [@Pastore08]: $$\begin{aligned}
{\bf j}^{\rm N^2LO}_{\Delta}&=& i\,\frac{ e\, C_\Delta}{9 \, m_N\, \Delta}
\frac{{\bm \sigma}_2 \cdot {\bf k}_2}{k_2^2+m_\pi^2}
\Big[ 4\, \tau_{2,z}\, {\bf k}_2-({\bm \tau}_1\times {\bm \tau}_2)_z
\,{\bm \sigma}_1 \times{\bf k}_2\Big]
\times {\bf q} + 1 \rightleftharpoons 2\ , \label{eq:jd}\\
{\bf j}^{\rm N^2LO}_{\Delta_c}&=& -i\,\frac{ e\, C_{\Delta_c}}{9 \, m_N\, \Delta}
\Big[ 4\, \tau_{2,z}\, {\bm \sigma}_2-
({\bm \tau}_1\times {\bm \tau}_2)_z\,
{\bm \sigma}_1 \times{\bm \sigma}_2 \Big]\times {\bf q} + 1 \rightleftharpoons 2\ ,
\label{eq:jdc}\end{aligned}$$ where $g_A$ and $F_\pi$ are the nucleon axial coupling and pion decay constants, $C_\Delta=\mu^*\, g_A\, h_A/F_\pi^2$ and $C_{\Delta_c}=\mu^*\, D_T$, $h_A$ and $\mu^*$ are the $N$-$\Delta$ transition axial coupling constant and magnetic moment, and $\Delta$ denotes the mass difference $m_\Delta -m_N$.
The configuration-space versions of the NLO and N$^2$LO operators have $1/r^2$ and $1/r^3$ singularities ($r$ is the interparticle separation), which need to be regularized in order to avoid divergencies in the matrix elements of these operators between nuclear wave functions. We adopt a simple regularization procedure, [*i.e.*]{} a momentum-space cutoff. While its precise functional form is arbitrary, the choice made here of a Gaussian cutoff function, $C_\Lambda(p) = {\rm e}^{-(p/\Lambda)^2}$, with the parameter $\Lambda \leq M$, is merely dictated by convenience, since it leads to analytical expressions for the Fourier transforms [@Pastore08]. It is expected that this arbitrariness be of little relevance, since the dependence of theoretical predictions on variations in the cutoff is (or should be, see next section) largely removed by a renormalization of the theory free parameters, which are fixed by reproducing a given set of observables.
Results and Conclusions
=======================
At N$^2$LO, the only isoscalar terms are from the (one-body) LO and N$^2$LO-RC operators, which are independent of the cutoff $\Lambda$.
\[tb:tab1\]
In Tables \[tb:tab1\] we list their contributions to the deuteron magnetic moment and isoscalar combination of the $^3$He and $^3$H magnetic moments. The N$^2$LO-RC correction is (in magnitude) about 1% of the LO contribution but of opposite sign, so that its inclusion increases the difference between the measured and calculated values. As a result the experimental deuteron and trinucleon isoscalar magnetic moments are underpredicted by theory at the (1.6–2.1)% and (3.0–4.7)% levels, respectively, depending on whether the CDB and CDB/UIX$^*$ or AV18 and AV18/UIX combinations are adopted in the $A$=2 and $A$=3 calculations. We note that a recent calculation of these same observables [@Song07], based on variational Monte Carlo (VMC) wave functions corresponding to the AV18/UIX Hamiltonian, finds the magnitude of the N$^2$LO-RC correction somewhat smaller in $A$=2 (–0.0069 n.m.) and significantly larger in $A$=3 (–0.012 n.m.) than obtained here. However, the expression for the magnetic dipole operator is different from that resulting from Eq. (\[eq:j1rc\]), and the VMC wave functions are less accurate than the hyperspherical harmonics (HH) wave functions [@Kievsky08] used in this work.
In the isovector sector, the NLO current involves the combination $g_A/F_\pi$, for which we adopt the value $(m_\pi\, g_A/F_\pi)^2/(4\pi)$=0.075 as inferred from an analysis of nucleon-nucleon elastic scattering data [@Stoks94]. In the N$^2$LO currents, the parameters $C_\Delta$ and $C_{\Delta_c}$ are determined by reproducing the $n$-$p$ radiative capture cross section and $^3$He/$^3$H isovector magnetic moment, respectively. We note that the $\Delta_c$ current in Eq. (\[eq:jdc\]) gives no contribution in the $n$-$p$ capture. It contributes in three-body matrix elements only because in the configuration-space version of this operator the $\delta$-function is replaced by a finite width Gaussian [@Pastore08]. It is for this reason that one can interpret the contributions resulting from the $\Delta_c$ current as representing corrections beyond N$^2$LO.
Results for the isovector combination $\mu_V$ of the trinucleon magnetic moments (without inclusion of the $\Delta_c$ current contribution) are presented in Table \[tb:tab5\]. The NLO contribution calculated
\[tb:tab5\]
in Ref. [@Song07] with VMC wave functions and a cutoff of 600 MeV is –0.205 n.m., which is 4% larger than obtained here. Of course, the parameter $C_{\Delta_c}$ is adjusted to reproduce, as function of $\Lambda$, the $\mu_V$ experimental value.
Predictions for the cross section $\sigma_T$ and photon circular polarization parameter $R_c$ measured in the reaction $^2$H($n,\gamma$)$^3$H (with unpolarized and polarized neutrons, respectively) are presented in Table \[tb:tab9\]. At thermal energies this process proceeds through S-wave capture predominantly via magnetic dipole transitions from the initial doublet $J$=1/2 and quartet $J$=3/2 $n$-$d$ scattering states. In addition, there is a small contribution due to an electric quadrupole transition from the initial quartet state [@Viviani96].
\[tb:tab9\]
At N$^2$LO the cross section is underpredicted by theory by (11–38)% as the cutoff is increased from 500 MeV to 800 MeV. This rather drastic cutoff dependence is mostly due to the contribution of the N$^2$LO-$\Delta_c$ current. Indeed removing it leads to a much weaker variation of the cross section—roughly $\pm 5$% about the value obtained with $\Lambda=600$ MeV (next to last row of Table \[tb:tab9\]). It will be interesting to see to what extent, if any, loop corrections at N$^3$LO will improve the present predictions, and in particular reduce the cutoff dependence.
The photon polarization parameter is very sensitive to contributions of NLO and N$^2$LO currents, which produce more than a sixfold increase, in absolute value, of the LO result, and bring it into much closer agreement with the measured value. All results listed in Table \[tb:tab9\] for $R_c$ (and $\sigma_T$) include the small $e_{44}$ RME, although it only has a significant effect for the LO prediction ($R_c$=–0.060 versus –0.072 depending on whether $e_{44}$ is retained or not).
We conclude by summarizing our results. Up to N$^2$LO, the only isoscalar terms are those generated in a non-relativistic expansion of the one-body current, and provide a (cutoff-independent) 1% correction—relative to LO—to the deuteron and isoscalar combination of the trinucleon magnetic moments. This correction is of opposite sign to the LO contribution, and therefore increases the underprediction of the corresponding experimental values from $(0.9\pm 0.3)$% for the deuteron and $(2.7 \pm 0.9)$% for the trinucleons at LO to, respectively, $(1.9 \pm 0.3)$% and $(3.8 \pm 0.8)$% at N$^2$LO. The spread reflects differences in the short-range behavior of the AV18 and CDB potentials, in particular the weaker tensor components of the latter relative to the former in this range.
At NLO, isovector terms arise from the pion seagull and in-flight contributions, while at N$^2$LO, in addition to the relativistic corrections mentioned above, isovector terms due to $\Delta$-isobar excitation are also obtained. The parameters $C_\Delta$ and $C_{\Delta_c}$ of the N$^2$LO two-body $\Delta$-excitation currents have been determined, as functions of the cutoff $\Lambda$ and for the Hamiltonian model of interest, by reproducing the cross section for the $n$-$p$ radiative capture at thermal neutron energies and the isovector combination of the trinucleon magnetic moments. This current has then been used to make predictions—with the AV18/UIX model only, since HH continuum wave functions are not yet available for the CDB/UIX$^*$ model—for the cross section $\sigma_T$ and photon circular polarization parameter $R_c$ measured in the capture of, respectively, unpolarized and polarized neutrons by deuterons. The experimental $\sigma_T$ ($|R_c|$) is found to be underestimated by 11% (overestimated by 4%) for $\Lambda$=500 MeV and 38% (underestimated by 21%) for $\Lambda$=800 MeV.
The results display a significant cutoff dependence, particularly so for the N$^2$LO contributions associated with $\Delta$ isobar degrees of freedom. Indeed these contributions are much larger than those at NLO. This is partly due to the fact that the two NLO (pion seagull and in-flight) terms interfere destructively. For example, the seagull (in-flight) contributions to doublet $m_{22}$ and quartet $m_{44}$ M1 matrix elements, in units of fm$^{3/2}$ and for $\Lambda$=500 MeV, are respectively –9.1 (+6.5) and –0.8 (+0.6). As a result $\sigma_T=0.425$ mb and $R_c=-0.425$ at LO+NLO (seagull only), which should be compared to $\sigma_T=0.272$ mb and $R_c=-0.218$ at LO+NLO (seagull+in-flight) from the second row of Table \[tb:tab9\]. The relatively large $\Delta$-excitation contributions also point to the need for including loop corrections at N$^3$LO, which these N$^2$LO currents, because of the procedure adopted here to determine their strength, are implicitly making up for.
The next stage in the research program we have undertaken is to incorporate the N$^3$LO operators derived in Ref. [@Pastore08] into the calculations of the captures and magnetic moments involving light nuclei (with mass number $A \leq 8$), and indeed to extend these calculations to also include $p$-$d$ capture at energies up to a few MeV’s, and possibly four-nucleon processes, in particular $^3$He($n,\gamma$)$^4$He at thermal energies. Of course, at N$^3$LO three-body currents also occur, and will need to be derived. Work along these lines is being pursued vigorously.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank E. Epelbaum, L. Girlanda, A. Kievsky, L.E. Marcucci, and M. Viviani for discussions. The work of R.S. is supported by the U.S. Department of Energy, Office of Nuclear Physics, under contract DE-AC05-06OR23177, while that of J.L.G. by NSF grant PHY-0555559. The calculations were made possible by grants of computing time from the National Energy Research Supercomputer Center.
[0]{} S. Pastore, R. Schiavilla, and J. Goity, in preparation. S. Weinberg, Phys. Lett. [**B251**]{}, 288 (1990); Nucl. Phys. [**B363**]{}, 3 (1991); Phys. Lett. [**B295**]{}, 114 (1992). U. van Kolck, Phys. Rev. C [**49**]{}, 2932 (1994); C. Ordónez, L. Ray, and U. van Kolck, Phys. Rev. C [**53**]{}, 2086 (1996). E. Epelbaum, W. Glöckle, and U.-G. Meissner, Nucl. Phys. [**A637**]{}, 107 (1998); Nucl. Phys. [**A747**]{}, 362 (2005). T.-S. Park, D.-P. Min, and M. Rho, Nucl. Phys. [**A596**]{}, 515 (1996). R.B. Wiringa, V.G.J. Stoks, and R. Schiavilla, Phys. Rev. C [**51**]{}, 38 (1995). R. Machleidt, Phys. Rev. C [**63**]{}, 024001 (2001). B.S. Pudliner [*et al.*]{}, Phys. Rev. C [**56**]{}, 1720 (1997). M. Viviani [*et al.*]{}. Phys. Rev. Lett. [**99**]{}, 112002 (2007). Y.-H. Song, R. Lazauskas, T.-S. Park, and D.-P. Min, Phys. Lett. [**B656**]{}, 174 (2007). A. Kievsky, S. Rosati, M. Viviani, L.E. Marcucci, and L. Girlanda, J. Phys. G [**35**]{}, 063101 (2008). V.G.J. Stoks, R.A.M. Klomp, C.P.F. Terheggen, and J.J. deSwart, Phys. Rev. C [**49**]{}, 2950 (1994). M. Viviani, R. Schiavilla, and A. Kievsky, Phys. Rev. C [**54**]{}, 534 (1996). E.T. Jurney, P.J. Bendt, and J.C. Browne, Phys. Rev. C [**25**]{}, 2810 (1982). M.W. Konijnenberg [*et al.*]{}, Phys. Lett. [**B205**]{}, 215 (1988).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we formulate a conjecture which partially generalizes the Gross-Kohnen-Zagier theorem to higher weight modular forms. For $f\in {S_{2k}(N)}$ satisfying certain conditions, we construct a map from the Heegner points of level $N$ to a complex torus, ${\mathbb{C}}/L_f$, defined by $f$. We define higher weight analogues of Heegner divisors on ${\mathbb{C}}/L_f$. We conjecture they all lie on a line, and their positions are given by the coefficients of a certain Jacobi form corresponding to $f$. In weight $2$, our map is the modular parametrization map (restricted to Heegner points), and our conjectures are implied by Gross-Kohnen-Zagier. For any weight, we expect that our map is the Abel-Jacobi map on a certain modular variety, and so our conjectures are consistent with the conjectures of Beilinson-Bloch. We have verified our map is the Abel-Jacobi for weight $4$. We provide numerical evidence to support our conjecture for a variety of examples.'
address: 'UT Austin, Department of Mathematics C1200, Austin, TX 78712 '
author:
- Kimberly Hopkins
bibliography:
- 'mybibliography.bib'
title: Higher Weight Heegner Points
---
Introduction {#S:intro}
=============
For integers $N, k\geq 1$, let $S_{2k}(N)$ denote the cusp forms of weight $2k$ on the congruence group ${\Gamma_0(N)}$. Let ${X_0(N)}$ be the usual modular curve and $J_0(N)$ its Jacobian. By $D$ we will always mean a negative fundamental discriminant which is a square modulo $4N$. For each $D$, one can construct a Heegner divisor $y_D$ in $J_0(N)$ and defined over ${\mathbb{Q}}$. Suppose $f\in S_2(N)$ is any normalized newform whose sign in the functional equation of $L(f,s)$ is $-1$. Then the celebrated theorem of Gross, Kohnen, and Zagier [@GZK Theorem C] says that, as $D$ varies, the $f$-eigencomponents of the Heegner divisors $y_D$ all ‘lie on a line[^1]’ in the quotient $J_0(N)_f$. Furthermore it says their positions on this line are given by the coefficients of a certain Jacobi form. In particular when $N$ is prime, the positions are the coefficients of a half-integer weight modular form in Shimura correspondence with $f$.
Now suppose $f\in{S_{2k}(N)}$ is a normalized newform of weight $2k$ and level $N$. In addition, assume the coefficients in its Fourier series are rational, and the sign in the functional equation of $L(f,s)$ is $-1$. Let ${\mathcal{H}_N}/{\Gamma_0(N)}\subset {X_0(N)}$ denote the Heegner points of level $N$. In this paper we construct a map, $$\alpha: {\mathcal{H}_N}/{\Gamma_0(N)}\rightarrow {\mathbb{C}}/L_f,$$ where ${\mathbb{C}}/L_f$ is a complex torus defined by the periods of $f$. Let $h(D)$ denote the class number of the imaginary quadratic field of discriminant $D$. For each $D$ and fixed choice of its square root $(\bmod2N)$, we get precisely $h(D)$ distinct representatives $\tau_1, \dots, \tau_{h(D)}$ of ${\mathcal{H}_N}/{\Gamma_0(N)}$. Define $(\mathcal{Y}_D)_f = \alpha(\tau_1) + \cdots + \alpha(\tau_{h(D)})$ and define $(y_D)_f = (\mathcal{Y}_D)_f + \overline{(\mathcal{Y}_D)_f}$ in ${\mathbb{C}}/L_f$. When $k=1$, $\alpha$ is the usual modular parametrization map restricted to Heegner points, and $(y_D)_f$ is equal to the $f$-eigencomponent of $y_D$ in $J_0(N)$ as described in the first paragraph. For $k\geq 1$ we formulate conjectures similar to Gross-Kohnen-Zagier. We predict the $(y_D)_f$ all lie in a line in ${\mathbb{C}}/L_f$, that is, there exists a point $y_f \in {\mathbb{C}}/L_f$ such that $$(y_D)_f = m_D y_f,$$ up to torsion, with $m_D \in {\mathbb{Z}}$. Furthermore we predict the positions $m_D$ on the line are coefficients of a certain Jacobi form corresponding to $f$. In the case when $N$ is prime and $k$ is odd, the $m_D$ should be the coefficients of a weight $(k+1/2)$ modular form in Shimura correspondence with $f$.
We expect our map is equivalent to the Abel Jacobi map on Kuga-Sato varieties in the following sense. Let $Y=Y^k$ be the Kuga-Sato variety associated to weight $2k$ forms on ${\Gamma_0(N)}$. (See [@Zh p.$117$] for details.) This is a smooth projective variety over ${\mathbb{Q}}$ of dimension $2k-1$. Set $\mathcal{Z}^k(Y)_\text{hom}$ to be the nullhomologous codimension $k$ algebraic cycles, and $\text{CH}^k(Y)_{\text{hom}}$ the group of $\mathcal{Z}^k(Y)_\text{hom}$ modulo rational equivalence. Let $\Phi^k$ be the usual $k$-th Abel-Jacobi map, $$\Phi^k: \text{CH}^k(Y)_{\text{hom}} \rightarrow J^k(Y),$$ where $J^k(Y)$ is the $k$-th intermediate Jacobian of $Y$. Given any normalized newform $f = \sum_{n\geq 1} a_n q^n \in {S_{2k}(N)}$ with rational coefficients, there exists an $f$-isotypical component $J_f^k(Y)$ of $J^k(Y)$, and thus an induced map, $$\xymatrix{
\text{CH}^k(Y)_{\text{hom}} \ar[r]^{\Phi^k} \ar@{-->}[rd]_{\Phi^k_f} & J^k(Y) \ar[d]\\
&J_f^k(Y)}$$ Our expectation is that the image of $\Phi_f^k$ on classes of CM cycles in $\text{CH}^k(Y)_{\text{hom}}$ is equal (up to a constant) to the image of our map $\alpha$ on Heegner points in ${X_0(N)}$. If we assume this is the case, then our conjectures are consistent with the conjectures of Beilinson and Bloch. In this setting they predict $$\text{rank}_{\mathbb{Z}}\: \text{CH}^k(Y_F)_{\text{hom}} = \text{ord}_{s=k} L_F(H^{2k-1}(Y),s).$$ If we assume $\text{ord}_{s=k}L(f,s)=1$, then a refinement of their conjecture predicts the image of $\Phi_f^k$ on CM divisors in $Y_{\mathbb{Q}}$ should have rank at most $1$ in $J_f^k(Y)$.
We have verified the equivalence of $\alpha$ and $\Phi^2_f$ in the case of weight $4$. For this we used an explicit description of $\Phi^2_f$ on CM cycles given by Schoen in [@Sc1]. In fact, in [@Sc2] Schoen uses this map to investigate a consequence of Beilinson-Bloch similar to the one described above. For a specific $Y=Y^4$ and $f$ he computes $\Phi_f$ on certain CM divisors in $Y$ defined over the quadratic number field ${\mathbb{Q}}(i)$. From this he finds numerical evidence that the images lie on a line and their positions are given by a certain weight $5/2$ form corresponding to $f$.
The sections of this paper are divided as follows. In Section \[S:mymap\] we describe our map and its lattice of periods. In Section \[S:conjectures\] we give explicit statements of our conjectures. In Section \[S:algorithm\] we describe the algorithm we created to numerically verify the conjectures in a variety of examples. Note our algorithm could be applied to compute coefficients of half-integer weight modular forms. In sections \[S:examples\] and \[S:moreexamples\] we compute some examples and use them to verify our conjectures in two different ways.
Higher Weight Heegner Points {#S:mymap}
============================
Let ${\mathfrak{h}}$ denote the upper half-plane. Suppose $f$ is a normalized newform in ${S_{2k}(N)}$ having a Fourier expansion of the form, $$f(\tau) = \sum_{n=1}^\infty a_n q^n, \qquad q = \exp(2\pi i \tau),\: \tau \in {\mathfrak{h}}$$ with $a_n \in {\mathbb{Q}}$.
Recall the $L$-function of $f$ is defined by the Dirichlet series, $$L(f,s) = \sum_{n=1}^\infty \frac{a_n}{n^s}, \qquad \Rea(s)>k+1/2,$$ and has an analytic continuation to all of ${\mathbb{C}}$. Moreover the function $\Lambda(f,s) = N^{s/2} (2\pi)^{-s} \Gamma(s) L(f,s)$ satisfies the functional equation, $$\Lambda(f,s) = {\varepsilon}\Lambda(f, 2k-s),$$ where ${\varepsilon}= \pm 1$ is the sign of the functional equation of $L(f,s)$.
For each prime divisor $p$ of $N$, let $q = p^\ell$, $\ell \in \mathbb{N}$ such that $\gcd(q,N/q)=1$ and set ${\omega}_q = \begin{smat}
{qx_0}{1}{Ny_0}{q}
\end{smat}$, for some $x_0,y_0\in {\mathbb{Z}}$, with $qx_0 - (N/q)y_0 =1$ . Define ${\Gamma^\ast_0(N)}$ to be the group generated by ${\Gamma_0(N)}$ and each ${\omega}_q$. Let $S$ be a set of generators for ${\Gamma^\ast_0(N)}$. Define the period integrals of $f$ for the set $S$ by, $$\mathcal{P} = \bigg\{ (2\pi i)^{k} {\int\limits_{{i}\infty}^{{\gamma}({i}\infty)} f(z)}z^m dz \::\: m\in \{0,\dots, 2k-2\},\: {\gamma}\in S \bigg\}\subseteq {\mathbb{C}}.$$ These are sometimes referred to as Shimura integrals. It is straightforward to see that every integral of the form, $$(2\pi i)^{k} {\int\limits_{{i}\infty}^{{\gamma}({i}\infty)} f(z)}z^m dz, \qquad {\gamma}\in {\Gamma^\ast_0(N)},\: 0\leq m \leq 2k-2$$ is in an integral linear combination of elements in $\mathcal{P}$. (See [@Sh Section $8.2$], for example). In fact, the ${\mathbb{Z}}$-module generated by $\mathcal{P}$ forms a lattice,
$L:=\text{Span}_{\mathbb{Z}}(\mathcal{P})$ is a lattice in ${\mathbb{C}}$.
By theorems of Razar [@Ra Theorem $4$] and Šokurov [@So Lemma $5.6$], the set $\mathcal{P} $ is contained in some lattice. Hence $L$ is of rank $\leq 2$. To show its rank is $2$, it suffices to show there exist nonzero complex numbers $u^+, u^- \in L$ with $u^+ \in {\mathbb R}$ and $u^-\in i{\mathbb R}$.
Suppose $m$ is a prime not dividing $N$, and $\chi$ a primitive Dirichlet character modulo $m$. Define $(f\otimes \chi) := \sum_{n\geq 1} \chi(n) a_n q^n$, and $L(f\otimes \chi, s)$ to be its Dirichlet series. Let $\Lambda(f\otimes \chi,s) = (2\pi)^{-s} (Nm^2)^{s/2} \Gamma(s) L(f\otimes \chi, s)$. Then for $\Rea(s)>k+1/2$, we have $$\label{E:lamfchi}
i^s (Nm^2)^{-s/2} \Lambda(f\otimes \chi, s) = \int\limits_0^{i\infty} (f\otimes\chi) (z) z^s \frac{dz}{z}.$$
Let $g(\chi)$ denote the Gauss sum associated to $\chi$. Then an expression for $\chi$ in terms of the additive characters is given by, $$\chi(n) = m^{-1} g(\chi) \sum_{u\bmod m} \bar{\chi}(-u) e^{2\pi i nu/m}.$$ So $$(f\otimes \chi)(\tau) = m^{-1} g(\chi) \sum_{u\bmod m} \bar{\chi}(-u) f(z + u/m).$$ Substituting this into gives $$i^s (Nm^2)^{-s/2} \Lambda(f\otimes \chi, s) = m^{-1} g(\chi) \sum_{u\bmod m} \bar{\chi}(-u) \int\limits_0^{i\infty} f(z+u/m) z^s \frac{dz}{z},$$ and replacing $z$ by $z-u/m$ and rearranging implies $$i^{-s} g(\chi)^{-1} N^{-s/2} \Lambda(f\otimes \chi, s) = (-1)^{s-1}\sum_{u\bmod m} \bar{\chi}(-u) \int\limits_{i\infty}^{u/m} f(z) (mz-u)^{s-1} dz.$$
Now let $s=2k-1$ in the above equation, and multiply both sides by $(2\pi i)^k$. In addition suppose $\chi$ is a quadratic Dirichlet character modulo $m$. If $m\equiv 3\bmod 4$, then $g(\chi) = i\sqrt{m}$, and if $m\equiv 1\bmod 4$ then $g(\chi) = \sqrt{m}$. Hence since $\Lambda(f\otimes \chi,2k-1)$ is real-valued and nonzero, the right hand side of this equation is either purely real or purely imaginary depending on the choice of $m$. Then this proves the lemma since the right hand side is in $L$ for any $m$.
Let $D<0$ be a fundamental discriminant, and assume $D$ is a square modulo $4N$. Fix a residue class $r\bmod 2N$ satisfying $D\equiv r^2\bmod 4N$. Then $${\mathcal{Q}_N^D(r)}:= \big\{ [A,B,C] : A>0, B, C \in {\mathbb{Z}}, D= B^2-4AC, A \equiv 0\bmod N, B\equiv r\bmod 2N \big\}.$$ corresponds to a subset of the positive definite binary quadratic forms of discriminant $D$. We define ${\mathcal{H}_N^D}(r)$ to be the roots in ${\mathfrak{h}}$ of ${\mathcal{Q}_N^D(r)}$, $${\mathcal{H}_N^D}(r) := \bigg\{ \tau = \frac{-B+\sqrt{D}}{2A} : [A,B, C] \in {\mathcal{Q}_N^D(r)}, C= \frac{|D|+B^2}{4A} \bigg\}.$$ ${\Gamma_0(N)}$ preserves ${\mathcal{H}_N^D}(r)$, and the classes of ${\mathcal{H}_N^D}(r)/{\Gamma_0(N)}$ are in bijection with the classes of reduced binary quadratic forms of discriminant $D$. We will call ${\mathcal{H}_N^D}(r)/{\Gamma_0(N)}$ the set of Heegner points of level $N$, discriminant $D$, and root $r$. Define ${\mathcal{H}_N}$ to be the union of ${\mathcal{H}_N^D}(r)$ over all $D, r$, and so ${\mathcal{H}_N}/{\Gamma_0(N)}$ are the Heegner points of level $N$.
For each $\tau = \frac{-B+\sqrt{D}}{2A} \in {\mathcal{H}_N^D}(r)$, set $Q_\tau(z) := Az^2 +Bz+C$. We now define a function $\alpha =\alpha_f: {\mathcal{H}_N}\rightarrow {\mathbb{C}}$ by $$\alpha(\tau) := (2\pi i )^{k} \int_{i\infty}^\tau f(z) Q_\tau(z)^{k-1}dz.$$
\[L:welldef\] The map $\alpha$ induces a well-defined map (which we will also denote by $\alpha$), $$\alpha: {\mathcal{H}_N}/{\Gamma_0(N)}\rightarrow {\mathbb{C}}/L.$$
For any $\tau \in {\mathcal{H}_N}$ of discriminant $D$ and ${\gamma}\in {\Gamma_0(N)}$, we will show $$\alpha({\gamma}\tau) - \alpha(\tau) = \:(2\pi i)^{k}\cdot\!\!\!{\int\limits_{{i}\infty}^{{\gamma}({i}\infty)} f(z)}Q_{{\gamma}\tau}(z)^{k-1} dz.$$ Since $Q_{{\gamma}\tau}(z)$ has integer coefficients, this will imply $\alpha({\gamma}\tau) - \alpha(\tau) \in L$ for all ${\gamma}\in {\Gamma_0(N)}$.
Let ${\gamma}= \begin{smat}
{a}{b}{c}{d}
\end{smat} \in {\Gamma_0(N)}$. Then $$\begin{aligned}
& \alpha({\gamma}\tau) - (2\pi i)^{k}\cdot\!\!\! {\int\limits_{{i}\infty}^{{\gamma}({i}\infty)} f(z)}\: Q_{{\gamma}\tau}(z)^{k-1} dz \\
&= (2\pi i)^{k}\cdot\!\!\! \int\limits_{{\gamma}(i\infty)}^{{\gamma}\tau} f( z) Q_{{\gamma}\tau}(z)^{k-1} dz\\
&= (2\pi i)^{k}\cdot\!\!\! \int\limits_{i\infty}^{\tau} f({\gamma}z) Q_{{\gamma}\tau}({\gamma}z)^{k-1} d({\gamma}z)\\
&= \alpha(\tau),\end{aligned}$$ where in the last equality we used $f({\gamma}z) = (cz+d)^{2k} f(z)$, $Q_{{\gamma}\tau}(z) = (-cz+a)^2 Q_\tau(z)$, and $d({\gamma}z) = (cz+d)^{-2} dz$.
Conjectures {#S:conjectures}
===========
Let $\{\tau_1, \dots, \tau_{h(D)}\} \in {\mathcal{H}_N^D}(r)$ be any set of distinct class representatives of ${\mathcal{H}_N^D}(r)/{\Gamma_0(N)}$. Define $$P_{D,r} := \sum_{i=1}^{h(D)} \tau_i \in \text{Div}({X_0(N)}),$$ where $\text{Div}({X_0(N)})$ denotes the group of divisors on ${X_0(N)}$. If $D=-3$ (resp. $D=-4$ ), scale $P_{D,r}$ by $1/3$ (resp. $1/2$). Extend $\alpha$ to $P_D$ by linearity and define $$(y_{D,r})_f = \alpha(P_{D,r}) + \overline{\alpha(P_{D,r})} \in {\mathbb{C}}/L.$$ Here, bar denotes complex conjugation in ${\mathbb{C}}$. We write $y_{D,r}$ or $y_D$ for $(y_{D,r})_f$, and $P_D$ for $P_{D,r}$ when the context of $f$, $r$ is clear.
By the actions of complex conjugation and Atkin-Lehner on ${\mathcal{H}_N}$, we have $$\overline{\alpha(P_{D,r})} = -{\varepsilon}\alpha(P_{D,r}),$$ where ${\varepsilon}$ is the sign of the functional equation of $L(f,s)$. Thus if ${\varepsilon}=+1$, then $y_{D,r}$ are in $L$ for all $D,r$. This is, in some sense, the trivial case. Hence we restrict our attention to the case when ${\varepsilon}=-1$.
Conjectures \[C:conj1\] and \[C:conj2\] give a partial generalization of the Gross-Kohnen-Zagier theorem to higher weights.
\[C:conj1\] Let $f = \sum_{n\geq 1} a_n q^n \in {S_{2k}(N)}$ be a normalized newform with rational coefficients, and assume ${\varepsilon}=-1$ and $L'(f,k) \neq 0$. Then for all fundamental discriminants $D<0$ and $r\bmod 2N$ with $D\equiv r^2 \bmod 4N$, there exist integers $m_{D,r}$ such that $$ty_{D,r} = m_{D,r} y_f \qquad \text{in}\: \: {\mathbb{C}}/L,$$ where $y_f \in {\mathbb{C}}/L$ and $t\in {\mathbb{Z}}$ are both nonzero and independent of $D$ and $r$.
Equivalently we could say $y_{D,r} = m_{D,r} y_f$ up to a $t$-torsion element in ${\mathbb{C}}/L$.
To state the second conjecture we will need to use Jacobi forms. (See [@Ei] for background). Let $J_{2k,N}$ denote the set of all Jacobi forms of weight $2k$ and index $N$. Then such a $\phi \in J_{2k,N}$ is a function $\phi: {\mathfrak{h}}\times {\mathbb{C}}\rightarrow {\mathbb{C}}$, which satisfies the transformation law $$\phi \bigg( \frac{a\tau+b}{c\tau +d}, \frac{z}{c\tau +d} \bigg) = (c\tau+d)^{2k} e^{2\pi i N \frac{cz^2}{c\tau+d}} \phi(\tau, z),$$ for all $\begin{smat}
{a}{b}{c}{d}
\end{smat} \in { SL_2({\mathbb{Z}})}$, and has a Fourier expansion of the form $$\label{E:jacfour}
\phi(\tau, z) = \sum_{\begin{subarray}{1} n, r \in {\mathbb{Z}}\\
r^2 \leq 4Nn \end{subarray} } c(n,r) q^n \zeta^r, \qquad q= e^{2\pi i \tau}, \:\zeta = e^{2\pi i z}.$$ The coefficient $c(n,r)$ depends only on $r^2 - 4Nn$ and on the class $r\bmod 2N$.
Suppose $f \in {S_{2k}(N)}$ is a normalized newform with ${\varepsilon}=-1$. Then by [@SkZa], there exists a non-zero Jacobi cusp form $\phi_f \in J_{k+1, N}$ which is unique up to scalar multiple and has the same eigenvalues as $f$ under the Hecke operators $T_m$ for $m, N$ coprime. We predict that the coefficients of $\phi_f$ are related to the $m_D$ from above in the following way,
\[C:conj2\] Let $f = \sum_{n\geq 1} a_n q^n \in {S_{2k}(N)}$ be a normalized newform with rational coefficients, and assume ${\varepsilon}=-1$ and $L'(f,k) \neq 0$. Assume Conjecture \[C:conj1\]. Then $$m_{D,r} = c(n,r)$$ where $n = \frac{|D|+r^2}{4N}$ and $c(n,r)$ is the $(n,r)$-th coefficient of the Jacobi form $ \phi_f \in J_{k+1,N}$.
When $k=2$, the points $(y_{D,r})_f$ and $y_f$ are the same as those defined in [@GZK], and both of our conjectures are implied by Theorem C of their paper. (Actually their theorem is only for $D$ coprime to $2N$ but they say the result remains ‘doubtless true’ with this hypothesis removed. See [@Ha] and [@Bo] for more details.) Particular to weight $2$ is the fact that ${\mathbb{C}}/L$ is defined over ${\mathbb{Q}}$ and so $y_D$ is a rational point on the elliptic curve $E_f \simeq {\mathbb{C}}/L$. In contrast, we should stress that for weight $k>2$, the elliptic curve $E \simeq {\mathbb{C}}/L$ is not expected to be defined over any number field. For instance, the $j$-invariants for our examples all appear to be transcendental over ${\mathbb{Q}}$.
For $N=1$ or a prime, and $k$ odd we can state Conjecture \[C:conj2\] in terms of modular forms of half-integer weight. Specifically, let $\phi\in J_{k+1}(N)$ be a Jacobi form with a Fourier expansion as in , and set $$g(\tau) = \sum_{M=0}^\infty c(M) q^M, \qquad q=e^{2\pi i \tau}$$ where $c(M)$ is defined by, $$c(M) := \left\{ \begin{array}{ll} c\bigg( \frac{M+r^2}{4N}, r\bigg) & \hbox{if $M\equiv -r^2 \bmod 4N$ for any $r\in {\mathbb{Z}}$;} \\ 0, & \hbox{otherwise.} \end{array} \right.$$ This function is well-defined because $c(n,r)$ depends only on $r^2-4nN$ when $N=1$ or a prime, and $k$ is odd. Then by [@Ei p.$69$], $g $ is in $M_{k+1/2}(4N)$, the space of modular forms of weight $k+1/2$ and level $4N$. In addition, if $f\in {S_{2k}(N)}$ is a normalized newform with ${\varepsilon}=-1$, then the form $g$ defined by $\phi_f$ is in Shimura correspondence with $f$.
Algorithm {#S:algorithm}
=========
Let $f =\sum_{n\geq 1} a_nq^n \in {S_{2k}(N)}$ be a normalized newform with rational Fourier coefficients. The sign ${\varepsilon}$ of the functional equation of $L(f,s)$ can be computed with the identity, $$f\bigg(\frac{-1}{Nz}\bigg) = (-1)^{k} {\varepsilon}N^{k} z^{2k} f(z)$$ given by the action of the Fricke involution of level $N$ on $f$. We will only consider $f$ such that ${\varepsilon}=-1$ and $L'(f,k)\neq 0$.
The first step is to find a basis of our lattice $L$, which is the ${\mathbb{Z}}$-module generated by the periods $\mathcal{P}$ as described above. Suppose $p_1, p_2, p_3$ are three periods in $\mathcal{P}$. Since $L$ has rank $2$, these are linearly dependent over ${\mathbb{Z}}$, that is $$a_1 p_1 + a_2 p_2 + a_3 p_3=0, \quad\text{for some}\quad a_i \in {\mathbb{Z}}.$$ We may assume $\gcd(a_1, a_2, a_3)=1$. Let $d = \gcd(a_1, a_2)$, then there exist integers $x,y \in {\mathbb{Z}}$ such that $xa_1 + ya_2 =d$. Similarly $\gcd(d,a_3)=1$ so there exist integers $u,v \in {\mathbb{Z}}$ such that $ud+ va_3 = 1$. Define the matrix $M$ by, $$M = \left(
\begin{array}{ccc}
a_1 & a_2 & a_3 \\
-y & x & 0 \\
-va_1/d & -va_2/d & u \\
\end{array}
\right).$$ Observe $M \in GL_3({\mathbb{Z}})$ and $M\cdot {}^T(p_1,p_2,p_3) = {}^T(0, -yp_1+xp_2, -va_1p_1/d - va_2p_2/d + up_3)$. Hence $-yp_1+xp_2$ and $ -va_1p_1/d - va_2p_2/d + up_3$ are a basis for the ${\mathbb{Z}}$-module generated by $p_1, p_2, p_3$.
We would also like our basis elements to have small norm. Given a basis ${\omega}_1,{\omega}_2$ of a lattice, its norm form is a real bilinear quadratic form defined by the matrix, $$B = \left(
\begin{array}{cc}
2|{\omega}_1|^2 & 2\Rea({\omega}_1 \bar{{\omega}}_2) \\
2\Rea({\omega}_1 \bar{{\omega}}_2) & 2|{\omega}_2|^2 \\
\end{array}
\right).$$ Thus it is equivalent to a reduced form of the same discriminant, that is, there exists $U \in { SL_2({\mathbb{Z}})}$ such that $${}^T U BU = \left(
\begin{array}{cc}
2\alpha & \beta \\
\beta & 2\gamma \\
\end{array}
\right), \qquad \alpha, \beta, \gamma \in {\mathbb R},$$ with $|\beta|\leq \alpha \leq \gamma$ and $\beta\geq 0$ if either $|\beta| = \alpha$ or $\alpha=\gamma$. Hence $({\omega}_1', {\omega}_2') := ({\omega}_1,{\omega}_2)U$ is a ‘reduced’ basis. For a basis of all of $L$ we simply apply this process iteratively on the elements of $\mathcal{P}$.
In fact it is not hard to see that $L$ is a real lattice, that is, $\bar{L} = L$. Thus given a basis ${\omega}_1, {\omega}_2$ of $L$, we may assume ${\omega}_1\in i{\mathbb R}$, and therefore $\tau:= {\omega}_2/{\omega}_1$ has real part equal to either $0$ or $1/2$. This implies $\Rea(L) = \Rea({\omega}_2)$ which will help simplify our computations.
To actually compute the elements in $\mathcal{P}$ we need to split the path from $(i\infty)$ to ${\gamma}(i\infty)$ of integration at some point $\tau \in {\mathfrak{h}}$ which gives, $${\int\limits_{{i}\infty}^{{\gamma}({i}\infty)} f(z)}z^m dz = \int\limits_{{i}\infty}^{{\gamma}(\tau)} f(z) z^m dz - \int\limits_{{i}\infty}^{\tau} f(z) (az+b)^m (cz+d)^{2k-2-m} dz,$$ for ${\gamma}= \begin{smat}
{a}{b}{c}{d}
\end{smat} \in{\Gamma_0(N)}$. We choose $\tau$ to be a point at which $f$ has good convergence. To compute integrals of the form, $$\int\limits_{{i}\infty}^{\tau} f(z) z^m dz,$$ we use repeated integration by parts to get the formula $$\label{E:compzpwr}
\int\limits_{{i}\infty}^{\tau} \!\! f(z) z^m dz = m! \,(-1)^m \sum_{j=-1}^{m-1} \frac{(-1)^{j+1}}{(j+1)!} \tau^{j+1} f_{m-j}(\tau),$$ where $f_\ell(\tau)$ is defined to be the $\ell$-fold integral of $f$ evaluated at $\tau \in {\mathfrak{h}}$, that is, $$f_\ell(\tau) = \frac{1}{(2\pi i)^\ell} \sum_{n\geq1} \frac{a_n}{n^\ell} q^n, \qquad q = \exp(2\pi i \tau)$$ which is well-defined for any $0\leq \ell \leq 2k-1$.
The next task is to compute $\alpha(\tau)$ for $\tau \in {\mathcal{H}_N}$. We could do this using , but it is computationally faster to use the following identity for $\alpha$. Recall the modular differential operator, $${\partial}_m := \frac{1}{2\pi i} \frac{d}{dz} - \frac{m}{4\pi y},\qquad z=x+iy\in {\mathfrak{h}},$$ for any integer $m$. Define ${\partial}_{m}^\ell(f) := {\partial}_{m+2(\ell-1)}\circ\cdots\circ {\partial}_{m+2}\circ{\partial}_m(f)$ to be the composition of the $\ell$ operators ${\partial}_m, {\partial}_{m+2}, \dots, {\partial}_{m+2(\ell-1)}$. Then a straightforward combinatorial argument yields the following identity, whose proof we will omit,
\[L:altalph\] Let $\tau$ be a Heegner point of level $N$ and discriminant $D$. Then $$\alpha(\tau) = {\kappa_D}\cdot{\partial}_{-2k+2}^{k-1}\circ f_{2k-1}(\tau),$$ where ${\kappa_D}=(k-1)! \, (2\pi i)^{k} (2\pi \sqrt{|D|} )^{k-1}$ is a constant depending only on $D$ and $2k$.
A closed formula for ${\partial}_m^\ell$ (see [@Vi] for example) allows us to write $\alpha$ as $$\label{E:compal}
\alpha(\tau) = {\kappa_D}(2\pi i) \bigg( \frac{-y}{\pi}\bigg)^{k} \sum_{n\geq 1} p\bigg(\frac{2k}{2}, \frac{1}{4\pi y n} \bigg) a_n q^n,$$ where $p(m,x)$, is the polynomial, $$p(m,x) = \sum_{\ell=m}^{2m-1} \binom{m-1}{2m-1-\ell}\frac{(\ell -1 )!}{(m-1)!} x^{\ell}, \qquad m\in {\mathbb{Z}},\: x\in {\mathbb R}.$$ We compute $\alpha(\tau)$ using . Also notice that Lemma \[L:altalph\] perhaps provides further insight into why the map $ {\mathcal{H}_N}\rightarrow {\mathbb{C}}/L$ inducing $\alpha$ is invariant under ${\Gamma_0(N)}$. Loosely speaking, this is because integrating $f$ $(2k-1)$-times lowers its weight by $2(2k-1)$ and ${\partial}_{-2k+2}^{k-1}$ increases its weight by $2(k-1)$ to get something morally of weight $0$.
Given a set of Heegner point representatives of level $N$, discriminant $D$, and root $r$, we can use the above to compute $y_{D,r}$. Verifying the first conjecture for each $D, r$ then amounts to choosing a complex number $y_f$, and an integer $t$, both non-zero, and showing the linear dependence, $$\label{E:testeq}
\Rea(y_{D,r})- m_{D,r} \Rea(y_f) + n_{D,r} \Rea({\omega}_2)/t = 0$$ for some integers $m_{D,r}, n_{D,r}$. The second conjecture consists of comparing the coefficients $m_{D,r}$ of $y_f$ we get above with the Jacobi form coefficients of the form $\phi_f$.
Examples {#S:examples}
========
The Fourier coefficients of the forms in these examples were computed using SAGE [@sage]. The rest of the calculations were done in PARI/GP [@PARI2].
We will always take a set of generators for ${\Gamma_0(N)}$ which includes the translation matrix $T = \begin{smat}
{1}{1}{0}{1}
\end{smat}$ but no other matrix whose $(2,1)$ entry is $0$. The period integrals for $T$ are always $0$ since $i\infty$ is its fixed fixed point, hence we can exclude it from our computations of $\mathcal{P}$. In addition the $(2\pi)^k$ factor in the definitions of $y_D$ and $L$ is left off from the computations, since it is just a scaling factor and requires unnecessary extra precision.
For each example below, we list the number of digits of precision and the number $M$ of terms of $f$ we used. Below that is a set of generators we chose for $\Gamma_0^\ast(N)$ and the bases, ${\omega}_1, {\omega}_2$, we got for $L$ from computing $\mathcal{P}$ and applying the lattice reduction algorithm explained in Section \[S:algorithm\]. We then provide a table listing the $m_D$ which satisfy equation for $t$, $y_f$ of our choosing, and $D$ less than some bound. Without getting into details, the precision we chose depended on the size of the $M$-th term of $f$ and on the a priori knowledge of the size of the coefficients satisfying .
$2k=10$, $N=3$. The space of cuspidal newforms of weight $10$ and level $3$ has dimension $2$, but only one form has ${\varepsilon}=-1$. The first few terms of it are $$f = q - 36q^2 - 81q^3 + 784q^4 - 1314q^5 + 2916q^6 - 4480q^7 - 9792q^8 + \cdots$$
$$\renewcommand{\arraystretch}{1.4}
\begin{array}{ll}
\text{Precision} & 60\\
\text{Number of terms} & 100\\
\Gamma^\ast_0(3) & \big< T, \begin{smat}
{-1}{1}{-3}{2}
\end{smat}, {\omega}_{3} = \begin{smat}
{0}{-1}{3}{0}
\end{smat} \big>\\
{\omega}_1 & -i\cdot0.00088850361439085\dots\\
{\omega}_2 & 0.00002189032158611\dots \\
y_f & y_{-8}/2 \\
t & 1 \\
\end{array}$$
[|rr | rr|]{}
------------------------------------------------------------------------
$|D|$ & $m_D$ &$|D|$ & $m_D$\
8 & 2 & 104 & 380\
11 & -5 & 107 & -507\
20 & 8 & 116 & -40\
23 & 8 & 119 & -560\
35 & 42 & 131 & 235\
47 & -48 & 143 & -376\
56 & 0 & 152 & -364\
59 & -155 & 155 & -64\
68 & 160 & 164 & -1440\
71 & 40 & 167 & 1528\
83 & 353 & 179 & 2635\
95 & 280 & 191 & -400\
The $m_D$ in Table \[Ta:T1\] give, up to scalar multiple, the coefficients of the weight $11/2$ level $12$ modular form found in [@Ei p. $144$]. Note we can use the theorems of Waldspurger to get information about the values $L(f,D,k)$ from this table. For example, $L(f,-56,5)=0$.
$2k=18$, $N=1$.
The weight $18$ level $1$ eigenform in $S_{18}(1)$ has the closed form $$f(z) = \frac{-E_6^3(z) + E_4^3(z)E_6(z)}{1728},$$ where $E_{2k}(z)$ is the normalized weight $2k$ Eisenstein series.
$$\renewcommand{\arraystretch}{1.4}
\begin{array}{ll}
\text{Precision} & 200 \\
\text{Number of terms} & 100 \\
\Gamma^\ast_0(1) = SL_2({\mathbb{Z}}) & \big< T,S = {\omega}_1 = \begin{smat}
{0}{-1}{1}{0}
\end{smat} \big>\\
{\omega}_1 & i\cdot0.001831876775870191761\dots\\
{\omega}_2 & 0.000000000519923858624\dots \\
y_f & y_{-3}/3 \\
t & 1
\end{array}$$
[|rr | rr|]{}
------------------------------------------------------------------------
$|D|$ & $m_D$ & $|D|$ & $m_D$\
3 & 1 &51 & 108102\
4 & -2 &52 & -93704\
7 & -16 &55 & -22000\
8 & 36 &56 & 80784\
11 & 99 &59 & -281943\
15 & -240 &67 & 659651\
19 & -253 &68 & 193392\
20 & -1800 &71 & -84816\
23 & 2736 &79 & -109088\
24 & -1464 &83 & -22455\
31 & -6816 &84 & -484368\
35 & 27270 &87 & 1050768\
39 & -6864 &88 & 143176\
40 & 39880 &91 & 195910\
43 & -66013 &95 & -370800\
47 & 44064 & &\
The $m_D$ in Table \[Ta:T2\] are identical to the coefficients of the weight $19/2$ level $4$ half-integer weight form in [@Ei p.$141$], which is in Shimura correspondence with $f$.
$2k=4$, $N=13$
The dimension of the new cuspidal subspace is $3$ in this case, but only one has integer coefficients in its $q$-expansion. $$f = q - 5 q^2 - 7 q^3 + 17 q^4 - 7 q^5 + 35 q^6 - 13 q^7 - 45 q^8 + 22 q^9 + \cdots$$
$$\renewcommand{\arraystretch}{1.4}
\begin{array}{ll}
\text{Precision} & 28 \\
\text{Number of terms} & 250 \\
\Gamma^\ast_0(13) & \big< T, \begin{smat}
{8}{-5}{13}{-8}
\end{smat}, \begin{smat}
{-3}{1}{-13}{4}
\end{smat}, \begin{smat}
{5}{-2}{13}{-5}
\end{smat} ,
\begin{smat}
{-9}{7}{ -13}{10}
\end{smat}, {\omega}_{13} = \begin{smat}
{0}{-1}{13}{0}
\end{smat} \big>\\
{\omega}_1 & i\cdot0.003124357726009878347400865279 \dots \\
{\omega}_2 & -0.04271662498543992056668379773\dots \\
& \quad - i\cdot0.001562178863004939178984383052\dots \\
y_f & y_{-3}/3 \\
t & 3
\end{array}$$
Notice this is the first example of a nonsquare lattice. In fact ${\omega}_2/{\omega}_1 = -0.5000\dots + i\cdot13.67212999\dots$ so $\Rea({\omega}_2/{\omega}_1)=1/2$ as explained earlier. This is also the first example where the choice of $r$ matters, since $k=2$ is not odd. For each $D$, we chose $r$ in the interval $0<r<13$. In addition this is our only example where $t>1$.
[|rr | rr|]{}
------------------------------------------------------------------------
$|D|$ & $m_{D,r}$ & $|D|$ & $m_{D,r}$\
3&1&107&4\
4&-1&116&-8\
23&2& 120&-13\
35&-7&127&14\
40&3&131&-3\
43&-17&139&29\
51&9&152&2\
55&-6&155&22\
56&1&159&-6\
68&-5&168&-21\
79&4&179&-17\
87&-6&183&-2\
88&10&191&-10\
95&4&199&4\
103&-8 & &\
A closed form expression for the weight $3$ index $13$ Jacobi form $\phi = \phi_f$ corresponding to $f$ was provided to us by Nils Skoruppa, $$\phi(\tau, z) = \vartheta_1^5 \vartheta_2^3 \vartheta_3 / \eta^3$$ Here $\eta$ is the usual Dedekind eta-function, $\eta = q^{1/24}\prod_{n\geq1} (1-q^n)$ with $q=e^{2\pi i \tau}$, and $\vartheta_a = \sum_{r \in {\mathbb{Z}}} \left( \frac{-4}r\right)
q^{\frac{r^2 } 8} \zeta^{\frac{ar} 2}$ for $a=1,2,3$, $\zeta = e^{2\pi iz}$. (This has a nice product expansion using Jacobi’s triple product identity.)
We verify that the $(n,r)$-th coefficient $c(n,r)$ in the Fourier expansion of $\phi$ is identically equal to the $m_{D,r}$ in Table \[Ta:T3\] for $|D|<200$.
More Examples {#S:moreexamples}
=============
The coefficients of Jacobi forms are difficult to compute, in particular for the cases when $N$ is composite or when $k$ is even. We chose the previous examples in part because the Fourier coefficients for their Jacobi forms already existed, thanks to the work of Zagier, Eichler, and Skoruppa mentioned above. However, given any weight and level, we can still provide convincing evidence for our conjecture without knowing the exact coefficients of its Jacobi form. This is done using a refinement of Waldspurger [@Wa] given in [@GZK p.$527$].
Specifically, let $f\in {S_{2k}(N)}$ be a normalized newform with ${\varepsilon}=-1$. Let $\phi = \phi_f \in J_{k+1,N}$, with Fourier coefficients denoted by $c(n,r)$, be the Jacobi form corresponding to $f$ as described in Section \[S:conjectures\]. For a fundamental discriminant $D$ with $\gcd(D,N)=1$ and square root $r$ modulo $4N$, [@GZK Corollary $1$] says $$|D|^{k-1/2} L(f,D,k) \doteq |c(n,r)|^2;$$ here $L(f,D,s)$ is $L$-series of $f$ twisted by $D$, and $n\in {\mathbb{Z}}$ satisfies $D=r^2 -4Nn$. By $\doteq$ we mean equality up to a nonzero factor depending on $N ,2k, f,$ and $\phi$, but independent of $D$. (Gross-Kohnen-Zagier give this constant explicitly in their paper, but for us it is unnecessary.)
Thus given two such discriminants $D_i = r_i^2-4Nn_i$, $i=1,2$, we have $$\frac{ |D_1|^{k-1/2} L(f,D_1,k)}{ |D_2|^{k-1/2} L(f,D_2,k)} = \frac{|c(n_1,r_1)|^2}{|c(n_2,r_2)|^2}.$$ Hence by computing central values of twisted $L$-functions of $f$, we can test if ratios of squares of our $m_{D_i,r_i}$ are equal to those of $c(n_i,r_i)$.
For the examples below we have the same format as the previous examples along with a fixed choice of discriminant $D_1$ for which we verified explicitly, $$\frac{ |D_1|^{k-1/2} L(f,D_1,k)}{ |D|^{k-1/2} L(f,D,k)} = \frac{m_{D_1,r}^2}{m_{D,r}^2}$$ for all $D$ coprime to $N$ less than a certain bound.
$2k=4$, $N=21$.
The dimension of the new cuspidal subspace of $S_4(21)$ is $4$. We chose $$f = q - 3 q^2 - 3 q^3 + q^4 - 18 q^5 + 9 q^6 + 7 q^7 +\cdots.$$
$$\renewcommand{\arraystretch}{1.4}
\begin{array}{ll}
\text{Precision} & 40 \\
\text{Number of terms} & 500 \\
\Gamma^\ast_0(21) & \big< T, \left(
\begin{smallmatrix}
-4 & 1 \\
-21 & 5 \\
\end{smallmatrix}
\right),
\left(
\begin{smallmatrix}
11 & -5 \\
42 & -19 \\
\end{smallmatrix}
\right),
\left(
\begin{smallmatrix}
13 & -9 \\
42 & -29 \\
\end{smallmatrix}
\right),
\left(
\begin{smallmatrix}
8 & -5 \\
21 & -13 \\
\end{smallmatrix}
\right),
\left(
\begin{smallmatrix}
26 &-19 \\
63& -46 \\
\end{smallmatrix}
\right),
\left(
\begin{smallmatrix}
-16 & 13 \\
-21 & 17 \\
\end{smallmatrix}
\right) \big>\\
{\omega}_1 & i\cdot 0.012130626847574141\dots \\
{\omega}_2 & -0.03257318919429172\dots \\
y_f & y_{-3} \\
t & 1 \\
D_1 & -20
\end{array}$$
For a consistent choice of each $r$ we chose the first positive residue modulo $2N$ which satisfies $D\equiv r^2\bmod 4N$ for each $D$.
[|rr | rr|]{}
------------------------------------------------------------------------
$|D|$ & $m_{D,r}$ & $|D|$ & $m_{D,r}$\
3 & 1 & 111 & 4\
20 & -1 & 119 & 0\
24 & -1 &131 & 3\
35 & 0 & 132 & 8\
47 & 2 & 143 & 2\
56 & 0 &152 & -7\
59 & 1 &159 & 0\
68 & -2 & 164 & -2\
83 & 5 & 167 & 4\
84 & 0 &168 & 0\
87 & -4 & 195 & 8\
104 & -3 & &\
$2k=12$, $N=4$.
The space of new cuspforms in $S_{12}(4)$ is spanned by one normalized newform whose Fourier series begins with, $$f = q - 516 q^3 - 10530 q^5 + 49304 q^7 + 89109 q^9 - 309420 q^11 + \cdots.$$
$$\renewcommand{\arraystretch}{1.4}
\begin{array}{ll}
\text{Precision} & 80 \\
\text{Number of terms} & 200\\
\Gamma^\ast_0(4) & \big< T, \begin{smat}
{1}{-1}{4}{-3}
\end{smat} \big>\\
{\omega}_1 & i\cdot 0.000960627675025996\dots\\
{\omega}_2 & -0.02998129737318938\dots \\
y_f & y_{-7} \\
t & 1 \\
D_1 & -7
\end{array}$$
Similar to the last example, we chose the first positive residue modulo $2N$ which satisfies $D\equiv r^2\bmod 4N$ for each $D$.
[|rr | rr|]{}
------------------------------------------------------------------------
$|D|$ & $m_{D,r}$ & $|D|$ & $m_{D,r}$\
7 & 1 &103 & 1649\
15 & 5 & 111 & -765\
23 & -3 & 119 & -90\
31 & -50 &127 & 2664\
39 & -35 & 143 & -3729\
47 & 186 & 151 & -505\
55 & 215 & 159 & -2825\
71 & -315 & 167 & 3819\
79 & -10 & 183 & 2539\
87 & -497 & 191 & 1830\
95 & 405 & 199 & -5755\
Acknowledgments {#acknowledgments .unnumbered}
===============
I am deeply grateful to my advisor, Fernando Rodriguez Villegas, for his continuing guidance and support and for sharing his ideas that have enriched this work. I would also like to thank Don Zagier, Winfried Kohnen, and Chad Schoen for their ideas and suggestions. Thanks to Jeffrey Stopple for his careful reading of the manuscript. The coefficients of modular forms were computed with the help of William Stein. This research was partially funded by the Donald D. Harrington Endowment Fellowship.
[^1]: We will say a subset $X$ of an abelian group $J$ lies on a line if $X \subseteq {\mathbb{Z}}\cdot x_0$ for some $x_0 \in J$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'During software front-end development, the work to convert *Graphical User Interface(GUI)* image to the corresponding front-end code is an inevitable tedious work. There have been some attempts to make this work to be automatic. However, the GUI code generated by these models is not accurate due to the lack of attention mechanism guidance. To solve this problem, we propose PixCoder based on an artificially supervised attention mechanism. The approach is to train a neural network to predict the style sheets in the input GUI image and then output a vector. PixCoder generate the GUI code targeting specific platform according to the output vector. The experimental results have shown the accuracy of the GUI code generated by PixCoder is over 95%.'
author:
- |
Xiaoling Huang$^a$, and Feng Liao$^a$\
$^a$ Dept. of Computer Science, Sun Yat-sen University, Guangzhou, China\
[email protected],
bibliography:
- 'ijcai18.bib'
title: Automatically Generating Codes from Graphical Screenshots Based on Deep Autocoder
---
Introduction
============
Using machine learning techniques to build an automatic programming system is a relatively new field. Automatic programming can be applied to many fields. For example, Automatic programming can help developers to implement some simple functions. Balog et al. and Riedel et al. have done related work[@DBLP:journals/corr/BalogGBNT16; @DBLP:journals/corr/RiedelBR16]. Balog et al. proposed DeepCoder, which take five input-output pairs as input, and output a program that satisfy these five input-output pairs. However, the program generated by DeepCoder is very simple, so Deepcoder is still far away from the actual application.
Moreover, automatic programming can prevent developers from time-consuming software front-end development. Beltramelli propose pix2code [@DBLP:journals/corr/Beltramelli17], which is a system that takes a GUI image as input and outputs GUI code. Pix2code consists encoding part and decoding part. The encoding part contains a *Convolutional Neural Network(CNN)* model and a RNN model. Pix2code accepts the GUI image and the corresponding DSLs as input. At this time, the CNN model is responsible for encoding the GUI image into a vector, and the RNN model is responsible for encoding a segment of the corresponding DSLs. And then input them to the decoding part. The decoding part is another RNN model. It generates the next character in the DSLs. By comparing the generated character with the character in the actual DSLs, a gradient can be obtained to optimize the parameters in pix2code. However, the result of pix2code is not satisfying since the code generation of pix2code lacks attention mechanism [@DBLP:journals/corr/LuongPM15]. As a result, the generation of every token is based on the whole image as shown in Figure \[contrast\]. In fact, each piece of the frond-end code corresponds to a specific block in the image. This led to low accuracy of the front-end code generated by pix2code.
{width="50.00000%"}
\[contrast\]
In this work, we focus on the problem in pix2code and propose PixCoder, which is a precise descriptive programming system using deep learning techniques. We apply the method named PixCoding, which applies an artificially supervised attention mechanism, to fix the problems in pix2code. In more detail, we first artificially encode the GUI image to a vector. As shown in Figure \[contrast\], Each bit in the vector is associated with some blocks in the image. In this way, we artificially establish the attention mechanism. As a result, the model generates code based on specific block of GUI image. And then We feed the images and vectors into a vision model based on CNN for supervised learning. During the training process, CNN learns the mapping between the style sheet in specific block of the GUI image and the specific bits in the vector. Finally, based on the predicted vector output by the vision model, we can directly generate the precise front-end code of GUI.
In summary, our concrete contributions are:
1. proposing an impressive automatic descriptive programming system. In some domains, the capacity of this system is comparable to humans.
2. defining a method in image recognition and classification. This method decomposes the original classification tasks into simple sub-classification tasks, which reduces the difficulty of the whole task and improves the classification accuracy.
We organize the paper as follows. We first review related work. After that, we present the formal details of our framework, and then give a detailed description of our algorithm. Finally, we evaluate PixCoder in three different datasets and conclude our work with a discussion on future work.
Related Work
============
As we mentioned, there has been prior work on automatic programming. In automatic functional programming field, Bunel et al. propose an adaptive neural-compilation framework to address the problem of efficient program learning [@DBLP:conf/nips/BunelDMKT16]. In the work of Riedel et al., through a neural implementation of the dual stack machine that underlies Forth, programmers can write program sketches with slots that can be filled with behaviour trained from program input-output data [@DBLP:journals/corr/RiedelBR16]. Ling et al. explored the generation of source code from a mixed natural language and structured program specification [@DBLP:conf/acl/LingBGHKWS16]. In the work of Gaunt et al., the source code is generated through a distinguishable interpreter that learns the relationship between input-output examples [@DBLP:journals/corr/GauntBSKKTT16]. Balog et al. propose DeepCoder, which leverages statistical predictions to generate computer programs [@DBLP:journals/corr/BalogGBNT16].
However, code generation with visual inputs is still an unexplored research area until Beltramelli propose pix2code [@DBLP:journals/corr/Beltramelli17]. The architecture of pix2code is similar to some models applied to other areas (c.f. [@DBLP:conf/cvpr/KarpathyL15; @DBLP:conf/cvpr/VinyalsTBE15; @DBLP:conf/icml/XuBKCCSZB15; @DBLP:conf/cvpr/DonahueHGRVDS15]). Due to the code generation of pix2code lacks attention mechanism [@DBLP:journals/corr/LuongPM15], the result of pix2code is far from expectation. For this problem, we apply an artificially supervised attention mechanism and adopt an approach similar to DeepCoder.
Most of these works rely on *Domain Specific Languages(DSLs)*. DSLs are programming languages that are designed for a specialized domain. Compare with full-featured programming languages, DSLs are more restrictive. As a result, DSLs limit the complexity of programming language, which makes automatic programming easier and makes special-purpose search algorithm efficient [@DBLP:conf/oopsla/PolozovG15]. In this work, we are only interested in the GUI layout, the different graphical components, and their relationships. Thus we choice a simple DSLs designed by Beltramelli [@DBLP:journals/corr/Beltramelli17].
Problem Definition
==================
The problem we focus on is designing an automatic descriptive programming system, which input is a GUI image ${\mathcal I}$ and the output is the front-end code ${\mathcal C}$ of the GUI. We denote $\alpha({\mathcal I})={\mathcal C}$ as the automatic descriptive programming process. So the problem can can be considered to find mapping $\alpha$ between ${\mathcal I}$ and ${\mathcal C}$.
According to different target platforms, ${\mathcal I}$ can be a web-based UI image, an iOS UI or an Android UI image. Figure \[imgdslcode\](a) is an input example of web-based UI image. Different UI images vary in size, so we resize the input image to 256$\times$256 pixels and the pixel values are normalized. After this process, we get standardized image ${\widetilde{\mathcal{I}}}$.
The output of the problem is the front-end code ${\mathcal C}$ of the GUI. However, as shown in Figure \[imgdslcode\](c), the front-end code is complicated. And it is difficult for the model to directly generate the front-end code. So our approach is designing DSLs ${\mathcal D}$ for target front-end code ${\mathcal C}$ and let model output ${\mathcal D}$ instead of ${\mathcal C}$. Figure \[imgdslcode\](b) is the DSLs of web-based UI image. The syntax of DSLs is relatively simple here, since we only need to use ${\mathcal D}$ to describe the outline of the GUI image.
Since the output of our model is ${\mathcal D}$ and the problem need ${\mathcal C}$. So we need to compile ${\mathcal D}$ to get ${\mathcal C}$. This process is denoted as $\beta({\mathcal D})={\mathcal C}$ and we denote $\delta({\widetilde{\mathcal{I}}})={\mathcal D}$ as the automatic descriptive programming process of model. The compile process is static so the key work is to find $\delta$, the mapping between ${\widetilde{\mathcal{I}}}$ and ${\mathcal D}$.
{width="42.50000%"}
\[imgdslcode\]
PixCoder
========
To implement such an automatic descriptive programming system, we have tried to add attentional mechanism to pix2code just like the work of Luong et al. [@DBLP:journals/corr/LuongPM15]. However, this did not improve the performance of the model. Therefore, we use a method named PixCoding to add a supervised attentional mechanism artificially. And in combination with the ideas of DeepCoder [@DBLP:journals/corr/BalogGBNT16], we propose PixCoder. The automatic descriptive programming process of PixCoder can be divided into the following phases: The first phase is a vision model based on CNN, which is used to understanding the given GUI image ${\mathcal I}$ and inferring the objects present. The vision model outputs a predicted vector ${\mathcal V}$ and each bit in ${\mathcal V}$ corresponds to a kind of style sheet in different block in ${\mathcal I}$. This can be seen as a vision model identifies each block in ${\mathcal I}$ and then record the result into ${\mathcal V}$. In the second phase, we exploit a parser in solving code generation problem. The parser leverages ${\mathcal V}$ generated by vision model to generate DSLs ${\mathcal D}$ describing ${\mathcal I}$. The last phase is a compiler. In this phase, with traditional compiler design techniques, ${\mathcal D}$ is compiled to the front-end code ${\mathcal C}$ targeting specific platform.
{width="40.00000%"}
\[architecture\]
PixCoding
---------
PixCoding is a universal image recognition method: we analyze the possible changes in different areas of the image and then devise a vector. Each bit in the vector corresponds to a different change in different areas of the image. And then we can make some rules based on our experience, according to which we are able to know that the changes in various combinations of different areas corresponds to which label. And then the task of image recognition can be completed. This image recognition method decomposes the image classification task into sub-classification tasks in different areas of the image, and then combines the results of the sub-classification tasks to obtain the final classification result. At this point, the accuracy of the classification depends on the accuracy of the sub-classification. Because sub-tasks are relatively simple, this method can improve the accuracy of the whole task.
In this work, we apply this method to automatic descriptive programming field. We design the corresponding vectors for a particular type of GUI image. Each bit in the vector corresponds to the changes of style sheet in different blocks in the GUI image. By combining each of the vectors we know all the style sheets in the GUI image. As a result, we can directly generate DSLs corresponding to the input GUI image. This process is no different from the process of combining the results of a sub-classification task to get the final classification result. In addition, PixCoding mentioned above can be applied to image text recognition, such as recognition of long strings. We only need to identify the different areas of the characters belong to which number, and then combined them to complete the string recognition task. This method can also be applied to motion recognition. We divide the hand, foot, torso and other parts of human into different areas and then encode the vectors so that the vectors imply all possible situations in these areas.
However, PixCoding has some limitations: it can only be used when the data in the dataset is very normative and has a clearly defined area.
**Input:** a GUI image ${\mathcal I}$.\
**Output:** the GUI’s front-end code ${\mathcal C}$.
resize ${\mathcal I}$ to 256$\times$256 pixels and normalized the pixel values. And get standardized image ${\widetilde{\mathcal{I}}}$: *Image\_Standardization*(${\mathcal I}$)=${\widetilde{\mathcal{I}}}$; input ${\widetilde{\mathcal{I}}}$ to vision model and get predicted vector ${\mathcal V}$: $\zeta({\widetilde{\mathcal{I}}})={\mathcal V}$; standardize ${\mathcal V}$ and get standardized vector ${\widetilde{\mathcal{V}}}$: *Vector\_Standardization*(${\mathcal V}$)=${\widetilde{\mathcal{V}}}$; generate DSLs ${\mathcal D}$ according to ${\widetilde{\mathcal{V}}}$: $\theta({\widetilde{\mathcal{V}}})={\mathcal D}$; compile ${\mathcal D}$ to get ${\mathcal C}$: $\beta({\mathcal D})={\mathcal C}$; ${\mathcal C}$;
Image To Vector
---------------
### Vector Design
First, we need to analyze and design a specific type of GUI image. An GUI image example from iOS UI dataset is shown in Figure \[blockinimg\]. After analyzing the iOS UI dataset, we learned that the GUI image here can be divided into two blocks: a stack block and a footer block (the area in the Figure \[blockinimg\] is surrounded by the black frame). There can be a maximum of eight row blocks in the stack block, and at least one row block (the area in the Figure \[blockinimg\] that is surrounded by the red frame). There are at most four controls in each row block, and there are five types of controls, and up to four controls in the footer block and four types of footer controls. If all possibilities are under consideration, we shall design a vector of 176 (4$\times$5$\times$8+4$\times$4) bits using one-hot encoding. In this case, the vector contains all the blocks and the changes of the style sheet of them and in the GUI image.
But if we do further analysis, we can prune the bits of the vector. After further analysis, we know that the number of row blocks is from 1 to 8; the number of controls in the row block is from 2 to 4. The number of controls in the footer block ranges from 2 to 4. The appearance of controls in the row block is somewhat regular: such that the slider does not appear alone, it always appears as a group like label, slider, label. According to the analysis results, we can prune the median of the vector, and finally get a 72-bit vector shown in Figure \[vector\].
Using the vector before and after pruning separately, we found that the use of pruned vectors can reduce training time and improve the accuracy of vision recognition. In addition, the pruned vector can also avoid some unusual circumstances. Such as the error of beyond the length: Due to the length of each control is different, so a row block cannot contain four longest slider controls. In the pruned vector, the situation which four sliders appear in a row block cannot happen. As a result, we can avoid this error of beyond the specified length, which in turn enhances our automatically descriptively programming system.
{width="40.00000%"}
\[blockinimg\]
### Vision Model
After designing the vector, we need a vision model to achieve the image-to-vector mapping $\zeta$. we employ a CNN to model and learn $\zeta$ due to it can learn rich latent representations from the image [@DBLP:conf/nips/KrizhevskySH12; @DBLP:journals/corr/SermanetEZMFL13]. CNN is one of the most widely used methods in vision problems. Compared with the traditional image processing algorithms, CNN avoids the complex pre-processing of images, which improves the efficiency of the experiment greatly.
Firstly, ${\mathcal I}$ are initially re-sized to 256$\times$256 pixels and normalized the pixel values. After this process, we get standardized image ${\widetilde{\mathcal{I}}}$ and fed it in CNN. To recognize different objects in input image, we exclusively used 3$\times$3 receptive fields which are convolved with stride 1 as VGGNet [@DBLP:journals/corr/SimonyanZ14a]. These operations are applied twice before to down-sample with max-pooling. The width of the first convolutional layer is 32, followed by a convolutional layer of width 64, and the width of the third convolutional layeris 128. Our CNN ends with two fully connected layers. The size of the first fully connected layer is 1024 and the size of the final connection layer is based on the complexity of the style sheets in given GUI images. Except for the final fully connected layer applying the softmax activation function, each layer of our model applying the rectified linear unit activation function.
The Training process of vision model is supervised. Since only ${\mathcal I}$ and ${\mathcal D}$ are provided in the dataset, after designing the vector, we need to generate ${\mathcal V}$ corresponding to ${\mathcal I}$ and use it as the label for the supervised training. After completing this pre-process, we can start to train the vision model.
-------------------- ----------------------- ---------------
Before pruning After pruning
web-based UI Image 1.08$\times$$10^{8}$ 3528
iOS UI Image 4.66$\times$$10^{25}$ 9072
Android UI Image 1.16$\times$$10^{28}$ 42768
-------------------- ----------------------- ---------------
: Patterns of GUI image statistics.[]{data-label="Number of all possible patterns"}
{width="40.50000%"}
\[varietyofvectorsize\]
Vector To DSLs
--------------
The central idea of our work is to use a vision model to guide code generation. The parser is responsible for generating DSLs based on vectors. The parser has two main functions: to standardize ${\mathcal V}$ output by the vision model, and to generate ${\mathcal D}$ based on the ${\widetilde{\mathcal{V}}}$($\theta({\widetilde{\mathcal{V}}})={\mathcal D}$).
As mentioned above, the vector we design contains all the stylistic changes in the GUI image. Therefore, we can know from the vector that the presence or absence of individual style sheets in individual block of the given input GUI image. In accordance with the previously designed DSLs syntax rules, we can generate the corresponding DSLs directly from the output vector.
However, ${\mathcal V}$ output by the vision model does not meet our expectations. First, the value of each bit in the output vector is not a standard value like 0 or 1. In addition, it is possible to have ambiguities in the output vectors, such as specifying conflicting patterns for a block, and so on. Therefore, before generating the corresponding DSLs based on vectors, we need to standardize ${\mathcal V}$ and get ${\widetilde{\mathcal{V}}}$. We can only generate ${\mathcal D}$ based on ${\widetilde{\mathcal{V}}}$.
{width="40.50000%"}
\[vector\]
### Vector Standardization
First, we need to use thresholds to process the value of each bit in ${\mathcal V}$ into a standard value like 0 or 1. Here we cannot use the method of setting the maximum value in each region of ${\mathcal V}$ directly to 1, and the rest of the value all to 0. Because this method does not consider the situation that ${\mathcal V}$ can have some regions of all 0. Using the iOS UI dataset as an example, with our method, for any input GUI image, the final output is a DSLs containing eight row blocks. Therefore, we need to use thresholds to deal with this problem.
In this work, we automatically get the threshold during the training of vision model. After each epoch after training halfway, we take a part of the samples from the training set and input them into the vision model to get the output vector. By comparing the output vector with the target vector, we can know which bit in the output vector correspond to 1 and which bit correspond to 0. The smallest one of the output values corresponding to 1 is recorded as a candidate for the threshold value. At the same time, we will also record the largest of the output values corresponding to 0. After training, We will apply a complex algorithm to get the threshold. For example, if the maximum value corresponding to 0 is smaller than the minimum one corresponding to 1, then we take the average of this interval as the threshold. Otherwise, We use other methods to get a reasonable threshold. According to Figure \[accthreshold\], The choice of threshold is important to the result of model. And the threshold generated by our algorithm is close to the best threshold.
For all possible ambiguities in ${\mathcal V}$, we have corresponding solution. For example, for the situation of specifying conflict patterns for a block, our approach is to take the larger one of the two models corresponding to the prediction and set it to 1, and the other set to 0. And for the situation that an empty row block in the middle of two row blocks, our approach is to swap the empty row block with next row block contains objects.
### DSLs Generation
After completing the process of standardizing ${\mathcal V}$, we can easily generate the corresponding DSLs according to ${\widetilde{\mathcal{V}}}$ based on the previously designed DSLs grammar rules($\theta({\widetilde{\mathcal{V}}})={\mathcal D}$). The normalized vector is error-free, which guarantees there will be no syntax error exists in the generated DSLs.
DSLs to Front-end code
----------------------
When we get ${\mathcal D}$ from parser, our work have not finished. We need a compiler to compile the ${\mathcal D}$ into ${\mathcal C}$ for various platforms(i.e. Android and iOS native mobile interfaces, and multi-platform web-based HTML/CSS interfaces). As mentioned, This process denotes as $\beta({\mathcal D})={\mathcal C}$. The compile process is static. Thanks to the parser guarantees ${\mathcal D}$ without syntax error, ${\mathcal C}$ generated by compiler is legal.
Experiments
===========
Dataset
-------
The dataset provided by Tony Beltramelli [@DBLP:journals/corr/Beltramelli17] has three types, including iOS GUI images and corresponding DSLs, Android GUI images and corresponding DSLs and web-based GUI images and corresponding DSLs. Each type of dataset is divided into two parts, including 1500 training set and 250 testing set. Each example contains an original image and the corresponding DSLs.
According to Figure \[varietyofvectorsize\], the complexity of these three datasets is increasing. The number of the style sheets of the GUI image are least in the Web-based UI dataset, the iOS UI dataset is the second, and the Android UI datasets is the most.
-------------- ------- ------ ------- ------
Image DSLs Image DSLs
web-based UI 1500 1500 250 250
iOS UI 1500 1500 250 250
Android UI 1500 1500 250 250
-------------- ------- ------ ------- ------
: Dataset statistics.[]{data-label="dataset"}
Criterion
---------
There are many ways to evaluate the quality of codes generated by the model. For example, screenshot the image showed by the generated GUI code, and then compare the similarities between this and the input image. Or, compare the similarities between the generated code and the target code. After experimental comparison, we evaluated the PixCoder with the second way: compare the similarities between the generated DSLs and the target DSLs.
In our experiment, every DSLs that in line with grammatical rules is in tree structure. We first convert the generated DSLs and the target DSLs to a tree, then using the similarities between tree structures to represent the similarities between DSLs. The classic algorithm used to match the similarity of traditional tree models is based on the similarity of edit distance [@DBLP:journals/jacm/Tai79], which allows cross-layer matching of tree nodes. Cross-layer matching and replacement may be useful when comparing the general tree similarity, but not suited to the tree model used in our experiment. The tree model of our experiment is similar to the HTML tree model. HTML tree model document label nodes will be read by the browser rendering to the screen, different root node corresponds to a different set of child nodes. Therefore, even if we replace the root node, the children will not be matched. So cross-layer matching and replacement cannot obfuscate the relationship between nodes in the HTML tree, which is rendered in the browser effect.
Therefore, we need another way to measure the similarity of the tree model in our experiment. Here we use the Simple Tree Matching(STM) algorithm [@DBLP:journals/spe/Yang91; @DBLP:journals/jcrd/He07]. STM bases on the principle of maximum matching, dynamic programming is used to calculate the maximum number of matching nodes of two trees, and then the similarity between two trees is obtained. This process does not allow cross-layer matching and node replacement, and requires that child nodes be completely ordered. If the root nodes of two subtrees do not match, the other nodes of the two subtrees are not considered, so as to achieve the effect of pruning. The time complexity of this algorithm is O($n^{2}$).
$$Similarity(T_1, T_2) = \frac{SimpleTreeMatching(T_1, T_2)}{(|T_1|+|T_2|)/2}$$
Above is the formula that calculates the similarity between the generated DSLs and the target DSLs. $T_{1}$ and $T_{2}$ represent the tree obtained by converting the generated DSLs and the target DSLs.
SimpleTreeMatching($T_{1}$, $T_{2}$) represents the maximum number of matched nodes for two trees. $|T_{1}|$ and $|T_{2}|$ represent the number of nodes in two trees. If and only if the maximum number of matching nodes of two trees is larger, the Similarity($T_{1}$, $T_{2}$) is larger, too. That is, SimpleTreeMatching($T_{1}$, $T_{2}$) is closer to 1. At this point the trees represent the generated DSLs and the target DSLs are more similar.
------------------ -------------- --------- ------------
web-based UI iOS UI Android UI
Baseline 62.882% 70.303% 65.825%
pix2code(beam 3) 76.905% 68.640% 54.644%
pix2code(greedy) 88.591% 87.621% 85.073%
PixCoder 98.699% 95.562% 98.177%
------------------ -------------- --------- ------------
: Experimental results in different dataset.[]{data-label="result"}
{width="40.50000%"}
\[accthreshold\]
Experimental Results
--------------------
We evaluate PixCoder in following aspects. We first compare the similarities between the target DSLs and the DSLs generated by PixCoder and Baseline. We then compare the similarities between the target DSLs and the DSLs generated by PixCoder and pix2code, specially, pix2code has two types, one in use of beam 3 method, and the other in use of greedy method. We use a randomly generated grammar-compliant DSLs with a similarity to the target DSLs of the test set as Baseline.
As shown in Figure \[accthreshold\], we can get the conclusion that the choice of threshold is important to the result of PixCoder. Threshold can not be too large or too small. In this experiment, we set threshold to 0.01, which is close to the best threshold according to the experimental results. As shown in Figure \[accthreshold\], the size of the dataset affects the result. As the dataset increases, the model can learn more knowledge about the style sheets of GUI image. As a result, the model is more capable. In this work, we feed all 1500 examples from dataset to the model while training.
The experimental results in the Table \[result\] are the average similarities of 250 samples in the test set obtained using baseline, pix2code and PixCoder respectively. As described above, the complexity of these three datasets is increasing. The average similarities of Baseline on these datasets do not reflect the complexity of the style sheet changes in the GUI image. We believe this is caused by the unevenly distributed style sheets of the GUI image in the test datasets. The main contrast model of PixCoder in this experiment is pix2code. Constrained by the experimental results, the DSLs generated by pix2code in use of beam 3 method has a lot of syntax errors. In the calculation of average similarities, our method is to write 0 for the similarity of DSLs that have syntax errors. This results in the similarities of pix2code in use of beam 3 method being particularly low. Pix2code in the use of the greedy method almost has no syntax errors, so its average similarities are relatively high. There is no syntax error in DSLs generated by PixCoder, and its accuracy is far better than pix2code. The difference in average similarities between Pix2code and PixCoder does not reflect the difference in the quality of the code they generate. In fact, the quality of the code generated by Pix2code is far from PixCoder. As shown in Figure \[result\], There is only one error(the area surrounded by the red frame) in the output of PixCoder and there are eight error in pix2code. But their difference in similarity is only 12.821%. In a word, to the best of our knowledge, PixCoder is the best model in automatic descriptive programming field.
{width="45.00000%"}
\[result\]
Conclusion
==========
In this paper, we present a method called PixCoding for image recognition and classification. PixCoding decomposes the image classification tasks into sub-classification tasks in different regions of the image. This process is an artificially supervised attention mechanism. We artificially design a vector to encode the image. The vector-image pairs guide CNN model to identify specific regions in the image and classify style sheets in these regions. In the end, following some rules, we can get the final classification results by integrating the results of sub-classification tasks. Our experimental results also show that PixCoding has an excellent effect on image recognition and classification tasks.
We apply the PixCoding to automatic programming field and propose PixCoder. PixCoder is an impressive automatic descriptive programming system, which takes an GUI image as input, and then generates the corresponding front-end code. Compared to other automatic descriptive programming systems, PixCoder produces extremely accurate GUI code that is close to human levels.
However, the GUI image in the dataset we used are simpler than the actual GUI image. Therefore, how to make PixCoder still perform excellent in more complex datasets is our future work. Moreover, we also interested in how to apply PixCoding in other image classification tasks.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Prospects of parametric x-rays (PXR) application for the development of a tuneable quasi-monochromatic x-ray source for medical imaging are discussed. Analysis of basic requirements for electron accelerator shows that it must be relatively low-energy and high-current linac. In comparison with known ultra-relativistic cases, at low energies PXR properties will be modified to a great extent by multiple scattering of the electrons. PXR intensity dependence on target thickness and beam energy are calculated taking multiple scattering into account. It is concluded that PXR source based on real medical accelerators is feasible and can provide x-ray flux needful for obtaining high quality medical images.'
author:
- |
Alexander Lobko\* and Olga Lugovskaya\
Institute for Nuclear Problems, Belarus State University\
11 Bobruiskaya Str., Minsk 220050, Belarus\
[*\* Corresponding author: [email protected]*]{}
title: 'X-RAY SOURCE BASED ON THE PARAMETRIC X-RAYS'
---
***Keywords*:** Parametric X-rays; PXR; monochromatic tuneable x-ray source; medical imaging
***PACS****:* 78.70.-g; 11.80.La; 87.56.By
Introduction
============
Parametric x-rays (PXR) produced by a relativistic charged particle uniformly moving through a single crystal were declared as a very attractive radiation mechanism for x-ray source since its theoretical prediction in 1972 [@1] - [@3] and first experimental observation in 1985 [@4]. Actually, it has a number of indisputable features, namely, high monochromaticity, energy tuneability, polarization, directivity, and possibility to be emitted to large angles relative to beam direction, i.e. provide monochromatic x-rays virtually free of background. PXR properties were studied at a number of accelerator facilities worldwide but a PXR-based x-ray source developments are still in progress.
Evidently, this is connected with the fact that PXR research is performed at existing linear and annular accelerators and limited by their technical specifications. So, we should recognize how to design a pxr-based source that will meet the consumer’s requirements for its use in industry and medicine.
Industry and science, as consumers, essentially yield to medicine, which represents the huge market for devices and methods for x-ray diagnostics. Statistically, x-ray devices count for about 60% of medical diagnostics devices [@5]. Despite the development of alternative diagnostics methods, e.g. ultrasound and magnetic resonance imaging, the number of studies in the field of x-ray diagnostics remains high. First of all, medicine demands decreased irradiation doses taken together with increased resolution and contrast. These requirements can be met if the customer will be offered a monochromatic tuneable source, providing x-radiation with exactly the energy required for examination of a specific organ. For example, in mammography the maximum spectral density lies in the range of 17-20 keV; in radiography of the chest, extremities and head, the range of 40-50 keV is optimal; while the 50-70 keV range is used for abdomen and pelvis radiography. Precise selection of monochromatic radiation energy will allow a reduction of a patient dose of tens times, even in comparison with digital radiography and computer tomography, with improvement of x-ray image quality. However, in comparison with scientific and industrial applications where signal accumulation is possible, in the case of medicine the source should emit number of quanta sufficient to provide necessary contrast of an image at the specific value of the signal-to-noise ratio. Estimations [@6; @7] give minimal value of x-ray quanta necessary for high-quality image equal to $\sim $10$^{7}$ photons/mm$^{2}$. PXR quantum yield in a thin crystal target even at high enough (e.g. 900 MeV) beam energy is not very large and amounts only $\sim $10$^{-5}$-10$^{-6}$ photons/e$^{-}\cdot $sr [@8]. Assuming the target to be imaged is 30 cm diameter at one meter distance from a source, it is necessary to apply $\sim $ 0.1 A current to get a required x-ray flux. Such currents can be achieved only in low energy accelerators. Though theoretically PXR may be emitted at any energy of charged particles [@9], there is a factor limiting the minimal value of beam energy for source under discussion. Assuming x-ray energy necessary for subtractive coronary angiography $\varpi _{B} $=33 keV then the angle of radiation $\theta _{B} $ in a diamond target will come to $\sim 4^{\circ}$ for (111) reflex and $\sim 7^{\circ}$ for (220) reflex that it is enough for arrangement of a patient at 1-2 meter distance from the target. Now let us estimate energy ${E}_{p}$ of an electron beam providing angular width of reflex [@10] $\vartheta _{ph}^{2} = \left|
{\chi _{0}^{'}} \right| + \gamma ^{ - 2} + \overline {\theta
_{s}^{2}} $ ($\left| {\chi _{0}^{'}} \right|$ is the Fourier component of a crystal dielectric constant, $\gamma=E_{p}/mc^{2}$, $\overline {\theta _{s}^{2}} $ is the mean-square angle of multiple scattering) equal $\sim 5^{\circ}$, i.e. energy at which x-ray reflex and forward background are still separated from each other and the radiation in which we are interested keeps all the properties necessary for a medical source. The dominant component here is the mean-square angle of multiple scattering $\overline
{\theta _{s}^{2}} $ and lowest estimation of beam energy gives a value $\sim $17 MeV. To provide more real evaluations for PXR source based on low-energy electron beam, we should look into multiple scattering (MS) of electrons more precisely, because MS appeared to be the most significant factor affecting PXR intensity and angular distribution at relatively low energies and/or thick targets [@8]. In low energy cases we shouldn’t use kinematics formulae [@10], where MS is considered phenomenologically. Let us describe the multiple scattering influence on x-ray reflex characteristics in more detail following [@11].
Calculation of x-radiation properties in the multiple scattering presence
=========================================================================
To get an expression for spectral-angular density of radiation with MS account, it is necessary to average the squared modulus of all possible particle trajectories in a crystal in the following expression [@12; @13]
$$\frac{{d^{2}N_{s}} }{{d\varpi d\Omega} } =
\frac{{e^{2}Q^{2}\varpi} }{{4\pi ^{2}\hbar c^{3}}} \left|
\int\limits_{ - \infty} ^{ + \infty} {\vec {v}\vec {E}_{\vec
{k}}^{\left( { -} \right)s}} \left( {\vec {r}\left( {t}
\right),\varpi } \right) exp\left( { - i\varpi t} \right)dt
\right|^{2} \label{1}$$
The obtained expression will describe all mechanisms of the radiation: PXR, *Bremsstrahlung*, and transition radiation in a single manner. The velocity vector is presented in the form: $\vec {v} = \vec {v}_{0} cos\theta + v_{0} \vec {\theta} $ (where $\vec {v}_{0} $ determines the direction of the initial velocity of a charged particle, $\theta \equiv \left| {\vec {\theta} }
\right|$ is the the angle of multiple scattering of a particle ($\theta < < 1$), $\vec {\theta}$ is the two-dimensional vector, $\vec {\theta } \bot \vec {v}_{0} $ and $\vec {\theta} = 0$ at $t
\le 0$).
Averaging is followed [@13], where the procedure was performed with the help of dispersion function, satisfying the Fokker-Planck kinetic equation in the amorphous medium. Averaging of the right part of the equation (\[1\]) was performed and the integral expressions for spectral-angular distributions of the radiation, taking into account multiple scattering for the lateral and forward PXR maxima were derived.
Selected functions of the particle distributions on coordinates and angles are usually applied in calculations of *Bremsstrahlung* intensity of ultra-relativistic electrons in amorphous media. It is applicable if direction of vector of a charged particle initial velocity does not coincide with directions of the basic crystallographic axes or planes. Also, one can ignore the influence of crystal structure on charged particles trajectories and apply dispersion functions of amorphous medium when value of initial angular divergence of a particle beam, falling on the crystal along the direction of main crystallographic axes or planes, satisfies the inequality $\left|
{\Delta \vec {v}_{0}} \right|/c > \sqrt {\left( {2U_{0} /E_{p}}
\right)} $, where $U_{0}$ is the potential of atomic axis or plane, $E_{p}$ is the particle energy. In this case during the passage of charged particles through the crystal, orientational effects will be expressed less strongly than in the case of a well-collimated beam, and the averaging of (\[1\]) over simple dispersion functions ensures a satisfactory approach.
It is necessary to note, that in the case, when the value $\left|
{\Delta \vec {v}_{0} /c} \right| < \left( {\gamma ^{ - 2} +
\frac{{\omega _{L}^{2} }}{{\omega _{B}^{2}} }} \right)^{ - 1/2}$ ($\omega _{B}^{\left( {n} \right)} = \frac{{\pi cn}}{{dsin\theta
_{B}} }$ is the Bragg frequency, $n=1,2,...$, $\theta _{B}$ is the angle between the particle velocity $\vec {v}_{0} $ and planes corresponding to the vector $\vec {\tau} $, $\omega _{L}$ is the Langmuir frequency of a crystal), it is not required to make additional averaging in (\[1\]) on the initial divergence of the vector direction $\vec {v}_{0} $.
There are two factors limiting the longitudinal size of quasi-Cherenkov radiation formation area: MS of charged particles on atoms of substance and x-ray absorption in the medium. As in case of radiation generation in an amorphous medium, the PXR characteristics essentially depend on the relation between crystal thickness along the direction of the charged particle movement $L_{0} $**** and *Bremsstrahlung* coherent length $L_{Br} $. The theory of PXR with the phenomenological account of MS influence on particle energy range exceeding threshold energy $E_{tr} = mc^{2}\left|{\chi} '_{0}\right|^{ - 1/2}$ is in good agreement with experimental results only at crystal thickness $L_{0} < < L_{Br} $. In this case MS influence appears only as a small addition in the intensity of generated x-ray caused by *Bremsstrahlung*. Also MS influence on PXR phase is taken into account and interference between PXR and *Bremsstrahlung* is neglected. In the inverse case MS essentially changes the parameters of the PXR itself.
PXR characteristics dependence on crystal target thickness
==========================================================
To demonstrate MS effect on PXR characteristics, numerical calculations were performed for following conditions: target Si (220) in Laue geometry, $2\theta _{B} = 19^{ \circ} $, electron beam energy 900 MeV.
PXR spectral-angular density as a function of crystal target thickness is shown in Fig.\[fig1\].
= 8 cm
It is evident, that already at $L_{0} \sim L_{Br} $ presence of MS results in PXR intensity distinctly less than intensity determined without MS account. During further increase of the crystal thickness, spectral-angular density achieves saturation at thickness of $\sim (2 \div 3) L_{Br}$ and then goes down. It is seen from this figure that PXR spectral-angular density without MS (curve 1) continues to grow with crystal thickness increase even up to $\sim 10L_{Br}$. The tendency shows that absorption still does not plays an essential role though $L_{0} $ becomes of the order of $L_{abs} $. Simultaneously with the variation of PXR spectral-angular density with crystal thickness increase, change of its spectral width appears (see Fig.\[fig2\]).
= 8 cm
As long as the thickness of the crystal target does not exceed $0.5L_{Br} $, spectral-angular density and spectral width of PXR reflex determined with MS account practically coincide with the values received without MS consideration. At the same time, spectral width of PXR peak decreases with $L_{0} $ increase. The width of PXR maximum is inversely proportional to radiation coherent length. As the target thickness increase (on thickness of $\sim L_{Br} $ and larger) PXR coherent length begins to be limited by $L_{Br} $. So, PXR peak width without MS continues to decrease with $L_{0}
$**** growth, but MS account results in stabilization of peak width at $L_{0} \sim L_{Br} $ (Fig.\[fig2\]).
Thus, calculation results, demonstrated in the Figures \[fig1\] and \[fig2\], show that MS of a charged particle results not only in radiation phase shift, but also in essential decrease of PXR spectral-angular density at angles of $\vartheta \sim
\vartheta _{ph} $ as well as in change of reflex width, even in case of $L_{0} \sim L_{Br} $.
As a result of joint processes of PXR spectral narrowing and spectral-angular density decrease, PXR integral characteristics do not vary essentially. As an example, in Fig.\[fig3\] there are given dependence of PXR angular distributions (at $\vartheta = 1.66 \cdot 10^{ - 3}$rad) on the crystal thickness without MS (1) and in the MS presence (2). One can see that for the target thickness $L_{0} = 10L_{Br} $ the maximal value of spectral-angular distribution (at $\vartheta
= 1.66 \cdot 10^{ - 3}$ rad) calculated without MS account exceeds the the maximal value obtained with MS account more than two orders of magnitude (compare with Fig.\[fig1\]), while amplitudes of angular distributions (for the same angle) differ less than by one order (Fig.\[fig3\]). MS also results in PXR angular distribution spreading, therefore if the angular aperture of a detector is big enough ($\theta _{D} > > \vartheta _{ph} $), the difference in results for integrated number of the quanta calculated with the MS account and without taking it into account will be even less, than for angular distributions. Thus, the most significant difference of the results obtained with and without MS account is observed for differential distributions.
= 8 cm
PXR intensity dependence on charged particle energy
===================================================
As well as for Cherenkov radiation in homogeneous media, PXR spectral-angular density has threshold behavior. However, due to essential dispersion of a crystal refraction index, dependence of integral PXR intensity on particle energy is relatively smooth. It is differs in $E_{p} < < E_{tr} $ and $E_{p}
> > E_{tr} $ energy regions, where threshold energy $E_{tr} =
mc^{2}\gamma _{tr} $, $\gamma _{tr} = \sqrt {\left| {{\chi} '_{0}}
\right|} \approx \frac{{\omega _{B}} }{{\omega _{L} }}$. So, PXR quantum yield at particle energy $E_{p} < < E_{tr} $ **** is $N^{PXR}\sim \left( {\frac{{E_{p}} }{{E_{tr}} }} \right)^{4}$ and $N^{PXR}\sim ln\left( {\frac{{E_{p}} }{{mc^{2}}}} \right)$ at $E_{p}
> > E_{tr} $. Note, such dependences of PXR intensity on energy were calculated without taking into account of MS influence on radiation characteristics.
The *Bremsstrahlung* coherent length is directly proportional to charged particle energy: $L_{Br} = \sqrt
{\frac{{4c}}{{\omega \overline {\theta _{s}^{2}} } }} $, where $\overline {\theta _{S}^{2}} = \frac{{1}}{{2}}\left(
{\frac{{E_{S}} }{{E_{P}} }} \right)^{2} \cdot \frac{{1}}{{L_{R}}
}$, *E*$_{S} $= 21.2 MeV - scale energy, *L*$_{R}$ - radiation length. Therefore, by fixing the thickness of a crystal target and changing the energy of charged particles (thus changing $L_{Br} $), it is possible to calculate dependences of spectral-angular density and angular distribution, and dependence of spectral maximum width on the $L_{0} /L_{Br} $ ratio (similar to shown in Fig.\[fig1\] - Fig.\[fig3\]). Certainly, the nature of these dependences is completely different. A change of a crystal target thickness results in the change of a trajectory length where particle radiates coherently. In fact, this is its cut-off because the PXR coherent length is equal to infinity. Besides, on particle energy decrease the PXR generation is realized at increasing values of $\left| {\alpha _{B}} \right|$ (detuning parameter from the exact value of the Bragg angle) that in turn results in decrease of radiation intensity.
PXR spectral-angular distributions calculated for the same geometry as Fig.\[fig1\] and Fig.\[fig2\] were plotted in Fig.\[fig4\].
= 8 cm
Crystal thickness was chosen equal to 0.01 cm and polar angle of radiation was $\vartheta $=1.8 mrad. It is evident from Fig.\[fig4\], which for 9 GeV electron beam energy distributions calculated without (curve 3) and with MS account (curve 4), practically coincides. In this case $L_{Br} = 1.22 \times 10^{ - 1}$cm, i.e. $L_{0} < <
L_{Br} $**,** so a situation of weak MS is realized here. Change of energy from 9 GeV down to 900 MeV yields in change of *Bremsstrahlung* coherent length to the order of magnitude, $L_{Br} = 1.22 \times 10^{ - 2}$cm**.** In this case target thickness along direction of charged particle movement $L_{0} $ becomes of one order with $L_{Br} $. Presence of MS results in essential decrease of height and spreading of PXR peak (curve 2) in comparison with the calculations, without taking MS into account (curve 1).
The spectrum corresponding to 900 MeV appeared shifted down for $\Delta \omega \sim 5 \times 10^{ - 6}\omega _{B} $ along frequency axis relative to spectrum corresponding to 9 GeV. In the same time, shift of curve 2 relative to curve 1 as a result of MS influence for 900 MeV appeared to be just $10^{ - 7}\omega _{B} $.
Amplitudes of spectral-angular distributions without MS (curves 1 and 3) has decreased less than 1.5 times, while MS account decrease it more than four times with energy decrease from 9 GeV down to 900 MeV. Comparing spectral-angular distributions for 900 MeV energy, it is possible to see that MS account results in decrease of height of spectral-angular distribution approximately 3.2 times and its spreading is $\sim $2.5 times. Thus, MS presence results in decrease of PXR angular intensity.
PXR angular distributions with MS account for the geometry described above in a crystal with thickness of 0.13 cm are plotted in Fig.\[fig5\]
= 8 cm
Detector resolution was considered equal to $\Delta \varpi /\varpi
_{B} = 10^{ - 3}$. It is evident from the figure that energy decrease to three times (from 900 down to 300 MeV) results in a decrease of maximum intensity almost five times. At the same time there is a shift of angular distribution maximum to the bigger angles from 2.9 mrad up to 4.8 mrad and spreading almost to one and half times more. Thus MS presence results in noticeable decrease of angular intensity, angular distribution spreading and to a shift of maximum in angular distribution towards the bigger angles in comparison with calculations without MS account. For example, for 300 MeV electron energy the maximum in PXR angular distribution without MS account is achieved at 2.3 mrad angle.
Study of radiation yield energy dependence is important for correct understanding of the PXR generation mechanism. This dependence, as mentioned above, has a threshold nature and its specific shape at experimental measurement in many respects is determined by the angular size of the detector. As energy decreases, the effective angle of radiation emission $\vartheta
_{ph} \approx \sqrt {\gamma ^{ - 2} - {\chi} '_{0} +
\frac{{\overline {\theta _{s}^{2}} L_{eff}} }{{2}}} $ increases. Value of *L*$_{eff}$ depends on crystal thickness, absorption length, Bragg angle, and radiation geometry [@8].
= 8 cm
Results of PXR quantum yield calculations for electron energy in a range from 300 up to 900 MeV are plotted in Fig.\[fig6\]; angular aperture of detector is 9.5$ \cdot $10$^{-3}$ rad. The angular aperture of detector is $\theta _{D} > \vartheta _{ph} $ even for 300 MeV, that is why the detector collects practically all generated radiation and we have behaviour of quantum yield energy dependence that is “natural” for proper PXR.
Presence of MS results in a steep decrease of PXR quantum yield. For example, for 700 MeV electron energy PXR quantum yield without MS exceeds the result received with MS consideration by up to 1.5 times, for 300 MeV this value already amounts $\sim $2.9 times.
Conclusion
==========
Summing up, multiple scattering of charged particles in a crystal target changes considerably the characteristics of x-rays originating due to a particle passage through a single crystal. MS effect is bigger at low energies and/or thick targets and demonstrates itself in x-rays distributions spreading, monochromaticity, and quantum yield decrease. Nevertheless, PXR at low energies can provide monochromaticity of 10$^{-3}$-10$^{-2}$ which is still applicable in medical imaging.
Using the approach described above, let us estimate the characteristics of PXR source, which can be produced on real medical accelerators [@14]. As a rule, maximal energy of electron beam in these accelerators is equal 20-28 MeV. We have evaluated angular density of radiation for next conditions: E$_{e}
$= 25 MeV, 33 keV x-rays, symmetrical Laue case for (111), (220), and (400). Silicon target of L=0.01 cm was chosen which is somewhat bigger, but more real than optimal for these conditions (about (1$\div$10)$L_{Br}$, i.e. $\sim 3 \div 30$ microns) target thickness. Angles between electron velocity direction and direction to diffraction reflex are $\sim $6.9, 11.2, and 15.9 degrees, respectively. Angular densities for discussed cases are plotted in Fig.\[fig7\].
= 8 cm
Integration on angular density gives us total amount of x-ray quanta on $20 \times 20$ cm area located at 1.5 m distance from the target. Appropriate amounts appeared equal to $\sim 3 \cdot
10^{-6}$, $\sim 5 \cdot 10^{-7}$, and $\sim 1 \cdot 10^{-7}$ photons/e$^{-}$ for (111), (220), and (400) reflexes, respectively. It is evident, that despite a decrease of radiation quantum yield at low energies, the integral amount of quanta do not decrease very drastically and the number of x-ray quanta needful for quality image can still be achieved at 0.1-0.2 A beam current.
However, to provide some margin of safety for PXR application for a medical x-ray source, it is necessary to find opportunities to increase its radiation yield. Let us consider some variants of radiation yield increase. One of the ways to increase x-rays spectral-angular density may be the generation of radiation in a multi-wave mode, in other words, when PXR emission goes simultaneously on several systems of the crystallographic planes. In [@15; @16] measurements of angular distributions of PXR generated in a GaAs crystal by relativistic electrons with energies 500 and 900 MeV were reported. In these experiments the anomalies (i.e. significant intensity increase in narrow angular range), which cannot be explained by the PXR two-wave theory [@10], were observed. The analysis of geometry has shown that conditions of multi-wave diffraction were realized for emitted photons. Thorough theoretical description for experiments [@15; @16] was performed in [@17]. In [@18] demonstration of multi-wave effects in low-energy range was numerically analysed for PXR generated by 7 MeV electrons in conditions of eight-wave diffraction. Calculations have shown that, despite of strong MS of electrons at low energies, multi-wave effects may appear in PXR angular distribution and also result in the formation of a strong narrow peak in the centre of two-wave angular distribution.
Next a rather promising way for radiation spectral-angular density increase is the application of stratified (multi-layered) crystal targets [@19]. PXR intensity increases up to 7-8 times in 10-layer target made of relatively thin silicon crystal foils in comparison with monolithic target of equivalent thickness have been reported in the paper. At last, targets made of mosaic crystals should be mentioned (for example, [@20]). They also can provide an increase in the intensity of x-rays.
Research of the behaviour of sophisticated targets at low energies, preferably at conditions that can provide existing medical accelerators, and determination of their optimal characteristics in this range, may lead to a practical design for a real medical monochromatic x-ray source. At the end, our results were obtained for Laue geometry, but for Bragg geometry MS influence can be somewhat weaker. Bragg geometry of PXR radiation can also provide more intensive x-ray yield. It is intended to be a subject of our further research.
Acknowledgements {#acknowledgements .unnumbered}
================
Authors are very grateful to Dr. O. Missevich for fruitful discussions on x-ray imaging issues.
[999999]{}
M.L. Ter-Mikaelian, High Energy Electromagnetic Processes in Condensed Media, New York: Wiley, 1972.
V.G. Baryshevsky and I.D. Feranchuk, Zh. Eksp. Teor. Fiz. 61 (1972) 944 (Sov. Phys. JETP 34 (1972) 502).
G.M. Garybyan and C. Yang, Zh. Eksp. Teor. Fiz. 61 (1972) 430 (Sov. Phys. JETP 34 (1972) 495).
Yu.N. Adishchev, V.G. Baryshevsky, S.A. Vorobiev, V.A. Danilov, S.D. Pak, A.P. Potylitsyn, P.F. Safronov, and I.D. Feranchuk, Pis’ma Zh. Eksp. Teor. Fiz. 41 (1985) 295 (Sov. Phys. JETP 41 (1985) 361).
V. Ingal and E. Beliaevskaya,
The Physics of Medical Imaging / S. Webb (Ed.), Bristol: Hilger, 1978.
R.B. Fiorito, D.W. Rule, M.A. Pierstrup et al., Nucl. Instrum. Meth. B 79 (1993) 758.
V.P. Afanasenko, V.G. Baryshevsky, A.S. Lobko, V.V. Panov, and R.F. Zuevsky, Nucl. Instrum. Meth. A334 (1993) 631.
I.D. Feranchuk, A. Ulyanenkov, J. Harada, and J.C.H. Spence, Phys. Rev. E 62, \#3 (2000) 4225.
I.D. Feranchuk and A.V. Ivashin, J. de Physique (Paris) 46 (1985) 1981.
O. Lugovskaya, Characteristics of parametric x-rays in conditions of dynamical diffraction and multiple scattering, Ph.D. Thesis, NAS Institute for Physics, Minsk, Belarus, 2003.
V.G. Baryshevsky, Channelling, Radiation and Reactions in Crystals at High Energies, Minsk: Belarussian University Publ., 1982 (in Russian).
V.G. Baryshevsky, A.O. Grubich, and Le Tien Hai, Sov. Phys. JETP 94 (1988) 51
E.A. Abramian, Industrial Electron Accelerators, Moscow: Energoatomizdat, 1986 (in Russian).
V.P. Afanasenko, V.G. Baryshevsky, O.T. Gradovsky et al., Phys. Lett. 141A (1989) 311.
V.P. Afanasenko, V.G. Baryshevsky, S.V. Gatsikha et al., Sov. JETP. Lett. 15 (1990) 242.
S.A. Stepanov, A.Ya. Silenko, A.P. Ulyanenkov, and I.Ya. Dubovskaya, Nucl. Instrum. Meth. B 117 (1996) 55.
I.Ya. Dubovskaya, In: Basic and Applied Physical Studies (1986-2001) / V. Baryshevsky (Ed.), Minsk: Belarussian University Publ., 2001 (in Russian).
Y. Takashima, K. Aramitsu, I. Endo et al., Nucl. Instrum. Meth. B 145 (1998) 25.
R.B. Fiorito, D.W. Rule, X.K. Maruyama et al., Phys. Rev. Lett. 71, \#5 (1993) 704.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The development of the Internet of Things (IoT) generates a significant amount of data that contains valuable knowledge for system operations and business opportunities. Since the data is the property of the IoT data owners, the access of the data requires the permission from the data owners, which gives rises to a potential market opportunity for the IoT data sharing and trading to create economic values and market opportunities for both data owners and consumers. In this work, we leverage optimal mechanism design theory to develop a monopolist matching platform for data trading over massive IoT networks. The proposed mechanism is composed of a pair of matching and payment rules for each side of the market to optimize the payoffs under welfare-maximization and the profit maximization schemes. We characterize the optimal mechanism with a class of threshold matching rules under welfare and profit-maximization and study three matching behaviors including separating, bottom pooling, and top pooling. We use `HealthyGo` as a case study to illustrate the optimal threshold matching rules and corroborate the analytical results.'
author:
- 'Tao Zhang and Quanyan Zhu$^\dagger$[^1]'
bibliography:
- 'acc\_DMT.bib'
title: ' **Optimal Two-Sided Market Mechanism Design for Large-Scale Data Sharing and Trading in Massive IoT Networks** '
---
Introduction
============
As the number of cyber and physical devices are connected to the existing Internet infrastructure, the Internet of Things (IoT) provides connectivity and network solutions to emerging applications such as smart homes, enterprises, and cities [@perera2015emerging]. Each application consists of a large number of IoT objects, and they are designed to perform heterogeneous tasks, including sensing, auction, data collection, storage, and processing. Data collected for fulfilling the individual tasks are often either deleted or stored and locked down in independent data silos, as shown in Fig. \[Fig\_IoT\_example\]. The data generated in the IoT system has a significant amount of valuable knowledge hidden, such as behaviors, life patterns, and habits, which can be used to improve the human lives as well as the IoT system itself. Enterprises can benefit from big data analytics to reduce cost, help product development, and optimize marketing strategies. The data owners, however, generally lack knowledge and techniques to conduct knowledge discovery from their data, and each owner can only gain access to her data which has little value with respect to knowledge discovery [@perera2017sensing]. On the other hand, the enterprises or other data consumers have abilities to discover the hidden knowledge accurately and efficiently, but they may not have a sufficient amount of targeted data or the rights to access them due to privacy concerns. As a result, data sharing over the IoT network is important to create economic values and market opportunities for both data owners and consumers.
![IoT example: Data is collected for primary usage. After that, the data is deleted or locked in independent data silos.[]{data-label="Fig_IoT_example"}](Figure/IoT_data){width="50.00000%"}
{width="100.00000%"}
Since the data is the property of the IoT data owners, post-primary usages of the data require the permission from the data owners, leading to a potential market opportunity for the IoT data trading, instead of the free data sharing. The difference in the roles in dealing with data can divide the market participants into two distinct classes, i.e., the data owners (sellers) and data consumers (buyers). Therefore, a two-sided market (e.g., [@rochet2006two]) is a suitable market model for IoT data trading. Data consumers have different requirements on the features of the data traded (including category, history, quality, and quantity of the data), the owner’s engagements with the IoT system (including the usage frequency of IoT products), and the user’s personal information (including life styles and income). On the other hand, the data owners have different preferences for the rewards, including cash and coupons, offered by the data consumers for trading the data and different tolerance levels of privacy violation that might occur in the data trading. The data consumers’ objective is to buy valuable data that satisfy their requirements while the data owners’ goal is to sell the data to those service providers who offer the highest rewards based on their preferences.
For example, `HealthyGo` is a healthy food provider, who is keen to know whether there are external factors that influence the healthy food consumption such as workout patterns, sleep quality, weather, and temperature through data analysis. Therefore, `HealthyGo` is looking for IoT data of household products that can collect such data. `HealthyGo` has barriers to obtain such data in an efficient and low-cost way. On the other hand, a family has a smart refrigerator with built-in sensors that can monitor the food consumption, a smart thermostat that controls the room temperature, and smart fitness trackers that provide heart rate monitoring, sleep tracking, and sports tracking. However, generally the data collected has little value after the primary usage. Therefore, a data trading platform is important to enable mechanisms to take advantages of both the data owner and the data consumers and provides motivations of participating while ensuring the interests of all parties in the IoT system.
Matching intermediary platforms are usually introduced to tie together multiple distinct classes of participants (i.e., users and service providers in this paper) in a market. The business of a matching intermediary is to match participants from multiple sides of a market. In the two-sided market, the platforms play a role to enable and encourage interactions between two sides of participants by appropriately charging each side for offering matching rules. There can be multiple platforms or a monopolist platform in a market. In this work, we focus on massive IoT networks with a sufficiently large number of owners and consumers. We develop a mechanism design approach and propose a two-sided matching market model for the IoT data trading as a sustainable business model that captures the potential market opportunity in the IoT data sharing mechanism. As illustrated in Fig. 2, data owners have IoT data generated during the primary usages of different IoT products to sell, and the data consumers have different requirements on the data demand. The data consumers offer rewards to the data owners for trading their data. A monopolist matching platform matches the data owners and the data consumers based on their information and preferences by charging both sides with an appropriate price. The price charged for offering matching rules generates profits for the platform. In the IoT data market framework, the matching rule determines the datasets that achieve the highest utility of the data consumers. Since the matching is reciprocal, the data traded (according to the matching rule) by the data owners, on the other hand, determines the level of rewards they can receive from the data consumers. In other words, the matching can generate welfare over the IoT network. Our framework considers the interactions between a large number of data buyers and sellers, and study the matching rules under profit and welfare maximization schemes [@hagiu2014information], respectively. We study a class of threshold matching rules and show that under certain assumptions both the profit-maximizing and the welfare-maximizing matching rules have optimal threshold matching rules.
Our main contribution of this work is summarized as follows:
- We leverage optimal mechanism design theory to develop an incentive-compatible optimal matching mechanism for data trading over massive IoT networks. The developed mechanism is composed of a pair of matching and payment rules for each side of the market to optimize the payoffs under profit and welfare maximization schemes.
- We develop a monopolist trading platform with two matching stages for massive IoT networks with a large number of devices. The framework captures the types of the data owners and consumers by their characteristics and valuations, which are private information unknown to the platform.
- We characterize the optimal mechanism with a class of threshold matching rules under welfare-maximization and the profit-maximization, and study three matching behaviors including separating, bottom pooling, and top pooling. We use a case study to show the numerical analysis of the optimal threshold matching rules.
Related Work
------------
There is a rich literature on the economic analysis and pricing schemes of the market model for data collection in IoT networks [@aazam2015fog; @perera2014sensing; @al2013priced] based on a variety of approaches, including smart data pricing scheme in [@sen2015smart] and auction-based pricing such as sealed-bid auctions [@sun2006wireless]. Also, utility maximization-based pricing schemes have also been studied [@lee2005non; @lee2006jointly; @farooq2018adaptive]. For example, in [@farooq2018optimal], an optimal dynamic spectrum reservation contract has been designed for mission-critical IoT systems. Devices can make an advance payment at the time of reservation and receive a rebate if the reservation is released. Devices are incentivized to reveal the true application type and it has been shown that the incentive compatible mechanism leads to an efficient utilization of the spectrum as well as a greater profitability of the IoT network operator. The IoT data market model in this work is a two-sided matching model with a monopolist intermediary matching platform. The trading is made directly between the original data owners and the final data consumers. The role of the platform is to provide matching rules to both sides of the market. The goal of the market model is to establish optimal matching rules under the profit-maximization for the platform and the welfare-maximization for the massive IoT network. Also, unlike the auction-based model, the model does not describe competitions on the side of data owners, thereby leading to a convenient monopolist matching rule that matches sellers and buyers in an efficient way. Furthermore, the reciprocal nature of the matching also prevents the monopolist platform from appropriating the entire surplus of the market to himself [@gomes2011price].
Organization of the Paper
-------------------------
The rest of the paper is structured as follows. Section \[Sec\_2\] describes the IoT data market framework. We first describe the market structure, followed by elaborating the mathematical market model. Section \[Sect\_3\] presents the matching mechanism by considering the welfare-maximizing and the profit-maximizing schemes. In Section \[Sect\_4\], we present optimal matching rules by considering a class of threshold matching rules with the theoretical analysis. A case study is conducted in Section \[Sect\_5\]. Finally, Section \[Sect\_6\] presents the concluding remarks and future research directions.
IoT Data Market {#Sec_2}
===============
In this section, we present the framework of the IoT data trading market, by starting with describing the market structure and followed by the mathematical model of the market.
Market Structure
----------------
In the scenario of IoT, the data market has two sides: the seller side ($S$) constituting of (market) participants that are the IoT data owners, and the buyer side ($B$) constituting of participants who are the IoT data consumers. We consider a massive IoT networks, in which there is a sufficiently large number of participants on both sides of the market, and each participant has no market power. As a result, the population of each side of the market is modeled by a unit-mass continuum of participants indexed by $i\in[0,1]$. A monopolistic market platform (platform) operates as a matching intermediary to providing matching rules with different prices that match the owners from side $S$ and consumers from side $B$. The pricing of matching rules in this market considers the second-degree price discrimination [@phillips2005pricing; @gomes2011price], in which the price of matching rules varies subject to the number of participants involved [@phillips2005pricing]. In this IoT data market model, the matching rule determines the benefits of the consumers in the trading. Since the matching is reciprocal, the data sold by the owners, on the other hand, determines the rewards they can receive from the data consumers for trading their data. Our model takes the two sides of the market into account and studies the matching rules under profit-maximization and welfare-maximization [@hagiu2014information], respectively.
Each participant $i$ from side $K\in\{S,B\}$ has private information about her *willingness to pay* for the quality of matching rules offered by the platform as well as her *market attractiveness* as perceived by the participants from the opposite side $L\neq K$. The platform has to internalize the existence of such private information and provide incentives in terms of trading matching quality and pricing rules for the participants when designing the matching rules. We consider two matching stages. In the first matching stage, the platform matches a group of owners and a group of consumers based on the data categories that the owners have and the consumers require. In the second matching stage, the responsibility of the platform is to design optimal matching rules under welfare-maximizing and profit-maximizing schemes, respectively, by considering the willingness-to-pay and the market attractiveness reported by the participants.
Market Model
------------
Each participant $i\in[0,1]$ from side $K\in\{S,B\}$ has information summarized as $\zeta =(\mu^K_i, \lambda^K_i)$, where $\mu^K_i$ represents the data categories, and $\lambda^K_i = (c^K_i, v^K_i) \in \Lambda^K = \mathbf{C}^K \times \mathbf{V}^K$ is the type of the participant $i$. The IoT data can be classified into different categories including household data (collected by household IoT products), fitness data (e.g., healthy food intake, workout information), experience data (e.g., information about road conditions collected by navigation system, environment of a neighborhood collected by outdoor surveillance cameras). Each type $\lambda^K_i = (c^K_i, v^K_i)$ includes the *characteristic* $c^K_i \in \mathbf{C}^K$ that includes the related information of the participant, and *valuation* $v^K_i \in\mathbf{V}^K = [\underline{v}_K, \bar{v}_K]\subset \mathbb{R}_+$, $K\in\{S,B\}$ that measures each participant’s satisfaction of the trading according to the matching rules offered. The type of each participant determines her base preferences over different matching rules in the IoT data trading market. The set $\mathbf{C}^K$ of characteristic $c^K_i$ can be discrete or continuous, while the support of valuation $\mathbf{V}^K $ is a real interval. For owner $i$ from side $S$, the data category describes the categories of the IoT data owned. The characteristic $c^S_i$ contains the data size and the statistics of the DO’s engagements with the IoT system (for example, the frequency of usage of IoT products, expense related to IoT, etc.). The valuation $v^S_i$ summarizes the owner $i$’s *willingness to pay* for the quality of data trading matching rules made by the platform based on her characteristics, the estimation of the rewards for selling her data, and her tolerance of privacy violation.
For consumer $j\in[0,1]$ from side $B$, the data category describes the categories of the IoT data wanted, and the characteristic $c^B_i$ includes the rewards offered to the matched owners. The reward $r_j(\lambda^S_i): \Lambda^S\rightarrow \mathbb{R}_+$ is a function of owner $i$’s type that is observed by the consumer $j$. A typical reward includes cash, vouchers, coupons, free products, loyalty points, and actionable advice, etc. The valuation $v^B_j$ summarizes consumer $j$’s willingness to pay for the quality of data trading matching rules made by the platform by taking into account the reward she offers to the data owners.
In our model, we assume that each type $\lambda^K_i$ of each participant $i$ from side $K\in\{S,B\}$ is drawn independently from a continuous distribution $F^K(\lambda^K_i) = F^K(c^K_i, v^K_i)$ with density $f^K(\lambda^K_i) = f^K( c^K_i, v^K_i )$. Let $F^K_v$ with density $f^K_v$ be the marginal distribution of $F^K$ with respect to the valuation $v^K_i$. A standard assumption in mechanism design is as follows.
\[Asmp\_1\] *The virtual valuation of participant $i$ from side $K$, $$v^K_i-\frac{1-F^K_v(v^K_i)}{f^K_v(v^K_i)}$$ which is a continuous and non-decreasing function of $v^K$ that measures the surplus that can be extracted from that participant [@myerson1981optimal]. For $F^K_i$ that satisfies the virtual valuation, we call it regular.*
Let $\alpha^K_i(\lambda^L_j): \Lambda^L \rightarrow \mathbb{R}_+$, $K\neq L \in \{S, B\}$ model the *interaction quality* that the participant $i$ from side $K$ obtains from being matched to a participant $j$ from the opposite side $L$. The interaction quality $\alpha^K_i( \lambda^L_j )$ measures the *market attractiveness* of the participant $j$ from side $L$ as perceived by the participant $i$ from side $K$. For example, owners with more frequent usages of IoT product induces relatively higher interaction quality (i.e., more attractive) than those with less frequent usage as seen by the consumer. The interaction quality $\alpha^S_i( \lambda^B_j )$ enjoyed by the owner $i$ models the profit generated by matching to consumer $j$ by measuring the monetary value of the non-monetary reward (i.e., vouchers, coupons, free products, loyalty points, and actionable advice) $r_i(\lambda^S_i)$ by taking into account the usefulness as seen by the owners.
![Data trading model: Each data owner and each data consumer report their data categories (true values) and types (reported values) to the platform. At the first matching stage, the platform matches the groups of owners and the groups of consumers according to their data category. At the second matching stage, the platform uses the reported types and returns the optimal matching rules to the market participants.[]{data-label="Fig_platform"}](Figure/Platform_match){width="50.00000%"}
The *trading matching quality* measures the cumulative interaction qualities of a set of participants. Let $g^L$ be a Borel measurable set of participants from side $L \in \{S,B \}$ interacted with by each participant from side $K\neq L \in \{S,B \}$. The trading matching quality of any set $g^L$ of participants from side $L\neq K$ with type profile $\{\lambda^L_j\}_{j\in g^L}^{L\neq K\in \{S,B\}}$, which is perceived by each participant $i$ from set $K$, is modeled as the sum of interaction qualities $\alpha^K_i(\lambda^L_j)$. Let $\beta^K(g^L)$, $K\neq L\in \{S, B\}$ denote the trading matching quality defined as $$\label{quality}
\beta^K(g^L) := \int_{j\in g^L} \alpha^K_i(\lambda^L_j) dF^L(\lambda^L_j),$$ for $K\neq L \in \{S,B\}$. In this model, all participants from side $K\neq L \in\{S,B\}$ agree on the value of $\alpha^K_i(\lambda^L_j)$, hence on the value of $\beta^K(g^L)$, for all $j$ in side $L$. The *marginal utility* of each participant $i$ from side $K\in\{S,B\}$ with $\beta^K(g^L)$ is modeled by a function denoted as $\gamma^K_i(\beta^K(g^L))$, which is a positive, strictly increasing, and continuously differentiable function. Given any set $g^L$ of participants from side $L\in\{S,B\}$, the payoff that a participant $i$ from side $K\neq L$ with type $\lambda^K_i$ obtains from being matched to the set $g^L$ is given by
- Buyer side $B$: $$\label{payoff_1}
\pi^B_i(g^S, p; \lambda^B_i):= v^B_i \gamma^B(\beta^B(g^S)) -r_i-p.$$
- Seller side $S$: $$\label{payoff_2}
\pi^S_i(g^B, p; \lambda^S_i):= v^S_i \gamma^S(\beta^S(g^B)) -p.$$
Here, $p$ is the price paid by each participant from each side to the platform. Each participant $i$ from side $K\in\{S,B\}$ is required to report her type $\lambda^K_i=(c^K_i, v^K_i)$ to the platform. Let $\hat{\lambda}^K_i = \hat{c}^K_i, \hat{v}^K_i)$ be the reported type by participant $i$ from side $K$, as shown in Fig. \[Fig\_platform\].
Matching Mechanism {#Sect_3}
==================
In this paper, we focus on the *anonymous* mechanism design. Specifically, the set of participants from side $L\neq K \in\{S,B\}$ that are matched to the participant $i$ from side $K$ and the corresponding payment $p$ depends only on the reported type $\hat{\lambda}^K_i$ by $i$. Also, if the reported types $\hat{\lambda}^K_i = \hat{\lambda}^K_k$, then the participant $i$ and $k$ are matched to the same set and the same payment $p$ is required; in other words, the individual identities in the same side are irrelevant if they have the same type. Thus, for simplicity, the subscript indicating the identity is suppressed in the rest of the paper. Let $h^L\subset \Lambda^L$ be a Borel measurable set of types $\lambda^L$ of participants from side $L\in\{S,B\}$. Then, the trading matching quality (\[quality\]) becomes $$\label{TMQ}
\beta^K(h^L) := \int_{\lambda^L\in h^L} \alpha^K(\lambda^L) dF^L(\lambda^L),$$ for $K\neq L \in \{S,B\}$.
Our model considers the *direct-revelation* [@myerson1979incentive; @myerson1981optimal] anonymous mechanisms $M = \{\tilde{h}^K(\cdot), \tilde{p}^K(\cdot) \}^{K\in\{S,B\}}$. The rule $\tilde{h}^K(\lambda^K)\in \Lambda^L$, $K\neq L\in\{S,B\}$, is the matching rule, which specifies the set of types of participants from side $L$ that is included in the matching set of all participants from side $K$ with the type $\lambda^K$. The rule $\tilde{p}^K(\lambda^K)$ specifies the payment required to all participants from side $K$ with the type $\lambda^K$. The feasible matching rule is defined as follows.
\[def\_1\] *A *matching rule* $\tilde{h}^K(\cdot)$, $K\in\{S,B \}$ is feasible if and only if the following reciprocity condition is satisfied $$\label{reciprocity}
\lambda^B\in \tilde{h}^S(\lambda^S) \leftrightarrow \lambda^S \in \tilde{h}^B(\lambda^B),$$ that is, if the type $\lambda^B$ of the consumer is in the matching set of the owner with type $\lambda^S$, then the type $\lambda^S$ is also in the matching set of the consumer with type $\lambda^B$.*
The payoff that a participant with type $\lambda^K = (c^K, v^K)$ from side $K\in\{S,B \}$ obtains when reporting type $\hat{\lambda}^K = (\hat{c}^K, \hat{v}^K)$ is defined as
- Buyer side $B$: $$\label{payoff_1}
\bar{\Pi}^B(\lambda^B, \bar{\lambda}^B; M):= v^B\cdot \gamma^B(\beta^B(\tilde{h}^B(\hat{\lambda}^B)))-r-\tilde{p}^B(\hat{\lambda}^B).$$
- Seller side $S$: $$\label{payoff_1}
\bar{\Pi}^S(\lambda^S, \bar{\lambda}^S; M):= v^S\cdot \gamma^S(\beta^S(\tilde{h}^S(\hat{\lambda}^S))) + r -\tilde{p}^S(\hat{\lambda}^S).$$
Denote by $\Pi(\lambda^K;M) = \bar{\Pi}^K(\lambda^K, \lambda^K; M)$ is the payoff that a participant with type $ \lambda^K$ obtains when reporting truthfully. Based on the direct revelation principle [@myerson1979incentive], we focus on the individual rationality and incentive compatibility constraints defined as follows.
\[def\_IR\] *A mechanism $M$ is individually rational (IR) if $\forall \lambda^K$, $ \bar{\lambda}^K\in \Lambda^K$, $$\Pi^K(\lambda^K;M) \geq 0 \;\;\;\forall \lambda^K\in \Lambda^K.$$*
\[def\_IC\] *A mechanism $M$ is incentive compatible (IC) if $\forall \lambda^K$, $ \bar{\lambda}^K\in \Lambda^K$ , $$\Pi^K(\lambda^K;M)\geq \bar{\Pi}^K(\lambda^K, \hat{\lambda}_K; M).$$*
A matching rule $\tilde{h}^K(\cdot)$, $K\in\{S,B \}$ is *implementable* if there is a payment rule $\{\tilde{p}^K\}_{K\in\{S,B\}}$ such that the mechanism $M$ is both IR and IC.
Welfare and Profit
------------------
Next, we introduce the welfare and the profit generated by market matching. The total welfare of the market participants from both sides generated by the mechanism $M$ is $$\label{welfare}
U^W(M) = \sum_{K = S,B} \int_{\Lambda^K} v^K \gamma^K(\beta^K( \tilde{h}^K(\lambda^K) )) d F^K(\lambda^K).$$ The profits generated by the mechanism $m$ for the platform is given by $$\label{profit}
U^P(M) = \sum_{K = S,B} \int_{\Lambda^K} \tilde{p}^K(\lambda^K) d F^K(\lambda ^K).$$ The mechanism $M$ is *efficient* (resp. *profit-maximizing*) if $M^W=\arg\max_{M} U^W(M)$ (resp. $M^P=\arg\max_{M} U^P(M)$), and the matching rule $\tilde{h}^K$, $K\in\{S,B \}$ of $M$ is feasible and implementable. We then have the following lemma.
\[lemma\_ICIR\]
We can see that in the profit-maximizing mechanisms, the payoff of the participants with the lowest valuations $\underline{v}^K$ satisfies $\Pi^K((c^K, \underline{v}^K); M)=0$, $\forall K\in\{S,B\}$, due to the IR constraints. Denote the valuation associated with welfare maximization by $\theta^K_W(v^K) := v^K$, and by $\theta^K_P(v^K):= v^K-\frac{1-F^K_v(v^K)}{f^K(v^K)}$ denote (virtual) valuation associated with profit maximization. Then, by the envelope formula (\[envelope\]), for $I = W, P$, (\[welfare\]) and (\[profit\]) can be rewritten as $$\label{max_WP}
U^I(M) = \sum_{K=S,B}\int_{\Lambda^K} \theta^K_I(v^K)\cdot\gamma^K(\beta^K(\tilde{h}^K_I( c^K,v^K ) ) ) dF^K_v(v^K)$$ Therefore, the platform’s responsibility includes finding an optimal matching rule $\{\tilde{h}_I^K\}^{K\in\{S,B\}}_{I\in\{ W,P\}}$, by maximizing (\[max\_WP\]) under the conditions in Lemma \[lemma\_ICIR\]. A matching rule ${\tilde{h}^{K*}_W}$ (resp. ${\tilde{h}^{K*}_P}$) is welfare-optimal (resp. profit-optimal) if $\tilde{h}^{K*}_W = \arg\min_{\tilde{h}^K_W} U^W(M)$ (resp. $\tilde{h}^{K*}_P = \arg\min_{\tilde{h}^K_P} U^P(M)$).
Optimal Matching Rules {#Sect_4}
======================
In this section, we present the optimal matching rules for the data market matching model. First, a fairly natural assumption is as follows.
\[asmp\_TP\] The interaction quality $\alpha^K$, the distribution of type $F^K$, and the marginal utility function $\gamma^K(\cdot)$, $\forall K \in\{S,B\}$, satisfy *one* of the following two:
- Under the distribution $F^K$, $(\alpha^K(\lambda^K), v^K)$, $\forall K\in\{S,B \}$, are weakly positively affiliated; the marginal utility function $\gamma^K(\cdot)$ is weakly concave and $\alpha^K$ is weakly increasing.
- Under the distribution $F^K$, $(\alpha^K(\lambda^K), v^K)$, $\forall K\in\{S,B \}$, are weakly negatively affiliated; the marginal utility function $\gamma^K(\cdot)$ is weakly convex and $\alpha^K$ is weakly decreasing.
We then have the following proposition.
\[prop\_1\] Under Assumption \[asmp\_TP\], both the welfare-maximizing and the profit-maximizing rules discriminate only along the valuation $v^K$, i.e., $\tilde{h}^K(c^K, v^K) = \tilde{h}^K({c^K}', v^K)$, $\forall v^K, c^K\neq {c^K}', K\in\{S,B\}$.
Based on Proposition \[prop\_1\], we re-define $\tilde{h}^K_I(v^K)\subset V^L$, $\forall K\neq L\in\{S,B\}$, and $\forall I\in\{W,P\}$.
Feasible Threshold Rules
------------------------
In this paper, we consider a class of feasible threshold rules [@gomes2011price] defined as follows.
*Let $y^K_I(\cdot):V^K\rightarrow V^K$ be a weakly decreasing threshold function such that $\forall K\in\{A,B \}$, $\forall I \in \{W,P\}$, $$\label{feasible_y}
y^K_I( v^K) = \min\{v^L: y^L_I(v^L)\leq v^K\}.$$ A matching rule $\tilde{h}^K$ is a feasible threshold rule if, $\forall$ $v^K\in V^K$, $K\in\{A,B \}$, $I \in \{W,P\}$, $$\label{threshold}
\tilde{h}^K_I(v^K)= \begin{cases}
[y^K_I(v^K), \bar{v}^L ], & \text{if } v^K\in [ \delta^K_I, \bar{v}^K]\\
\emptyset, & \text{otherwise}
\end{cases},$$ where $\delta^K \in [\underline{v}^K, \bar{v}^K]$ is the threshold value. We say that $(y^K_I(v^K), \delta^K_I)$ is the feasible threshold structure associated with the threshold rule $\tilde{h}^K((c^K, v^K))$ .*
Under the threshold rule, the participant with valuation below $\delta^K$ are excluded, while the participant from side $K\neq L \in\{S,B\}$ with valuation $v^K\geq \delta^K$ is matched to any participant from side $L$ whose valuation is within $[y^K_I(v^K), \bar{v}^L ]$. The condition (\[feasible\_y\]) of $y^K$ means that the weakly decreasing threshold function $y^K_I(v^K)$, $\forall K\in \{S,B \}$, coincides with the inverse of the threshold function $y^L_I(v^L)$, $L\neq K$. Thus, the threshold rule $\tilde{h}^K$, i.e., the reciprocity condition \[reciprocity\] in Definition \[def\_1\] is satisfied. Moreover, the participants of side $K\in\{S,B\}$ with low $v^K$ are matched only to those participants from side $L\neq K$ with sufficiently high value of $v^L$. Also, the matching sets are ordered across valuations, i.e., $\tilde{h}^K_I(v^K)\subseteq \tilde{h}^K_I({v^K}')$ for $v^K< {v^K}'$. We summarize the optimal matching rule in the sense of feasible threshold rules in the following theorem.
\[proposition\_exist\] Under Assumption \[asmp\_TP\], both the welfare-maximizing and profit-maximzing rules $\tilde{h}^K$ have a feasible threshold structure $(y^K_I, \delta^K)$, $\forall K\in\{S,B\}$, $\forall I\in\{S,B\}$.
Analysis of Optimal Threshold Rules
-----------------------------------
In this subsection, we analyze the property of the optimal threshold rules given the result in Proposition \[proposition\_exist\]. We first define the separation, and pooling in the sense of two-sided market in the following definitions.
\[separation\] The $I$-optimal matching rule $\tilde{h}^{K*}_I$, $K\neq L\in\{S,B\}$ and $I\in\{W,P\}$, shows *separation* if $\exists \tilde{V}^K\subset V^K$, $y^K_I(v^K)\neq y^K_I({v^K}')$, $\forall v^K\neq {v^K}' \in \tilde{V}^K$, and $[\delta^K_I, y^L_I(\delta^L_I)]$ is the *separating range*. The matching rule is *maximally separating* if $y^K_I(\cdot)$ is strictly increasing in the separating range.
\[EB\] The $I$-optimal matching rule $\tilde{h}^{K*}_I$, $K\neq L\in\{S,B\}$ and $I\in\{W,P\}$, shows
- *bottom pooling* if $\delta^K_I>\underline{v}^K$, i.e., all participants with valuation in a neighborhood of $\underline{v}^K$ are assigned *empty* matching sets;
- *top pooling* if $y^L_I(\delta^L_I) < \bar{v}^K$, i.e., all participants in a neighborhood of $\bar{v}^K$ are assigned the same matching sets.
Let $\Gamma^K(\cdot): V^L \rightarrow \mathbb{R}_+$ be defined as $$\begin{aligned}
\Gamma^K(v^L):= &\gamma^K(\int^{\bar{v}^L}_{v^L}\alpha^K((c^L, x))d F^L_v(x)).
\end{aligned}$$ Therefore, the platform’s responsibility includes choosing a bottom pooling threshold rule with a feasible threshold structure $(y^K_I(v^K), \delta^K_I)^{K\in\{S,B\}}_{I\in\{W,P\}}$ by maximizing, $\forall$ $K\in \{S,B\}$ and $I\in\{W,P\}$, [$$\label{max_WP_2}
\begin{aligned}
U^I(M) =& \sum_{K=S,B}\int_{\delta^K_I}^{ \bar{v}^K } \theta^K_I(v^K)\cdot\Gamma^K(y^K_I(v^K)) dF^K_v(v^K)\\
\text{s.t.\;\;}& y^K_I( v^K) = \min\{v^L: y^L_I(v^L)\leq v^K, \}, \forall v^K\in [\delta^K_I, \bar{v}^K]
\end{aligned}$$]{} The following assumption is based on Myerson’s regularity condition [@myerson1981optimal] that is used to guarantee that the optimal matching rule $\tilde{h}^{K*}_I$ is maximally separating.
\[asp\_matching\_regularity\] The function $\Theta^K_I: V^K\rightarrow \mathbb{R}$ defined as, $K\neq L\in \{S,B\}$, $I\in\{W,P\}$, $$\label{matching_regularity}
\begin{aligned}
\Theta^K_I(v^K) :=& \frac{ f^K_v(v^K) \theta^K_I(v^K) }{ -\frac{d }{dv^K} \Gamma^L(v^K) }\\
=& \frac{ \theta^K_I(v^K) }{ \frac{d}{d v^K}\gamma^L(\mathbf{C}^K\times [v^K, \bar{v}^K] ) \mathbb{E}(\alpha^L({c^K}', v_k)) }
\end{aligned}$$ is strictly increasing.
Assumption \[asp\_matching\_regularity\] is interpreted as follows. Consider the case of profit maximization, i.e., $I=P$. Let $\theta^K_P(v^K):= v^K-\frac{1-F^K_v(v^K)}{f^K(v^K)}$ be the virtual valuation of the participant of side $K$ with valuation $v^K$, and the denominator $\frac{d}{d v^K}\gamma^L(\mathbf{C}^K\times [v^K, \bar{v}^K] ) \mathbb{E}(\alpha^L({c^K}', v_k))$ is proportional to the expected interaction quality. Thus, Assumption \[asp\_matching\_regularity\] requires that the contribution from the willingness-to-pay (i.e., valuation) of a participant from side $K$ to the platform’s revenue grows faster than her contribution from the attractiveness (i.e., interaction quality).
{width="100.00000%"}
{width="100.00000%"}
{width="80.00000%"}
{width="80.00000%"}
Let $D^K_I(v^K,v^L):V^K\times V^L \rightarrow \mathbb{R}$ be the *marginal surplus* defined as [$$\begin{aligned}
D^K_I(v^K,v^L)=&-\frac{d}{dv_L} \Gamma^K(v^L)\cdot \theta^K_I(v^K)\cdot f^K_v(v^K) \\
&- \frac{d}{dv_k} \Gamma^L(v^K)\cdot \theta^L_I(v^L)\cdot f^L_v(v^L).
\end{aligned}$$]{} The following theorem summarizes the optimal threshold matching rules.\
\[theorem\_1\] Under Assumption \[asmp\_TP\] and \[asp\_matching\_regularity\], we have the following, $\forall K\neq L\in\{S,B\}$, $\forall I\in\{W,P\}$,
- If $D^K_I(\underline{v}^K, \underline{v}^L)\geq 0$, then the $I$-optimal matching rule is $\tilde{h}^{L*}_I(v^K) = V^L$, $\forall v^K\in V^K$.
- If $D^K_I(\underline{v}^K, \underline{v}^L)< 0$, then the $I$-optimal matching rule is maximally separating with
- If $D^K_I(\bar{v}^K, \underline{v}^L)>0$, then it shows top pooling at side $K$ and no bottom pooling at side $L$.
- If $D^K_I(\bar{v}^K, \underline{v}^L)<0$, then it shows bottom pooling at side $L$ and no top pooling at side $K$.
- The threshold function $y^K_I$ satisfies $
D^K_I(v^K, y^K_I(v^K)) = 0,
$ $\forall v^K\in[\delta^K_I, y^L_I(\delta^L_I) ]$.
Theorem \[theorem\_1\] indicates that when the marginal surplus determined on the lowest valuations of any data owners and consumers, the optimal threshold matching rule shows separation. When such marginal surplus is non-negative, any data owner is matched to any data consumers.
Numerical Analysis {#Sect_5}
==================
In this section, we present numerical analysis for the second matching stage to illustrate the optimal matching rules under the welfare and the profit maximization.
*Mr. and Ms. Smith have a daughter. Their house has several household IoT devices. A smart refrigerator monitors the category, quantity, and freshness of food stored inside. A smart thermostat controls the room temperature based on the preference of the family members. They also have a fitness tracker for each family member that monitors their workout patterns and tracks their sleeping. `HealthyGo` is a health supplement provider. It aims to know the patterns of healthy food (e.g., a variety of vegetables, low-calorie food, etc.) consumption of families like the Smith. `HealthyGo` wants to figure out whether there are external factors such as workout patterns, sleeping quality, weather, and temperature that influence the healthy food consumption. Traditionally, these studies are done via focus group interview and customer survey. By trading the IoT data from thousands of families like the Smith, `HealthyGo` can understand the consumption patterns better through data analysis. `HealthyGo` can benefit from such data analysis and help new product developments and optimize the supply chain to increase the sales and reduce the cost. These advantages motivate `HealthyGo` to participate in the IoT data trading. As a data owner, the Smith family receives rewards from trading the IoT data with `HealthyGo`.*
In the numerical analysis, consider the case when $\mathbf{V}^K\in \mathbb{R}_+$, $\alpha^K((c^K,v^K)) = v^K$, $\gamma^K(x) = x$, $\forall K\in\{S,B\}$. Suppose that all valuations $v^K$ are drawn independently from a uniform distribution over $V^K = [1, 10]$ for $K\in\{S,B\}$, i.e., $F^K_v = \frac{v^K - 1}{9}$. Then, $\Gamma^K(v^L) = \frac{1}{9} \int^{10}_{v^L} xd F^L_v(x)=\int^{10}_{v^L} x d x$. For $I=W$, we have $$D^K_W(\underline{v},\underline{v})= \frac{2}{ 81} >0,$$ and by Theorem \[theorem\_1\], the welfare-optimal matching rule is $\tilde{h}^{L*}_W(c^K,v^K) = V^L$, $\forall v^K\in \mathbf{V}^K$, a shown in Fig. \[welfare\_fig\]. Fig. \[welfare\_opt\] shows the change of welfare when $\delta^K$ changes. As can be seen, the welfare is maximized when $\delta^K=1$, which coincides with the optimal $\tilde{h}^{L*}_W((c^K,v^K))$. For $I=P$, we have $$\begin{aligned}
D^K_P(\underline{v},\underline{v})=& \frac{-16}{81}<0,
\end{aligned}$$ and $$D^K_P(\bar{v},\underline{v}) = \frac{-70}{81}<0,$$ and by Theorem \[theorem\_1\], the profit-optimal matching rule is maximally separating and shows bottom pooling at side $L$ and no top pooling at side $K$. The threshold function is found as $$y^K_P(v^K) = \frac{10v^K}{4v^K- 10},$$ and thus from $y^K_P(\delta^K_P) = 10 $, we have $\delta^K_P = \frac{10}{3}$, $\forall K\in\{S,B\}$. Then, the corresponding profit-maximizing threshold matching rule, as shown in Fig. \[profit\_fig\], is $$\tilde{h}^K_P((c^K,v^K) ) =\begin{cases}
[ \frac{10v^K}{4v^K- 10}, 10], &\text{if } v^K\in [\frac{10}{3}, 10]\\
\emptyset, & \text{otherwise}
\end{cases}.$$ Fig. \[profit\_opt\] shows the change of profit when $\delta^K$ varies in $[1,10]$. As shown in \[profit\_opt\], the profit is non-negative only when $\delta^K \geq \frac{10}{3}$, which coincides with the optimal $\tilde{h}^K_P((c^K,v^K) )$ obtained above.
Conclusion {#Sect_6}
==========
In this paper, we have proposed a two-sided matching market framework for IoT data trading as a sustainable pricing model that incentivizes both the data owners and the data consumers. A monopolist platform has been introduced to match the data owners and the data consumers based on their knowledge and preferences. We have established a quantitative framework to model the IoT data trading market for welfare and profit maximizations. This work has characterized a class of feasible threshold matching rules. Under mild assumptions of the utility functions and the distributions of the valuations, there exist optimal threshold matching rules that maximize the welfare and the profit. We have used HealthGo IoT networks as a case study to understand the behavior of participants from both sides of the market under the optimal threshold matching rules. Our future work would extend the market design framework by introducing multiple platforms among heterogeneous IoT networks and understanding competitive behaviors among platforms.
[^1]: $^\dagger$ Both authors are with the Department of Electrical and Computer Engineering, New York University, Brooklyn, New York, USA, 11201, Email: [{tz636, qz494}@nyu.edu]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study global variational properties of the space of solutions to $-{\varepsilon}^2\Delta u + W''(u)=0$ on any closed Riemannian manifold $M$. Our techniques are inspired by recent advances in the variational theory of minimal hypersurfaces and extend a well-known analogy with the theory of phase transitions. First, we show that solutions at the lowest positive energy level are either stable or obtained by min-max and have index 1. We show that if ${\varepsilon}$ is not small enough, in terms of the Cheeger constant of $M$, then there are no interesting solutions. However, we show that the number of min-max solutions to the equation above goes to infinity as ${\varepsilon}\to 0$ and their energies have sublinear growth. This result is sharp in the sense that for generic metrics the number of solutions is finite, for fixed ${\varepsilon}$, as shown recently by G. Smith. We also show that the energy of the min-max solutions accumulate, as ${\varepsilon}\to 0$, around limit-interfaces which are smooth embedded minimal hypersurfaces whose area with multiplicity grows sublinearly. For generic metrics with ${\rm Ric}_M>0$, the limit-interface of the solutions at the lowest positive energy level is an embedded minimal hypersurface of least area in the sense of Mazet-Rosenberg. Finally, we prove that the min-max energy values are bounded from below by the widths of the area functional as defined by Marques-Neves.'
address: |
Instituto de Matemática Pura e Aplicada (IMPA)\
Estrada Dona Castorina 110\
22460-320 Rio de Janeiro\
Brazil
author:
- 'Pedro Gaspar and Marco A. M. Guaraco'
bibliography:
- 'references.bib'
title: 'The Allen-Cahn equation on closed manifolds'
---
Introduction
============
Low energy levels of E\_e {#sec:least}
=========================
Multiparameter min-max for the energy functional {#sec:min-max}
================================================
Upper Bound {#sec:upperbound}
===========
Lower bound {#sec:bounds}
===========
Comparison with Marques-Neves p-widths {#sec:comp}
======================================
A cohomological index theory for free Z\_2 actions {#ap:index}
==================================================
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a theory based on Green’s function formalism to study magnetism in disordered Heisenberg systems with long range exchange integrals. Disordered Green’s function are decoupled within Tyablicov scheme and solved with a CPA method. The CPA method is the extension of Blackmann-Esterling-Beck approach to system with environmental disorder term which uses cumulant summation of the single-site non crossing diagrams. The crucial point is that we are able to treat simultaneously and self-consistently the RPA and CPA loops. It is shown that the summation of s-scattering contribution can always be performed analytically. While the p,d,f .. contributions are difficult to handle in the case of long-range coupling. To overcome this difficulty we propose and provide a test of a simplified treatment of these terms. In the case of 3D disordered nearest-neighbor Heisenberg system, a good agreement between the simplified treatment and the full calculation is achieved. Our theory allows in particular to calculate the Curie temperature, the spectral functions and the temperature dependence of the magnetization of each constituant as a function of concentration of impurity. Additionally it is shown that a virtual crystal treatment fails even at low impurity concentration.'
author:
- |
G. Bouzerar and P. Bruno\
Max-Planck-Institute für Mikrostrukturphysik, Weinberg 2, D-06120 Halle, Germany
title: 'RPA-CPA theory for magnetism in disordered Heisenberg binary systems with long range exchange integrals.'
---
PACS numbers: 75.10.-b, 75.25.+z, 71.10.-w, 75.50.Cc
Introduction.
=============
The coherent potential approximation (CPA) is widely used to study the effect of disorder in crystals (for reviews see [@Elliott; @Morigaki]). The CPA was initially developed independently by Soven [@Soven] and Taylor [@Taylor] to study systems with only [*diagonal*]{} disorder. Using a $2 \times 2$ formulation, a generalization to the presence of [*off-diagonal*]{} disorder was provided by Blackmann Esterling and Berck (BEB) [@BEB; @Gonis]. In these approaches the main idea is to replace the system by an effective medium which is determined by the condition that the averaged T-matrix of a single impurity immersed in the effective medium is zero. An alternative approach is based on cumulant expansion [@Yonezawa; @Leath]. This latter method has the advantage that it can handle the [*environmental*]{} disorder term which is characteristic of the Goldstone’s systems (phonons, magnons). The first proper treatment of the environmental disorder term, by using the cumulant expansion method is due to Lage and Stinchcombe [@LS] who studied the diluted Ising problem (S=1/2). Later, using the 2x2 matrix method of BEB, the method was extended by Whitelaw [@Whitelaw] to the phonon problem. In their calculations the coupling and locator are fixed quantities and restricted to nearest neighbor exchange couplings. It is well known that magnetism in clean ferromagnetic systems can be tackled with Green’s function formalism using Tyablicov decoupling procedure (RPA). This method goes beyond a simple mean field since it includes quantum fluctuations. Additionally, it fulfills the Goldstone and Mermin-Wagner theorems which is not the case of a mean field treatment. In the case of clean systems, combining first principle calculations to evaluate the exchange integrals and RPA method it was shown that one can provide satisfactory Curie temperature for Co and Fe [@Pajda]. Whilst, a simple mean field calculation largely overestimate the Curie temperature. It is our objective to provide in this paper a generalization of the RPA method to the disordered systems. We show that by combining in a self-consistent manner the RPA method and the CPA treatment of the disorder we are able to calculate Curie temperature, magnetization of the different constituants, spectral weights.... The CPA treatment is done in a similar way as done by Lage and Stinchcombe and by Whitelaw. However, due to Tyablicov decoupling scheme for the disordered Green’s functions, the locators and the effective exchange integrals are temperature dependent and have to be determined self-consistently for a given temperature.
The paper is organized as follows. In the first section we derive after Tyablicov decoupling scheme the disordered binary alloy Green’s function which includes [*diagonal*]{}, [*off-diagonal*]{} and [*environmental*]{} disorder. In section II, we perform the calculation of the averaged Green’s functions for A (respectively B) atom. In section III, by generalizing Callen’s formula we derive the equations for the magnetizations $m_{A}$, $m_{B}$ and for the Curie Temperature. In section IV, we propose an alternative simplified treatment of the p,d,..scattering contribution to the self-energy to the case of system with long-range exchange coupling. Finally in section V we present some numerical results and proceed to a test of our approximation of the self-energy contribution of the higher scattering terms.
Disordered Green’s function and RPA decoupling scheme.
======================================================
We study the magnetism in a binary alloy $A_{1-c}B_{c}$, A and B can be either magnetic ions or paramagnetic. We denote their spin respectively $S_{A}$ and $S_{B}$. The total Hamiltonian reads,
$$\hat {H}=\sum_{ij}-J_{ij}{\bf S}_{i} \cdot {\bf S}_{j} - \sum_{i}D_{i}(S^{z}_{i})^2
-B \sum_{i} g \mu_{i}(S^{z}_{i})$$
where the $J_{ij}$ and $D_{i}$ are random variables:$J_{ij}=J^{\lambda \lambda'}_{ij}$ with the probability $P_{i}^{\lambda} P_{j}^{\lambda'}$ where $P_{i}^{\lambda}$ is the probability that the site i is occupied by a $\lambda$-atom. Similarly $D_{i}=D_{\lambda}$ with probaility $P_{i}^{\lambda}$. The exchange integrals are assumed to be long range, our study is not restricted to the nearest neighbor Heisenberg model. The second term which describes anisotropy is only relevant in the case of 2D systems to get a non zero Curie temperature $T_{c}$ (Mermin-Wagner theorem). However in the case of 3D systems the contribution of this term can be neglected. We also include the effect of an external magnetic field.
Let us consider the following retarded Green’s function,
$$G^{+-}_{ij}(t)=-i\theta (t) \langle [S^{+}_{i}(t), S^{-}_{j}(0)] \rangle$$
where $\langle ..\rangle $ denotes the statistical average at temperature T, $$\langle \hat{O}\rangle=\frac{1}{Z} Tr(e^{-\beta \hat{H}} \hat{O})$$ where $Z=Tr(e^{-\beta \hat{H}})$.
$G^{+-}_{ij}(t)$’s Fourier transform in Energy space is, $$\ll S^{+}_{i};S^{-}_{j} \gg =G^{+-}_{ij}(\omega)= \int_{-\infty}^{+\infty} G^{+-}_{ij}(t)e^{i \omega t} dt$$ Its equation of motion reads, $$\omega G^{+-}_{ij}(\omega)= 2 m_{i}\delta_{ij} + \ll [S^{+}_{i},H];S^{-}_{j} \gg$$ where $m_{i}=<S_{i}^{z}>$, or $m_{i}=m_{A}$ (resp. $m_{B}$) if $i=A$ (resp. $i=B$).
After expanding the second term on the right side of the equality we obtain,
$$\begin{aligned}
(\omega -g\mu_{i}B) G^{+-}_{ij}(\omega)= 2m_{i}\delta_{ij}\nonumber \\
-\sum_{l}J_{il}\ll S^{z}_{i}S^{+}_{l}-S^{+}_{i}S^{z}_{l};S^{-}_{j} \gg\nonumber \\
+D_{i} \ll S^{z}_{i}S^{+}_{i}+S^{+}_{i}S^{z}_{i} \gg\end{aligned}$$
The next step consists in decoupling the higher order Green’s function. For the second term we use the standard Tyablicov decoupling [@Tyablicov] (equivalent to RPA). The last term due to anisotropy is somehow more complicate since on-site correlation are involved. Following the approach discussed in Ref. [@Jensen] we adopt for this term the Anderson-Callen decoupling scheme [@Anderson-Callen]:
$$\begin{aligned}
D_{i} \ll S^{+}_{i}S^{z}_{i}+S^{z}_{i}S^{+}_{i} \gg = 2D_{i}\gamma_{i} m_{i}\end{aligned}$$
where, $$\gamma_{i}=1-\frac{1}{2S^2}(S_i(S_i+1)-<(S_{i}^{z})^2>)
\label{aniso}$$ After simplification we find,
$$G^{+-}_{ij}=g_i \delta_{ij}+g_i \sum_{l} \Phi_{il}G^{+-}_{lj}
-\epsilon g_i(\sum_{l} \Psi_{il})G^{+-}_{ij}
\label{propag}$$
where $\phi_{il}=-1/2 J_{il}$ and $\Psi_{il}=-1/2 J_{il} \frac{m_{l}}{m_{i}}$ and $g_i$ denotes the locator: $g_{i}=g^{0}_{A}$ (resp. $g^{0}_{B}$) if $i=A$ (resp. $i=B$). $$g^{0}_{\lambda} (E)=\frac{\frac{m_{\lambda}}{m}}{E -g\mu_{\lambda}B/2m - D_{\lambda} \gamma_{\lambda} \frac{m_{\lambda}}{m}}$$
where $\lambda = A$ or B. For convenience, we have also introduced the reduced variable $E=\frac{\omega}{2m}$, m denotes the averaged magnetization: $m=\sum_{\lambda} c_{\lambda} m_{\lambda}$. The term which is proportional to $\epsilon$ comes from the environmental disorder term. This term is crucial to recover the Goldstone mode and requires to be treated very carefully. We have introduced the coefficient $\epsilon$ which is in principle equal to 1, in order to follow the influence of the environmental disorder term during the calculations. Note also that this term appears because of RPA decoupling. If $\epsilon =0$ Eq. \[propag\] is analogous to the propagator of an electron in a disordered medium with [*on-site*]{} potential and random long-range hopping terms $ t_{il}=\Phi_{il}$ ([*off-diagonal*]{} disorder). In this case the problem can be solved just within the BEB formalism. However, one should stress that the BEB formalism does not apply when the environmental term is present. Note also that in our model the locator $g^{0}_{\lambda}$, $\Psi_{il}$ and $\gamma_{i}$ are all temperature dependent, thus CPA and RPA loops have to be treated simultaneously in a self-consistent manner. It is also interesting to note that $\Psi_{il} \ne \Psi_{li}$ in the case where the sites i and l are occupied by different type of atoms.
Cumulant expansion method for the averaged Green’s functions.
=============================================================
As it is done in Ref. [@Whitelaw], the basic idea is to write Eq.\[propag\] as a locator expansion in BEB manner [@BEB]. We define the random variable $p_{i}$: $p_{i}=1$ if A is at site i or $p_{i}=0$ if i is occupied by a B ion. Therefore the locator reads, $$g_i=p_i g_{A}^{0} + (1-p_i) g_{B}^{0}= g^A_i+g^B_i$$ and, $$\begin{aligned}
\phi_{il}=p_i J^{AA}_{il}p_l +p_i J^{AB}_{il} (1-p_l)+ \nonumber \\
(1-p_i)J^{AB}_{il}p_l+(1-p_i)J^{BB}_{il}(1-p_l)\end{aligned}$$ similarly, $$\begin{aligned}
\Psi_{il}=p_i J^{AA}_{il}p_l +p_i J^{AB,1}_{il} (1-p_l)+\nonumber \\
(1-p_i)J^{AB,2}_{il} p_l+(1-p_i)J^{BB}_{il}(1-p_l)\end{aligned}$$ where $J^{AB,1}_{il} = \frac{m_{B}}{m_{A}}J^{AB}_{il} $ and $J^{AB,2}_{il} = \frac{m_{A}}{m_{B}}J^{AB}_{il}$.
The Green’s function are expressed in term of a 2x2 matrix and one gets for the equation of motion, $$\begin{aligned}
{\bf G}_{ij}=\left(
\begin{array}{cc}
g^{A}_{i} & 0 \\ 0 & g^{B}_{i}
\end{array}
\right)\delta_{ij}+
\left(
\begin{array}{cc}
g^{A}_{i} & 0 \\ 0 & g^{B}_{i}
\end{array}
\right)
\sum_{m}\left( \begin{array}{cc}
J^{AA}_{im} & J^{AB}_{im} \\ J^{AB}_{im} & J^{BB}_{im}
\end{array}
\right)
\left(
\begin{array}{cc}
G^{AA}_{mj} & G^{AB}_{mj} \\ G^{BA}_{mj} & G^{BB}_{mj}
\end{array}
\right)\nonumber \\
-\epsilon \left(
\begin{array}{cc}
g_{A}^{0} & 0 \\ 0 & g_{B}^{0}
\end{array}
\right)
\left(
\begin{array}{cc}
J^{AB,1} +\sum_{l}(J^{AA}_{il}-J^{AB,1}_{il})p_l & 0 \\ 0 & J^{BB} +\sum_{l}(J^{AB,2}_{il}-J^{BB}_{il})p_l
\end{array}
\right)
\left(
\begin{array}{cc}
G^{AA}_{ij} & G^{AB}_{ij} \\ G^{BA}_{ij} & G^{BB}_{ij}
\end{array}
\right)\end{aligned}$$
We have defined the variables $J^{AB,1}=\sum_l J^{AB,1}_{il}$ and $J^{BB}=\sum_l J^{BB}_{il}$ .
The aim is to expand this expression into a product of the p factors, which can then be averaged over disorder by expanding into cumulants. For that purpose we separate out the factors and introduce a new variable $\rho_i$ by $p_i=\rho_i+c$ (where $c_{A}=c$) . The idea is to separate out the virtual crystal part.
$$\begin{aligned}
{\bf g}_i= \rho_i
\left(
\begin{array}{cc}
g_{A}^{0} & 0 \\ 0 & -g_{B}^{0}
\end{array}
\right)+
\left(
\begin{array}{cc}
cg_{A}^{0} & 0 \\ 0 & (1-c)g_{B}^{0}
\end{array}
\right)\end{aligned}$$
There is still the environmental term which is more difficult to handle. As it was done by Lage and Stinchcombe [@LS], by converting into [**k**]{}-space the calculations become easier to perform.
We define the Fourier transform by,
$${\bf G}_{\bf kk'}=\sum_{ij} \exp(i{\bf k} \cdot {\bf r}_{i})\exp(-i{\bf k'} \cdot {\bf r}_j){\bf G}_{ij}$$
After some manipulation one gets,
$$\begin{aligned}
{\bf G}_{{\bf kk'}}= {\bf G}^{vc}_{\bf k}\rho_{\bf k-k'}+{\bf G}^{vc}_{\bf k}
\left(
\begin{array}{cc}
c & 0 \\ 0 & c-1
\end{array}
\right) \delta_{\bf k-k'}+
{\bf G}^{vc}_{\bf k} \frac{1}{N} \sum_{\bf q}\rho_{\bf k-q}
{\bf V}_{\bf kq}{\bf G}_{\bf qk'}
\label{propag2}\end{aligned}$$
where the $2 \times 2$ matrix ${\bf V}_{\bf kq}$ is defined by: $$\begin{aligned}
{\bf V}_{\bf kq}=\left(
\begin{array}{cc}
J^{AA}_{\bf q} -\epsilon (J^{AA}_{\bf k-q}-J^{AB,1}_{\bf k-q}) & J^{AB}_{\bf q} \\ J^{AB}_{\bf q} & J^{BB}_{\bf q} -\epsilon (J^{BB}_{\bf k-q}-J^{AB,2}_{\bf k-q})
\end{array}\right)\end{aligned}$$
and the virtual-crystal Green’s function ${\bf G}^{vc}_{\bf k}$, $$\begin{aligned}
[{\bf G}^{vc}_{\bf k}]^{-1}={\bf M}_{0}-c{\bf M}_{1}\end{aligned}$$
where the matrices ${\bf M}_{0}$ and ${\bf M}_{1}$ are, $$\begin{aligned}
{\bf M}_{0}=
\left(
\begin{array}{cc}
(g_{A}^{0})^{-1} & 0 \\ 0 & -(g_{B}^{0})^{-1}
\end{array}\right)+
\left(
\begin{array}{cc}
\epsilon J^{AB,1} & 0 \\ J^{AB}_{\bf k} & J^{BB}_{\bf k}-\epsilon J^{BB}
\end{array}\right)\end{aligned}$$ and,
$$\begin{aligned}
{\bf M}_{1}=
\left(
\begin{array}{cc}
J^{AA}_{\bf k}-\epsilon(J^{AA}- J^{AB,1}) & J^{AB}_{\bf k} \\ J^{AB}_{\bf k} & J^{BB}_{\bf k}-\epsilon (J^{BB}- J^{AB,2})
\end{array}\right)\end{aligned}$$
The equation (\[propag2\]) can be expanded into 2 sub-series. $$\begin{aligned}
{\bf G}_{\bf kk'}={\bf G}^{(1)}_{\bf kk'}+{\bf G}^{(2)}_{\bf kk'}\end{aligned}$$ where the sub-series are respectively, $$\begin{aligned}
{\bf G}^{(1)}_{\bf kk'}=
{\bf G}^{vc}_{\bf k}\rho_{\bf k-k'}+
\frac{1}{N}\sum_{\bf q}{\bf G}^{vc}_{\bf k}{\bf V}_{\bf kq}{\bf
G}^{vc}_{\bf q}\rho_{\bf k-q}\rho_{\bf q-k'}+....\end{aligned}$$ and, $$\begin{aligned}
{\bf G}^{(2)}_{\bf kk'}=
\left({\bf G}^{vc}_{\bf k}\delta_{\bf k-k'}+{\bf G}^{vc}_{\bf k}{\bf V}_{\bf kk'}{\bf
G}^{vc}_{\bf k'}\rho_{\bf k-k'}+
\frac{1}{N}\sum_{\bf q}{\bf G}^{vc}_{\bf k}{\bf V}_{\bf kq}{\bf
G}^{vc}_{\bf q}{\bf V}_{\bf qk'}{\bf G}^{vc}_{\bf k'}
\rho_{\bf k-q}\rho_{\bf q-k'}+...\right)
\left(
\begin{array}{cc}
c & 0 \\ 0 & c-1
\end{array}\right)\end{aligned}$$
The averaged Green’s function is obtained by averaging over products of $\rho$ by expanding into cumulants $P_{i}(c)$. For instance, $$\langle \rho_{\bf k_1}\rho_{\bf k_2} \rangle =\frac{P_{2}(c)}{N}
\delta({\bf k_1}+ {\bf k_2})$$ $$\langle \rho_{\bf k_1}\rho_{\bf k_2}\rho_{\bf k_3} \rangle =\frac{P_{3}(c)}{N^{2}}
\delta({\bf k_1}+ {\bf k_2} +{\bf k_3} )$$ and, $$\begin{aligned}
\langle \rho_{\bf k_1}\rho_{\bf k_2}\rho_{\bf k_3} \rho_{\bf k_4}\rangle =\frac{P_{4}(c)}{N^{3}}
\delta({\bf k_1}+ {\bf k_2} +{\bf k_3} ) + (\frac{P_{2}(c)}{N})^{2}[
\delta({\bf k_1} +{\bf k_2})\delta({\bf k_3} +{\bf k_4})+\nonumber \\
\delta({\bf k_1} +{\bf k_3})\delta({\bf k_2} +{\bf k_4})+
\delta({\bf k_1} +{\bf k_4})\delta({\bf k_2} +{\bf k_3})]\end{aligned}$$
The cumulants are systematically obtained the generating function, $$g(x,c)=\text{ln}(1-c+ce^{x})=\sum_{i=1}^{\infty} P_{i}(c) \frac{x^{i}}{i!}$$ From this equation one gets $P_{1}(c)=c$, $P_{2}(c)=c(1-c)$, $P_{3}(c)=c(1-c)(1-2c)$....
In order to get a closed form for the series we have to make the usual CPA approximation which consists in keeping only the diagrams with no crossings of external lines. As it is was shown by Yonezawa et al. [@Yonezawa; @Leath], the self-consistency requires a modification of the semi-invariants to be attributed to each vertex. In other words it means that the cumulants $P_{i}(c)$ have to be replaced by a new set of coefficients $Q_{i}(c)$ which satisfies the relation, $$Q_{1}(c)+Q_{2}(c)x+Q_{3}(c)x^2 .....=\sigma_c(x)=\frac{c}{1-x(1-\sigma_c(x))}
\label{sigm}$$
where the modified cumulants are, $$Q_{i}(c)=\sum_{m=1}^{i} [(-1)^{m-1} \frac{(i+m-2)!}{m!(i-m)!(m-1)!}] c^{m}$$
In the single site approximation, after averaging, one gets for the averaged $2\times2$ Green’s function matrix,
$${\bf \bar{G}}_{\bf kk'}={\bf \bar{G}}_{\bf k}\delta_{\bf k-k'}={\bf \tilde{G}}_{\bf k}
\left[\left(
\begin{array}{cc}
c & 0 \\ 0 & c-1
\end{array}\right) + {\bf \Delta}_{\bf k}
\right]
\label{fullg}$$
where,
$${\bf \tilde{G}}_{\bf k}=\left[({\bf G}^{vc}_{\bf k})^{-1}-{\bf \Sigma}_{\bf k}
\right]^{-1}
\label{ghat}$$
${\bf \Sigma}_{\bf k}$ denotes the self-energy, it is given by,
$${\bf \Sigma}_{\bf k}=Q_{2} \frac{1}{N} \sum_{\bf q}{\bf V}_{\bf kq}{\bf \tilde{G}}_{\bf q}{\bf V}_{\bf qk}+
Q_{3} \frac{1}{N^2} \sum_{\bf q,t}{\bf V}_{\bf kq}{\bf \tilde{G}}_{\bf q}{\bf V}_{\bf qt}
{\bf \tilde{G}}_{\bf t}{\bf V}_{\bf tk}+...
\label{self}$$
and,
$${\bf \Delta}_{\bf k}=Q_{2} \frac{1}{N} \sum_{\bf q}{\bf V}_{\bf kq}{\bf \tilde{G}}_{\bf q}+
Q_{3} \frac{1}{N^2} \sum_{\bf q,t}{\bf V}_{\bf kq}{\bf \tilde{G}}_{\bf q}{\bf V}_{\bf qt}
{\bf \tilde{G}}_{\bf t}+...
\label{delta}$$
The term ${\bf \Delta}_{\bf k}$ which is very similar to the self-energy is called end correction [@LS]. Note that, inside the CPA loop, Eq.(\[ghat\]) and Eq.(\[self\]) are the only 2 equations which have to be solved self-consistently. To summarize, in Fig. \[fig1\] we show a diagrammatic representation of the previous set of equations.
Evaluation of ${\bf \Delta}_{\bf k}$.
-------------------------------------
It is convenient for the calculations to start by defining, $$\gamma_{i}({\bf q})=\frac{1}{z_i} \sum_{{\bf r}^i_l} \exp(i{{\bf qr}^i_l})$$
The sum, ${\bf r}^{i}_{l}$ runs over the i-th type of neighbors of the i-th shell from a given site 0 and $z_i$ is the total number of neighbors in the shell. Note that, from now $\sum_{i}$ will correspond to a summation over the different shells. With this definition it follows immediately,
$$J^{AA}({\bf q})= \sum_{i} J^{AA}_{i} z_{i} \gamma_{i}({\bf q})$$
We get similar expression for $J^{BB}({\bf q})$ and $J^{AB}({\bf q})$...
It is convenient to decompose the matrix ${\bf V}_{\bf kq}$ into two terms,
$${\bf V}_{\bf kq}={\bf V}^{(1)}_{\bf kq}+{\bf V}^{(2)}_{\bf kq}$$
where $${\bf V}^{(1)}_{\bf kq}=\sum_{i}{\bf V}^{(1),i}_{\bf kq}=
\sum_{i} \left[{\bf A}_{i}-\epsilon {\bf D}_{i}\gamma_{i}({\bf k})\right]\gamma_{i}({\bf q})
\label{vpot}$$ and, $${\bf V}^{(2)}_{\bf kq}=\sum_{i}{\bf V}^{(2),i}_{\bf kq}=\epsilon {\bf D}_{i}\left[\gamma_{i}({\bf k})\gamma_{i}({\bf q}) - \gamma_{i}({\bf k-q})\right]$$
${\bf A}_{i}$ and ${\bf D}_{i}$ are the following 2x2 matrices:
$${\bf A}_{i}=\left(
\begin{array}{cc}
J^{AA}_i & J^{AB}_i \\ J^{AB}_i & J^{BB}_i
\end{array}\right)z_{i}$$
$${\bf D}_{i}=\left(
\begin{array}{cc}
J^{AA}_i-J^{AB,1}_i & 0 \\ 0 & J^{BB}_i-J^{AB,2}_i
\end{array}\right)z_{i}$$
By using the following very useful property [@Callen]: if $f(r)$ is a function which is equivaluated at each site $r_i$ of $E_i$ then, $$\frac{1}{N} \sum_{\bf q}\gamma_{i}({\bf k-q})f({\bf q})=\gamma_{i}({\bf k})\frac{1}{N}\sum_{\bf q} \gamma_{i}({\bf q})f({\bf q})
\label{prop}$$
By using Eq.(\[prop\]), we find significant simplifications in the calculations. Indeed all the terms of the serie involving at least one factor ${{\bf V}}^{(2)}$ reduces to zero. Thus the end correction term does not explicitely depend on the environmental disorder term.
After calculation we finally get,
$${\bf \Delta}_{\bf k}= \sum_{ij}{{\bf V}}^{(1),i}_{\bf k,0} \left[ Q_{2}{\bf I}+Q_{3} {\bf M}+Q_{4}
{\bf M}^{2}+....
\right]_{ij}{{\bf F}^{j}}
\label{delt}$$
Like ${{\bf V}^{(1),i}}$, ${{\bf F}^{j}}$ is a $2 \times 2$ matrix, and ${\bf M}$ a $N_{s} \times N_{s}$ matrix, where each matrix element ${\bf M}_{ij}$ is a $2 \times 2$ matrix. $N_{s}$ denotes the number of considered shells. ${{\bf V}^{(1),i}}$ is given in Eq. (\[vpot\]) and ${{\bf F}^{j}}$ and ${\bf M}_{ij}$ are defined by $${\bf F}^i=\frac{1}{N} \sum_{\bf q}\gamma_{i}({\bf q}){\bf \tilde{G}}_{\bf q}$$ and, $${\bf M}_{ij}=\frac{1}{N} \sum_{\bf q}\gamma_{i}({\bf q}){\bf \tilde{G}}_{{\bf q}}{\bf V}^{(1),j}_{\bf q,0}$$
The sum in Eq. \[delt\] is obtained after diagonalization of the $2N_{s} \times 2N_{s}$ matrix ${\bf M}={\bf P}^{-1} {\bf M}_{diag} {\bf P}$, $$Q_{2}{\bf I}+Q_{3}{\bf M}+Q_{4}{\bf M}^{2}+....={\bf P}^{-1}
[(\sigma_{c}({\bf M}_{diag})-Q_{1}{\bf I}){\bf M}_{diag}^{-1}]{\bf P}$$ The function $\sigma_{c}$ was previously defined in Eq. (\[sigm\]), and $[\sigma_{c}({\bf M}_{diag})]_{ij} =
\sigma_{c}(\lambda_{i}) \delta_{ij}$ where $\lambda_{i}$ are the eigenvalues of ${\bf M}$. Hence, we get for the end correction
$${\bf \Delta}_{\bf k}= \sum_{ij}{{\bf V}}^{(1),i}_{\bf k,0}
[{\bf P}^{-1}
[(\sigma_{c}({\bf M}_{diag})-Q_{1}{\bf I}){\bf M}_{diag}^{-1}]{\bf P}]_{ij}
{{\bf F}^{j}}$$
Let us now proceed further and evaluate the self-energy ${\bf \Sigma}_{\bf k}$.
Evaluation of ${\bf \Sigma}_{\bf k}$.
-------------------------------------
Using the remarks made in the previous section, we find that the self-energy can be written, $${\bf \Sigma}_{\bf k}={\bf \Sigma}^{(1)}_{\bf k}+{\bf \Sigma}^{(2)}_{\bf k}$$
where ${\bf \Sigma}^{(1)}_{\bf k}$ (resp.${\bf \Sigma}^{(2)}_{\bf k}$) is obtained by replacing ${\bf V}_{\bf k,q}$ by ${\bf V}^{(1)}_{\bf k,q}$ (resp.${\bf V}^{(2)}_{\bf k,q}$). Indeed we find that each term of the serie containing both ${\bf V}^{(1)}$ and ${\bf V}^{(2)}$ reduces to zero. After simplifications we obtain for ${\bf \Sigma}^{(1)}_{\bf k}$, $${\bf \Sigma }^{(1)}({\bf k})=\sum_{i,j}{{\bf V}^{i}}_{\bf k,0} \left[ Q_{1}{\bf I}+Q_{2}{{\bf M}}+Q_{3}
{\bf M}^{2}+....
\right]_{ij}{\bf \Gamma}^{j}({\bf k})$$ where ${\bf \Gamma}^{j}({\bf k})=\gamma_{j}({\bf k}) \left(
\begin{array}{cc}
1 & 0 \\ 0 & 1
\end{array}\right)$.
As previously done for the end correction, using the function $\sigma_{c}(z)$ defined in Eq. \[sigm\] we obtain immediately, $$Q_{1}{\bf I}+Q_{2}{\bf M}+Q_{3}
{\bf M}^{2}+....={\bf P}^{-1} [\sigma({\bf M}_{diag})]{\bf P}$$
Note that, we have included in the sum the first order term depending on [*c*]{} ($Q_{1}$) which comes from the virtual crystal Green’s function ${\bf G}^{vc}_{q}$.
In general, the evaluation of the second term ${\bf \Sigma }^{(2)}({\bf k})$ is much more complicated. One can get an analytical form only for simple cases. For example if the exchange integrals are restricted to only nearest neighbor, the complete summation of the serie can be performed by using the space group symmetry of the lattice [@Izyumov; @LS]. In the case of nearest neighbor Heisenberg system one gets, $${\bf \Sigma}^{(2)}_{\bf k}(E)={\bf C}_{p}(1-\gamma(2{\bf k}))+{\bf C}_{d}(1+\gamma(2{\bf k})-
2\gamma({\bf k})^{2})
\label{sigm2}$$ where, $${\bf C}_{p,d}=-\frac{\epsilon}{2}(Q_{1}{\bf I}+Q_{2}{\bf M}_{p,d}+Q_{3}{\bf M}_{p,d}^{2}+...){\bf D}_{1}
\label{cpd}$$ ${\bf C}_{p,d}$ are evaluated in the same way it was done for ${\bf \Sigma}^{(1)}_{\bf k}(E)$ and ${\bf \Delta}_{\bf k}(E)$. The matrices ${\bf D}_{1}$, ${\bf M}_{p}$ and ${\bf M}_{d}$ are respectively, $${\bf D}_{1}=\left(
\begin{array}{cc}
J^{AA}-J^{AB,1} & 0 \\ 0 & J^{BB}-J^{AB,2}
\end{array}\right)z$$
$${\bf M}_{p}=-\frac{\epsilon}{6}{\bf D}_{1}{\bf \tilde{G}}_{p}$$
$${\bf M}_{d}=-\frac{\epsilon}{4}{\bf D}_{1}{\bf \tilde{G}}_{d}$$
where ${\bf \tilde{G}}_{p}=\frac{1}{N} \sum_{\bf q} (1-\gamma(2{\bf q})){\bf \tilde{G}}({\bf q})$ and ${\bf \tilde{G}}_{d}=\frac{1}{N} \sum_{\bf q} (1+\gamma(2{\bf q})-2\gamma({\bf q})^{2}){\bf \tilde{G}}({\bf q})$.
Note that the virtual crystal approximation for ${\bf \Sigma}^{(2)}_{\bf k}(E)$ consists in taking in Eq. (\[cpd\]) the first term only. Then it follows immediately, $$C_{p}^{VCA}=C_{d}^{VCA}=-\frac{\epsilon c}{2} {\bf D}_{1}$$
which substituted in Eq.(\[sigm2\]) leads to, $${\bf \Sigma}^{(2),VCA}_{\bf k}(E)=-\epsilon c {\bf D}_{1}(1-\gamma({\bf k})^{2})
\label{sig2vca}$$ Note that ${\bf \Sigma}^{(2),VCA}_{\bf k}$ is energy independent. It is also important to stress that at the lowest order the self-consistency for ${\bf \Sigma}^{(2)}$ is not required.
Most of the ferromagnetic materials are of itinerant type, which means that the exchange integrals between different localized magnetic ions are long range and driven by the polarization of the conduction electrons gas as it is for the RKKY mechanism [@rkky]. Analytically, the generalization of the previous calculations to the more interesting case where $J_{ij}$ are long ranged is not an easy task. However by truncating the serie, the summation can be performed numerically. It is important to note that ${\bf \Sigma }^{(2)}({\bf k})$ is (i) proportional to $\epsilon$ which means that it originates only from the environmental disorder term and (ii) each term of the serie vanishes in the long wave length limit ${\bf \Sigma }^{(2)}({\bf k=0})=0$. This implies that even after truncation of the serie at any order, the Goldstone theorem remains fulfilled. Thus the long wave length magnons are always treated properly. Furthermore, since ${\bf \Sigma }^{(2)}({\bf k})$ corresponds to higher order scattering terms (p,d,f,...) it is natural to expect that these terms should not affect the Curie temperature in a dramatic way. In other words we expect that a truncation of ${\bf \Sigma }^{(2)}({\bf k})$ serie to the first few term should already provide a good approximation of Curie temperature compared to the one one would get with the complete series. However, it is crucial to consider at least the lowest order term (the virtual crystal contribution) otherwise even in the clean limit one would not recover the correct result and the Goldstone’s theorem would be violated . If we consider the lower approximation ${\bf \Sigma }^(2) \approx {\bf \Sigma }^{(2)}_{VCA}$, we get the expected results in the limit $c=0$ and $c=1$. It is not a priori clear whether such an approximation of ${\bf \Sigma }^{(2)}({\bf k})$ to the lowest order provides satisfying results for the Curie temperature at moderate impurity concentration. Such an approximation will be tested later on.
To conclude this section the complete averaged $2 \times 2$ Green’s function is obtained after solving self-consistently the set of equations Eqs. (\[ghat\]) and (\[self\]) within the CPA loop and then using Eqs. \[fullg\] and \[delta\] to get ${\bf \Delta}_{\bf k}$ and ${\bf {\bar G}_k}$. However, as was already mentioned in the introduction, the problem is not solved until we are able to calculate the locators $g_{\lambda}^{0}$ and the exchanged integrals $\Psi_{il}$ which depend on the averaged magnetization $m_{\lambda}$. The determination of $m_{\lambda}$ has to be done self-consistently in an additional external loop (RPA loop).
Magnetization and Curie temperature.
====================================
We assume that the averaged $2 \times 2 $ Green’s function matrix ${\bf \bar G} ({\bf k},E)$ is calculated according to the previous section within the CPA loop. We show how from ${\bar G}_{\lambda} ({\bf k},E)$, $\lambda =A$ or B we can get the missing self-consistent equations (RPA loop) to get the temperature dependent locator $g_{\lambda}^{0}$ and the exchange integrals $\Psi_{il}$. This will allow us to calculate the element-resolved magnetizations $m_{\lambda}=<S_{\lambda}^{z}>$ as function of temperature and the Curie temperature. It was shown by Callen, in the case of a clean system (pure A or B) that the magnetization can be expressed in the following way [@Callen], $$m_{\lambda}=\frac{(S_{\lambda}-\Phi_{\lambda})(1+\Phi_{\lambda})^{2S_{\lambda}+1}+(S_{\lambda}+1+\Phi_{\lambda})\Phi_{\lambda}^{2S_{\lambda}+1}}{(1+\Phi_{\lambda})^{2S_{\lambda}+1}-\Phi_{\lambda}^{2S_{\lambda}+1}}
\label{Callen-form}$$ where $\Phi_{\lambda}= \frac{1}{N} \sum_{\bf q}\Phi_{\lambda}({\bf q})$ and $\Phi_{\lambda}({\bf q})$ is defined as, $$\Phi_{\lambda}({\bf q})=\int_{-\infty}^{+\infty} dE \frac{A_{\lambda}({\bf q},E)}{e^{2mE/kT}-1
}
\label{eqfi}$$
where, $$A_{\lambda}({\bf q},E)=\frac{-1}{\pi} \text {Im} G_{\lambda}^{+-}({\bf q},E)$$ is the spectral function.
Note also that the Callen’s approach to get the magnetization allows to derive a lot of local spin-spin correlation, they are only expressed as a function of $\Phi_{\lambda}$. For instance, $$\langle (S_{\lambda}^{z})^{2} \rangle = S(S+1) - m_{\lambda}(1+2\Phi_{\lambda})
\label{s2}$$ which is needed to determine the anisotropy parameters $\gamma_{\lambda}$ given in Eq. (\[aniso\]).
In the case of clean systems, the normalized spectral function $A_{\lambda}({\bf q},E)$ is given by $$A_{\lambda}({\bf q},E)= \delta(E -E({\bf q}))$$ $E({\bf q})=\omega({\bf q})/2m$ and $\omega({\bf q})$ denotes the magnon dispersion.
In the case of a binary (or multi-component) alloy this formula can be generalized in the following way,
$$A_{\lambda}({\bf q},E)=\frac{-1}{\pi} \frac{\text{Im} \bar{G}_{\lambda}^{+-}({\bf q},E)}
{c_{\lambda} x_{\lambda}}
\label{spectral}$$
where $c_{\lambda}$ is the concentration of $\lambda$-ion and we have for convenience introduced a T-dependent reduced variable $x_{\lambda}=\frac{m_{\lambda}}{m}$.
Note that in the presence of impurities the spectral function is not anymore a $\delta$ function, but because of the finite imaginary part of the self-energy will consists of peaks of finite width with more or less a Lorentzian shape. In the case of binary alloy we expect for a given ${\bf q}$, 2 peaks, more generally [*n*]{} peaks for an [*n*]{}-component alloy.
For a given temperature the complete self-consistency is obtained by (i) providing good starting values for $m_{\lambda}$ then (ii) performing the CPA loop which provide ${\bf \bar G} ({\bf k},E)$ and finally (ii) going into the RPA loop by using Eq. (\[Callen-form\]), (\[s2\]) and (\[spectral\]) one gets the new values of $m_{\lambda}$ and $\langle (S_{\lambda}^{z})^{2} \rangle$ which are re-injected in the locators $g_{\lambda}^{0}$, the exchange integrals $\Psi_{il}$ and $\gamma_{\lambda}$.
Let us now show how to get the Curie temperature of a disordered Heisenberg binary alloy. We start by expanding Eq. (\[eqfi\]) in the limit $T \rightarrow T_{c}$ (i.e, $m_{\lambda} \rightarrow 0$). We immediately get,
$$\Phi_{\lambda} \approx \frac{kT_{c}}{2m} F_{\lambda}$$
where, $$F_{\lambda}=\frac{1}{N}\sum_{\bf q}
\int_{-\infty}^{+\infty} dE \frac{A_{\lambda}({\bf q},E)}{E}$$
After expanding Eq. \[Callen-form\] as a function of $\frac{1}{\Phi_{\lambda}}$ one obtains,
$$m_{\lambda}=\frac{S_{\lambda}(S_{\lambda}+1)}{3}\frac{2m}{kT_{c}}\frac{1}{F_{\lambda}}$$
Since the averaged magnetization m is defined by, $m=\sum_{\lambda}c_{\lambda}m_{\lambda}$, combining the two previous equations one finds for the Curie-Temperature,
$$k_{B}T_{c}=\frac{2}{3}\sum_{\lambda}c_{\lambda} \frac{S_{\lambda}(S_{\lambda}+1)}{F_{\lambda}}
\label{eqtc}$$
Eq. (\[eqtc\]) is the RPA generalization of the Curie Temperature to a multi-component disordered alloy. The previous equation provides a direct measure of the weight $w_{\lambda}=\frac{1}{k_{B}T_{c}}[c_{\lambda}\frac{S_{\lambda}(S_{\lambda}+1)}{F_{\lambda}}]$ of each $\lambda$-element to the Curie Temperature.
Numerical results.
==================
In this section we provide an illustration of the RPA-CPA theory and a test for the approximation suggested above for the higher order scattering contribution of the self-energy. For simplicity, we consider the case of a 3D disordered binary alloy on a simple cubic lattice. Additionally we restrict the exchange integrals to nearest neighbor only which allows us to test the validity of the approximation scheme suggested in Sec III Bfor estmating $\Sigma^{2}$. For further simplifications of the calculations we consider the cas of a zero external field and neglect the anisotropy term which is reasonable for a 3D systems.
In Fig. \[fig2\], we have plotted the Curie Temperature as a function of $c$ obtained with the full CPA treatment, the $\Sigma^{(2)}$ part of the self-energy is calculated exactly (full summation of the serie). Note that pure A (resp. B) corresponds to $c=1$ (resp. $c=0$). Depending on the chosen set of parameters $T_{c}$ shows (i) a minimum ($J_{AB} S_{A} S_{B} \le \text{min}(J_{AA} S_{A}^{2},J_{BB} S_{B}^{2})$, (ii) a maximum ($J_{AB} S_{A} S_{B} \ge \text{max}(J_{AA} S_{A}^{2},J_{BB} S_{B}^{2})$ or (iii) is monotonic ($ \text{min}(J_{AA} S_{A}^{2},J_{BB} S_{B}^{2}) \le J_{AB} S_{A} S_{B} \le \text{max}(J_{AA} S_{A}^{2},J_{BB} S_{B}^{2})$. These three different cases are shown in the figure.
As already mentioned in section III, it is difficult to perform the full summation of $\Sigma^{(2)}$ for the case of long-range exchange integrals which is the case of many realistic and interesting systems, for example permalloy. As it was discussed previously the simplest approximation consists in keeping only the lowest order term of the serie ( virtual crystal approximation). In the case of nearest neighbor Heisenberg system $\Sigma^{(2)}$ and $\Sigma^{(2),VCA}$ are respectively given in Eqs (\[sigm2\]) and (\[sig2vca\]). In Fig. \[fig3\] we have plotted the Curie Temperature calculated with (i) a full CPA treatment, (ii) the one performed with the approximation discussed previously ${\bf \Sigma}^{(2)}={\bf \Sigma}^{(2)}_{VCA}$ and (iii) the one obtained with virtual crystal approximation. In the case (iii), the averaged Green’s function is, $${\bf \bar{G}}_{\bf k}={\bf G}^{vc}_{\bf k}
\left(
\begin{array}{cc}
c & 0 \\ 0 & c-1
\end{array}\right)$$ since in VCA ${\bf \Delta_{k}}=0$.
The comparison between the full CPA and the virtual crystal approximation shows that the Curie temperature differs significantly. Even, very close to the clean limit the VCA appears to be inappropriate, for instance for $c=0.1$, we observe that $T^{VCA}_{c}$ is about $35 \%$ larger than the full CPA calculated one. Note that the disagreement is even more pronounced in the vicinity of $c=0$ than $c=1$. This can be understood in the following way: since $J_{AB}=3 J_{BB}=1.5 J_{AA}$ and $S_{A}=S_{B}$ a substitutition of a B site by a A site (close to $c=0$) introduces a change of energy (with respect to the pure case) 2 times larger than a substitution of a B site by a A site near $c=1$. As discussed previously it is interesting to compare the Curie temperature where the VCA is only done on ${\bf \Sigma}^{(2)}$ ($T^{2,VCA}_{c}$). We observe a good agreement between the full CPA calculated $T_{c}$ and $T^{2,VCA}_{c}$, in the whole range of concentration, the agreement is even excellent for $c \ge 0.6$. A comparison between $T^{VCA}_{c}$ and $T^{2,VCA}_{c}$ in the vicinity of $c=0$ and $c=1$ shows that the reason why the VCA approximation breaks down is essentially because of the crude approximation of the s-part of the scattering. Thus this figure validates a simple treatment of ${\bf \Sigma}^{(2)}$. It is also expected that including only few additional terms of the serie will lead to an excellent agreement in the whole range of concentration. Additionally, the approximation ${\bf \Sigma}^{(2)} \approx {\bf \Sigma}^{(2),VCA}$ will get better in the case of long range exchange integrals.
In Fig. \[fig4\] we show the temperature dependence of the element-resolved magnetizations. In order to demonstrate the versatility of our approach, we have chosen a set of parameters which mimics a ferrimagnetic behavior with compensation point. Additionally, the parameters are such that $T_{c}^{A} \gg T_{c}^{B}$. While the temperature dependence of $m_{A}$ follows a standard behavior, $m_{B} (T)$ start to strongly decrease even at low temperature. For example at $T \approx 2.5$, $m_{A}$ has reduced by less that $20 \%$ while $m_{B}=0.5 \,m_{B}(0)$. As a result of our choice of the parameters we see that the averaged magnetization $m_{av}=|c_{A} m_{A}+c_{B} m_{B}|$ is non monotonic and vanishes for an intermediate temperature value (compensation point). It is found that the function $\frac{m_{B}}{m_{A}}(T)$ decreases monotonically with temperature. As a result and since at $T=0$, $\frac{m{B}}{m_{A}}=\frac{S_{B}}{S_{A}}$, thus if $\frac{S_{B}}{S_{A}} \le \frac{c_{A}}{c_{B}}$ then $m_{av}$ will not have a compensation point . However, the condition that $\frac{S_{B}}{S_{A}} \ge \frac{c_{A}}{c_{B}}$ is not sufficient to get one, it also required that $\frac{m_{B}}{m_{A}}(T_{c}) \le \frac{c_{A}}{c_{B}}$.
In Fig. \[fig5\] we now show the magnon spectral density (MSD) $\rho_{\lambda} (E)=\frac{\text{Im} G^{\lambda} (E)}{x_{\lambda} c_{\lambda}}$ as a function of E. We consider 3 different cases: almost clean A and B ( (a) and (c)), and the intermediate situation $c_{A}=c_{B}=0.5$. In both, Fig. \[fig5\]a and Fig. \[fig5\]c we observe that the MSD is very similar to the clean case. This is clearer in case (c) than (a), it is easy to understand that when doping A with B the difference in energy with the undoped case is only of order $10 \%$ ($J_{AA} (S_{A})^2= 0.8$ and $J_{AB}S_{A}S_{B}= 0.9$) whilst doping B with A the change is more drastic (about $100 \%$). To get a similar MSD to Fig.\[fig5\]c for a weakly doped B sample, one should take $c \approx 0.005$.
In Fig. \[fig6\] we show the spectral function $S_{\lambda} ({\bf q},E)$ as afunction of energy for different values of the momentum ${\bf q}$. This quantity is more interesting that the integrated MSD since it provides direct information about the elementary excitation dispersions and their spectral weight. Additionally it is directly related to inelastic neutron scattering measurements. Let us now briefly discuss Fig. \[fig6\]. At precisely ${\bf q}={\bf 0}$ momentum, in both $S_{\lambda=A,B}$, we observe 2 peaks structure (i) a well defined peak [@footnote1] at $E=0$ as expected since our theory fulfills the Goldstone theorem and (ii) a very broad one at intermediate energy $E \approx 1$. For intermediate values of the momentum, it is difficult to separate the peak and one get a single broad peak. We see clearly that the peaks are crossing each other at ${\bf q} \approx \frac{\pi}{2} (1,1,1)$. Note that due to (i) the different spectral weight of the peaks and to the closeness of their location, the single peak-structure which is observed at $ q =\frac{\pi}{2}$ is located at different energy for A and B. From this figure we see also that the dispersion of the second peak is almost flat $E_{2} (q) \approx 1$, while the Goldstone mode $E_{1}({\bf q})$ [@footnote2] goes from $E=0$ to $E_{max} \approx 2$ when moving in the $(1,1,1)$-direction.
Conclusion.
===========
In conclusion, we have presented in this paper a theory based on Green’s function formalism to study magnetism in disordered Heisenberg systems with long range exchange integrals. The disordered Green’s function are decoupled within Tyablicov procedure and the disorder ([*diagonal, off-diagonal*]{} and [*environmental*]{}) is treated with a $2 \times 2$ modified cumulant CPA approach. The crucial point is that we are able to treat simultaneously and self-consistently the RPA and CPA loops. Our theory allows in particular to calculate Curie temperature, spectral functions and temperature dependence of the magnetization for each element as a function of concentration of impurity. Additionally, we have proposed a simplified treatment of the p,d,f .. contribution of the self-energy which is difficult to handle in the case of long range exchange integrals. The approximation was tested successfully on 3D disordered nearest-neighbor Heisenberg systems. Combined with first principle calculations which can provide the exchange integrals this method appears to be very promising to study magnetism in disordered systems.
R.J. Elliott, B.R. Leath and J.A. Krumhansl Rev. Mod. Phys., [**46**]{}, 465 (1974). F. Yonezawa and K. Morigaki Prog. Theor. Phys. Suppl. [**53**]{}, 1 (1973).
P. Soven Phys. Rev. [**156**]{},809 (1967).
D. W. Taylor [**156**]{}, 1017 (1967). J. A. Blackman, D.M. Esterling and N.F. Berk Phys. rev. B [**4**]{} 2412 (1971). A.Gonis and J.W. Garland, Phys. Rev. B [**16**]{}, 1495 (1977). F. Yonezawa Prog. theo. phys. [**40**]{} 734 (1968). P.L. Leath Phys. rev. [**171**]{}, 725 (1968). E.J.S. Lage and R.B. Stinchcombe J. Phys. C: Solid State Phys. [**10**]{} 295 (1977). D.J. Whitelaw J. Phys. C: Solid State Phys. [**14**]{} 2871 (1981). M. Pajda, J. Kudrnovský, I. Turek, V. Drchal and P. Bruno, Phys. Rev. lett. [**85**]{}, 5424 (2000), Phys. Rev. B [**64**]{}, 174402 (2001). S.V. Tyablicov, [*Methods in quantum theory of magnetism*]{} (Plenum Press, New York, 1967). P. Fröbrich, P.J. Jensen and P.J. Kuntz, Eur. Phys. J. B [**13**]{}, 477 (2000). F.B. Anderson and H.B. Callen Phys. Rev. [**136**]{}, A 1068 (1964). H. B. Callen Phys. Rev. [**130**]{}, 890 (1963); see also H.B. Callen and S. Shtrikman, Solid State Comm. [**5**]{}, 5 (1965). Y. Izyumov, Proc. Phys. Soc. [**87**]{} 505 , (1966). A.A. Rudermann and C. Kittel, Phys. Rev. [**96**]{}, 99 (1954); T. Kasuya, Prog. Theor. Phys. [**16**]{}, 45 (1956); K. Yosida, Phys. Rev.[**106**]{}, 893 (1957). For convenience a small broadening has been introduced to make the figure easier for the reader. At exactly $q=0$ we find that the peak at $E=0$ is a $\delta$-peak consistent with the fact that $\text{Im} G_{\lambda} (E=0)=0$. As expected, analyzing closely $S_{\lambda}({\bf q},E)$ in the vicinity of $q=0$, we find that $E_{1} ({\bf q})= D q^{2}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider non-monotone DR-submodular function maximization, where DR-submodularity (diminishing return submodularity) is an extension of submodularity for functions over the integer lattice based on the concept of the diminishing return property. Maximizing non-monotone DR-submodular functions has many applications in machine learning that cannot be captured by submodular set functions. In this paper, we present a $\frac{1}{2+\epsilon}$-approximation algorithm with a running time of roughly $O(\frac{n}{\epsilon}\log^2 B)$, where $n$ is the size of the ground set, $B$ is the maximum value of a coordinate, and $\epsilon > 0$ is a parameter. The approximation ratio is almost tight and the dependency of running time on $B$ is exponentially smaller than the naive greedy algorithm. Experiments on synthetic and real-world datasets demonstrate that our algorithm outputs almost the best solution compared to other baseline algorithms, whereas its running time is several orders of magnitude faster.'
author:
- |
Tasuku Soma\
The University of Tokyo\
tasuku\[email protected] Yuichi Yoshida\
National Institute of Informatics, *and*\
Preferred Infrastructure, Inc.\
[email protected]
bibliography:
- 'main.bib'
title: |
Non-monotone DR-Submodular Function Maximization\
(Full version)
---
Introduction
============
Submodular functions have played a key role in various tasks in machine learning, statistics, social science, and economics. A set function $f : 2^E \to {\mathbb{R}}$ with a ground set $E$ is *submodular* if $$\begin{aligned}
f(X \cup {\{e\}}) - f(X) \geq f(Y \cup {\{e\}}) - f(Y)\end{aligned}$$ for arbitrary sets $X,Y \subseteq E$ with $X \subseteq Y$, and an element $e \in E \setminus Y$. The importance and usefulness of submodularity in these areas are due to the fact that submodular functions naturally capture the *diminishing return property*. Various important functions in these areas such as the entropy function, coverage function, and utility functions satisfy this property. See, e.g., [@Krause2014survey; @Fujishige2005].
Recently, *maximizing* (non-monotone) submodular functions has attracted particular interest in the machine learning community. In contrast to *minimizing* submodular functions, which can be done in polynomial time, maximizing submodular functions is NP-hard in general. However, we can achieve a constant factor approximation for various settings. Notably, [@Buchbinder:2012hc] presented a very elegant *double greedy* algorithm for (unconstrained) submodular function maximization, which was the first algorithm achieving $\frac{1}{2}$-approximation, and this approximation ratio is tight [@Feige2011]. Applications of non-monotone submodular function maximization include efficient sensor placement [@Krause2008a], privacy in online services [@Krause2008utility], and maximum entropy sampling [@Ko:1995hk].
The models and applications mentioned so far are built upon submodular *set functions*. Although set functions are fairly powerful for describing problems such as variable selection, we sometimes face difficult situations that cannot be cast with set functions. For example, in the budget allocation problem [@Alon2012], we would like to decide how much budget should be set aside for each ad source, rather than whether we use the ad source or not. A similar issue arises when we consider models allowing multiple choices of an element in the ground set.
To deal with such situations, several generalizations of submodularity have been proposed. [@Soma:2014tp] devised a general framework for maximizing *monotone submodular functions on the integer lattice*, and showed that the budget allocation problem and its variant fall into this framework. In their framework, functions are defined over the integer lattice ${\mathbb{Z}}_+^E$ and therefore effectively represent discrete allocations of budget. Regarding the original motivation for the diminishing return property, one can naturally consider its generalization to the integer lattice: a function $f : {\mathbb{Z}}_+^E \to {\mathbb{R}}$ satisfying $$\begin{aligned}
f({{\bm x}}+ {{\bm \chi}}_e) - f({{\bm x}}) \geq f({{\bm y}}+ {{\bm \chi}}_e) - f({{\bm y}})\end{aligned}$$ for ${{\bm x}}\leq {{\bm y}}$ and $e \in E$, where ${{\bm \chi}}_e \in {\mathbb{R}}^E$ is the vector with ${{\bm \chi}}_e(e) = 1$ and ${{\bm \chi}}_e(a) = 0$ for every $a \neq e$. Such functions are called *diminishing return submodular (DR-submodular) functions* [@sfcover:nips2015] or *coordinate-wise concave submodular functions* [@Milgrom2009]. DR-submodular functions have found various applications in generalized sensor placement [@sfcover:nips2015] and (a natural special case of) the budget allocation problem [@Soma:2014tp].
As a related notion, a function is said to be *lattice submodular* if $$f({{\bm x}}) + f({{\bm y}}) \geq f({{\bm x}}\vee {{\bm y}}) + f({{\bm x}}\wedge {{\bm y}})$$ for arbitrary ${{\bm x}}$ and ${{\bm y}}$, where $\vee$ and $\wedge$ are coordinate-wise max and min, respectively. Note that DR-submodularity is *stronger* than *lattice submodularity* in general (see, e.g., [@Soma:2014tp]). Nevertheless, we consider the DR-submodularity to be a “natural definition” of submodularity, at least for the applications mentioned so far, because the diminishing return property is crucial in these real-world scenarios.
Our contributions
-----------------
We design a novel polynomial-time approximation algorithm for maximizing (non-monotone) DR-submodular functions. More precisely, we consider the optimization problem $$\begin{aligned}
\label{eq:max-DRsfm}
\begin{array}{ll}
\text{maximize}& f({{\bm x}})\\
\text{subject to} & {{\bm 0}}\leq {{\bm x}}\leq {{\bm B}},
\end{array}\end{aligned}$$ where $f : {\mathbb{Z}}_+^E \to {\mathbb{R}}_+$ is a non-negative DR-submodular function, ${{\bm 0}}$ is the zero vector, and ${{\bm B}}\in {\mathbb{Z}}_+^E$ is a vector representing the maximum values for each coordinate. When ${{\bm B}}$ is the all-ones vector, this is equivalent to the original (unconstrained) submodular function maximization. We assume that $f$ is given as an evaluation oracle; when we specify ${{\bm x}}\in {\mathbb{Z}}_+^E$, the oracle returns the value of $f({{\bm x}})$.
Our algorithm achieves $\frac{1}{2+\epsilon}$-approximation for any constant $\epsilon > 0$ in $O(\frac{|E|}{\epsilon} \cdot \log (\frac{\Delta}{\delta})\log B \cdot (\theta + \log B))$ time, where $\delta$ and $\Delta$ are the minimum positive marginal gain and maximum positive values, respectively, of $f$, $B = {\lVert{{\bm B}}\rVert}_{\infty} := \max_{e \in E}{{\bm B}}(e)$, and $\theta$ is the running time of evaluating (the oracle for) $f$. To our knowledge, this is the first polynomial-time algorithm achieving (roughly) $\frac{1}{2}$-approximation.
We also conduct numerical experiments on the revenue maximization problem using real-world networks. The experimental results show that the solution quality of our algorithm is comparable to other algorithms. Furthermore, our algorithm runs several orders of magnitude faster than other algorithms when $B$ is large.
DR-submodularity is necessary for obtaining polynomial-time algorithms with a meaningful approximation guarantee; if $f$ is only lattice submodular, then we cannot obtain constant-approximation in polynomial time. To see this, it suffices to observe that an *arbitrary* univariate function is lattice submodular, and therefore finding an (approximate) maximum value must invoke $O(B)$ queries. We note that representing an integer $B$ requires ${\lceil \log_2 B \rceil}$ bits. Hence, the running time of $O(B)$ is pseudopolynomial rather than polynomial.
Fast simulation of the double greedy algorithm {#fast-simulation-of-the-double-greedy-algorithm .unnumbered}
----------------------------------------------
Naturally, one can reduce the problem to maximization of a submodular set function by simply duplicating each element $e$ in the ground set into ${{\bm B}}(e)$ distinct copies and defining a set function over the set of all the copies. One can then run the double greedy algorithm [@Buchbinder:2012hc] to obtain $\frac{1}{2}$-approximation. This reduction is simple but has one large drawback; the size of the new ground set is $\sum_{e\in E} {{\bm B}}(e)$, a pseudopolynomial in $B$. Therefore, this naive double greedy algorithm does not scale to a situation where $B$ is large.
For scalability, we need an additional trick that reduces the pseudo-polynomial running time to a polynomial one. In *monotone* submodular function maximization on the integer lattice, [@sfcover:nips2015; @sfm:arxiv2015] provide such a speedup trick, which effectively combines the *decreasing threshold technique* [@Badanidiyuru2013] with binary search. However, a similar technique does not apply to our setting, because the double greedy algorithm works differently from (single) greedy algorithms for monotone submodular function maximization. The double greedy algorithm examines each element in a *fixed* order and marginal gains are used to decide whether to include the element or not. In contrast, the greedy algorithm chooses each element in *decreasing* order of marginal gains, and this property is crucial for the decreasing threshold technique.
We resolve this issue by splitting the set of all marginal gains into polynomially many small intervals. For each interval, we approximately execute multiple steps of the double greedy algorithm at once, as long as the marginal gains remain in the interval. Because the marginal gains do not change (much) within the interval, this simulation can be done with polynomially many queries and polynomial-time overhead. To our knowledge, this speedup technique is not known in the literature and is therefore of more general interest.
Very recently, [@DBLP:journals/corr/EneN16] pointed out that a DR-submodular function $f:{\{0,1,\ldots,B\}}^E \to {\mathbb{R}}_+$ can be expressed as a submodular set function $g$ over a polynomial-sized ground set, which turns out to be $E \times {\{0,1,\ldots,k-1\}}$, where $k={\lceil \log_2(B+1) \rceil}$. Their idea is representing ${{\bm x}}(e)$ in binary form for each $e \in E$, and bits in the binary representations form the new ground set. We may want to apply the double greedy algorithm to $g$ in order to get a polynomial-time approximation algorithm. However, this strategy has the following two drawbacks: (i) The value of $g(E \times {\{0,1,\ldots,k-1\}})$ is defined as $f({{\bm x}})$, where ${{\bm x}}(e) = 2^k-1$ for every $e \in E$. This means that we have to extend the domain of $f$. (ii) More crucially, the double greedy algorithm on $g$ may return a large set such as $E \times {\{0,1,\ldots,k-1\}}$ whose corresponding vector ${{\bm x}}\in {\mathbb{Z}}_+^E$ may violate the constraint ${{\bm x}}\leq {{\bm B}}$. Although we can resolve these issues by introducing a knapsack constraint, it is not a practical solution because existing algorithms for knapsack constraints [@Lee:2009tc; @Chekuri:2014ed] are slow and have worse approximation ratios than $1/2$.
#### Notations
For an integer $n \in {\mathbb{N}}$, $[n]$ denotes the set ${\{1,\ldots,n\}}$. For vectors ${{\bm x}}, {{\bm y}}\in {\mathbb{Z}}^E$, we define $f({{\bm x}}\mid {{\bm y}}) := f({{\bm x}}+ {{\bm y}}) - f({{\bm y}})$. The $\ell_1$-norm and $\ell_{\infty}$-norm of a vector ${{\bm x}}\in {\mathbb{Z}}^E$ are defined as ${\lVert{{\bm x}}\rVert}_1 := \sum_{e \in E} {\lvert{{\bm x}}(e)\rvert}$ and ${\lVert{{\bm x}}\rVert}_{\infty} := \max_{e \in E} {\lvert{{\bm x}}(e)\rvert}$, respectively.
Related work
============
As mentioned above, there have been many efforts to maximize submodular functions on the integer lattice. Perhaps the work most related to our interest is [@Gottschalk:2015fq], in which the authors considered maximizing lattice submodular functions over the bounded integer lattice and designed $1\over3$-approximation pseudopolynomial-time algorithm. Their algorithm was also based on the double greedy algorithm, but does not include a speeding up technique, as proposed in this paper. In addition there are several studies on the *constrained* maximization of submodular functions [@Feige2011; @Buchbinder2014; @Buchbinder2015], although we focus on the unconstrained case. Many algorithms for maximizing submodular functions are randomized, but a very recent work [@Buchbinder2015] devised a derandomized version of the double greedy algorithm. [@Gotovos2015] considered maximizing non-monotone submodular functions in the *adaptive* setting, a concept introduced in [@Golovin:2011cn].
A continuous analogue of DR-submodular functions is considered in [@DBLP:journals/corr/BianMB016].
Algorithms {#sec:algorithm}
==========
In this section, we present a polynomial-time approximation algorithm for maximizing (non-monotone) DR-submodular functions. We first explain a simple adaption of the double greedy algorithm for (set) submodular functions to our setting, which runs in pseudopolynomial time. Then, we show how to achieve a polynomial number of oracle calls. Finally, we provide an algorithm with a polynomial running time (details are placed in Appendix \[sec:proof-truely-polynomial\]).
Pseudopolynomial-time algorithm {#subsec:pseudo-poly}
-------------------------------
Algorithm \[alg:pseudo-poly\] is an immediate extension of the double greedy algorithm for maximizing submodular (set) functions [@Buchbinder:2012hc] to our setting. We start with ${{\bm x}}= {{\bm 0}}$ and ${{\bm y}}= {{\bm B}}$, and then for each $e \in E$, we tighten the gap between ${{\bm x}}(e)$ and ${{\bm y}}(e)$ until they become exactly the same. Let $\alpha = f({{\bm \chi}}_e \mid {{\bm x}})$ and $\beta = f(-{{\bm \chi}}_e \mid {{\bm y}})$. We note that $$\alpha + \beta = f({{\bm x}}+ {{\bm \chi}}_e) - f({{\bm x}}) - (f({{\bm y}}) - f({{\bm y}}- {{\bm \chi}}_e)) \geq 0$$ holds from the DR-submodularity of $f$. Hence, if $\beta < 0$, then $\alpha > 0$ must hold, and we increase ${{\bm x}}(e)$ by one. Similarly, if $\alpha < 0$, then $\beta > 0$ must hold, and we decrease ${{\bm y}}(e)$ by one. When both of them are non-negative, we increase ${{\bm x}}(e)$ by one with probability $\frac{\alpha}{\alpha+\beta}$, or decrease ${{\bm y}}(e)$ by one with the complement probability $\frac{\beta}{\alpha+\beta}$.
${{\bm x}}\leftarrow {{\bm 0}}$, ${{\bm y}}\leftarrow {{\bm B}}$. \[line:pseudo-poly-e\] $\alpha \leftarrow f({{\bm \chi}}_{e} \mid {{\bm x}})$ and $\beta \leftarrow f(-{{\bm \chi}}_{e} \mid {{\bm y}})$. \[line:pseudo-poly-a-b\] ${{\bm y}}(e) \leftarrow {{\bm y}}(e) - 1$. \[line:pseudo-poly-else\] ${{\bm x}}$.
Algorithm \[alg:pseudo-poly\] is a $\frac{1}{2}$-approximation algorithm for with time complexity $O(\|{{\bm B}}\|_1 \cdot \theta + \|{{\bm B}}\|_1)$, where $\theta$ is the running time of evaluating $f$.
We omit the proof as it is a simple modification of the analysis of the original algorithm.
Algorithm with polynomially many oracle calls {#subsec:poly-oracle-call}
---------------------------------------------
In this section, we present an algorithm with a polynomial number of oracle calls.
Our strategy is to simulate Algorithm \[alg:pseudo-poly\] without evaluating the input function $f$ many times. A key observation is that, at Line \[line:pseudo-poly-a-b\] of Algorithm \[alg:pseudo-poly\], we do not need to know the exact value of $f({{\bm \chi}}_e \mid {{\bm x}})$ and $f(-{{\bm \chi}}_e \mid {{\bm y}})$; good approximations to them are sufficient to achieve an approximation guarantee close to $\frac{1}{2}$. To exploit this observation, we first design an algorithm that outputs (sketches of) approximations to the functions $g(b) := f({{\bm \chi}}_e \mid {{\bm x}}+ b{{\bm \chi}}_e)$ and $h(b) := f(-{{\bm \chi}}_e \mid {{\bm y}}-b{{\bm \chi}}_e)$. Note that $g$ and $h$ are non-increasing functions in $b$ because of the DR-submodularity of $f$.
To illustrate this idea, let us consider a non-increasing function $\phi:{\{0,1,\ldots,B-1\}}\to {\mathbb{R}}$ and suppose that $\phi$ is non-negative ($\phi$ is either $g$ or $h$ later on). Let $\delta$ and $\Delta$ be the minimum and the maximum positive values of $\phi$, respectively. Then, for each $\delta \leq \tau \leq \Delta$ of the form $\delta (1+\epsilon)^k$, we find the minimum $b_\tau$ such that $\phi(b_\tau) < \tau$ (we regard $\phi(B) = -\infty$). From the non-increasing property of $\phi$, we then have $\phi(b) \geq \tau$ for any $b < b_\tau$. Using the set of pairs ${\{(\tau,b_\tau)\}}_{\tau}$, we can obtain a good approximation to $\phi$. The details are provided in Algorithm \[alg:approximate\].
$S \leftarrow \emptyset$. We regard $\phi(B) = -\infty$. Find the minimum $b_0 \in {\{0,1,\ldots,B\}}$ with $\phi(b_0) \leq 0$ by binary search. \[line:approximate-b\] $\Delta \leftarrow \phi(0)$ and $\delta \leftarrow \phi(b_0-1)$. Find the minimum $b_\tau \in {\{0,1,\ldots,B\}}$ with $\phi(b_\tau) < \tau$ by binary search. $S \leftarrow S \cup {\{(b_\tau,\tau)\}}$ $S \leftarrow S \cup {\{(B,0)\}}$. $S$.
\[lem:approximate\] For any $\phi:{\{0,1,\ldots,B-1\}}\to {\mathbb{R}}$ and $\epsilon > 0$, Algorithm \[alg:approximate\] outputs a set of pairs ${\{(b_\tau,\tau)\}}_\tau$ from which, for any $b \in {\{0,1,\ldots,B-1\}}$, we can reconstruct a value $v$ in $O(\log B)$ time such that $v \leq \phi(b) < (1+\epsilon)v$ if $\phi(b) > 0$ and $v = 0$ otherwise. The time complexity of Algorithm \[alg:approximate\] is $O(\frac{1}{\epsilon}\log (\frac{\Delta}{\delta}) \log B \cdot \theta)$ if $\phi$ has a positive value, where $\delta$ and $\Delta$ are the minimum and maximum positive values of $\phi$, respectively, and $\theta$ is the running time of evaluating $\phi$, and is $O(\log B \cdot \theta)$ otherwise.
Let $S = {\{(b_\tau,\tau)\}}_\tau$ be the set of pairs output by Algorithm \[alg:approximate\]. Our reconstruction algorithm is as follows: Given $b \in {\{0,1,\ldots,B-1\}}$, let $(b_{\tau^*},\tau^*)$ be the pair with the minimum $b_{\tau^*}$, where $b < b_{\tau^*}$. Note that such a $b_{\tau^*}$ always exists because a pair of the form $(B,\cdot)$ is always added to $S$. We then output $\tau^*$. The time complexity of this reconstruction algorithm is clearly $O(\log B)$.
We now show the correctness of the reconstruction algorithm. If $\phi(b) > 0$, then, in particular, we have $\phi(b) \geq \delta$. Then, $\tau^*$ is the maximum value of the form $\delta(1+\epsilon)^k$ at most $\phi(b)$. Hence, we have $\tau^* \leq \phi(b) < (1+\epsilon)\tau^*$. If $\phi(b) \leq 0$, $(b_{\tau^*},\tau^*) = (B,0)$ and we output zero.
Finally, we analyze the time complexity of Algorithm \[alg:approximate\]. Each binary search requires $O(\log B)$ time. The number of binary searches performed is $O(\log_{1+\epsilon}\frac{\Delta}{\delta} ) = O(\frac{1}{\epsilon}\log\frac{\Delta}{\delta})$ when $\phi$ has a positive value and 1 when $\phi$ is non-positive. Hence, we have the desired time complexity.
We can regard Algorithm \[alg:approximate\] as providing a value oracle for a function $\tilde{\phi}:{\{0,1,\ldots,B-1\}}\to {\mathbb{R}}_+$ that is an approximation to the input function $\phi:{\{0,1,\ldots,B-1\}}\to {\mathbb{R}}$.
We now describe our algorithm for maximizing DR-submodular functions. The basic idea is similar to Algorithm \[alg:pseudo-poly\], but when we need $f({{\bm \chi}}_e \mid {{\bm x}})$ and $f(-{{\bm \chi}}_e \mid {{\bm y}})$, we use approximations to them instead. Let $\alpha$ and $\beta$ be approximations to $f({{\bm \chi}}_e \mid {{\bm x}})$ and $f(-{{\bm \chi}}_e \mid {{\bm y}})$, respectively, obtained by Algorithm \[alg:approximate\]. Then, we increase ${{\bm x}}(e)$ by one with probability $\frac{\alpha}{\alpha+\beta}$ and decrease ${{\bm y}}(e)$ by one with the complement probability $\frac{\beta}{\alpha+\beta}$. The details are given in Algorithm \[alg:polynomial-oracle\].
${{\bm x}}\leftarrow {{\bm 0}}$, ${{\bm y}}\leftarrow {{\bm B}}$. Define $g,h:{\{0,1,\ldots,B-1\}} \to {\mathbb{R}}$ as $g(b) = f({{\bm \chi}}_e \mid {{\bm x}}+b{{\bm \chi}}_e)$ and $h(b) = f(- {{\bm \chi}}_e \mid {{\bm y}}-b {{\bm \chi}}_e)$. Let $\tilde{g}$ and $\tilde{h}$ be approximations to $g$ and $h$, respectively, given by Algorithm \[alg:approximate\]. \[line:polynomial-oracle-while\] $\alpha \leftarrow \tilde{g}({{\bm x}}(e))$ and $\beta \leftarrow \tilde{h}({{\bm B}}(e) - {{\bm y}}(e))$. \[line:alpha-beta\] ${{\bm x}}$.
We now analyze Algorithm \[alg:polynomial-oracle\]. An *iteration* refers to an iteration in the while loop from Line \[line:polynomial-oracle-while\]. We have $\|{{\bm B}}\|_1$ iterations in total. For $k \in {\{1,\ldots,\|{{\bm B}}\|_1\}}$, let ${{\bm x}}_k$ and ${{\bm y}}_k$ be ${{\bm x}}$ and ${{\bm y}}$, respectively, right after the $k$th iteration. Note that ${{\bm x}}_{\|{{\bm B}}\|_1} = {{\bm y}}_{\|{{\bm B}}\|_1}$ is the output of the algorithm. We define ${{\bm x}}_0 = {{\bm 0}}$ and ${{\bm y}}_0 = {{\bm B}}$ for convenience.
Let ${{\bm o}}$ be an optimal solution. For $k \in {\{0,1,\ldots,\|{{\bm B}}\|_1\}}$, we then define ${{\bm o}}_{k} = ({{\bm o}}\vee {{\bm x}}_{k}) \wedge {{\bm y}}_{k}$. Note that ${{\bm o}}_0 = {{\bm o}}$ holds and ${{\bm o}}_{\|{{\bm B}}\|_1}$ equals the output of the algorithm. We have the following key lemma.
\[lem:key-lemma\] For every $k \in [\|{{\bm B}}\|_1]$, we have $$\begin{aligned}
&{\mathop{\mathbf{E}}}[f({{\bm o}}_{k-1}) - f({{\bm o}}_{k})] \nonumber\\
&\leq \frac{1+\epsilon}{2}{\mathop{\mathbf{E}}}[f({{\bm x}}_{k}) - f({{\bm x}}_{k-1}) + f({{\bm y}}_{k}) - f({{\bm y}}_{k-1}) ] \label{eq:key-inequality}
\end{aligned}$$
Fix $k \in [\|{{\bm B}}\|_1]$ and let $e$ be the element of interest in the $k$th iteration. Let $\alpha$ and $\beta$ be the values in Line \[line:alpha-beta\] in the $k$th iteration. We then have $$\begin{aligned}
& {\mathop{\mathbf{E}}}[f({{\bm x}}_{k}) - f({{\bm x}}_{k-1}) + f({{\bm y}}_{k}) - f({{\bm y}}_{k-1})] \nonumber \\
&=
\frac{\alpha}{\alpha+\beta}f({{\bm \chi}}_{e} \mid {{\bm x}}_{k-1})
+
\frac{\beta}{\alpha+\beta}f(-{{\bm \chi}}_{e} \mid {{\bm y}}_{k-1}) \nonumber \\
&\geq
\frac{\alpha}{\alpha+\beta}\alpha
+
\frac{\beta}{\alpha+\beta}\beta
= \frac{\alpha^2+\beta^2}{\alpha+\beta}, \label{eq:marginal-of-alg}
\end{aligned}$$ where we use the guarantee in Lemma \[lem:approximate\] in the inequality.
We next establish an upper bound of ${\mathop{\mathbf{E}}}[f({{\bm o}}_{k-1}) - f({{\bm o}}_{k})]$. As ${{\bm o}}_k = ({{\bm o}}\vee {{\bm x}}_k) \wedge {{\bm y}}_k$, conditioned on a fixed ${{\bm o}}_{k-1}$, we obtain $$\begin{aligned}
&{\mathop{\mathbf{E}}}[f({{\bm o}}_{k-1}) - f({{\bm o}}_{k})] \nonumber \\
&=
\frac{\alpha}{\alpha+\beta} \Bigl(f({{\bm o}}_{k-1})-f({{\bm o}}_{k-1} \vee {{\bm x}}_{k}(e) {{\bm \chi}}_e)\Bigr) \nonumber \\
& \qquad +
\frac{\beta}{\alpha+\beta} \Bigl(f({{\bm o}}_{k-1})-f({{\bm o}}_{k-1} \wedge {{\bm y}}_{k}(e) {{\bm \chi}}_e)\Bigr). \label{eq:marginal-of-opt}
\end{aligned}$$
\[cla:key-lemma\] $\eqref{eq:marginal-of-opt} \leq \frac{(1+\epsilon)\alpha\beta}{\alpha+\beta}$.
We show this claim by considering the following three cases.
If ${{\bm x}}_{k}(e) \leq {{\bm o}}_{k-1}(e) \leq {{\bm y}}_{k}(e)$, then is zero.
If ${{\bm o}}_{k-1}(e) < {{\bm x}}_{k}(e)$, then ${{\bm o}}_{k}(e) = {{\bm o}}_{k-1}(e) + 1$, and the first term of is $$\begin{aligned}
& f({{\bm o}}_{k-1})-f({{\bm o}}_{k-1} \vee {{\bm x}}_{k}(e) {{\bm \chi}}_e) \\
& = f({{\bm o}}_{k-1}) - f({{\bm o}}_{k-1}+{{\bm \chi}}_e)\\
& \leq f({{\bm y}}_{k-1} - {{\bm \chi}}_e) - f({{\bm y}}_{k-1}) \\
& = f(-{{\bm \chi}}_e \mid {{\bm y}}_{k-1}) \\
& \leq (1+\epsilon)\beta.
\end{aligned}$$ Here, the first inequality uses the DR-submodularity of $f$ and the fact that ${{\bm o}}_{k-1} \leq {{\bm y}}_{k-1}-{{\bm \chi}}_e$, and the second inequality uses the guarantee in Lemma \[lem:approximate\]. The second term of is zero, and hence we have $\eqref{eq:marginal-of-opt} \leq \frac{(1+\epsilon)\alpha\beta}{\alpha+\beta}$.
If ${{\bm y}}_{k}(e) < {{\bm o}}_{k-1}(e)$, then by a similar argument, we have $\eqref{eq:marginal-of-opt} \leq \frac{(1+\epsilon)\alpha\beta}{\alpha+\beta}$.
We now return to proving Lemma \[lem:key-lemma\]. By Claim \[cla:key-lemma\], $$\eqref{eq:marginal-of-opt}
\leq
\frac{(1+\epsilon)\alpha\beta}{\alpha+\beta} \leq \frac{1+\epsilon}{2}\Bigl( \frac{\alpha^2+\beta^2}{\alpha+\beta} \Bigr)
\leq
\frac{1+\epsilon}{2}\cdot \eqref{eq:marginal-of-alg},$$ which indicates the desired result.
\[the:polynomial-oracle\] Algorithm \[alg:polynomial-oracle\] is a $\frac{1}{2+\epsilon}$-approximation algorithm for with time complexity $O(\frac{|E|}{\epsilon} \cdot \log (\frac{\Delta}{\delta}) \log \|{{\bm B}}\|_\infty \cdot \theta + \|{{\bm B}}\|_1\log \|{{\bm B}}\|_\infty)$, where $\delta$ and $\Delta$ are the minimum positive marginal gain and the maximum positive value, respectively, of $f$ and $\theta$ is the running time of evaluating $f$.
Summing up for $k \in [\|{{\bm B}}\|_1]$, we get $$\begin{aligned}
& \sum_{k = 1}^{\|{{\bm B}}\|_1} {\mathop{\mathbf{E}}}[f({{\bm o}}_{k-1}) - f({{\bm o}}_{k})] \\
&\leq
\frac{1+\epsilon}{2}\sum_{k = 1}^{\|{{\bm B}}\|_1}{\mathop{\mathbf{E}}}[f({{\bm x}}_{k}) - f({{\bm x}}_{k-1}) + f({{\bm y}}_{k}) - f({{\bm y}}_{k-1}) ].
\end{aligned}$$ The above sum is telescopic, and hence we obtain $$\begin{aligned}
& {\mathop{\mathbf{E}}}[f({{\bm o}}_0) - f({{\bm o}}_{\|{{\bm B}}\|_1})] \\
&\leq
\frac{1+\epsilon}{2}{\mathop{\mathbf{E}}}[ f({{\bm x}}_{\|{{\bm B}}\|_1}) - f({{\bm x}}_0) + f({{\bm y}}_{\|{{\bm B}}\|_1}) - f({{\bm y}}_0)] \\
&\leq
\frac{1+\epsilon}{2} {\mathop{\mathbf{E}}}[f({{\bm x}}_{\|{{\bm B}}\|_1})+f({{\bm y}}_{\|{{\bm B}}\|_1})] \\
& = (1+\epsilon) {\mathop{\mathbf{E}}}[f({{\bm x}}_{\|{{\bm B}}\|_1})].
\end{aligned}$$ The second inequality uses the fact that $f$ is non-negative, and the last equality uses ${{\bm y}}_{\|{{\bm B}}\|_1} = {{\bm x}}_{\|{{\bm B}}\|_1}$. Because ${\mathop{\mathbf{E}}}[f({{\bm o}}_0) - f({{\bm o}}_{\|{{\bm B}}\|_1})] = f({{\bm o}}) - {\mathop{\mathbf{E}}}[f({{\bm x}}_{\|{{\bm B}}\|_1})]$, we obtain that ${\mathop{\mathbf{E}}}[f({{\bm x}}_{\|{{\bm B}}\|_1})] \geq \frac{1}{2+\epsilon}f({{\bm o}})$.
We now analyze the time complexity. We only query the input function $f$ inside of Algorithm \[alg:approximate\], and the number of oracle calls is $O(\frac{|E|}{\epsilon} \log (\frac{\Delta}{\delta}) \log B)$ by Lemma \[lem:approximate\]. Note that we invoke Algorithm \[alg:approximate\] with $g$ and $h$, and the minimum positive values of $g$ and $h$ are at least the minimum positive marginal gain $\delta$ of $f$. The number of iterations is $\|{{\bm B}}\|_1$, and we need $O(\log B)$ time to access $\tilde{g}$ and $\tilde{h}$. Hence, the total time complexity is as stated.
\[rem:non-negative\] We note that even if $f$ is not a non-negative function, the proof of Theorem \[the:polynomial-oracle\] works as long as $f({{\bm x}}_0) \geq 0$ and $f({{\bm y}}_0) \geq 0$, that is, $f({{\bm 0}}) \geq 0$ and $f({{\bm B}}) \geq 0$. Hence, given a DR-submodular function $f:{\mathbb{Z}}_+^E \to {\mathbb{R}}$ and ${{\bm B}}\in {\mathbb{Z}}_+^E$, we can obtain a $\frac{1}{2+\epsilon}$-approximation algorithm for the following problem: $$\begin{aligned}
\begin{array}{ll}
\text{maximize} & f({{\bm x}}) - \min{\{f({{\bm 0}}),f({{\bm B}})\}} \\
\text{subject to} & {{\bm 0}}\leq {{\bm x}}\leq {{\bm B}},
\end{array}\label{eq:max-DRsfm-g}
\end{aligned}$$ This observation is useful, as the objective function often takes negative values in real-world applications.
Polynomial-time algorithm {#subsec:poly}
-------------------------
In many applications, the running time needed to evaluate the input function is a bottleneck, and hence Algorithm \[alg:polynomial-oracle\] is already satisfactory. However, it is theoretically interesting to reduce the total running time to a polynomial, and we show the following. The proof is deferred to Appendix \[sec:proof-truely-polynomial\].
\[the:truely-polynomial\] There exists a $\frac{1}{2+2\epsilon}$-approximation algorithm with time complexity $\widetilde{O}(\frac{|E|}{\epsilon} \log (\frac{\Delta}{\delta})\log \|{{\bm B}}\|_\infty \cdot (\theta + \log \|{{\bm B}}\|_\infty))$, where $\delta$ and $\Delta$ are the minimum positive marginal gain and the maximum positive value, respectively of $f$ and $\theta$ is the running time of evaluating $f$. Here $\widetilde{O}(T)$ means $O(T \log^c T)$ for some $c \in {\mathbb{N}}$.
Experiments {#sec:experiments}
===========
In this section, we show our experimental results and the superiority of our algorithm with respect to other baseline algorithms.
Experimental setting
--------------------
We conducted experiments on a Linux server with an Intel Xeon E5-2690 (2.90 GHz) processor and 256 GB of main memory. All the algorithms were implemented in C\# and were run using Mono 4.2.3.
We compared the following four algorithms:
- Single Greedy (): We start with ${{\bm x}}= {{\bm 0}}$. For each element $e \in E$, as long as the marginal gain of adding ${{\bm \chi}}_e$ to the current solution ${{\bm x}}$ is positive, we add it to ${{\bm x}}$. The reason that we do not choose the element with the maximum marginal gain is to reduce the number of oracle calls, and our preliminary experiments showed that such a tweak does not improve the solution quality.
- Double Greedy (, Algorithm \[alg:pseudo-poly\]).
- Lattice Double Greedy (): The $1/3$-approximation algorithm for maximizing non-monotone lattice submodular functions [@Gottschalk:2015fq].
- Double Greedy with a polynomial number of oracle calls with error parameter $\epsilon>0$ ($_{\epsilon}$, Algorithm \[alg:polynomial-oracle\]).
We measure the efficiency of an algorithm by the number of oracle calls instead of the total running time. Indeed, the running time for evaluating the input function is the dominant factor of the total running, because objective functions in typical machine learning tasks contain sums over all data points, which is time consuming. Therefore, we do not consider the polynomial-time algorithm (Theorem \[the:truely-polynomial\]) here.
Revenue maximization
--------------------
In this application, we consider revenue maximization on an (undirected) social network $G = (V, W)$, where $W = (w_{ij})_{i,j\in V}$ represents the weights of edges. The goal is to offer for free or advertise a product to users so that the revenue increases through their word-of-mouth effect on others. If we invest $x$ units of cost on a user $i \in V$, the user becomes an advocate of the product (independently from other users) with probability $1-(1-p)^x$, where $p \in (0,1)$ is a parameter. This means that, for investing a unit cost to $i$, we have an extra chance that the user $i$ becomes an advocate with probability $p$. Let $S \subseteq V$ be a set of users who advocate the product. Note that $S$ is a random set. Following a simplified version of the model introduced by [@Hartline:2008fw], the revenue is defined as $\sum_{i \in S} \sum_{j \in V \setminus S} w_{ij}$. Let $f:{\mathbb{Z}}_+^E \to {\mathbb{R}}$ be the expected revenue obtained in this model, that is, $$\begin{aligned}
f({{\bm x}}) & = {\mathop{\mathbf{E}}}_{S} \Bigl[\sum_{i \in S} \sum_{j \in V \setminus S} w_{ij}\Bigr] \\
& = \sum_{i \in S} \sum_{j \in V \setminus S} w_{ij}(1-(1-p)^{{{\bm x}}(i)})(1-p)^{{{\bm x}}(j)}.\end{aligned}$$ It is not hard to show that $f$ is non-monotone DR-submodular function (see Appendix \[sec:proof-of-DRsubmod\] for the proof).
In our experiment, we used three networks, Adolescent health (2,539 vertices and 12,969 edges), Advogato (6,541 vertices and 61,127 edges), and Twitter lists (23,370 vertices and 33,101 edges), all taken from [@KONECT]. We regard all the networks as undirected. We set $p = 0.0001$, and set $w_{ij}=1$ when an edge exists between $i$ and $j$ and $w_{ij}=0$ otherwise. We imposed the constraint $0 \leq {{\bm x}}(e) \leq B$ for every $e \in E$, where $B$ is chosen from ${\{10^2,\ldots,10^6\}}$.
Table \[tab:ov\] shows the objective values obtained by each method. As can be seen, except for , which is clearly the worst, the choice of a method does not much affect the obtained objective value for all the networks. Notably, even when $\epsilon$ is as large as $0.5$, the objective values obtained by are almost the same as and .
Figure \[fig:oc-ml\] illustrates the number of oracle calls of each method. The number of oracle calls of and is linear in $B$, whereas that of slowly grows. Although the number of oracle calls of also slowly grows, it is always orders of magnitude larger than that of with $\epsilon=0.5$ or $\epsilon=0.05$.
In summary, $_{0.5}$ achieves almost the best objective value, whereas the number of oracle calls is two or three orders of magnitude smaller than those of the other methods when $B$ is large.
Conclusions {#sec:conclusions}
===========
In this paper, we proposed a polynomial-time $\frac{1}{2+\epsilon}$-approximation algorithm for non-monotone DR-submodular function maximization. Our experimental results on the revenu maximization problem showed the superiority of our method against other baseline algorithms. Maximizing a submodular set function under constraints is well studied [@Lee:2009tc; @Gupta:2010wj; @Chekuri:2014ed; @Mirzasoleiman:2016vp]. An intriguing open question is whether we can obtain polynomial-time algorithms for maximizing DR-submodular functions under constraints such as cardinality constraints, polymatroid constraints, and knapsack constraints.
### Acknowledgments {#acknowledgments .unnumbered}
T. S. is supported by JSPS Grant-in-Aid for Research Activity Start-up. Y. Y. is supported by JSPS Grant-in-Aid for Young Scientists (B) (No. 26730009), MEXT Grant-in-Aid for Scientific Research on Innovative Areas (No. 24106003), and JST, ERATO, Kawarabayashi Large Graph Project.
| {
"pile_set_name": "ArXiv"
} |
[ ]{}4.3in
[Noncommutative Differential Geometry and Twisting of Quantum Groups]{}\
\
[ ]{}\
Shahn Majid\
[ ]{}\
School of Mathematical Sciences, Queen Mary and Westfield College\
University of London, Mile End Rd, London E1 4NS[^1]
> [**Abstract**]{} We outline the recent classification of differential structures for all main classes of quantum groups. We also outline the algebraic notion of ‘quantum manifold’ and ‘quantum Riemannian manifold’ based on quantum group principal bundles, a formulation that works over general unital algebras.
Introduction
============
There have been many attempts in the last decades to arrive at a theory of noncommutative geometry applicable to ‘coordinate’ algebras that are not necessarily commutative, notably that of A. Connes coming out of abstract $C^*$-algebra theory in the light of the Gelfand-Naimark and Serre-Swan theorems. One has tools such as cyclic cohomology and examples such as the noncommutative torus and other foliation $C^*$-algebras. Another ‘bottom up’ approach, which we outline, is based on the idea that the theory should be guided by the inclusion of the large vein of ‘naturally occuring’ examples, the coordinate algebras of the quantum groups $U_q(g)$ in particular, and Hopf algebras in general, whose validity for several branches of mathematics has already been established. This is similar to the key role that Lie groups played in the development of modern differential geometry. Much progress has been made in recent years and there is by now (at least at the algebraic level) a more or less clear formulation of ‘quantum manifold’ suggested by this approach. After being validated on the q-deformation examples such as quantum groups, quantum homogeneous spaces etc, one can eventually apply the theory quite broadly to a wide range of unital algebras. The approach will be algebraic, although not incompatible with $C^*$ completions at a later stage. In particular, as a bonus, one can apply the theory to finite-dimensional algebra, i.e. to discrete classical and quantum systems.
An outline of the paper is the following. We start with the lowest level structure which (in our approach) is the choice of differential structure. This is the topic of Section 2 where we outline the recently achieved more or less complete classification results. In Section 3 we develop the notion of ‘quantum manifold’[@Ma:rie] based on noncommutative frame bundles with quantum group fibre. Usual ideas of ‘sheaf theory’ and ‘local trivialisations’ do not work in this setting, but from [@BrzMa:gau] one has global algebraic replacements. There is also an appropriate notion of automorphism or diffeomorphism quantum groups[@Ma:dif].
Quantum differential forms
==========================
Let $M$ be a unital algebra, which we consider as playing the role of ‘co-ordinates’ in algebraic geometry, except that we do not require the algebra to be commutative. The appropriate notion of cotangent space or differential 1-forms in this case is
1\. $\Omega^1$ an $M$-bimodule
2\. ${{\rm d}}:M\to \Omega^1$ a linear map obeying the Leibniz rule ${{\rm d}}(ab)=a{{\rm d}}b+({{\rm d}}a)b$ for all $a,b\in M$.
3\. The map $M{\mathop{\otimes}}M\to \Omega^1$, $a{\mathop{\otimes}}b\mapsto a{{\rm d}}b$ is surjective.
Differential structures are not unique even classically, and even more non-unique in the quantum case. There is, however, one universal example of which others are quotients. This is [$$\Omega^1_{\rm univ}=\ker\cdot\subset M{\mathop{\otimes}}M,\quad {{\rm d}}a=a{\mathop{\otimes}}1-1{\mathop{\otimes}}a.\label{univ}$$]{} This is common to more or less all approaches to noncommutative geometry.
The main feature here is that, in usual algebraic geometry, the multiplication of forms $\Omega^1$ by ‘functions’ $M$ is the same from the left or from the right. However, if $a{{\rm d}}b=({{\rm d}}b)a$ then by axiom 2. we have ${{\rm d}}(ab-ba)=0$, i.e. we cannot naturally suppose this when $M$ is noncommutative. We say that a differential calculus is noncommutative or ‘quantum’ if the left and right multiplication of forms by functions do not coincide.
When $M$ has a Hopf algebra structure with coproduct $\Delta:M\to
M{\mathop{\otimes}}M$ and counit ${{\epsilon}}:M\to k$ ($k$ the ground field), we say that $\Omega^1$ is [*bicovariant*]{} if[@Wor:dif]
4\. $\Omega^1$ is a bicomodule with coactions $\Delta_L:\Omega^1\to
M{\mathop{\otimes}}\Omega^1,\Delta_R:\Omega^1\to \Omega^1{\mathop{\otimes}}M$ bimodule maps (with the tensor product bimodule structure on the target spaces, where $M$ is a bimodule by left and right multiplication).
5\. ${{\rm d}}$ is a bicomodule map with the left and right regular coactions on $M$ provided by $\Delta$.
A morphism of calculi means a bimodule and bicomodule map forming a commuting triangle with the respective ${{\rm d}}$ maps. One says[@Ma:cla] that a calculus is [*coirreducible*]{} if it has no proper quotients. Whereas the translation-invariant calculus is unique classically, in the quantum group case we have at least complete classification results in terms of representation theory[@Ma:cla]. The dimension of a calculus is that of its space of (say) left-invariant 1-forms, which can be viewed as generating the rest of the calculus as a right $M$-module. Similarly with left and right interchanged.
We note also that in the bicovariant case there is a natural extension[@Wor:dif] from $\Omega^1$ to $\Omega^n$ with ${{\rm d}}^2=0$. This is defined as the tensor algebra over $M$ generated by $\Omega^1$ modulo relations defined by a braiding which acts by a simple transposition on left-invariant and right-invariant forms. Other extensions are also possible and in general the differential structure can be specified order by order. Given the extension, one has a quantum DeRahm cohomology defined in the usual way as closed forms modulo exact ones. Apart from cohomology one can also start to do ‘$U(1)$’ gauge theory with trivial bundles, where a gauge field is just a differential form $\alpha\in \Omega^1$ and its curvature is $F={{\rm d}}\alpha+\alpha\wedge\alpha$, etc. A gauge transform is [$$\alpha^\gamma=\gamma^{-1}\alpha\gamma
+\gamma^{-1}{{\rm d}}\gamma,\quad F^\gamma=\gamma^{-1} F\gamma\label{trivialgauge}$$]{} for any invertible ‘function’ $\gamma\in M$, and so on. One can define then the space of flat connections as those with $F=0$ modulo gauge transformation. This gives two examples of ‘geometric’ invariants which work therefore for general algebras equipped with differential structure.
$M=k[x]$
--------
For polynomials in one variable the coirreducible calculi have the form[@Ma:fie] $$\Omega^1=k_\lambda[x],\quad {{\rm d}}f(x)=\frac{f(x+\lambda)-f(x)}{\lambda},\quad f(x)\cdot
g(\lambda,x)=f(x+\lambda)g(\lambda,x),\quad g(\lambda,x)\cdot
f(x)=g(\lambda,x)f(x)$$ for functions $f$ and one-forms $g$. Here $k_\lambda$ is a field extension of the form $k[\lambda]$ modulo $m(\lambda)=0$ and $m$ is an irreducible monic polynomial. The dimension of the calculus is the order of the field extension or the degree of $m$.
For example, the calculi on ${\mathbb{C}}[x]$ are classified by $\lambda_0\in
{\mathbb{C}}$ (here $m(\lambda)=\lambda-\lambda_0$) and one has $$\Omega^1={{\rm d}}x{\mathbb{C}}[x],\quad {{\rm d}}f={{\rm d}}x\frac{f(x+\lambda_0)-f(x)}{\lambda_0},\quad x{{\rm d}}x=({{\rm d}}x)x+\lambda_0.$$ We see that the Newtonian case $\lambda_0=0$ is only one special point in the moduli space of quantum differential calculi. But if Newton had not supposed that differentials and forms commute he would have had no need to take this limit. What one finds with noncommutative geometry is that there is no need to take this limit at all. It is also interesting that the most important field extension in physics, ${\mathbb{R}}\subset{\mathbb{C}}$, can be viewed noncommutative-geometrically with complex functions ${\mathbb{C}}[x]$ the quantum 1-forms on the algebra of real functions ${\mathbb{R}}[x]$. There is nontrivial quantum DeRahm cohomology in this case.
$M={\mathbb{C}}[G]$
-------------------
For the coordinate algebra of a finite group $G$ (for convenience we work over ${\mathbb{C}}$) the coirreducible calculi correspond to nontrivial conjugacy classes ${\hbox{{$\cal C$}}}\subset G$ and have the form $$\Omega^1={\hbox{{$\cal C$}}}\cdot{\mathbb{C}}[G],\quad {{\rm d}}f
=\sum_{g\in{\hbox{{$\cal C$}}}}g\cdot (L_g(f)-f),\quad f\cdot g=g\cdot L_g(f)$$ where $L_g(f)=f(g\cdot)$ is the translate of $f\in {\mathbb{C}}[G]$. The dimension of the calculus is the order of the conjugacy class.
For the coordinate algebra ${\mathbb{C}}[G]$ of a Lie group with Lie algebra ${\mathfrak{g}}$ the coirreducible calculi correspond to maximal ideals in $\ker{{\epsilon}}$ stable under the adjoint coaction. Or in a natural reformulation[@Ma:cla] in terms of quantum tangent spaces the correspondence is with irreducible $Ad$-invariant subspaces of the enveloping algebra $\ker{{\epsilon}}\subset U({\mathfrak{g}})$ which are stable under the coaction $\Delta_L=\Delta-{{\rm id}}{\mathop{\otimes}}1$ of $U({\mathfrak{g}})$. For example ${\mathfrak{g}}$ itself defines the standard translation-invariant calculus and this is coirreducible when ${\mathfrak{g}}$ is semisimple.
$M={\mathbb{C}}G$
-----------------
For the group algebra of a nonAbelian finite group $G$, we definitely need the machinery of noncommutative geometry since $M$ itself is noncommutative. We regard these group algebras ‘up side down’ as if coordinates, i.e. we describe the geometry of the noncommutative space $\hat G$ in some sense. The above definitions make sense and differential structures abound. The coirreducible calculi correspond to pairs $(V,\rho,\lambda)$ where $(V,\rho)$ is a nontrivial irreducible representation and $\lambda\in V/{\mathbb{C}}$[@Ma:cla]. They have the form $$\Omega^1=V\cdot {\mathbb{C}}G,\quad {{\rm d}}g
=((\rho(g)-1)\lambda)\cdot g,\quad g\cdot v=(\rho(g)v)\cdot g$$ where $g\in G$ is regarded as a ‘function’. The dimension of the calculus is that of $V$. The minimum assumption for merely a differential calculus is that $\lambda$ should be cyclic.
For $M=U({\mathfrak{g}})$ (the Kirillov-Kostant quantisation of ${\mathfrak{g}}^*$) one has a similar construction for any irreducible representation $V$ of the Lie algebra ${\mathfrak{g}}$ and choice of ray $\lambda$ in it. Then $$\Omega^1=V\cdot
U({\mathfrak{g}}),\quad {{\rm d}}\xi=\rho(\xi)\lambda,\quad \xi\cdot
v=\rho(\xi)v+v\cdot\xi$$ where $\xi\in{\mathfrak{g}}$ is regarded as a ‘function’. The dimension is again that of $V$.
For example, let ${\mathfrak{g}}=b_+$ be the 2-dimensional Lie algebra with $[x,t]=x$. Let $V$ be the 2-dimensional representation with matrix and ray vector $$\rho(t)=\begin{pmatrix}0&0\\ 0&1\end{pmatrix},\quad
\rho(x)=\begin{pmatrix}0&1\\ 0&0\end{pmatrix},\quad
\lambda=\begin{pmatrix}0\\ 1\end{pmatrix}.$$ Then ${{\rm d}}t=\lambda$ and ${{\rm d}}x$ are the usual basis of $V$ and obey $$[t,{{\rm d}}x]=[x,{{\rm d}}x]=0,\quad [t,{{\rm d}}t]={{\rm d}}t,\quad
[x,{{\rm d}}t]={{\rm d}}x.$$ Replacing $x$ trivially by a vector $x_i$, $i=1,2,3$ one obtains similarly a natural candidate for noncommutative Minkowski space along with its differential structure. It has measurable astronomical predictions[@AmeMa:wav].
This covers the classical objects or their duals viewed ‘up side down’ as noncommutative spaces. For a finite-group bicrossproduct ${\mathbb{C}}[M]{{\blacktriangleright\!\!\!\triangleleft}}{\mathbb{C}}G$ the classification is a mixture of the two cases above and is given in [@BegMa:dif]. The Lie version remains to be worked out in detail. The important example of the Planck scale quantum group ${\mathbb{C}}[x]{{\blacktriangleright\!\!\!\triangleleft}}{\mathbb{C}}[p]$, however, is a twisting by a cocycle of its classical limit ${\mathbb{C}}[{\mathbb{R}}{{\triangleright\!\!\!<}}{\mathbb{R}}]$ and is therefore covered by a later subsection.
Proofs
------
The above cases are all sufficiently elementary that they can be easily worked out using the following simple observations known essentially (in some form or other) since [@Wor:dif]. We suppose for convenience that $H$ has invertible antipode.
1\. $\ker{{\epsilon}}\subset M$ is an object in the braided category of left crossed $M$-modules (i.e. modules over the quantum double $D(M)$ in the finite-dimensional case) by multiplication and the left adjoint coaction.
2\. The isomorphism $M{\mathop{\otimes}}M{{\cong}}M{\mathop{\otimes}}M$ given by $a{\mathop{\otimes}}b\mapsto
(\Delta a)b$ restricts to an isomorphism $\Omega^1_{\rm Univ}{{\cong}}\ker{{\epsilon}}{\mathop{\otimes}}M$ of bimodules and of bicomodules, where the right hand side is a right (co)module by the (co)product of $M$ and a left (co)module by the tensor product of the (co)action on $\ker{{\epsilon}}$ and the (co)product of $M$.
This implies that every other bicovariant $\Omega^1$ is of the form $\Omega^1{{\cong}}\Omega_0{\mathop{\otimes}}M$ where $\Omega_0$ is a quotient object of $\ker{{\epsilon}}$ in the category of crossed $M$-modules. I.e. the calculi correspond to ideals in $\ker{{\epsilon}}$ stable under the adjoint coaction. Given $\Omega^1$ the space $\Omega_0$ is given by the right-invariant differentials. In categorical terms the braided category of bicovariant $M$-bimodules as featuring above (i.e. bimodules which are also bicomodules with structure maps being bimodule maps) can be identified with that of crossed $M$-modules, under which the Hopf module for the universal calculus corresponds to $\ker{{\epsilon}}$.
When this is combined with the notion of coirreducibility and with the Peter-Weyl decomposition of an appropriate type for $\ker{{\epsilon}}$, one obtains the classification results above. These latter steps have been introduced by the author[@Ma:cla] (before that one found only sporadic examples of calculi on particular quantum groups, usually close to the unique classical calculus.)
We also note that for any finite-dimensional bicovariant calculus the map ${{\rm d}}:M\to \Omega_0{\mathop{\otimes}}M$ can be viewed as a ‘partial derivative’ ${\partial}_x:M\to M$ for each $x\in \Omega_0^*$. The space $\Omega_0^*$ is called the invariant ‘quantum tangent space’ and is often more important than the 1-forms in applications. These ${\partial}_x$ are not derivations but together form a braided derivation in the braided category of $M$-crossed modules (there is a braiding as $x\in
\Omega_0^*$ passes $a\in M$) [@Ma:cla].
Cotriangular quantum groups and twisting of calculi
---------------------------------------------------
We recall[@Ma:book] that if $M$ is a quantum group and $\chi:M{\mathop{\otimes}}M\to k$ a cocycle in the sense $$\chi(b{{{}_{\scriptscriptstyle(1)}}}{\mathop{\otimes}}c{{{}_{\scriptscriptstyle(1)}}})\chi(a{\mathop{\otimes}}b{{{}_{\scriptscriptstyle(2)}}}c{{{}_{\scriptscriptstyle(2)}}})=\chi(a{{{}_{\scriptscriptstyle(1)}}}{\mathop{\otimes}}b{{{}_{\scriptscriptstyle(1)}}})
\chi(a{{{}_{\scriptscriptstyle(2)}}}b{{{}_{\scriptscriptstyle(2)}}}{\mathop{\otimes}}c),\quad \chi(1{\mathop{\otimes}}a)={{\epsilon}}(a),\quad\forall a,b,c\in M$$ then there is a ‘twisted’ quantum group $M^\chi$ with product $$a\cdot_\chi b=\chi(a{{{}_{\scriptscriptstyle(1)}}}{\mathop{\otimes}}b{{{}_{\scriptscriptstyle(1)}}})a{{{}_{\scriptscriptstyle(2)}}}b{{{}_{\scriptscriptstyle(2)}}}\chi^{-1}(a{{{}_{\scriptscriptstyle(3)}}}{\mathop{\otimes}}b{{{}_{\scriptscriptstyle(3)}}})$$ and unchanged unit, counit and coproduct. Here $\chi^{-1}$ is the inverse in $(M{\mathop{\otimes}}M)^*$, which we assume, and $\Delta a=a{{{}_{\scriptscriptstyle(1)}}}{\mathop{\otimes}}a{{{}_{\scriptscriptstyle(2)}}}$, etc., is a notation.
[@MaOec:twi] The bicovariant differentials $\Omega^1(M^\chi)$ are in 1-1 correspondence with those of $M$.
In fact the entire exterior algbera in the bicovariant case is known to be a super-Hopf algebra (Brzezinski’s theorem) and that of $M^\chi$ is the twist of that of $M$ when $\chi$ is trivially extended to a cocycle on the latter. The more direct proof involves the following:
[@MaOec:twi] There is an equivalence ${\hbox{{$\cal F$}}}$ of braided monoidal categories from left $M$-crossed modules to left $M^\chi$-crossed modules given by the functor $${\hbox{{$\cal F$}}}(V,{{\triangleright}},\Delta_L)=(V,{{\triangleright}}^\chi,\Delta_L),\quad a{{\triangleright}}^\chi v
=\chi(a{{{}_{\scriptscriptstyle(1)}}}{\mathop{\otimes}}v{{{}^{\scriptscriptstyle(\bar 1)}}})(a{{{}_{\scriptscriptstyle(2)}}}{{\triangleright}}v{{{}^{\scriptscriptstyle(\bar 2)}}}){{{}^{\scriptscriptstyle(\bar 2)}}}\chi^{-1}((a{{{}_{\scriptscriptstyle(2)}}}{{\triangleright}}v{{{}^{\scriptscriptstyle(\bar 2)}}})
{{{}^{\scriptscriptstyle(\bar 1)}}}{\mathop{\otimes}}a{{{}_{\scriptscriptstyle(3)}}}),\quad \forall a\in M,\ v\in V,$$ where ${{\triangleright}}$ denotes the action and $\Delta_Lv=v{{{}^{\scriptscriptstyle(\bar 1)}}}{\mathop{\otimes}}v{{{}^{\scriptscriptstyle(\bar 2)}}}$ is a notation. There is an associated natural transformation $$c_{V,W}: {\hbox{{$\cal F$}}}(V){\mathop{\otimes}}{\hbox{{$\cal F$}}}(W){{\cong}}{\hbox{{$\cal F$}}}(V{\mathop{\otimes}}W),\quad c_{V,W}(v{\mathop{\otimes}}w)
=\chi(v{{{}^{\scriptscriptstyle(\bar 1)}}}{\mathop{\otimes}}w{{{}^{\scriptscriptstyle(\bar 1)}}})v{{{}^{\scriptscriptstyle(\bar 2)}}}{\mathop{\otimes}}w{{{}^{\scriptscriptstyle(\bar 2)}}}.$$
As a corollary we deduce by Tannaka-Krein reconstruction arguments:
[@MaOec:twi] When $M$ is finite-dimensional the dual of the Drinfeld double $D(M^\chi)$ is isomorphic to a twist of the dual of the Drinfeld double $D(M)$ by $\chi^{-1}$ viewed on $D(M){\mathop{\otimes}}D(M)$.
Here we use the theorem[@Ma:tan] that an equivalence of comodule categories respecting the forgetful functor corresponds to a twist of the underlying quantum groups. The corollary itself can then be verified directly at an algebraic level once the required (nontrivial) isomorphism has been found in this way. Of course one can state it also in terms of Drinfeld’s coproduct twists.
Starting with a classical (commutative) Hopf algebra such a twist yields a cotriangular one and (from recent work of Etingof and Gelaki[@EtiGel]) every finite-dimensional cotriangular Hopf algebra in the (co)semisimple case over $k$ algebraically closed is of this form. Hence the differential calculus in this case reduces by the above theorem to the classification in the classical cases considered in previous sections. There are many other instances where an important quantum group is a twisting of another – the theorem provides its differential calculus from that of the other.
Factorisable quantum groups $M={\mathbb{C}}_q[G]$
-------------------------------------------------
Finally, we come to the standard quantum groups ${\mathbb{C}}_q[G]$ dual to the Drinfeld-Jimbo $U_q({\mathfrak{g}})$. Here[@Ma:cla] the coirreducible calculi are essentially provided by nontrivial finite-dimensional irreducible right comodules $V$ of the quantum group (i.e. essentially by the irreducible representations of the Lie algebra) and have the form [$$\Omega^1={{\rm End}}(V)\cdot{\mathbb{C}}_q[G],\quad {{\rm d}}a
=\rho_+(a{{{}_{\scriptscriptstyle(1)}}})\circ \rho_-(Sa{{{}_{\scriptscriptstyle(2)}}})\cdot a{{{}_{\scriptscriptstyle(3)}}}-{{\rm id}}\cdot a,\quad
a\cdot\phi=\rho_+(a{{{}_{\scriptscriptstyle(1)}}})\circ\phi\circ \rho_-(S a{{{}_{\scriptscriptstyle(2)}}})\cdot a{{{}_{\scriptscriptstyle(3)}}}\label{diffqG}$$]{} for all $\phi\in{{\rm End}}(V)$, where $$\rho_+(a)v=v{{{}^{\scriptscriptstyle(\bar 1)}}}{\hbox{{$\cal R$}}}(a{\mathop{\otimes}}v{{{}^{\scriptscriptstyle(\bar 2)}}}),\quad \rho_-(a)v
=v{{{}^{\scriptscriptstyle(\bar 1)}}}{\hbox{{$\cal R$}}}^{-1}(v{{{}^{\scriptscriptstyle(\bar 2)}}}{\mathop{\otimes}}a)$$ and ${\hbox{{$\cal R$}}}:{\mathbb{C}}_q[G]{\mathop{\otimes}}{\mathbb{C}}_q[G]\to {\mathbb{C}}$ is the dual-quasitriangular structure[@Ma:book]. $S$ denotes the antipode. The construction works for any dual-quasitriangular Hopf algebra with factorisable ${\hbox{{$\cal R$}}}$ (the minimum one needs for a differential calculus is that ${\hbox{{$\cal Q$}}}_\rho(a)=\rho_+(a{{{}_{\scriptscriptstyle(1)}}})\rho_-(Sa{{{}_{\scriptscriptstyle(2)}}})$ is surjective) and gives a classification of calculi if $M$ has in addition the Peter-Weyl property that $M=\oplus_V V{\mathop{\otimes}}V^*$ as left and right $M$-comodules. This is the result in [@Ma:cla Thm. 4.3] cast in a purely comodule form. Or in the original formulation in terms of quantum tangent spaces the correspondence is given in more familiar terms with representations $\rho$ of the quantum enveloping algebra and $$\rho_+(a)=(a{\mathop{\otimes}}\rho)({\hbox{{$\cal R$}}}),\quad \rho_-(a)=(\rho{\mathop{\otimes}}a)({\hbox{{$\cal R$}}}^{-1}),$$ where ${\hbox{{$\cal R$}}}$ is the quasitriangular structure or universal R-matrix of $U_q({\mathfrak{g}})$. For finite-dimensional representations only a finite number of terms in its powerseries contribute here, i.e. there are no convergence problems.
The factorisability etc. hold formally for ${\mathbb{C}}_q[G]$ so that although one has one natural calculus for each irreducible representation there are also some ‘shadows’ or technical variants allowed according to the precise formulation of the relevant quantum groups and their duality (this is more a deficit in the technical definitions than anything else). The latter aspect has been subsequently clarified in [@BauSch][@HecSch] following our basic result in [@Ma:cla].
For the sake of a sketch of the proof of the basic result assume that $M$ is strictly factorisable dual-quasitriangular and has Peter-Weyl decomposition in terms of irreducible representations $(V,\rho)$ of a suitable dual Hopf algebra $U$. Classifying the quotient $M$-crossed modules of $\ker{{\epsilon}}$ is equivalent essentially to classifying the subobjects of $\ker{{\epsilon}}\subset U$ as $U$-crossed modules (the quantum tangent spaces). When $U$ is strictly factorisable its quantum double $D(U)$ is isomorphic to $U{{\blacktriangleright\!\!\blacktriangleleft}}U$ which, as an algebra, is a tensor product (the coproduct is twisted). Hence $U$-crossed modules are equivalent to two $U$-modules. Next, under the isomorphism of linear spaces $U{{\cong}}M$ assumed under strict factorisability, this is the same as classifying subobjects of $\ker{{\epsilon}}\subset M$. The $U$-crossed module structure on this final $\ker{{\epsilon}}\subset M$ under this chain of reasoning is simply evaluation against $M$ coacting independently from the left and the right (viewed from the left via the antipode). This is just the action with respect to which the assumed Peter-Weyl decomposition $M=\oplus_V {{\rm End}}(V)$ is a decomposition into irreducibles as $V$ runs over the irreducible representations of $U$. One may make a similar proof working only with $M$-crossed modules and $M$-comodules throughout and the corresponding comodule Peter-Weyl decomposition.
[@Ma:cla] The quantum tangent spaces $\Omega_0^*=V^*{\mathop{\otimes}}V$ for the above differential calculi on ${\mathbb{C}}_q[G]$ are braided-Lie algebras in the sense of [@Ma:lie]. The action of basis element $f^i{\mathop{\otimes}}e_i$ is $${\partial}_{x^i{}_j} (a)={\langle}x^i{}_j,a{{{}_{\scriptscriptstyle(1)}}}{\rangle}a{{{}_{\scriptscriptstyle(2)}}},\quad x^i{}_j
={\hbox{{$\cal Q$}}}(\rho^i{}_j)-\delta^i{}_j$$ where ${\hbox{{$\cal Q$}}}:{\mathbb{C}}_q[G]\to U_q({\mathfrak{g}})$ is defined by ${\hbox{{$\cal R$}}}_{21}{\hbox{{$\cal R$}}}$ and $\rho^i{}_j\in {\mathbb{C}}_q[G]$ are the matrix elements in the representation $V$ with basis $\{e_i\}$ and dual basis $\{f^i\}$.
Recall that the dual of any quantum group acts on the quantum group by the ‘coregular representation’ in the manner shown, in our case by the $x^i{}_j$. These and $1$ together form a braided-Lie algebra. This is described by a system of axioms in any braided category including a pentagonal ‘braided-Jacobi’ identity. Moreover, such objects have braided enveloping algebras which, for usual Lie algebras ${\mathfrak{g}}$, returns a homogenized version of $U({\mathfrak{g}})$. In our case it returns a quadratic and braided version of $U_q({\mathfrak{g}})$, i.e. this solved (some years ago[@Ma:lie]) the Lie problem for such quantum groups. The above gives its geometric interpretation.
For the example of ${\mathbb{C}}_q[SU_2]$ there is basically one bicovariant calculus for each spin $j$ with dimension $(2j+1)^2$. The lowest corresponds to the 4-dimensional braided-Lie algebra $gl_{q,2}$ spanned by $$x=\begin{pmatrix}q^H-1& q^{-{{\scriptstyle\frac{1}{2}}}}(q-q^{-1})q^{H\over 2}X_-\\
q^{-{{\scriptstyle\frac{1}{2}}}}(q-q^{-1})X_+q^{H\over 2}&
q^{-H}-1+q^{-1}(q-q^{-1})^2X_+X_-\end{pmatrix}$$ in the usual notations for $U_q(su_2)$. This braided-Lie algebra is irreducible for generic $q$ but as $q\to 1$ it degenerates into $su_2\oplus u(1)$. The partial derivatives degenerate into the usual invariant vector fields on $SU_2$ and an additional 2nd order operator which turns out to be the Casimir or Laplacian.
Discrete manifolds
------------------
To close with one non-quantum group example, consider any actual manifold with a finite good cover $\{U_i\}_{i\in I}$. Instead of building geometric invariants on a manifold and studying them modulo diffeomorphisms we can use the methods above to first pass to the skeleton of the manifold defined by its open set structure and do differential geometry directly on this indexing set $I$. Thus we take $M={\mathbb{C}}[I]$ which just means collections $\{f_i\in {\mathbb{C}}\}$. The universal $\Omega^1$ is just matrices $\{f_{ij}\}$ vanishing on the diagonal. We use the intersection data for the open sets to set some of these to zero. Similarly for higher forms. Thus[@BrzMa:dif] $$\Omega^1=\{f_{ij}|\ U_i\cap U_j\ne \emptyset\},\quad
\Omega^2=\{f_{ijk}|\ U_i\cap U_j\cap U_k\ne\emptyset\}$$ $$({{\rm d}}f)_{ij}=f_i-f_j,\quad ({{\rm d}}f)_{ijk}
=f_{ij}-f_{ik}+f_{jk}$$ and so on. Then one has that the quantum cohomology is just the additive Cech cohomology of the original manifold. Similarly, one has that the zero curvature gauge fields modulo gauge transformations recovers again the first Cech cohomology, but now in a multiplicative form.
Bundles and connections
=======================
The next layer of differential geometry is bundles, connections, etc. Usually in physics one needs only the local picture with trivial bundles in each open set – but for a general noncommutative algebra $M$ there may be no reasonable ‘open sets’ and one has therefore to develop the global picture from the start. This is needed for example to describe the frame bundle of a topologically nontrivial ‘manifold’. It also turns out to be rather easier to do the gauge theory beyond the ‘$U(1)$’ case (i.e. with a nontrivial quantum structure group and nonuniversal calculus on it) if one takes the global point of view, even if the bundle itself is trivial. We take a Hopf algebra $H$ in the role of ‘functions’ on the structure group of the bundle. To keep things simple we concentrate on the universal differential calculus but it is important that the general case is also covered by making suitable quotients. Recall that a classical bundle has a free action of a group on the total space $P$ and a local triviality property. In our algebraic terms we need [@BrzMa:gau]:
1\. An algebra $P$ and a coaction $\Delta_R:P\to P{\mathop{\otimes}}H$ of the quantum group $H$ such that the fixed subalgebra is $M$, [$$M=P^H=\{p\in P|\ \Delta_R p=p{\mathop{\otimes}}1\}.\label{fixed}$$]{} We assume that $P$ is flat as an $M$-module.
2\. The sequence [$$0\to P(\Omega^1M)P\to\Omega^1P{\buildrel{\rm ver}
\over\longrightarrow} P{\mathop{\otimes}}\ker{{\epsilon}}\to 0\label{exactness}$$]{} is exact, where ${\rm ver}=(\cdot{\mathop{\otimes}}{{\rm id}})\Delta_R$.
The map ${\rm ver}$ plays the role of generator of the vertical vector fields corresponding classically to the action of the group (for each element of $H^*$ it maps $\Omega^1P\to P$ like a vector field). Exactness on the left says that the one-forms $P(\Omega^1M)P$ lifted from the base are exactly the ones annihilated by the vertical vector fields. In the universal calculus case this can be formulated as a Hopf-Galois extension, a condition arising in other contexts in Hopf algebra theory also. The differential geometric picture is more powerful and includes general calculi when we use the right-handed version of $\Omega_0$ in place of $\ker{{\epsilon}}$.
One can then define a connection as an equivariant splitting [$$\Omega^1P=P(\Omega^1 M)P\oplus {\rm complement}\label{connection}$$]{} i.e. an equivariant projection $\Pi$ on $\Omega^1P$. One can show[@BrzMa:gau] the required analogue of the usual theory, i.e. that such a projection corresponds to a connection form such that [$$\omega:\ker{{\epsilon}}\to\Omega^1P,\quad {\rm ver}\circ
\omega=1{\mathop{\otimes}}{{\rm id}}\label{conform}$$]{} where $\omega$ intertwines with the adjoint coaction of $H$ on itself. Finally, if $V$ is a vector space on which $H$ coacts then we define the associated ‘bundles’ $E^*=(P{\mathop{\otimes}}V)^H$ and $E=\hom^H(V,P)$, the space of intertwiners. The two bundles should be viewed geometrically as ‘sections’ in classical geometry of bundles associated to $V$ and $V^*$. Given a suitable (so-called strong) connection one has a covariant derivative[@BrzMa:gau] [$$D_\omega:E\to \Omega^1 M{\mathop{\otimes}}_M E,\quad D_\omega
=({{\rm id}}-\Pi)\circ {{\rm d}}.\label{covderiv}$$]{}
All of this can be checked out for the example of the $q$-monopole bundle over the $q$-sphere[@BrzMa:gau]. Recall that classically the inclusion $U(1)\subset SU_2$ in the diagonal has coset space $S^2$ and defines the $U(1)$ bundle over the sphere on which the monopole lives. In our case the coordinate algebra of $U(1)$ is the polynomials ${\mathbb{C}}[g,g^{-1}]$ and the classical inclusion becomes the projection $$\pi:{\mathbb{C}}_q[SU_2]\to {\mathbb{C}}[g,g^{-1}],\quad
\pi\left(\begin{matrix}
a&b\\ c&d
\end{matrix}\right)=\left(\begin{matrix}
g&0\\ 0&g^{-1}
\end{matrix}\right)$$ Its induced coaction $\Delta_R=({{\rm id}}{\mathop{\otimes}}\pi)\Delta$ is by the degree defined as the number of $a,c$ minus the number of $b,d$ in an expression. The quantum sphere ${\mathbb{C}}_q[S^2]$ is the fixed subalgebra i.e. the degree zero part. Explicitly, it is generated by $b_3=ad$, $b_+=cd$, $b_-=ab$ with $q$-commutativity relations $$b_\pm b_3=q^{\pm 2}b_3b_\pm+(1-q^{\pm 2})b_\pm,
\quad q^{2}b_-b_+=q^{-2}b_+b_-+(q-q^{-1})(b_3-1)$$ and the sphere equation $b_3^2=b_3+qb_-b_+$. When $q\to 1$ we can write $b_\pm=\pm(x\pm\imath y)$, $b_3=z+{{\scriptstyle\frac{1}{2}}}$ and the sphere equation becomes $x^2+y^2+z^2=\frac{1}{4}$ while the others become that $x,y,z$ commute. One may verify that we have a quantum bundle in the sense above and that there is a connection $\omega(g-1)=d{{\rm d}}a-qb{{\rm d}}c$ which, as $q\to 1$, becomes the usual Dirac monopole constructed algebraically. If we take $V=k$ with coaction $1\mapsto 1{\mathop{\otimes}}g^n$, the sections of the associated vector bundles $E_n$ for each charge $n$ are just the degree $n$ parts of ${\mathbb{C}}_q[SU_2]$. The associated covariant derivative acts on these.
This example demonstrates compatibility with the more traditional $C^*$-algebra approach of A. Connes[@Con:geo] and others. Traditionally a vector bundle over any algebra is defined as a finitely generated projective module. However, there was no notion of quantum principal bundle before quantum groups.
[@HajMa:pro] The associated bundles $E_n$ for the $q$-monopole bundle are finitely generated projective modules, i.e. there exist $$e_n\in M_{|n|+1}({\mathbb{C}}_q[S^2]),\quad e_n^2=e_n,\quad E_n
=e_n{\mathbb{C}}_q[S^2]^{|n|+1}.$$ The covariant derivative for the monopole has the form $e_n{{\rm d}}e_n$. The classes $[e_n]$ are elements of the noncommutative $K$-theory $K_0({\mathbb{C}}_q[S^2])$ and have nontrivial duality pairing with cyclic cohomology, hence the $q$-monopole bundle is nontrivial.
The potential applications of quantum group gauge theory hardly need to be elaborated. For example, for a classical manifold [$$\left\{ {{{\rm Flat\ connections\ on}\ G-{\rm bundle}}
\atop{\rm modulo\ gauge}}\right\}{{\cong}}\hom(\pi_1,G)/G\label{pi1}$$]{} using the holonomy. One can view this as a functor from groups to sets and the homotopy group $\pi_1$ as more or less the representing object in the category of groups. The same idea with quantum group gauge theory essentially defines a homotopy quantum group $\pi_1(M)$ for any algebra $M$ as more or less the representing object of the functor that assigns to a quantum group $H$ the set of zero-curvature gauge fields with this quantum structure group. This goes somewhat beyond vector bundles and $K$-theory alone. Although in principle defined, this idea has yet to be developed in a computable form.
Finally we mention that one needs to make a slight generalisation of the above to include other noncommutative examples of interest. In fact (and a little unexpectedly) the general theory above can be developed with only a coalgebra rather than a Hopf algebra $H$. Or dually it means only an algebra $A$ in place of the enveloping algebra of a Lie algebra. This was achieved more recently, in [@BrzMa:coa][@BrzMa:geo], and allows us to include the full 2-parameter quantum spheres as well as (in principle) all known $q$-deformed symmetric spaces. This setting of gauge theory based on inclusions of algebras could perhaps be viewed as an algebraic analogue of the notion of ‘paragroup’ in the theory of operator algebras. Also, in a different direction, one may do the quantum group gauge theory in any braided monoidal category at the level of braids and tangles[@Ma:diag] so that one has braided group gauge theory and in principle gauge theory for quasiassociative algebras such as[@AlbMa:qua] the octonions.
Quantum soldering forms and metrics
===================================
We are finally ready to take the plunge and offer at least a first definition of a ‘quantum manifold’. The approach we take is basically that of the existence of a bein or, in global terms, a soldering form. The first step is to define a generalised frame bundle or [*frame resolution*]{} of our algebra $M$ as[@Ma:rie]
1\. A quantum principal bundle $(P,H,\Delta_R)$ over $M$.
2\. A comodule $V$ and an equivariant ‘soldering form’ $\theta:V\to P\Omega^1M\subset\Omega^1P$ such that the induced map [$$E^*\to \Omega^1M,\quad p{\mathop{\otimes}}v\mapsto p\theta(v)\label{frame}$$]{} is an isomorphism.
What this does is to express the cotangent bundle as associated to a principal one. Other tensors are then similarly associated, for example vector fields are $E{{\cong}}\Omega^{-1}M$. Of course, all of this has to be done with suitable choices of differential calculi on $M,P,H$ whereas we have been focusing for simplicity on the universal calculi. There are some technicalities here but more or less the same definitions work in general. The working definition[@Ma:rie] of a [*quantum manifold*]{} is simply this data $(M,\Omega^1,P,H,\Delta_R,V,\theta)$. The definition works in that one has analogues of many usual results. For example, a connection $\omega$ on the frame bundle induces a covariant derivative $D_\omega$ on the associated bundle $E^*$ which maps over under the soldering isomorphism to a covariant derivative [$$\nabla:\Omega^1M\to \Omega^1M{\mathop{\otimes}}_M\Omega^1M.\label{nabla}$$]{} Its torsion is defined as corresponding similarly to $\bar D_\omega\theta$, where we use a suitable (right-handed) version of the covariant derivative.
Defining a Riemannian structure can be done in a ‘self-dual’ manner as follows. Given a framing, a ‘generalised metric’ isomorphism $\Omega^{-1}M{{\cong}}\Omega^1M$ between vector fields is equivalent to[@Ma:rie]
3\. Another framing $\theta^*:V^*\to(\Omega^1M)P$, which we call the [*coframing*]{}, this time with $V^*$.
The associated quantum metric is [$$g= \theta^*(f^a)\theta(e_a)\in\Omega^1M
{\mathop{\otimes}}_M\Omega^1M\label{quametric}$$]{} where $\{e_a\}$ is a basis of $V$ and $\{f^a\}$ is a dual basis.
Now, this self-dual formulation of ‘metric’ as framing and coframing is symmetric between the two. One could regard the coframing as the framing and vice versa. From our original point of view its torsion tensor corresponding to $D_\omega\theta^*$ is some other tensor, which we call the [*cotorsion tensor*]{}. A natural proposal for a generalised Levi-Civita connection on a quantum Riemannian manifold is therefore[@Ma:rie]
4\. A connection $\omega$ such that the torsion and cotorsion tensors both vanish.
There is a corresponding covariant derivative $\nabla$. The Riemannian curvature of course corresponds to the curvature of $\omega$, which is ${{\rm d}}\omega+\omega\wedge\omega$, via the soldering form. I would not say that the Ricci tensor and Einstein tensor are understood abstractly enough in this formalism but of course one can just write down the relevant contractions and proceed blindly.
[@Ma:rie] Every quantum group $M$ has a framing by $H=M$, $P=M{\mathop{\otimes}}M$, $V=\ker{{\epsilon}}$ and $\theta$ induced from the quantum group Maurer-Cartan form $e(v)=Sv{{{}_{\scriptscriptstyle(1)}}}{\mathop{\otimes}}v{{{}_{\scriptscriptstyle(2)}}}$. Likewise for all $M$ equipped with a bicovariant differential calculus, with $V=\Omega_0$.
In this construction one builds the framing from a $V$-bein $e$ inducing the isomorphism $\Omega^1(M)=M{\mathop{\otimes}}\Omega_0$ as in Section 2 (in a right-handed setting). Moreover, for quantum groups such as ${\mathbb{C}}_q[SU_2]$ there is an Ad-invariant non-degenerate braided Killing form[@Ma:lie] on the underlying braided-Lie algebra, which provides a coframing from a framing – so that quantum groups such as ${\mathbb{C}}_q[SU_2]$ with the corresponding bicovariant differential calculi are quantum Riemannian manifolds in the required sense. The existence of a generalised Levi-Civita connection in such cases remains open and may require one to go beyond strong connections.
At least with the universal calculus every quantum homogeneous space is a quantum manifold too. That includes quantum spheres, quantum planes etc. In fact, there is a notion of comeasuring or quantum automorphism bialgebra[@Ma:dif] for practically any algebra $M$ and when this has an antipode (which typically requires some form of completion) one can write $M$ as a quantum homogeneous space. So almost any algebra $M$ is more or less a quantum manifold for some principal bundle (at least rather formally). This is analogous to the idea that any classical manifold is, rather formally, a homogeneous space of diffeomorphisms modulo diffeomorphisms fixing a base point.
Finally, to get the physical meaning of the cotorsion tensor and other novel ideas coming out of this noncommutative Riemannian geometry, let us consider the semiclassical limit. What we find is that noncommutative geometry forces us to slightly generalise conventional Riemannian geometry itself [@Ma:rie]:
1\. We should allow any group $G$ in the ‘frame bundle’, hence the more general concept of a ‘frame resolution’ $(P,G,V,\theta_\mu^a)$ or [generalised manifold]{}.
2\. The [generalised metric]{} $g_{\mu\nu}
=\theta^*_\mu{}^a\theta_{\nu a}$ corresponding to a coframing $\theta^*_{\mu}{}^a$ is nondegenerate but need not be symmetric.
3\. The [generalised Levi-Civita]{} connection defined as having vanishing torsion and vanishing cotorsion respects the metric only in a skew sense [$$\nabla_\mu g_{\nu\rho}-\nabla_\nu g_{\mu\rho}=0\label{genlev}$$]{} and need not be uniquely determined.
This generalisation of Riemannian geometry includes special cases of symplectic geometry, where the generalised metric is totally antisymmetric. It is also remarkable that metrics with antisymmetric part are exactly what are needed in string theory to establish T-duality. In summary, one has on the table a general noncommutative Riemannian geometry to play with. It can be applied to a variety of algebras far removed from conventional geometry. Some finite dimensional examples will be presented elsewhere.
Acknowledgements {#acknowledgements .unnumbered}
----------------
It is a pleasure to thank the organisers of the LMS symposium in Durham for a thoroughly enjoyable conference.
[10]{}
S. Majid. Quantum and braided group [R]{}iemannian geometry. , 30:113–146, 1999.
T. Brzeziński and S. Majid. Quantum group gauge theory on quantum spaces. , 157:591–638, 1993. Erratum 167:235, 1995.
S. Majid. Quantum and braided diffeomorphism groups. , 28:94–128, 1998.
S.L. Woronowicz. Differential calculus on compact matrix pseudogroups (quantum groups). , 122:125–170, 1989.
S. Majid. Classification of bicovariant differential calculi. , 25:119–140, 1998.
S. Majid. Quantum geometry of field extensions. , 40:2311–2323, 1999.
G. Amelino-Camelia and S. Majid. Waves on noncommutative spacetime and gamma-ray bursts. (In press.)
E. Beggs and S. Majid. Quasitriangular and differential structures on bicrossproduct hopf algebras. , 219:682–727, 1999.
S. Majid. . Cambridge Univeristy Press, 1995.
S. Majid and R. Oeckl. Twisting of quantum differentials and the [P]{}lanck scale [H]{}opf algebra. , 205:617–655, 1999.
S. Majid. Tannaka-[K]{}rein theorem for quasi[H]{}opf algebras and other results. , 134:219–232, 1992.
P. Etingof and S. Gelaki. The classification of triangular semisimple and cosemisimple Hopf algebras over an algebraically closed field. , 1999.
P. Baumann and F. Schmitt. Classification of bicovariant differential calculi on quantum groups (a representation-theoretic approach). , 194:71–86, 1998.
I. Heckenberger and K. Schmudgen. Classification of bicovariant differential calculi on the quantum groups SLq(n+1) and Sp(q)(2n). , 502:141-162, 1998.
S. Majid. Quantum and braided [L]{}ie algebras. , 13:307–356, 1994.
T. Brzeziński and S. Majid. Quantum differentials and the [$q$]{}-monopole revisited. , 54:185–232, 1998.
A. Connes. . Academic Press, 1994.
P. Hajac and S. Majid. Projective module description of the [$q$]{}-monopole. , 206:246–464, 1999.
T. Brzezinski and S. Majid. Coalgebra bundles. , 191, 1998. 467-492.
T. Brzeziński and S. Majid. Quantum geometry of algebra factorisations and coalgebra bundles. (In press.)
S. Majid. Diagrammatics of braided group gauge theory. , 8:731–771, 1999.
H. Albuquerque and S. Majid. Quasialgebra structure of the octonions. , 220:188–224, 1999.
[^1]: Reader and Royal Society University Research Fellow
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The first optical proposal for the realization of the two-bit version of the Deutsch-Jozsa algorithm \[D. Deutsch and R. Jozsa, Proc. R. Soc. London A [**493**]{}, 553 (1992)\] is presented. The proposal uses Stark shifts in an ensemble of atoms and degenerate sources of photons. The photons interact dispersively with an atomic ensemble, leading to an effective Hamiltonian in atom-field basis, which is useful for performing the required two-qubit operations. Combining these with a set of one-qubit operations, the algorithm can be implemented. A discussion of the experimental feasibility of the proposal is given.'
address:
- '$^1$Department of Chemistry, University of Southern California, Los Angeles, CA 90089, USA'
- '$^2$Department of Physics, Oklahoma State University, Stillwater, OK - 74078, USA'
author:
- Shubhrangshu Dasgupta$^1$
- 'G. S. Agarwal$^2$[^1]'
title: 'Two-bit Deutsch-Jozsa algorithm using an atomic ensemble'
---
It is known that the quantum computers are able to perform certain specific jobs like factorization, searching a database etc. much faster than its classical counterparts. Several quantum algorithms have been developed to demonstrate the power of quantum computers. For example, Grover has proposed an algorithm to search a quantum state from an unsorted database of $2^n$ states [@grover]. This requires $O(2^{n/2})$ repetitions of certain unitary operations, demonstrating quadratic speed-up than the classical computers. Deutsch-Jozsa algorithm [@dj] can be used to identify a certain function of $n$ binary variables as constant or balanced by a single enquiry, whereas classically it would require up to $2^{n-1}+1$ enquiries. The key of this enormous speed of quantum computers lies in the so-called quantum parallelism, which enables one to apply the same unitary operation simultaneously on a number of basis states.
There have been several proposals and experiments on implementation of these algorithms. For example, Grover’s algorithm has been implemented using nuclear magnetic resonance (NMR) in bulk systems [@groverNMR]. NMR [@djNMR1; @djNMR2], ion trap [@djION], and linear optical [@mohseni] techniques have been employed to implement the Deutsch-Jozsa algorithm. Recently, an optical scheme to realize the one-bit version of the Deutsch-Jozsa algorithm was proposed [@chandan]. To date, most of the works on Deutsch-Jozsa algorithm is for the case of a single qubit and we need to realize this algorithm for larger number of qubits, so that the full power of this algorithm can be realized. In this paper, we show how this can be achieved using Stark shifts in an ensemble of atoms. In this model, two freely propagating photonic qubits encoded in their polarization modes interact dispersively with an atomic ensemble. This leads to an effective Hamiltonian which along with several one-qubit operations enables us to implement the Deutsch-Jozsa algorithm for two qubits. We further discuss the experimental feasibility of our scheme using current technology. We start with the main features of the algorithm. This algorithm helps to identify whether a bivalued function $f_i(x)$ (which can take only the values 0 or 1) of a variable $x$ is constant or balanced. Here $x$ is the decimal equivalent of the bit string $(x_1,x_2,...,x_n)$ for $n$ bits, i.e., $x=\sum_{i=1}^nx_i2^{n-i}$. Thus, $x$ can take any of the $2^n$ values between 0 and $2^n-1$, for different binary combinations. If for half the values of $x$, the function takes the value $0$ and for the other half, it takes the value 1, then the function is called balanced. The function will be called constant if it assumes either the value 0 or 1 for all values of $x$. In classical sense, it is easy to verify that at most $2^{n-1}+1$ evaluations of the function $f_i(x)$ for different values of $x$ are necessary to determine whether the function is constant or balanced \[see Table \[table1\]\]. The Deutsch-Jozsa algorithm, on the contrary, requires only a single “evaluation" of the function for the same. To demonstrate such an immense power of this algorithm, we show in Fig. 1 the basic circuit [@chuang] to perform the $n$-bit version of the Deutsch-Jozsa algorithm. It can be shown that if all the qubits of the system A are in $|0\rangle$ state, then the function $f_i(x)$ is a constant function. But if at least one of these $n$ qubits is in the state $|1\rangle$, then the function can be identified as a balanced function. In what follows, we will demonstrate the two-bit version of the Deutsch-Jozsa algorithm in an ensemble of atoms interacting with a quantized field.
Input
------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ----------
$x$ $f_1(x)$ $f_2(x)$ $f_3(x)$ $f_4(x)$ $f_5(x)$ $f_6(x)$ $f_7(x)$ $f_8(x)$
0 0 1 0 1 0 1 0 1
1 0 1 0 1 1 0 1 0
2 0 1 1 0 0 1 1 0
3 0 1 1 0 1 0 0 1
: \[table1\]Different possible functions showing their characterizations for $n=2$. Here $x$ is the decimal equivalent of the bit string $(x_1,x_2)$, e.g., $x=2$ for two-bit input (1,0).
In order to implement the Deutsch-Jozsa algorithm for two qubits, we need two distinguishable photonic qubits. Each photonic qubit is of course formed from two states of polarization. We would thus use two single photons propagating in opposite directions so as to keep the qubits distinguishable. Note that single photons have been used extensively as qubits for quantum computation based on linear optics [@fransonQC; @dowling]. Further, the implementation of the algorithm requires a number of Hadamard transformations of photons which can be done by using quarter-wave and half-wave plates. We show in Fig. \[fig3\] a possible experimental set-up for the interaction of distinguishable photonic qubits. We use a degenerate optical parametric amplifier to produce two photons simultaneously [@kwait95; @franson; @teich]. These photons interact with an ensemble of atoms. The experimental scheme shows how to make the photons distinct. The Deutsch-Jozsa algorithm requires that the polarization of two photons are measured simultaneously. For this purpose, we propose use of single-photon detectors [@milburn]. The photon “1" (“2") can be detected in either of the detectors HD1 (HD2) and VD1 (VD2), which detects it in either horizontal or vertical polarization. The polarization beam splitter before the detectors preselects the polarization. Clearly, if the two detectors (HD1 or VD1 and HD2 or VD2) click simultaneously, the polarization of photons can be measured and thus the experiment leads to a conclusion for the algorithm.
We consider the interaction of two distinct photons with an ensemble of $N$ atoms where only four levels are relevant as shown in Fig. \[fig2\]. Note that this kind of configuration can be found in the clock transitions in $^{133}$Cs atoms and has been studied very extensively in [@expts1; @expts2]. Each photon has two circular polarization modes $\epsilon_\pm$. We designate these modes as two orthogonal states $|1,0\rangle_k$ and $|0,1\rangle_k$, for the $k$th photon, where $|1,0\rangle$ ($|0,1\rangle$) refers to a single photon in $\epsilon_+$ ($\epsilon_-$) mode and no photon in $\epsilon_-$ ($\epsilon_+$) mode. The photons in the states $|1,0\rangle$ and $|0,1\rangle$ interact respectively with the $|e\rangle\leftrightarrow |g\rangle$ and $|e'\rangle\leftrightarrow |g'\rangle$ transitions of each atom. We assume that the common frequency difference $\Delta$ between the photon polarization mode and atomic transition is much larger than the atom-photon coupling constant $g$, which is the same for both the modes. In this regime, the ground states $|g\rangle$ and $|g'\rangle$ of each atom get Stark-shifted. Using second-order perturbation theory, the respective shifts can be calculated as $\hbar|g|^2n_+/\Delta$ and $\hbar |g|^2n_-/\Delta$, where $n_\pm$ are the total number of photons in $\epsilon_\pm$ modes. As the excited states $|e\rangle$ and $|e'\rangle$ of the atoms remain unpopulated by interaction with the photons in large detuning regime, the effective Hamiltonian can be written as $$\label{effH1}H_{\mathrm{eff}}=\hbar\lambda\sum_{j=1}^N\sum_{k=1}^2\left[|g_j\rangle\langle
g_j||1,0\rangle_k\langle 1,0|+|g'_j\rangle\langle g'_j||0,1\rangle_k\langle
0,1|\right]\;,\;\lambda=|g|^2/\Delta\;.$$ We rewrite the above Hamiltonian in terms of the horizontal ($|0\rangle_k$)and vertical ($|1\rangle_k$) polarization states of each $k$th photon as $$\begin{aligned}
H_{\mathrm{eff}}&=&\frac{1}{2}\hbar\lambda\sum_{j=1}^N\sum_{k=1}^2\left[|g_j\rangle\langle
g_j|(|0\rangle_k\langle 0|-|1\rangle_k\langle 1|+i|0\rangle_k\langle 1|+i|1\rangle_k\langle
0|)\right.\nonumber\\ \label{effH}&+&\left.|g'_j\rangle\langle g'_j|(|0\rangle_k\langle
0|-|1\rangle_k\langle 1|-i|0\rangle_k\langle 1|-i|1\rangle_k\langle 0|)\right]\;,\end{aligned}$$ where $|0\rangle_k=\frac{1}{\sqrt{2}}(|1,0\rangle_k+|0,1\rangle_k)$ and $|1\rangle_k=\frac{1}{i\sqrt{2}}(|1,0\rangle_k-|0,1\rangle_k)$ are identified as two basis states of the $k$th photonic qubit. As we deal with only two states of each atom, the ensemble of $N$ atoms can have $2^N$ states. We use two of them as the states of the atomic qubit, namely $\prod_{j=1}^N|g\rangle_j (\equiv |0\rangle_A) $ and $\prod_{j=1}^N|g'\rangle_j (\equiv
|1\rangle_A)$. Note that as the atomic qubit consists of only the ground states of the atoms, decoherence due to spontaneous emission does not affect the process. Further, as the photons in free space can have long decay time, they do not lead to any effective decoherence as well.
Under the action of the Hamiltonian (\[effH\]), the following evolution occurs: $$\begin{aligned}
|0\rangle_A|0,0\rangle_F &\longrightarrow& -e^{-i\lambda
Nt}|0\rangle_A(-i|0\rangle+|1\rangle)^{\otimes 2}\;,\nonumber\\
|0\rangle_A|1,1\rangle_F &\longrightarrow&e^{-i\lambda Nt}
|0\rangle_A(-i|0\rangle+|1\rangle)^{\otimes 2}\;,\nonumber\\
|0\rangle_A|0,1\rangle_F &\longrightarrow& -ie^{-i\lambda Nt}
|0\rangle_A(-i|0\rangle+|1\rangle)^{\otimes 2}\;,\nonumber\\
\label{phases}|0\rangle_A|1,0\rangle_F &\longrightarrow&-ie^{-i\lambda Nt}
|0\rangle_A(-i|0\rangle+|1\rangle)^{\otimes 2}\;,\end{aligned}$$ where the state $|0,1\rangle_F$ represents the photon 1 in the horizontal polarization state $|0\rangle_F$ and the photon 2 in the vertical polarization state $|1\rangle_F$ and so on. Similar evolution occurs when the atomic qubit is initially in the state $|1\rangle_A$. Note that the Hamiltonian (\[effH\]) changes the polarization of each photon from linear to circular polarization, i.e., the atomic ensemble acts as a polarizer for each photon. We next show how this change in polarization helps us to implement the Deutsch-Jozsa algorithm for two qubits. Now onwards, we choose the interaction time $T$ of the atomic ensemble with the photons such that $\lambda NT=\pi/2$. As shown previously [@chandan], this interaction time is achievable with the available technology [@expts1; @expts2]. For example, in an ensemble of $^{133}$Cs atoms in a cell of dimensions 10$\times$10$\times$200 $\mu$m$^3$ at room temperature, the coupling constant $g$ of each photon with the ensemble becomes $2.91\times 10^8$ s$^{-1}$ and time $T$ of interaction of the photon with the ensemble becomes $6.67\times 10^{-13}$ s. Then for an atomic density of the order of $5\times 10^{12}$ cm$^{-3}$ (i.e., for $N=10^5$), the required detuning of each photon can be calculated from the above relation as $\Delta=3.59$ GHz $\approx 12.35g$. On the other hand, for $D_1$ transitions (5$^2S_{1/2}\leftrightarrow 5^2P_{1/2}$, transition wavelength $\sim$ 795 nm) in an ensemble of $^{87}$Rb atoms in a magneto-optical trap (with diameter $\sim$ 0.5 mm) [@lin], we calculate $g\sim 3.53 \times 10^6$ s$^{-1}$ and $T\sim 1.67 \times 10^{-12}$ s. Thus for $N\sim
2.5\times 10^6$ in the trap, we find $\Delta\sim 9g$ Clearly, the condition $\Delta\gg g$ is well satisfied in both cases.
We now introduce the following four inequivalent Hadamard operations for each photon in $(|0\rangle_k, |1\rangle_k)$ basis and for each atom in the ensemble in ($|g\rangle_j,
|g'\rangle_j$) basis: $$\begin{aligned}
h_1=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}1&-1\\
1&1\end{array}\right)\;&;&
h_2=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}1&i\\ i&1\end{array}\right)\;;\nonumber\\
\label{Hadamards}h_3=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}1&1\\
-1&1\end{array}\right)\;&;&
h_4=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}1&-i\\
-i&1\end{array}\right)\;.\end{aligned}$$ These operations can be implemented by using a resonant microwave field coupling the levels $|g\rangle_j$ and $|g'\rangle_j$ of the $j$th atom. The corresponding interaction Hamiltonian is given by $$\label{microH}
H^j_\mu=-\hbar\Omega\left[e^{i\phi}|g'\rangle_j\langle
g|+\textrm{H.c.}\right]\;,$$ where $\Omega$ is the coupling constant of the microwave field with each atom in the ensemble and $\phi$ is the phase of the field with respect to the matrix element between the levels $|g\rangle_j$ and $|g'\rangle_j$. Under the action of this Hamiltonian, the rotations (\[Hadamards\]) for each atom can be obtained by choosing $\Omega t=\pi/4$ and $\phi=-\pi/2, 0,
\pi/2, \pi$, respectively. Note that we do not provide the above rotations for the collective atomic states $|0\rangle_A$ and $|1\rangle_A$. Alternatively, one could use two classical fields in Raman resonance to provide the equivalent rotation for each atom in the ensemble. For the photonic qubits, it is easy to verify that the operations (\[Hadamards\]) refer to rotation of linear polarizations to linear or circular polarizations, which can be performed using different sequences of quarter-wave and half-wave plates. Note that the matrix forms for transformations through quarter-wave plates and half-wave plates are given by $$\begin{aligned}
Q_\phi &=&
\frac{i}{\sqrt{2}}\left(\begin{array}{cc}\cos(2\phi)-i&\sin(2\phi)\\
\sin(2\phi)&-\cos(2\phi)-i\end{array}\right)\;;\nonumber\\
H_\phi &=& i\left(\begin{array}{cc}\cos(2\phi)&\sin(2\phi)\\
\sin(2\phi)&-\cos(2\phi)\end{array}\right)\;,\end{aligned}$$ in $(|0\rangle_k, |1\rangle_k)$ basis, where $\phi$ is the angle of alignment of the wave-plate with the axis perpendicular with its plane. The single-qubit operations (\[Hadamards\]) for photons can be obtained using a series of these wave-plates of different orientations \[“the SU(2) gadget"\] as follows [@gsa_jmo; @simon_mukunda]: $$\begin{aligned}
h_1 \equiv Q_{\pi/4}Q_{\pi/4}H_{-3\pi/8}\;\;;\;\;h_2\equiv Q_{\pi/4} %Q_{\pi/2}Q_{3\pi/4}H_{\pi/8}
\nonumber\\
h_3\equiv Q_{\pi/4}Q_{\pi/4}H_{-\pi/8}\;\;;\;\;h_4\equiv Q_{-\pi/4}\;.\end{aligned}$$ Note that the half-wave plate rotates the polarization state from $|0\rangle_k$ to $|1\rangle_k$ and vice versa for $\phi=\pi/4$.
Further, as required by the Deutsch-Jozsa algorithm, two photonic qubits should be initially in the same polarization state $|0\rangle_k$. This can be done by using a half-wave plate \[see Fig. \[fig3\]\]. The degenerate parametric amplifier produces two photons in horizontal and vertical polarization states. One of them (“2") is in vertical polarization $|1\rangle$ and thus is reflected by a polarization beam splitter. This photon is then sent through a half-wave plate (HWP in Fig. \[fig3\]) that rotates its polarization from $|1\rangle$ to $|0\rangle$, while the other photon (“1"), which is already in horizontal polarization state $|0\rangle$, passes through the polarization beam splitter. Thus both the photons are initialized to the state $|0\rangle_k$.
We now discuss how one can implement the Deutsch-Jozsa algorithm for two qubits using the evolution (\[phases\]) and (\[Hadamards\]). We assume that two photons (the atomic ensemble) serve the purpose of the system A (B) in Fig. \[fig1\]. We start with the state $|\psi_0\rangle_{F+A}=|0,0\rangle_F|1\rangle_A$ of the photons+ensemble system. Applying the Hadamard rotation $h_1$ on each photon and the atomic qubit, we obtain $$|\psi_1\rangle_{F+A}=\left[\frac{1}{\sqrt{2}}(|0\rangle+|1\rangle)\right]_F^{\otimes
2}\prod_{j=1}^N\frac{1}{\sqrt{2}}(|g_j\rangle-|g'_j\rangle)\;.$$ Next the $U_{f_i}$ operations are applied. We show that in the present case, $U_{f_i}=h_i^{\mathrm{eq}}U_{\mathrm{eff}}h_1^A$, where $U_{\mathrm{eff}}=\exp[-iH_{\mathrm{eff}}t]$ is the unitary operator corresponding to the Hamiltonian (\[effH\]). Here the operations $h_i^{\mathrm{eq}}$ correspond to different functions $f_i(x)$.
In course of the operator sequence $U_{f_i}$, the operation $h_1^A$ first prepares the atoms in $|0\rangle_A$ state: $$|\psi'_1\rangle_{F+A}=\left[\frac{1}{\sqrt{2}}(|0\rangle+|1\rangle)\right]_F^{\otimes
2}|0\rangle_A\;.$$ Then the operation $U_{\mathrm{eff}}$ for an interaction time $T$ defined by $\lambda NT=\pi/2$ yields the following \[see Eqs. (\[phases\])\]: $$\label{psi1DP}|\psi''_1\rangle_{F+A}\equiv\left[\frac{1}{\sqrt{2}}(-i|0\rangle+|1\rangle)\right]_F^{\otimes
2}|0\rangle_A\;.$$ We next identify the operations $h_i^{\mathrm{eq}}$ for different possible balanced functions $f_i(x)$ \[see Table \[table1\]\] as follows: $$h_{3,4}^{\mathrm{eq}}\equiv
h''_{F_1}h'_{F_2}\;,~~h_{5,6}^{\mathrm{eq}}\equiv
h'_{F_1}h''_{F_2}\;,~~h_{7,8}^{\mathrm{eq}}\equiv
h''_{F_1}h''_{F_2}\;.$$ Here $h'=h_1h_4h_3$ and $h''=h_1h_2h_3$ are composite sequences of different Hadamard rotations defined in Eq. (\[Hadamards\]) and yield the following: $$\begin{aligned}
&&h'|0\rangle=e^{i\pi/4}|0\rangle\;,\;h'|1\rangle=e^{-i\pi/4}|1\rangle\;;\nonumber\\
&&h''|0\rangle=e^{-i\pi/4}|0\rangle\;,\;h''|1\rangle=e^{i\pi/4}|1\rangle\;.\end{aligned}$$ Note that $h'_{F_{i}}$ refers to $h'$ operations on $i$th photon. The operators $h_i^{\mathrm{eq}}$ lead to the following results, when applied on the state $|\psi''_1\rangle$ in Eq. (\[psi1DP\]): $$\begin{aligned}
|\psi_2\rangle_{f_3,f_4}&=&h_{3,4}^{\mathrm{eq}}|\psi''_1\rangle\nonumber\\
&=&\pm\frac{1}{2}(|0,0\rangle+|0,1\rangle-|1,0\rangle-|1,1\rangle)_F|0\rangle_A\nonumber\\
&=&\pm\frac{1}{\sqrt{2}}(|0\rangle-|1\rangle)_{F_1}\frac{1}{\sqrt{2}}(|0\rangle+|1\rangle)_{F_2}|0\rangle_A\\
|\psi_2\rangle_{f_5,f_6}&=&h_{5,6}^{\mathrm{eq}}|\psi''_1\rangle\nonumber\\
&=&\pm\frac{1}{2}(|0,0\rangle-|0,1\rangle+|1,0\rangle-|1,1\rangle)_F|0\rangle_A\nonumber\\
&=&\pm\frac{1}{\sqrt{2}}(|0\rangle+|1\rangle)_{F_1}\frac{1}{\sqrt{2}}(|0\rangle-|1\rangle)_{F_2}|0\rangle_A\\
|\psi_2\rangle_{f_7,f_8}&=&h_{7,8}^{\mathrm{eq}}|\psi''_1\rangle\nonumber\\
&=&\pm\frac{1}{2}(|0,0\rangle-|0,1\rangle-|1,0\rangle+|1,1\rangle)_F|0\rangle_A\nonumber\\
&=&\pm\frac{1}{\sqrt{2}}(|0\rangle-|1\rangle)_{F_1}\frac{1}{\sqrt{2}}(|0\rangle-|1\rangle)_{F_2}|0\rangle_A\;.\end{aligned}$$
The final step of the algorithm is to apply the Hadamard rotation $h_1^F$ on each photon, which leads to the following outcomes:$
|\psi_3\rangle_{f_3,f_4}$$\equiv$ $|0,1\rangle_F|0\rangle_A$, $
|\psi_3\rangle_{f_5,f_6}$$\equiv$ $|1,0\rangle_F|0\rangle_A$, and $|\psi_3\rangle_{f_7,f_8}$$\equiv$ $|0,0\rangle_F|0\rangle_A$. Thus, after the final Hadamard rotations, either at least one of the photons remains in $|0\rangle_F$ state, if the functions were balanced.
On the other hand, in case of constant function $f_1$, the $U_{f_1}$ operation is equivalent to the identity operation, while for the function $f_2$, the $U_{f_2}$ operation is equivalent to the NOT operation on the atomic qubit. This can be implemented using microwave field for $\Omega t=\pi/2$ and $\phi=\pi/2$ \[see the Hamiltonian (\[microH\])\]. Then, the output states $|\psi_3\rangle_{f_1,f_2}$ of the photons+ensemble system become $|1,1\rangle_F\prod_{j=1}^N\frac{1}{\sqrt{2}}(|g\rangle_j\pm |g'\rangle_j)$. This means that if the functions were constant, both the photons can be detected in the state $|1\rangle_F$. Clearly, by measuring the polarization states of two photons at the end of the algorithm, one can characterize whether the functions were constant or balanced.
To this end, we discuss possibility of decoherence in the entire process. As the photons interact with the ensemble dispersively, the excited states of the atoms get hardly populated. So the effect of spontaneous emission during the process is minimal. However, there could be collisional relaxation of the atomic ground states, which leads to loss of coherence in these states. But the time-scales of both these relaxation processes are of the order of $10^{-6}$ s, whereas, interaction time $T$ of the photon with the ensemble is of the order of $10^{-13}$ s. Clearly, this is much less than the time at which the effects of different relaxation processes set in significantly. Thus, our model with photons and atomic ensemble is virtually decoherence-free.
In conclusion, we demonstrated how the Deutsch-Jozsa algorithm can be implemented for two qubits using an atomic ensemble. One requires two distinguishable photons for this. We show how this can be achieved by using the output of a degenerate parametric amplifier. These photons dispersively interact with an ensemble of $N$ atoms. The resulting Stark shifts of the atomic states lead to an effective Hamiltonian, which along with several Hadamard transformations, are used to implement the algorithm. We provide all the relevant operations to implement the algorithm and discussed the expected outcomes. One of us (G.S.A.) gratefully acknowledges the support from NSF grant No CCF 0524673.
References {#references .unnumbered}
==========
[999]{} Grover L K 1997 [*Phys. Rev. Lett.*]{} [**79**]{} 325 Deutsch D 1985 [*Proc. R. Soc. London*]{} A [**400**]{} 97 Deutsch D and Jozsa R 1992 [*Proc. R. Soc. London*]{} A [**493**]{} 553
Chuang I L, Gershenfeld N and Kubinec M 1998 [*Phys. Rev. Lett.*]{} [**80**]{} 3408
Chuang I L, Vandersypen L M K, Zhou X, Leung D W and Lloyd S 1998 [*Nature*]{} [**393**]{} 143 Jones T F and Mosca M 1998 [*J. Chem. Phys.*]{} [**109**]{} 1648
Dorai K, Arvind and Kumar A 2000 *Phys. Rev.* A [**61**]{} 042306 Das R and Kumar A 2003 *Phys. Rev.* A [**68**]{} 032304
Gulde S, Riebe M, Lancaster G P T, Becher C, Eschner J, Häffner H, Schmidt-Kaler F, Chuang I L and Blatt R 2003 *Nature* [**421**]{} 48
Mohseni M, Lundeen J S, Resch K J and Steinberg A M 2003 [**]{} **91** 187903
Dasgupta S, Biswas A and Agarwal G S 2005 *Phys. Rev.* A [**71**]{} 012333
Nielsen M A and Chuang I L 2002 [*Quantum Computation and Quantum Information*]{} (Cambridge University Press, Cambridge)
Franson J D, Donegan M M, Fitch M J, Jacobs B C and Pittman T B 2002 [**]{} **89** 137901 Pittman T B, Fitch M J, Jacobs B C and Franson J D 2003 *Phys. Rev.* A **68** 032316 Franson J D, Jacobs B C and Pittman T B 2004 [*Phys. Rev.*]{} A **70** 062302 Pittman T B, Jacobs B C and Franson J D 2005 [*Phys. Rev.*]{} A [**71**]{} 032307
Gingrich R M, Kok P, Lee H, Vatan F and Dowling J P 2003 ** **91** 217901 Kok P, Williams C P and Dowling J P 2003 *Phys. Rev.* A **68** 022301
Kwiat P G, Mattle K, Weinfurter H, Zeilinger A, Sergienko A V and Shih Y 1995 [**]{} [**75**]{} 4337 Kwiat P G, Waks E, White A G, Appelbaum I and Eberhard P H 1999 [*Phys. Rev.*]{} A **60** 773
Pittman T B, Jacobs B C and Franson J D 2002 *Phys. Rev.* A **66** 042303
Tsegaye T, Soderholm J, Atature M, Trifonov A, Bjork G, Sergienko A V, Saleh B E A and Teich M C 2000 [**]{} **85** 5013 Giuseppe G D, Atature M, Shaw M D, Sergienko A V, Saleh B E A and Teich M C 2002 [*Phys. Rev.*]{} A **66** 013801 Walton Z D, Sergienko A V, Saleh B E A and Teich M C 2004 [*Phys. Rev.*]{} A **70** 052317
Knill E, Laflamme K and Milburn G J 2001 *Nature* **409** 46
Kuzmich A, Bigelow N P and Mandel L 1998 *Europhys. Lett.* [**42**]{} 481 Kuzmich A, Mandel L and Bigelow N P 2000 *Phys. Rev. Lett.* [**85**]{} 1594
Takahashi Y, Honda K, Tanaka N, Toyoda K, Ishikawa K and Yabuzaki T 1999 *Phys. Rev.* A [**60**]{} 4974
Schappe R S, Walker T, Anderson L W and Lin C C 1996 [*Phys. Rev. Lett.*]{} **76** 4328
Agarwal G S and Chaturvedi S 2003 [*J. Mod. Opt.*]{} [**50**]{} 711
Simon R and Mukunda N 1990 [*Phys. Lett.*]{} A [**138**]{} 474 Simon R and Mukunda N 1990 [*Phys. Lett.*]{} A [**143**]{} 165
[^1]: On leave of absence from Physical Research Laboratory, Navrangpura, Ahmedabad - 380 009, India
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider an implementation of the Elitzur-Vaidman bomb experiment in a DC-biased electronic Mach-Zehnder interferometer with a leakage port on one of its arms playing the role of a “lousy bom”. Many-body correlations tend to screen out manifestations of interaction-free measurement. Analyzing the correlations between the current at the interformeter’s drains and at the leakage port, we identify the limit where the originally proposed single-particle effect is recovered. Specifically, we find that in the regime of sufficiently diluted injected electron beam and short measurement times, effects of quantum mechanical wave-particle duality emerge in the cross-current correlations.'
author:
- Oded Zilberberg
- Alessandro Romito
- Yuval Gefen
title: 'Many-body manifestation of interaction-free measurement: the Elitzur-Vaidman bomb'
---
Introduction
============
Since its introduction, quantum mechanics has kindled the imagination of scholars due to the interplay of its non-local character and particle-wave duality. Using recent advances in technological control over coherent systems, demonstration of these treats are still at the forefront of contemporary research [@haroche_nobel_2013]. In other words, a measurement of a quantum particle (the latter may be described as a wave packet) unveils its discrete nature, when it collapses to reside at a single point. The same particle, before “collapsing”, had assumed a non-local character. The compatibility of particle collapsing at a point and non-locality has been discussed and demonstrated in the context of the so-called Elitzur-Vaidman (EV) bomb \[aka “interaction free measurement” (IFM)\]: the wave-like interference of a single quantum particle is modified by the onset of a measurement (bomb) performed at one of an interferometer’s arms, which could (but may not) destroy the particle [@Elitzur:1993].
The interferometer at hand is tuned such that when the “bomb” is absent, wave-like destructive interference renders one of its output ports dark. One then introduces the bomb (hidden in a black box) in one of the interferometer’s arms. The bomb being “lousy” implies that even when a particle goes through that arm, there is a finite probability (possibly close to 1) that it will not explode. If the bomb eventually explodes, one knows *a posteriori* that the bomb was there. But there is a probability that the bomb does not go off, yet one detects a particle at the interferometer’s dark port. That would definitely indicate that the black box has modified the interference pattern, hence a bomb has been introduced inside the black box. The detection of the presence of the bomb occurs when no interaction with it took place. Notably, there is another possible [inconclusive]{} outcome: the bomb does not go off, and the interfering particle exits at the bright port. In that case one does not know whether the bomb was there or not. No matter how lousy the bomb is, within the many-body context of quantum physics, as the signal in the interferometer is collected over an ensemble of injected particles, there is a vanishing probability that the bomb would remain unexploded at asymptotically long times. Rather than a bomb, the realization of this EV experimental setup requires the construction of an interferometer with an absorber positioned on one of the interfering paths, as well as, the introduction of a single-particle source [@Kwiat:1995; @Voorthuysen:1996; @Hafner:1997; @Tsegaye:1998; @Kwiat:1999; @Jang:1999; @Hosten:2006; @Wolfgramm:2011]. As such, this topic has remained mostly in the realm of quantum optics where IFM experiments have been proposed and demonstrated in various systems [@Kwiat:1995; @Voorthuysen:1996; @Hafner:1997; @Tsegaye:1998; @Kwiat:1999; @Jang:1999; @Hosten:2006; @Wolfgramm:2011] with a variety of applications including imaging [@Kwiat:1998], quantum computing [@Hosten:2006; @Vaidman:2007], and single-photon generation [@Wolfgramm:2011].
Interestingly, several theoretical studies of the realization and utilization of IFM in electronic solid-state devices were recently pursued by considering, for example, superconducting quantum-bits (qubits) [@Paraoanu:2006]. Additionally, an earlier study of electronic Mach Zehnder interferometers (e-MZI) [@Strambini:2010; @Chirolli:2010], has focused on the employment of a wave-like picture, and the influence on the interference signal of a local perturbation in the interferometer. As such, the particle facet of the EV picture was missing. Indeed, e-MZI are realized using chiral edge modes of quantum Hall bars [@Halperin1982; @Wen1990], which are 1D channels well described as collective many-body plasmonic waves [@vonDelft:1998; @PhysRevB.62.7454; @ji2003electronic]. Typically, these devices are operated at constant voltage bias leading to the injection of numerous electrons that would eventually, with certainty, trigger the EV-bomb. We note, additionally, that single-particle excitations on top of the electron sea in quantum Hall edges have recently been obtained [@Dubois:2013]. All this implies that the topic of non-locality along with wave-particle duality in complex many-electron systems is amenable to experimental studies.
![ Illustration of the Mach-Zehnder interferometers (MZIs) under study. Chiral channels are represented by full lines leading from the sources ($S_1$ — biased and $S_2$ — grounded) to the drains ($D_1$ and $D_2$). Inter-edge tunneling takes place at intersection points. (a) A standard MZI with arms 1 and 2 of lengths $l_1$ and $l_2$, respectively. (b) The dangling end at $C$ (leading to $D_3$) serves as an absorber replacing the “lousy” bomb. []{data-label="setup"}](uncollapse3.pdf){width="0.8\columnwidth"}
Here we analyze the correlations of transport through an e-MZI with a leaking edge. This is an electronic manifestation of a variant of the EV-bomb where the leaky-edge corresponds to an absorber instead of a bomb [@Mitchison:2001]. In the particle-like limit of this device, the probability of a particle being absorbed and transmitted to the drains at the same time is zero. Such correlations in the case of many-particles will yield a non-vanishing result. This signifies the fact that the bomb may “explode” even if a signal is detected at the interferometer’s dark port. Employing a wave-like scattering matrix formulation, we compute the experimentally measurable many-body correlator and compare to two limiting cases (single particle impinging vs. a large influx of particles). Subsequently, we find the conditions for manifesting the wave-particle duality, and specifically obtaining the EV physics, in the context of many-body electronic system.
System
======
We consider a standard e-MZI geometry where particles are injected from the source $S_1$ and eventually detected at the drains, $D_1$ and $D_2$ \[see Fig. \[setup\](a)\]. Note that all channels are chiral, i.e. particles may move only in the direction of the arrow. The evolution of an injected wave packet through the setup is described by considering incoming scattering states from the various sources that are labeled by their quantum number $k$. Schematically, the state of a particle injected from $S_1$, after passing through beam-splitter $A$ at position $x=0$, is described by ${\left|{i}\right>}=r^{\phantom{\dagger}}_A{\left|{1}\right>}+t^{\phantom{\dagger}}_A{\left|{2}\right>}$, where $r^{\phantom{\dagger}}_A$ and $t^{\phantom{\dagger}}_A$ are the reflection and transmission amplitudes [@Note1] corresponding to beam-splitter $A$, and ${\left|{1}\right>},{\left|{2}\right>}$ are the scattering states corresponding to the upper and lower e-MZI arms. Similarly, the beam splitter $B$ is characterized by reflection and transmission amplitudes $r_B$ and $t_B$, respectively. Between the beam splitters $A$ and $B$, orbital phases are accumulated along arm 1 and arm 2, i.e. $e^{ik
l_1}$ and $e^{ik l_2}$, respectively. Additionally, for charged particles in the presence of a magnetic field, the relative phase of the two respective trajectories includes an Aharonov-Bohm phase $\Phi_{\rm AB}\equiv 2 \pi \frac{\Phi}{\Phi_0}$, where $\Phi_0$ is a quantum of flux. With a proper gauge choice, we reabsorb these phases in an extra phase shift of the transmission coefficient of $t_B \to t_B e^{i\phi}$, with the interference phase $\phi=\phi^{\phantom{\dagger}}_B\equiv
k(l_2-l_1) +\Phi_{\rm AB}$.
We incorporate a semi-transparent absrober on the arm-1 of the e-MZI using an additional beam-splitter $C$ at position $0<\tilde{l} < l_1$ \[see Fig.\[setup\](b)\]. The propagation of an impinging particle is thus modified: the particle may exit the MZI through arm $3$ and reach drain $D_3$. The effect of this extra beam splitter evolves the scattering state component in arm-1, ${\left|{1}\right>} \to r^{\phantom{\dagger}}_C{\left|{1}\right>} + t^{\phantom{\dagger}}_C{\left|{3}\right>}$. This process is commonly referred to as *partial-collapse* and has been studied in the context of qubit-uncollapse [@Korotkov:2007; @Katz08] and null weak values [@Zilberberg:2013; @zilberbergCrypta; @zilberberg2014standard].
This schematic evolution through the e-MZI can be conveniently recast in a scattering matrix formulation, i.e., we can write the state of a particle in the interferometer in second quantization, with an annihilation operator $$\begin{gathered}
\label{solution}
\psi_{km}(x)=e^{ikx} \left\{
\begin{array}{ll}
a^{\phantom{\dagger}}_{km}, & \hbox{$x<0;$} \\
b^{\phantom{\dagger}}_{km}, & \hbox{$0<x<\tilde{l};$} \\
c^{\phantom{\dagger}}_{km}, & \hbox{$\tilde{l}<x<l_2;$} \\
d^{\phantom{\dagger}}_{km}, & \hbox{$l_m<x.$} \\
\end{array}
\right.\end{gathered}$$ Here $m=1,2,3$ labels the different device arms and we assume arbitrarily that $l_2<l_1$. The operators $a_{km}$, $b_{km}$, $c_{km}$, $d_{km}$ are the annihilation operators of momentum eigenstates in the different sectors of the interferometer. They can be arranged in vectors $\mathbf{a}_k$, $\mathbf{b}_k$, $\mathbf{c}_k$, $\mathbf{d}_k$, labeled by the arm-index $m$, and are related by scattering matrices describing the effects of beam splitters via $$\begin{aligned}
\mathbf{b}^{\phantom{\dagger}}_k=\mathcal{S}^{\phantom{\dagger}}_A
\mathbf{a}^{\phantom{\dagger}}_k,\, \mathbf{c}^{\phantom{\dagger}}_k=\mathcal{S}^{\phantom{\dagger}}_C
\mathbf{b}^{\phantom{\dagger}}_k,\, \mathbf{d}^{\phantom{\dagger}}_k=\mathcal{S}^{\phantom{\dagger}}_B
\mathbf{c}^{\phantom{\dagger}}_k, \label{scatterer}\end{aligned}$$ with $$\begin{aligned}
\mathcal{S}^{\phantom{\dagger}}_i &= \left(\begin{array}{ccc}
r^{\phantom{\dagger}}_i & t^{\phantom{\dagger}}_i & 0 \\
-t_i^* & r^{\phantom{\dagger}}_i & 0 \\
0 & 0 & 1
\end{array}\right) \,\, ; \,\, i=A,B, \\
\mathcal{S}^{\phantom{\dagger}}_C &= \left(\begin{array}{ccc}
r^{\phantom{\dagger}}_C & 0 & -t^{\phantom{\dagger}}_C \\
0 & 1 & 0 \\
t^{\phantom{\dagger}}_C & 0 & r_C^*
\end{array} \right). \nonumber\end{aligned}$$
Single-particle limit
=====================
As a first step we analyze the the effect of the extra beam splitter $C$ using a schematic single-particle formulation. We assume that the incoming state is labeled by the quantum number $k$, which, for clarity we omit in the notation below. In the absence of the leakage port, the probability to measure the particle in drain $D_1$ is $P_0(i\rightarrow D_1)=|{\left<{D_1}\right|}{\left|{i}\right>}|^2$, where ${\left|{i}\right>}=r^{\phantom{\dagger}}_A{\left|{1}\right>}+t^{\phantom{\dagger}}_A{\left|{2}\right>} $ includes the effect of beam splitter $A$, and we have defined ${\left|{D_1}\right>}=r^{\phantom{\dagger}}_B{\left|{1}\right>}+t^{\phantom{\dagger}}_B
e^{i\phi}{\left|{2}\right>}$ to include the effect of beam splitter $B$ and the subsequent detection in $D_1$. We have used the subscript ${\cdot}_0$ to denote the probability in the absence of a leakage port. We obtain for the setup of Fig. \[setup\](a), $
P_0(i\rightarrow
D_1)=|r^{\phantom{\dagger}}_A|^2|r^{\phantom{\dagger}}_B|^2+|t^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_B|^2
+2|r^{\phantom{\dagger}}_A r^{\phantom{\dagger}}_B t^{\phantom{\dagger}}_A t^{\phantom{\dagger}}_B|\cos(\phi+\phi^{\phantom{\dagger}}_T)$, where $\phi^{\phantom{\dagger}}_T=\arg(r^{\phantom{\dagger}}_A
r^{\star}_B t^{\phantom{\dagger}}_A t^{\star}_B)$. We think of the state of the propagating electron as a superposition of quibit states, ${\left|{1}\right>}$, ${\left|{2}\right>}$.
Introducing the beam-splitter $C$ on arm 1, allows the state ${\left|{1}\right>}$ to “leak out” (partial-collapse) to through branch 3 with probability $|t^{\phantom{\dagger}}_C|^2$ \[cf. Fig. \[setup\](b)\]. The probability to reach drain $D_3$ is therefore, $$\begin{aligned}
P(i\rightarrow D_3)=|r^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_C|^2\,.
\label{probD3}\end{aligned}$$ Upon detection of the injected electron in $D_3$, we declare the interference experiment void. In such a “partial collapse” the state ${\left|{1}\right>}$ is projected out of the space spanned by ${\left|{1}\right>}$ and ${\left|{2}\right>}$. If such a projection-out does not take place (i.e. the electron is *not* detected in $D_3$), the original qubit state is rotated by the measurement’s back-action into ${\left|{i_C}\right>}=(1/\tilde{\mathcal{N}})\left(r^{\phantom{\dagger}}_A
r^{\phantom{\dagger}}_C{\left|{1}\right>}+t^{\phantom{\dagger}}_A{\left|{2}\right>}\right)$ with normalization $\tilde{\mathcal{N}}=\sqrt{1-P(i\rightarrow D_3)}$. Consequently, the probability for the particle to subsequently arrive in drain $D_1$ is $P(i_C \rightarrow D_1)P(\overline{i\rightarrow
D_3})$, where by overline we denote the complementary event, i.e. $P(\overline{i\rightarrow D_3})=1-P(i\rightarrow D_3)$. Note that $P(i_C \rightarrow D_1)$ can be written using the conditional probability $P(i \rightarrow D_1 \,|\, \overline{i\rightarrow D_3})$. As a result we obtain that the particle would reach drain $D_1$ with the joint probability $$\begin{aligned}
\label{collapseProb}
&P(i\rightarrow D_1) = P(i\rightarrow D_1 , \overline{i\rightarrow D_3})
=|r^{\phantom{\dagger}}_A|^2|r^{\phantom{\dagger}}_B|^2|r^{\phantom{\dagger}}_C|^2
\\ &+|t^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_B|^2
+2|r^{\phantom{\dagger}}_C||r^{\phantom{\dagger}}_A r^{\phantom{\dagger}}_B t^{\phantom{\dagger}}_A t^{\phantom{\dagger}}_B|\cos(\phi+\phi^{\phantom{\dagger}}_T+\phi^{\phantom{\dagger}}_C)\,,\nonumber \end{aligned}$$ where $\phi^{\phantom{\dagger}}_C=\arg(r^{\phantom{\dagger}}_C)$. Note that due to causality $P(i\rightarrow D_1)=P(i\rightarrow D_1 , \overline{i\rightarrow D_3})$ and similarly $$\begin{aligned}
P(i\rightarrow D_1 , i\rightarrow D_3)=0\,.
\label{holyGrail}\end{aligned}$$ The fact that $P(i \rightarrow D_1 )\neq P_0(i \rightarrow D_1 )$ can be used to detect the presence of the leakage port. Specifically, if the MZI is tuned to have $P_0(i \rightarrow D_1
)=0$, the detection of a particle at $D_1$ in any single realization of the experiment indicates the presence of the leakage port without the particle having leaked out. If the particle is not detected at $D_1$, no conclusion on the presence of a leakage channel can be drawn. This is a manifestation of the EV-bomb detection scheme.
It is instructive to recover the results of this single particle analysis in the scattering matrix formalism, which provides the basis to analyze the statistical many-body effects in the following section. In the scattering matrix formalism we consider the injection of a [*single particle*]{} (in the scattering state $k$) into the system, i.e. ${\left|{i_k}\right>}=a^\dagger_{k,1}{\left|{0}\right>}$. The detection of the particle in $D_{1(3)}$ is described by the projection operator $\Pi_{D_1(D_3)} \equiv
d^{\dag}_{k,1(3)}d^{\phantom{\dagger}}_{k,1(3)}$. From Eq. (\[scatterer\]), the probabilities for the injected particle to reach $D_1$ or $D_3$ are $$\begin{aligned}
\label{currentD1}
P(i_k \rightarrow D_1) ={\left<{0}\right|}a^{\phantom{\dagger}}_{k,1}
d^{\dag}_{k,1}d^{\phantom{\dagger}}_{k,1}
a^\dagger_{k,1}{\left|{0}\right>}&
=\mathcal{A}_{11},\\
P(i_k \rightarrow D_3)= {\left<{0}\right|}a^{\phantom{\dagger}}_{k,1}
d^{\dag}_{k,3}d^{\phantom{\dagger}}_{k,3}
a^\dagger_{k,1}{\left|{0}\right>} &
=\mathcal{B}_{11}\,,
\label{currentD3}\end{aligned}$$ where we have introduced the quantities $\mathcal{A}_{ij} \equiv
(\mathcal{S}_A^{\dag}\mathcal{S}_B^{\dag}\mathcal{S}_C^{\dag})_{i1}(\mathcal{S}_C\mathcal{S}_B\mathcal{S}_A)_{1j}$, $ \mathcal{B}_{ij}\equiv (\mathcal{S}_A^{\dag}\mathcal{S}_B^{\dag}\mathcal{S}_C^{\dag})_{i3}(\mathcal{S}_C\mathcal{S}_B\mathcal{S}_A)_{3j}$. Indeed, an explicit evaluation of $\mathcal{A}_{11}$ and $\mathcal{B}_{11}$ yields, for Eqs. and exactly the same expressions as Eqs. and , respectively.
Additionally, the joint probability of detecting a particle at $D_1$ and $D_3$ is given by $$\begin{gathered}
P(i\rightarrow D_1 , i\rightarrow D_3) {\left<{0}\right|}a^{\phantom{\dagger}}_{k,1}
d^{\dag}_{k,3}d^{\phantom{\dagger}}_{k,3}d^{\dag}_{k,1}d^{\phantom{\dagger}}_{k,1}
a^\dagger_{k,1}{\left|{0}\right>}\\
= \sum_{\beta=1}^3
\mathcal{A}_{1\beta}\mathcal{B}_{\beta 1}\equiv 0\,,
\label{crosscurrent}\end{gathered}$$ where, when the incoming state is of a single particle, we recover the result in Eq. .
The results of this section describe experiments where a single particle is injected into the interferometer. While this is possible in quantum optics, it does not represent the typical experimental conditions of electronic devices. Single-particle sources have been only recently reported in some specifically designed experimental architectures [@Dubois:2013]. Since many-electron physics is an essential part of quantum reality, we next analyze this limit.
Many-body conditional correlations
==================================
In a typical experiment with e-MZI, particles are injected into the source from a voltage biased reservoir, and are collected in the drain over a macroscopically long time. This being the case, only statistical quantities averaged over a many-particle ensemble are accessible, and the signals at the detector correspond to statistical averages of the source-drain transition probabilities computed in the previous section. Specifically, for an e-MZI with a voltage bias $eV$ at $S_1$, the measured current at $D_1$ is given by the rate of electrons reaching this drain out of the total rate, $eV/\hbar$, of electrons impinging from the source. The currents through the device are therefore statistical probabilities for an impinging electron to reach the various drains, and are precisely given in terms of the probabilities calculated in the single-particle picture above: the current at drain $j$ will be given by $I_j=
(e^2/h) P(i\rightarrow D_j) V$. When the signal in $D_1$ is collected over a large number of particles, any outcome of the IFM-experiment would have a macroscopic leakage of particles in $D_3$ even if the e-MZI is tuned to have a vanishing current in the absence of the port $D_3$. Hence, in the original formulation of the problem with the bomb, the bomb would necessarily explode. In short, under the above conditions the detection of the current at $D_3$ does not constitute an uncontested manifestation of IFM.
Can, and under what conditions, an electronic MZI setup reproduce the original EV bomb measurement scheme? In order to clarify this we focus on the difference between the single-particle results and the many-particle statistical averages relevant for experiments, which appears when dealing with joint probabilities.
This is clearly demonstrated considering, e.g., the statistical joint probability of detecting particles at drain $D_1$ and $D_3$. In order to relate such a joint probability with a quantity directly accessible in experiments, we next study the current-current correlations in a [*many-body*]{} (albeit non-interacting) system. We assume that a voltage bias $V$ is applied to the source $S_1$, which is held at temperature $T$. For a system with linear dispersion relation, the current operator is $\hat{I}_i(x,t)=ev:\psi_i^{\dag}(x,t)\psi_i(x,t):$, where $\psi_i(x,t)$ is the annihilation operator in the $i$-th arm, and the normal order operator, $:\,:$, indicates the subtraction of the mean equilibrium contribution.
We consider the cross-current correlation defined by $$F_{1,3} \equiv \frac{h^2}{e^4 V^4 \tau} \int_{-\tau/2}^{\tau/2}dt
{\left\langle \hat{I}_3(x_0,t) \hat{I}_1(x,0)\right\rangle},$$ where $\tau$ is an infrared cut-off, $\tau\gg\frac{L}{v}$, and $x_0,x>l_1$. Importantly, since the average current is related to the electron transfer probability by the factor $e^2V/h$, the prefactor in the definition of $F_{1,3}$ allows us to directly compare this correlator with the averaged joint probability of detecting electrons at drain $D_1$ and $D_3$ \[cf. Eq. \].
Using Wick’s theorem, the fact that all ohmic contacts are grounded apart from $S_1$ which is at $eV$, the identity $f_\alpha(1-f_\beta)=\frac{1}{2}\left[f_\alpha(1-f_\alpha)+f_\beta(1-f_\beta)+(f_\alpha-f_\beta)+(f_\alpha-f_\beta)^2\right]$ where $f$ is the Fermi-Dirac distribution, and the limit of $\tau \gg L/v$, we obtain
$$\begin{aligned}
&F_{1,3} \equiv F_\infty (\alpha,\Delta \tilde{L}) - F_N (\alpha,\Delta \tilde{L}) = \frac{1}{ \alpha}|r_A|^2|t_C|^2\Bigg[\alpha \left(|t_A|^2 |t_B|^2+|r_A|^2 |r_B|^2 |r_C|^2\right)+2K(\alpha,\Delta \tilde{L})|t_A t_B r_A r_B r_C|\cos[\Phi(\alpha,\Delta \tilde{L},\Phi_{\rm AB})] \Bigg]\nonumber\\
&-\frac{1}{\alpha N} |r_A|^2|t_C|^2\Bigg[L(\alpha)\left(|t_A|^2 |t_B|^2+|r_A|^2 |r_B|^2 |r_C|^2\right)+2M(\alpha,\Delta \tilde{L})|t_A t_B r_A r_B r_C|\cos[\Phi(\alpha,\Delta \tilde{L},\Phi_{\rm AB})] \Bigg]\, .
\label{condEq}\end{aligned}$$
where $F_\infty$ and $F_N$ are functions of the dimensionless parameters $\alpha=eV \beta/(2\pi)$, $\Delta \tilde{L}=\pi (l_2-l_1)/(\hbar \beta v)$, and $N=eV\tau/(2\pi\hbar)$. Here $\beta=1/(k_B T)$ is the inverse temperature. We have also introduced the functions $K(\alpha,\Delta \tilde{L})=\sin[\alpha\Delta \tilde{L}]/\sinh[\Delta \tilde{L}]$, $L(\alpha)=(\pi \alpha \coth[\pi
\alpha]-1)/\pi $, $M(\alpha,\Delta \tilde{L})=(\pi \sin[\alpha \Delta
L]\coth[\pi \alpha]-\Delta \tilde{L} \cos[\alpha\Delta \tilde{L}])/ (\pi \sinh[\Delta
L])$, and $\Phi(\alpha,\Delta \tilde{L},\Phi_{\rm AB})=\Phi_{\rm AB}+\phi^{\phantom{\dagger}}_T+\phi^{\phantom{\dagger}}_C+\alpha \Delta \tilde{L}$.
![ The many-body cross-current correlator as a function of $\alpha$ \[cf. \]. Here, we have taken $|t_C|^2=0.3$,$|r_A|^2=|r_B|^2=0.5$ and $\Phi_{\rm AB}+\phi^{\dagger}_T+\phi^{\dagger}_C=\pi$. For a temperature of $T=10$mK the parameter $\alpha$ corresponds to realistic bias values of up to $\sim 54\mu$V. The different plots correspond to different values of $N=1\ldots 10$. As a function of $\alpha$ for a fixed $T$, $\tau$ should be changed in order to keep $N$ constant, i.e. $\tau= N\frac{2\pi\hbar}{eV}=N\frac{\hbar\beta}{\alpha}\sim 7.63823\times 10^{-10}(N/\alpha)$\[second\]. We mark by circles the point $\alpha\equiv N$ as the threshold for which our assumption $\tau \gg L/v$ breaks for existing electronic interferometers [@Neder:2007]. (a) The case of $\Delta \tilde{L} \rightarrow 0$. (b) The case of $\Delta \tilde{L} =0.03$ where dephasing affects both the classical and quantum correlators (due to varying interference lengths per wavenumber). Nonetheless, the single-particle limit remains unaffected as expected from Eq. . []{data-label="fig2"}](fig2.pdf){width="\columnwidth"}
Before discussing the implication of this result, it is instructive to contrast the many-body conditional correlator to purely classical correlations of an ensemble of statistically independent impinging electrons. In the latter case, we obtain the statistical average of a joint signal at port $D_1$ and $D_3$: $$\begin{gathered}
\label{statisticalMeans}
\tilde{P}(i\rightarrow D_1 , i\rightarrow D_3)=P(i\rightarrow D_1)P(i\rightarrow D_3)\\
=|r^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_C|^2\Big[|r^{\phantom{\dagger}}_A|^2|r^{\phantom{\dagger}}_B|^2|r^{\phantom{\dagger}}_C|^2+|t^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_B|^2\\
+2|r^{\phantom{\dagger}}_C||r^{\phantom{\dagger}}_A r^{\phantom{\dagger}}_B t^{\phantom{\dagger}}_A t^{\phantom{\dagger}}_B|\cos(\phi+\phi^{\phantom{\dagger}}_T+\phi^{\phantom{\dagger}}_C)\Big]\,. \end{gathered}$$ For a better comparisson with the full many-body results that include the effect of averaging over a statistical ensamble due to termal fluctuations, as well as out-of equilibrium voltage bias, one can further average over a density matrix, $\rho$, that describes and ensemble of initial states. For example, assuming that a voltage bias $V$ is applied to the source $S_1$, which is held at temperature $T$, the state of the impinging electrons is described by $\rho=(1/L)
\sum_k [f(\hbar vk-eV)-f(\hbar vk)]a_{k1}^{\dag}a_{k1}$, where $f(x)$ is the Fermi-Dirac distribution, and the system length, $L$, is taken to be la largest length scale in the problem. When averaged over the initial density matrix, the “classical” correlations in Eq. (\[statisticalMeans\]) yield $$\begin{gathered}
\label{MBstatMean}
\tilde{P}(i\rightarrow D_1 , i\rightarrow D_3)=|r^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_C|^2\left[|r^{\phantom{\dagger}}_A|^2|r^{\phantom{\dagger}}_B|^2|r^{\phantom{\dagger}}_C|^2+|t^{\phantom{\dagger}}_A|^2|t^{\phantom{\dagger}}_B|^2\right]\\
+2\frac{K(\alpha) }{\alpha}|r^{\phantom{\dagger}}_C||r^{\phantom{\dagger}}_A r^{\phantom{\dagger}}_B t^{\phantom{\dagger}}_A t^{\phantom{\dagger}}_B|\cos\left[\Phi(\alpha,\Delta \tilde{L},\Phi_{\rm AB})\right]\,.\end{gathered}$$
Comparing the statistical probability analysis in Eq. with the many-body joint correlation in Eq. (\[condEq\]), we obtain that $\tilde{P} (i\rightarrow D_1 ,
i\rightarrow D_3)=F_\infty$, which is the dominant contribution of $F_{1,3}$ in the zero-frequency DC-limit. Indeed, this represents the well-known fact that $\lim_{\tau \to \infty} \int_{-\tau/2}^{\tau_2} dt \langle \hat{I}_3(x_0,t) \hat{I}_1(x,0) \rangle
=\langle \hat{I}_3(x_0,t) \rangle \langle \hat{I}_1(x,0) \rangle \tau$. Similarly, a standard analysis of current-current correlations [@Blanter:2000] singles out the non-trivial correlations in the cross-current noise $S_{1,3} \equiv \lim_{\tau \to \infty} \int_{-\tau/2}^{\tau_2} dt ( \langle \hat{I}_3(x_0,t) \hat{I}_1(x,0) \rangle -
\langle \hat{I}_3(x_0,t) \rangle \langle \hat{I}_1(x,0) \rangle )$. These non-trivial contributions are encoded in the term $F_N =
S_{1,3}/ ( \hat{I}_0^2 \tau )$ of the many-body cross-current correlation in Eq. (\[condEq\]). Technically it corresponds to a particle-hole loop contribution.
While at low frequencies, $F_\infty$ is the dominant contribution to cross-current correlations, Eq. (\[condEq\]) clearly shows how, for measurements averaged over a finite time, the effects of $F_\infty$ and $F_N$ are competing. In fact, they become of the same order for short measurements times, such that the average currents are comparable with their fluctuations, i.e., $\langle \hat{I}_3(x_0,t) \rangle \langle \hat{I}_1(x,0) \rangle \tau\sim S_{1,3}$. In particular, one expects that in the limit where the average number of particles in the interferometer is $\sim 1$ during the measurement time $\tau$, these two terms cancel each other, and we can recover the single-particle result of Eq. (\[holyGrail\]). By estimating the average number of electrons impinging on the e-MZI during the measurement time by $N=\frac{eV\tau}{2\pi\hbar}$, we are in the position of interpreting the cross-current correlator in terms of a crossover between single-particle quantum-mechanical correlations and classical statistical correlations.
Fig. \[fig2\](a) depicts the cross-current correlations as function of the voltage bias, $\alpha$, measured in units of temperature, for different values of $N$. For any value of $N$, at $\alpha \lesssim 4$, thermal fluctuations dominate over the quantum ones, and the correlations will ultimately reduce to those of classical waves. For large $\alpha$, upon decreasing $N$, $F_{1,3}$ decreases, and for $N \sim 1$ it is essentially vanishing, i.e., we obtain $F_{1,3} \ll F_\infty$ which signals that quantum correlations are important. Note that Eq. , depicted in Fig. \[fig2\], is valid for $\tau \gg L/v$. Recall that as a function of $\alpha$ for a fixed temperature $T$, $\tau$ changes in order to keep a constant $N$, i.e. $\tau=
N\frac{2\pi\hbar}{eV}=N\frac{\hbar\beta}{\alpha}\sim 7.63823\times
10^{-10}(N/\alpha)$\[second\], where we considered $T=10$mK. Taking experimental values of existing electronic interferometers, $L\sim 10\mu$m and $v\sim 2-6\times 10^{-4}$m/s [@Neder:2007], we mark the point $\alpha\equiv N$ as a threshold beyond which our prediction no longer holds. As such, in order to reach the limit of single-particle demonstration of IFM, one should construct smaller interferometers or generate higher edge mobility. Alternatively, one could consider single-particle injection on top of a Fermi sea [@Dubois:2013], but this is beyond the scope of our analysis.
In Fig. \[fig2\](b), we see the effect of a finite $\Delta
\tilde{L}$. As each wavenumber experiences a slightly different interference path, both the classical and quantum many-body correlations are affected by averaging over many wavenumbers. As a result, when many particles are considered \[Eq. \] the result moves further away from the single-particle limit of Eq. reflecting this effective dephasing. Nonetheless, in the limit of short pulses, $N=1$, the correlator yields an outcome that agrees with the single-particle picture.
Conclusions
===========
The main focus of this study is the assessment of feasible detection of IFM in a genuine many-body electronic system. To this goal, we have analyzed an electronic MZI with a leakage port located on one of the interferometer arms, which servs as an experimentally viable implementation of the EV-bomb gedanken experiment. We considered the typical experimental settings when an ensemble of particles is injected in the interferometer, i.e., the current in the interferometer yields a statistically averaged signal. We analyzed the cross-current correlation at the dark and leakage ports, which is vanishing in the single-particle original proposal of the experiment, but remains generally finite in the many-particle statistical implementation. This has allowed us to identify the parameters’ regime (voltage bias, temperature) for which the many-body correlations approach the single-particle result. We find the regime where the wave-particle duality emerges is lies just at the frontiers of actual experiments with electronic MZIs, where the main limitations are due to the size of the interferometer and the mobility of the electrons at the edges of a Hall bar.
In summary, our results show that the detection of IFM in a many-body electronic system seems to involve two competing facets that need to be dealt with: IFM a-la Elitsur-Vaidman requires to deal with particles (that, in principle, can be pin-pointed to a specific spatial coordinate); at the same time, the setup employed is an interferometer, which invokes the wave character of the quantum object. One thus needs to fine-tune the system to zoom on a regime where particle-wave duality is manifest. Our analysis might trigger experiments with single-electron biased MZIs, where this physics may be elucidated.
This work has been supported by the Swiss National Science Foundation, the German-Israel Foundation (GIF), Deutsche Forschungsgemeinschaft (DFG) grants and RO 2247/8-1 and RO 4710/1-1, and the Israel Science Foundation (ISF).
[30]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, , , , , ****, ().
, **** ().
, ****, (), ISSN .
, , , , , , ****, ().
, , , , , , , ****, ().
, ****, ().
, , , , , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , , , , ****, ().
, , , , , , , , ****, ().
, ****, (), .
, ****, ().
, ****, ().
, ****, ().
, , , , , , ****, ().
, , , , , , , , , ****, ().
, ****, ().
, ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , ****, ().
, , , ****, ().
, , , in ** (, ), pp. .
, , , , , , ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the present paper we study the thermodynamic properties of superionic conductor $SrCl_2$ at high temperatures by means of molecular dynamics method. Firstly, we calculate the melting line. Then we compute the equations of state and the response functions (heat capacity, thermal expansion coefficient, etc) at the temperatures up to the melting. We show that the response functions show maxima or minima at the temperatures well above the temperature of transition into the conductive state, and therefore are not related to this transition.'
author:
- 'Yu. D. Fomin'
title: 'Melting line and thermodynamic properties of a supeionic compound SrCl$_2$ by molecular dynamics simulation'
---
Superionic conductors are crystalline ionic salts which demonstrate high electrical conductivity [@book1]. Their high electrical conductivity is related to very different mobilities of two types of ions in the salt: while one type of the ions remains mostly fixed at the lattice sites, another type demonstrates high diffusion. Depending on the substance the mobile type of particles can be either cation or anion. Importantly, superionic compounds are widely used in modern energetics as solid electrolytes in the batteries. Because of this deep understanding of their properties is of great importance for both fundamental research and technological applications [@book2; @book3].
A wide class of superionic compounds is the one of salts with the structure of fluorite [@chadwick; @disorder; @fl1; @fl2]. These substances have a chemical formula $CA_2$ where $C$ means the cation and $A$ - the anion. In the fluorite structure the cations occupy the cites of an Face Centered Cubic (FCC) lattice, while the anions are located in the tetrahedral holes. Numerous salts with composition $CA_2$ demonstrate the fluorite structure, for instance, fluorite itself $CaF_2$, $BaCl_2$, $PbO_2$, $SrCl_2$, etc. The mobile specie of the superionic compounds with fluorite structure is the anion.
At low temperatures the supeionic compounds do not demonstrate high conductivity. However, at certain temperature they experience a phase transition into the conducting state. Typically phase transitions are related to some thermodynamic peculiarities, such as the maximum of the heat capacity. The peak of heat capacity in the vicinity of the transition into superionic state was reported in a number of experimental [@cp1; @cp2] and theoretical [@cp3; @cp4; @cp5; @cp6; @gillan1] works. Several theoretical models were proposed to describe the behavior of this peak. However, other thermodynamic response functions were studied much less and therefore they still require a careful investigation.
In the present paper we study the thermodynamic properties of a superionic conductor $SrCl_2$ by means of molecular dynamics method. We calculate the melting line and thermodynamic response functions (heat capacity, thermal expansion, etc). We show that close to the melting line this substance demonstrates maxima or minima of several thermodynamic properties. The temperatures of these peculiarities are rather far from the temperature of transition into conductive state. Because of this we assume that these maxima and minima are not related to this transition, but are induced by some other mechanisms, which means that the behavior of fast ions conductors is rather complex even far from the point of transition into the conductive state.
We simulate a system of $4116$ particles (i.e. 1372 atoms of Sr and 2744 atoms of Cl) in a cubic box with periodic boundary condition by means of molecular dynamics methods. We use the Born-Mayer-Huggins potential to describe the interaction between the species. It has the form
$$\label{pot}
U_{\alpha \beta} (r) = \frac{Z_{\alpha}Z_{\beta} e^2}{r} +A_{\alpha \beta} e^{-r/\rho_{\alpha \beta}} -\frac{C_{\alpha \beta}}{r^6},$$
where $\alpha$ and $\beta$ mark different species (Sr or Cl), $Z$ is the charge of the nuclei and $e$ is the elementary charge. Other parameters of the potential are taken from Ref. [@gillan]. They are given in Table I. Importantly, Refs. [@gillan1; @gillan] themselves report a computational study of the behavior of $SrCl_2$. In particular, the authors also discuss the anomaly of the isochoric heat capacity. However, they do not discuss the behavior of isobaric heat capacity and thermal expansion coefficient. They also do not calculate the melting line.
All simulations of the present work are performed using lammps simulation package [@lammps].
$A_{Sr-Sr}$ (eV) 0.0
---------------------------- --------
$A_{Sr-Cl}$ (eV) 774.14
$A_{Cl-Cl}$ (eV) 1227.2
$\rho_{Sr-Sr}$ -
$\rho_{Sr-Cl}$ ($\AA$) 0.3894
$\rho_{Cl-Cl}$ ($\AA$) 0.3214
$C_{Sr-Sr}$ (($ eV\AA^6$)) 0.0
$C_{Sr-Cl}$ (($ eV\AA^6$)) 0.0
$C_{Cl-Cl}$ (($ eV\AA^6$)) 1.69
: Parameters of potential of Eq. \[pot\] used in the present study. Note, that since $A_{Sr-Sr}$ is zero, the parameter $\rho_{Sr-Sr}$ is not required.
We start our discussion from simulation of $SrCl_2$ at ambient pressure and different temperatures. The system is simulated at fixed pressure (P=1 bar) and different temperatures for 10 ns with the time step 1 fs. In these simulations we evaluate the average density of the system and monitor the structure. The dependence of the density on the temperature is shown in Fig. \[p1\]. One can see that the density experiences a jump at $T=1560$ where it is $\rho=2.64$ $g/cm^3$. After the jump at $T=1600$ K the density is $\rho=2.2$ $g/cm^3$. From investigation of the structural properties we observe that at low temperatures the system is solid, while at the high ones it is liquid. Although this method does not give accurate estimation of the melting point, it allows us to estimate the lowerst density of solid $SrCl_2$ to be studied.
![\[p1\] Equation of state of $SrCl_2$ at ambient pressure and different temperatures. The jump of the density signalizes that melting of the system takes place.](p1.pdf){width="6cm" height="6cm"}
Having estimated the melting density at ambient pressure, we calculate the melting line on a more precise ground. For doing this we employ a so called Z-method [@zmethod]. The density is varied from $\rho_{min}=2.6$ $g/cm^3$ up to $\rho_{max}=3.05$ $g/cm^3$. The system is simulated for 10 ns with the time step 1 fs in microcanonical ensemble (constant number of particles N, energy E and volume V). According to the Z-method, the dependence of pressure on the temperature has a Z-like shape at the melting point. Fig. \[melting-line\] (a) shows an example of Z-method calculations for $\rho=2.7$ $g/cm^3$. One can see that the melting temperature is $T_m=1817$ K and the melting pressure is $13483$ bar. The melting line is given in Fig. \[melting-line\] (b).
![\[melting-line\] (a) Example of calculations of the melting point by Z-method. $\rho=2.7$ $g/cm^3$. (b) Melting line of $SrCl_2$ obtained by Z-method calculations.](z-method.pdf "fig:"){width="6cm" height="6cm"} ![\[melting-line\] (a) Example of calculations of the melting point by Z-method. $\rho=2.7$ $g/cm^3$. (b) Melting line of $SrCl_2$ obtained by Z-method calculations.](melting.pdf "fig:"){width="6cm" height="6cm"}
We calculate equation of state of $SrCl_2$ along several isochores from $T_{min}=1000$ K up to $T_{max}=2600$ K. Examples of these equations of state are given in Fig. \[isov\] (a). At some points the pressure has negative values. These points are discarded from the plot. A jump of equation of state (shown only for $\rho=2.9$ $g/cm^3$) means crossing of the melting line. The points corresponding to liquid are also discarded since the focus of the present study is related to the crystalline state only. These equations of state do not demonstrate any unusual features: pressure monotonously increases with temperature.
![\[isov\] Equation of state of $SrCl_2$ along several isochores. Jump at the isochore $\rho=2.9$ $g/cm^3$ signalizes that melting of the system takes place. The points at the temperature above the jump correspond to the liquid state while below to solid. The data for liquid for other isochores are removed from the plot to make it clearer. ](eos-isov.pdf){width="6cm" height="6cm"}
In our study we simulate almost 200 data points in ($\rho$,T) plane along several isochores. It allowed us to find the properties of the system along other thermodynamic pathes like isotherms and isobars by interpolation of our data. Simple linear interpolation was enough to obtain the data along isotherms and isobars. Numerical differentiation of the pressure, internal energy or enthalpy was used to calculate the thermodynamic response functions. Fig. \[isopt\] shows the equation of state of solid $SrCl_2$ along several isotherms and isobars.
![\[isopt\] (a) Equation of state of $SrCl_2$ along several isotherms. (b) Equation of state along several isobars.](eos-isot.pdf "fig:"){width="6cm" height="6cm"} ![\[isopt\] (a) Equation of state of $SrCl_2$ along several isotherms. (b) Equation of state along several isobars.](eos-isop.pdf "fig:"){width="6cm" height="6cm"}
Although the equation of state does not demonstrate any peculiarities neither along isobars, not isotherms and isochores, the thermodynamic response functions do demonstrate some maxima and minima. We calculate the thermal expansion coefficient $\alpha_P= \frac{1}{V} \left( \frac{\partial V}{\partial T} \right)_P$, the isochoric heat capacity $c_V=\left( \frac{\partial U}{\partial T} \right)_V$, where $U$ is the internal energy of the system, and the isobaric heat capacity $c_P=\left( \frac{\partial H}{\partial T} \right)_V$, where $H=U+PV$ is the enthalphy of the system. The results of these calculations are given in Fig. \[response\] (a)-(c). One can see that all three functions demonstrate maxima and minima.
![\[response\] (a) Thermal expansion coefficient of $SrCl_2$ along several isobars. (b) Isobaric heat capacity along several isobars. (c) Isochoric heat capacity along several isotherms.](al-p-isop.pdf "fig:"){width="6cm" height="6cm"} ![\[response\] (a) Thermal expansion coefficient of $SrCl_2$ along several isobars. (b) Isobaric heat capacity along several isobars. (c) Isochoric heat capacity along several isotherms.](cp-isop.pdf "fig:"){width="6cm" height="6cm"} ![\[response\] (a) Thermal expansion coefficient of $SrCl_2$ along several isobars. (b) Isobaric heat capacity along several isobars. (c) Isochoric heat capacity along several isotherms.](cv-isov.pdf "fig:"){width="6cm" height="6cm"}
The location of maxima of response functions is widely discussed in the vicinity of liquid-gas or liquid-liquid critical point [@w1; @w2; @w3; @w4; @w5; @w6; @w7; @w8; @w9]. In these cases the locations of maxima in density-temperature (or, equivalently, pressure-temperature) plane should tend to the critical point as the temperature tends to the critical one. However, the curves of maxima of different quantities rapidly diverge as the temperature goes away from the critical point.
In the case of $SrCl_2$ the situation is different and one does not expect coincidence of the lines of maxima of different functions. However, the location of the maxima and minima of the response functions is of a great interest for characterisation of the thermodynamic properties of the system. Fig. \[max-pd\] shows the points of maxima and minima of $\alpha_P$, $c_P$ and $c_V$ on the phase diagram of the system. One can see that the lines of maxima of $c_V$ along isochores and minima of $\alpha_P$ along isobars are almost isotherms with $T_{c_V-max}=1300$ K and $T_{\alpha_P-min}=1180$ K. The line of minima of $c_P$ along isobars is located close to the melting line and goes almost parallel to it.
![\[max-pd\] Maxima and minima of the response functions of $SrCl_2$ placed in the phase diagram of the system.](max-pd.pdf){width="6cm" height="6cm"}
Finally we evaluate the diffusion coefficients of the species at different temperatures. Fig. \[diff\] (a) and (b) show the mean square displacement of chlorine and strontium at the density $\rho=2.8$ $g/cm^3$ and the temperatures from 1000 K to 2000 K, i.e. up to the melting point. One can see that even at $T=1000$ K the chlorine has finite diffusion coefficient, the diffusion coefficient of strontium at this temperature is apparently zero. Therefore, the system is in conducting state at this temperature. This is consistent with Ref. [@gillan], where the transition from insulating state into the conducting one is estimated to be about 1000 K. The mean square displacement of strontium remains shoulder-like up to the temperature $T=1900$ K where a slight deviation at long times appears. This temperature is close to the melting point and therefore moderate diffusion of $Sr$ due to the formation of defects can appear in the system. Fig. \[diff1\] shows the diffusion coefficient of chlorine at $\rho=2.8$ $g/cm^3$ and different temperatures. The diffusion of Sr is nearly zero at all these temperatures.
![\[diff\] Mean square displacement of (a) clorine and (b) strontium at $\rho=2.8$ $g/cm^3$ and different temperatures. In both cases the temperatures from $T=1000$ K (the lowest curve) to $2000$ K (the highest curve) with the step $\Delta T=100$ K are shown](r2t-cl.pdf "fig:"){width="6cm" height="6cm"} ![\[diff\] Mean square displacement of (a) clorine and (b) strontium at $\rho=2.8$ $g/cm^3$ and different temperatures. In both cases the temperatures from $T=1000$ K (the lowest curve) to $2000$ K (the highest curve) with the step $\Delta T=100$ K are shown](r2t-sr.pdf "fig:"){width="6cm" height="6cm"}
![\[diff1\] Diffusion coefficient of chlorine at the isochore $\rho=2.8$ $g/cm^3$.](diff-cl.pdf){width="6cm" height="6cm"}
One can see that the transition from insulating to conducting state appears at the temperatures much lower then the ones where we observe maxima and minima of the response functions. At the same time some of them are also far from the melting line. Because of this it is not evident whether one can relate the maxima and minima of the response functions to one of these transitions. The exact nature of these maxima and minima is still under question and requires further investigation.
In conclusion, we performed a molecular dynamic study of thermodynamic properties of the melting line and thermodynamic properties of superionic conductor $SrCl_2$. We calculated the melting line and the location of maxima and minima of thermal expansion coefficient, isobaric and isochoric heat capacities. We find that the peculiarities of the thermodynamic functions are located far from both phase transitions in the density-temperature (or pressure-temperature) plane. Because of this they cannot be directly attributed to these transitions. The reasons for these peculiarities to appear require further investigation.
This work was carried out using computing resources of the federal collective usage center “Complex for simulation and data processing for mega-science facilities” at NRC “Kurchatov Institute”, http://ckp.nrcki.ru, and supercomputers at Joint Supercomputer Center of the Russian Academy of Sciences (JSCC RAS). The work was supported by the Russian Foundation of Basic Research (Grant No 18-02-00981).
[99]{}
Topics in Current Physics, Physics of Superionic Conductors, Edited by M.B. Salamon, Springer-Verlag Berlin Heidelberg (1979)
Superionic Conductors, Edited by G. D. Mahan and W. L. Roth, Plenum Press, New York and London (1976)
The Physics of Superionic Conductors and Electrode Materials, Edited by J. W. Perram, Plenum Press New York and London (1980).
A.V. Chadwick, Solid State Ionics, 8, 209-220 (1983)
M. J. Gillan and D. D. Richardson, J. Phys. C: Solid State Phys., 12, L61 (1979)
St. Hull, St. T. Norberg , I. Ahmed, St. G. Eriksson, Ch. E. Mohn, J. Sol. State Chem. 184, 2925–2935 (2011).
B.M. Voronin, S.V. Volkov, J. Physics and Chemistry of Solids 62, 1349-1358 (2001).
R. Shaviv, J. Chem. Thermodynamics 21, 631-651 (1989).
W. Schroter and J. Nolting, Physique, 41 C6-20 (1980).
Yu. Ya. Gurevich, Yu. I. Kharkats, Sov. Phys. Usp. 25 257–276 (1982).
V. A. Mezrin, Phys. Stat. Sol. (a) 114, 145 (1989).
J. Oberschmidt, Phys. Rev. B 23, 5038 (1981).
S. Dasgupta, Phys. Lett. A 133, 501 (1988).
M. J. Gillan and M. Dixon, J. Phys. C: Solid St. Phys., 13, 1919-29 (1980).
M. J. Gillan and M. Dixon, J. Phys. C: Solid St. Phys., 13, 1901-17 (1980).
S. Plimpton, Fast Parallel Algorithms for Short-Range Molecular Dynamics, J. Comp. Phys. 117, 1-19 (1995). http://lammps.sandia.gov/
A. B. Belonoshko, N. V. Skorodumova, A. Rosengren, and B. Johansson, Phys. Rev. B 73, 012201 (2006).
L. Xu, P. Kumar, S. V. Buldyrev, S.-H. Chen, P. H. Poole, F. Sciortino, and H. E. Stanley, Proc. Natl. Acad. Sci. U.S.A. 102, 16558 (2005). V. V. Brazhkin and V. N. Ryzhov, J. Chem. Phys. 135, 084503 (2011). V. V. Brazhkin, Yu. D. Fomin, A. G. Lyapin, V. N. Ryzhov, and E. N. Tsiok, J. Phys. Chem. B 115, 14112 (2011). H.-O. May and P. Mausbach, Phys. Rev. E 85, 031201 (2012). V. V. Brazhkin, Yu. D. Fomin, V. N. Ryzhov, E. E. Tareyeva, and E. N. Tsiok, Phys. Rev. E 89, 042136 (2014). A. Dey, P. Roy, and T. Sarkar, Phys. A 392, 6341 (2013). A. R. Imre, U. K. Deiters, T. Kraska, and I. Tiselj, Nucl. Eng. Des. 252, 179 (2012). G. Franzese and H. Eugene Stanley, J. Phys.: Condens. Matter 19, 205126 (2007). J. L. F.Abascal and C.Vega, J. Chem. Phys. 133, 234502 (2010).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The relativistic quantum dynamics of an electrically charged particle subject to the Klein-Gordon oscillator and the Coulomb potential is investigated. By searching for relativistic bound states, a particular quantum effect can be observed: a dependence of the angular frequency of the Klein-Gordon oscillator on the quantum numbers of the system. The meaning of this behaviour of the angular frequency is that only some specific values of the angular frequency of the Klein-Gordon oscillator are permitted in order to obtain bound state solutions. As an example, we obtain both the angular frequency and the energy level associated with the ground state of the relativistic system. Further, we analyse the behaviour of an electrically charged particle subject to the Klein-Gordon oscillator, the Coulomb potential and a linear scalar potential.'
author:
- 'R. L. L. Vitória'
- 'C. Furtado'
- 'K. Bakke'
title: 'On a relativistic scalar particle subject to the Klein-Gordon oscillator, the Coulomb potential and a linear scalar potential'
---
Introduction
============
Relativistic effects on quantum systems where the motion of a particle is governed by harmonic oscillations have been the subject of several works in the literature [@bah; @bah2; @bah3; @bah4]. Examples of these studies are the binding of heavy quarks [@qui; @chai], the vibrational spectrum of diatomic molecules [@ikh] and oscillations of atoms in crystal lattices by mapping them as a position-dependent mass system [@pdm; @pdm2; @pdm3; @pdm4]. In recent decades, the relativistic generalization of the harmonic oscillator has attracted a great deal of attention whose model proposed by Moshinsky and Szczepaniak [@osc1] is the best known for Dirac particles. This model is called the Dirac oscillator and has a great interest in studies of Jaynes-Cummings model [@jay2; @osc3], quantum phase transitions [@extra2; @extra3] and the Ramsey-interferometry effect [@osc6].
Another interesting model was proposed by Bruce and Minning [@kgo] where an analogous coupling to the Dirac oscillator coupling [@osc1] is introduced into the Klein-Gordon equation in such a way that one can recover the Schrödinger equation for a harmonic oscillator in the nonrelativistic limit. This coupling is called the Klein-Gordon oscillator [@kgo; @kgo2; @kgo8; @kgo7; @kgo9]. In recent years, the Klein-Gordon oscillator has been investigated in noncommutative space [@kgo3; @kgo4], in noncommutative phase space [@kgo5] and in $\mathcal{PT}$-symmetric Hamiltonian [@kgo6]. In particular, the isotropic Klein-Gordon oscillator in $\left(2+1\right)$ dimensions allows us the write the Klein-Gordon equation in the form: $$\begin{aligned}
\left[\mathcal{E}^{2}-m^{2}\right]\phi=\left[\hat{p}+im\omega\rho\,\hat{\rho}\right]\cdot\left[\hat{p}-im\omega\rho\,\hat{\rho}\right]\phi,
\label{kgo}\end{aligned}$$ where $m$ is the rest mass of the scalar particle, $\omega$ is the angular frequency of the Klein-Gordon oscillator, $\rho=\sqrt{x^{2}+y^{2}}$ is the radial coordinate and $\hat{\rho}$ is a unit vector in the radial direction.
In recent years, the relativistic quantum dynamics of a scalar particle subject to different confining potentials has been investigated in several areas of physics [@alvaro; @qian; @castro; @alhardi; @adame; @xu]. A special case is the confinement of a relativistic scalar particle to the Coulomb potential [@kg; @kg2; @kg3; @kg4; @greiner]. The procedure in introducing a scalar potential into the Klein-Gordon equation occurs by modifying the momentum operator $\hat{p}_{\mu}=i\partial_{\mu}$ in the form: $\hat{p}_{\mu}\rightarrow \hat{p}_{\mu}-q\,A_{\mu}\left(x\right)$ [@greiner]. On the other hand, one can also introduce a scalar potential (non-electromagnetic potential) in the Klein-Gordon equation by making a modification in the mass term in the form: $m\rightarrow m+S\left(\vec{r},\,t\right)$, where $S\left(\vec{r},\,t\right)$ is the scalar potential [@scalar; @greiner]. This modification in the mass term has been explored in recent decades, for instance, by analysing the behaviour of a Dirac particle in the presence of a static scalar potential and a Coulomb potential [@scalar2], and a relativistic scalar particle in the cosmic string spacetime [@eug]. Another example is the quark-antiquark interaction discussed in Ref. [@bah], where this interaction is mapped into a problem of a relativistic spinless possessing a position-dependent mass, where the mass term acquires a contribution given by a interaction potential that consists of a linear and a harmonic confining potential plus a Coulomb potential term.
In this work, we investigate the relativistic quantum dynamics of an electrically charged particle under the influence of the Klein-Gordon oscillator and the Coulomb potential. The Coulomb potential is introduced into the Klein-Gordon equation via minimal coupling, that is, as the component $A_{0}$ of the gauge potential. Then, we show that relativistic bound state solutions to the Klein-Gordon equation can be obtained for both attractive and repulsive Coulomb potentials. Moreover, a quantum effect characterized by a restriction of the values of the angular frequency of the Klein-Gordon oscillator is investigated. We show that this restriction is imposed by the quantum numbers of the system. From the mathematical point of view, this dependence of the angular frequency of this relativistic oscillator on the quantum numbers results from obtaining a polynomial solution to the biconfluent Heun equation. From the quantum mechanics point of view, this is an effect which arises from the influence of the Coulomb potential on the Klein-Gordon oscillator, whose meaning is that not all values of the angular frequency are allowed. Besides, we analyse the relativistic quantum dynamics of an electrically charged particle under the influence of the Klein-Gordon oscillator, the Coulomb potential and a linear scalar potential.
The structure of this paper is as follows: in section II, we study the relativistic quantum dynamics of a scalar particle under the influence of the Klein-Gordon oscillator and the Coulomb potential in the Minkowski spacetime in $\left(2+1\right)$ dimensions. We obtain relativistic bound state solutions to the Klein-Gordon equation for both attractive and repulsive Coulomb potentials. Besides, we analyse a quantum effect characterized by a dependence of the angular frequency of this relativistic oscillator on the quantum numbers; in section III, we analyse the relativistic quantum dynamics of an electrically charged particle under the influence of the Klein-Gordon oscillator, the Coulomb potential and a linear scalar potential; in section IV, we present our conclusions.
Klein-Gordon oscillator under the influence of a Coulomb potential
==================================================================
In this section, we investigate the relativistic quantum dynamics of a scalar particle under the influence of the Klein-Gordon oscillator and the Coulomb potential in (2+1) dimensions. We analyse the influence of the Coulomb potential on the spectrum of energy of the Klein-Gordon oscillator. By following Ref. [@greiner], the Coulomb potential is introduced into the Klein-Gordon equation via minimal coupling, that is, as the component $A_{0}$ of the gauge potential. Therefore, the Klein-Gordon equation with an electromagnetic field is written by introducing the minimal coupling $\hat{p}_{\mu}-q\,A_{\mu}$, where $q$ is the electric charge and $A_{\mu}=\left(-A_{0},\,\vec{A}\right)$ is the electromagnetic 4-vector potential. Therefore, we have [@greiner]: $$\begin{aligned}
\left[\hat{p}^{\mu}-q\,A^{\mu}\right]\left[\hat{p}_{\mu}-q\,A_{\mu}\right]\phi-m^{2}\,\phi=0,
\label{1.1}\end{aligned}$$ where the line element of the Minkowski spacetime is given in the form: $$\begin{aligned}
ds^{2}=-dt^{2}+d\rho^{2}+\rho^{2}\,d\varphi^{2},
\label{1.2}\end{aligned}$$ and the Coulomb potential is given by $$\begin{aligned}
q\,A_{0}=\frac{f}{\rho}=\pm\frac{\left|f\right|}{\rho},
\label{1.3}\end{aligned}$$ where $f$ is a constant and $\rho=\sqrt{x^{2}+y^{2}}$ is the radial coordinate.
Henceforth, let us consider a relativistic scalar particle subject to the Coulomb potential (\[1.3\]) and the Klein-Gordon oscillator given in Eq. (\[kgo\]). Thereby, the Klein-Gordon equation becomes $$\begin{aligned}
m^{2}\phi=\left[i\frac{\partial}{\partial t}+q\,A_{0}\right]^{2}\phi-\left[\hat{p}+im\omega\rho\,\hat{\rho}\right]\cdot\left[\hat{p}-im\omega\rho\,\hat{\rho}\right]\phi,
\label{1.4}\end{aligned}$$ where $m$ is the rest mass of the particle, $\omega$ is the angular frequency of the Klein-Gordon oscillator and $\hat{\rho}$ is a unit vector in the radial direction. Thereby, the Klein-Gordon equation (\[1.4\]) in the Minkowski spacetime, in (2+1) dimensions, becomes $$\begin{aligned}
m^{2}\phi=-\frac{\partial^{2}\phi}{\partial t^{2}}+i\frac{2f}{\rho}\frac{\partial\phi}{\partial t}+\frac{f^{2}}{\rho^{2}}\,\phi+\frac{\partial^{2}\phi}{\partial\rho^{2}}+\frac{1}{\rho}\frac{\partial\phi}{\partial\rho}+\frac{1}{\rho^{2}}\frac{\partial^{2}\phi}{\partial\varphi^{2}}+m\omega\,\phi-m^{2}\omega^{2}\rho^{2}\,\phi.
\label{1.5}\end{aligned}$$
In what follows, let us consider a particular solution to Eq. (\[1.5\]) which is an eigenfunction of the operator $\hat{L}_{z}=-i\partial_{\varphi}$. Therefore, we can write a particular solution to Eq. (\[1.5\]) in terms of the eigenvalues of the $z$-component of the angular momentum, $\hat{L}_{z}=-i\partial_{\varphi}$, as follows: $$\begin{aligned}
\phi\left(t,\,\rho,\,\varphi\right)=e^{-i\mathcal{E}t}\,e^{il\varphi}\,R\left(\rho\right),
\label{5}\end{aligned}$$ where $l=0,\pm1,\pm2,\ldots$ and $R\left(\rho\right)$ is a function of the radial coordinate. Then, substituting (\[5\]) into Eq. (\[1.5\]), we obtain $$\begin{aligned}
\frac{d^{2}R}{d\rho^{2}}+\frac{1}{\rho}\frac{dR}{d\rho}-\frac{\gamma^{2}}{\rho^{2}}\,R+\frac{2\mathcal{E}f}{\rho}\,R-m^{2}\omega^{2}\rho^{2}\,R+\beta\,R=0,
\label{6}\end{aligned}$$ where we have defined the following parameters: $$\begin{aligned}
\beta&=&\mathcal{E}^{2}-m^{2}+m\omega;\nonumber\\
[-2mm]\label{7}\\[-2mm]
\gamma^{2}&=&l^{2}-f^{2}.\nonumber\end{aligned}$$
From now on, let us consider $r=\sqrt{m\omega}\,\rho$, thus, we rewrite the radial equation (\[6\]) in the form: $$\begin{aligned}
\frac{d^{2}R}{dr^{2}}+\frac{1}{r}\frac{dR}{dr}-\frac{\gamma^{2}}{r^{2}}\,R+\frac{\delta}{r}\,R-r^{2}\,R+\frac{\beta}{m\omega}\,R=0,
\label{8}\end{aligned}$$ where we have defined a new parameter $$\begin{aligned}
\delta=\frac{2\mathcal{E}f}{\sqrt{m\omega}}.
\label{9}\end{aligned}$$
Let us discuss the asymptotic behaviour of the possible solutions to Eq. (\[8\]), which are determined for $r\rightarrow0$ and $r\rightarrow\infty$. From Refs. [@eug; @mhv; @vercin], the behaviour of the possible solutions to Eq. (\[8\]) at $r\rightarrow0$ and $r\rightarrow\infty$ allows us to write the function $R\left(r\right)$ in terms of an unknown function $H\left(r\right)$ as follows: $$\begin{aligned}
R\left(r\right)=e^{-\frac{r^{2}}{2}}\,r^{\left|\gamma\right|}\,H\left(r\right).
\label{10}\end{aligned}$$ Substituting Eq. (\[10\]) into Eq. (\[8\]), we obtain $$\begin{aligned}
\frac{d^{2}H}{dr^{2}}+\left[\left(2\left|\gamma\right|+1\right)\frac{1}{r}-2r\right]\frac{dH}{dr}+\left[\frac{\beta}{m\omega}-2-2\left|\gamma\right|+\frac{\delta}{r}\right]H=0.
\label{11}\end{aligned}$$
The second order differential equation (\[11\]) corresponds to the biconfluent Heun equation [@heun; @eug; @b50] and the function $H\left(r\right)$ is the biconfluent Heun function: $H\left(r\right)=H_{\mathrm{B}}\left(2\left|\gamma\right|,\,0,\,\frac{\beta}{m\omega},\,2\delta,-r\right)$. In order to proceed with our discussion about bound states solutions, let us use the Frobenius method [@arf; @f1]. Thereby, the solution to Eq. (\[11\]) can be written as a power series expansion around the origin: $$\begin{aligned}
H\left(r\right)=\sum_{j=0}^{\infty}\,a_{j}\,r^{j}.
\label{13}\end{aligned}$$
Substituting the series (\[13\]) into (\[11\]), we obtain the recurrence relation: $$\begin{aligned}
a_{j+2}=-\frac{\delta}{\left(j+2\right)\,\left(j+1+\alpha\right)}\,a_{j+1}-\frac{\left(g-2j\right)}{\left(j+2\right)\,\left(j+1+\alpha\right)}\,a_{j},
\label{14}\end{aligned}$$ where $$\begin{aligned}
\alpha=2\left|\gamma\right|+1;\,\,\,\,g=\frac{\beta}{m\omega}-2-2\left|\gamma\right|.
\label{14a}\end{aligned}$$
By starting with $a_{0}=1$ and using the relation (\[14\]), we can calculate other coefficients of the power series expansion (\[13\]). For instance, $$\begin{aligned}
a_{1}&=&-\frac{\delta}{\alpha}\nonumber\\
&=&-\frac{2\mathcal{E}f}{\sqrt{m\omega}}\frac{1}{\left(2\left|\gamma\right|+1\right)};\nonumber\\
[-2mm]\label{15}\\[-2mm]
a_{2}&=&\frac{\delta^{2}}{2\alpha\left(1+\alpha\right)}-\frac{\theta}{2\left(1+\alpha\right)}\nonumber\\
&=&\frac{2\mathcal{E}^{2}f^{2}}{m\omega}\frac{1}{\left(2\left|\gamma\right|+1\right)\left(2\left|\gamma\right|+2\right)}-\frac{g}{2\left(2\left|\gamma\right|+2\right)}.\nonumber\end{aligned}$$
The quantum theory requires that the wave function (\[5\]) must be normalizable, then, we assume that the function $R\left(r\right)$ vanishes at $r\rightarrow0$ and $r\rightarrow\infty$. In this way, bound state solutions can be obtained because there is no divergence of the wave function at $r\rightarrow0$ and $r\rightarrow\infty$. On the other hand, we have written the function $H\left(r\right)$ as a power series expansion around the origin in Eq. (\[13\]). Thereby, bound state solutions can be achieved by imposing that the power series expansion (\[13\]) or the biconfluent Heun series becomes a polynomial of degree $n$. In this way, we guarantee that $R\left(r\right)$ behaves as $r^{\left|\gamma\right|}$ at the origin and vanishes at $r\rightarrow\infty$ [@vercin; @mhv]. Through the recurrence relation (\[14\]), we can see that the power series expansion (\[13\]) becomes a polynomial of degree $n$ by imposing two conditions [@f1; @bb2; @bb4; @eug; @b50; @vercin; @mhv; @bf]: $$\begin{aligned}
g=2n\,\,\,\,\,\,\mathrm{and}\,\,\,\,\,\,a_{n+1}=0,
\label{16}\end{aligned}$$ where $n=1,2,3,\ldots$ and $g$ is given in Eq. (\[14a\]). From the condition $g=2n$, we can obtain the expression for the energy levels for bound states: $$\begin{aligned}
\mathcal{E}_{n,\,l}^{2}=m^{2}+m\,\omega_{n,\,l}\left[2n+2\left|\gamma\right|+1\right].
\label{17}\end{aligned}$$
Hence, by introducing the scalar potential via minimal coupling, then, we can see in Eq. (\[17\]) that the relativistic energy levels of the Klein-Gordon oscillator is modified by the influence of the Coulomb potential. This influence yields the ground state of the Klein-Gordon oscillator to be defined by the quantum number $n=1$ in contrast to the quantum number $n=0$ as obtained in Ref. [@kgo].
Let us now analyse the condition $a_{n+1}=0$ imposed in Eq. (\[16\]) in order to obtain a polynomial of degree $n$ the power series expansion given in Eq. (\[13\]). Thereby, let us assume that the angular frequency $\omega$ of the Klein-Gordon oscillator can be adjusted in such a way that the condition $a_{n+1}=0$ is satisfied. As a consequence, the quantum numbers of the system $\left\{n,\,l\right\}$ restrict the possible values of the angular frequency. Therefore, there are values of the angular frequency which are not allowed in the system. For this reason, we have labelled $\omega=\omega_{n,\,l}$ in Eq. (\[17\]). In this way, the conditions established in Eq. (\[16\]) are satisfied and a polynomial solution to the function $H\left(r\right)$ given in Eq. (\[13\]) is achieved [@eug]. As an example, let us consider the ground state $n=1$ and analyse the condition $a_{n+1}=0$. For $n=1$, thus, the condition $a_{n+1}=0$ yields $a_{2}=0$. By using Eq. (\[15\]), then, the condition $a_{2}=0$ yields $$\begin{aligned}
\omega_{1,\,\,l}=\frac{2\,\mathcal{E}_{1,\,l}^{2}\,\,f^{2}}{m}\frac{1}{\left(2\left|\gamma\right|+1\right)},
\label{18}\end{aligned}$$ which corresponds to the possible values of the angular frequency of the Klein-Gordon oscillator in the ground state. By substituting Eq. (\[18\]) into Eq. (\[17\]), the energy levels of the ground state is given by $$\begin{aligned}
\mathcal{E}_{1,\,l}=\pm \frac{m}{\sqrt{1-2f^{2}\frac{\left(3+2\left|\gamma\right|\right)}{\left(2\left|\gamma\right|+1\right)}}}.
\label{19}\end{aligned}$$
In what follows, let us consider the simplest case of the function $H\left(r\right)$ which corresponds to a polynomial of first degree. In this way, for $n=1$, we can write $H_{1,\,l}\left(r\right)=1+\frac{\delta}{\alpha}\,r$. Thereby, the radial wave function (\[10\]) associated with the ground state is given in the form: $$\begin{aligned}
R_{1,\,l}\left(r\right)=e^{-r^{2}/2}\,r^{\left|\gamma\right|}\,\left(1+\frac{\delta}{\alpha}\,r\right).
\label{19}\end{aligned}$$
Therefore, from the introduction of the scalar potential via minimal coupling in the Klein-Gordon equation, we have that the effects of the Coulomb potential on the the spectrum of energy of the Klein-Gordon oscillator is given by a change of the energy levels, where the ground state is defined by the quantum number $n=1$. Moreover, the values of the angular frequency of the Klein-Gordon oscillator are restricted to a set of values in which allow us to obtain a polynomial solution to the biconfluent Heun series. From the the quantum mechanics point of view, this is an effect characterized by the dependence of angular frequency of the Klein-Gordon oscillator on the quantum numbers $\left\{n,l\right\}$ of the system [@eug; @b50].
Klein-Gordon oscillator under the influence of a Coulomb potential and a linear scalar potential
================================================================================================
Let us extend our discussion by introducing a scalar potential into the Klein-Gordon equation by modifying in the mass term in the form: $m\rightarrow m+V\left(\vec{r},\,t\right)$, where $V\left(\vec{r},\,t\right)$ is the scalar potential [@scalar; @greiner]. Let us consider a linear scalar potential given by: $$\begin{aligned}
V\left(\rho\right)=\nu\,\rho,
\label{3.1}\end{aligned}$$ where $\nu$ is a constant that characterizes the linear confining potential. It has been proposed to describe the confinement of quarks [@linear; @linear1] due to experimental data show a behaviour of the confinement to be proportional to the distance between the quarks [@linear4; @linear4a; @linear4b; @linear4c]. It has also been explored in studies of the quark-antiquark interaction as a problem of a relativistic spinless particle which possesses a position-dependent mass, where the mass term acquires a contribution given by a interaction potential that consists of a linear and a harmonic confining potential plus a Coulomb potential term [@bah]. Furthermore, the linear scalar potential has attracted a great interest in atomic and molecular physics as pointed out in Refs. [@linear3a; @linear3b; @linear3c; @linear3d; @linear3e; @linear3f] and in several discussions of relativistic quantum mechanics [@linear2; @linear2a; @linear2b; @linear2c; @linear2d; @linear2e; @linear2f; @eug; @scalar2; @vercin; @mhv].
Hence, the general form of the Klein-Gordon equation describing the interaction of the Klein-Gordon oscillator with the static scalar potential (\[3.1\]) and the Coulomb potential (\[1.3\]) is given by (with $c=\hbar=1$) $$\begin{aligned}
\left[m+V\right]^{2}\phi=\left[i\frac{\partial}{\partial t}+q\,A_{0}\right]^{2}\phi-\left[\hat{p}+im\omega\rho\,\hat{\rho}\right]\cdot\left[\hat{p}-im\omega\rho\,\hat{\rho}\right]\phi.
\label{3.2}\end{aligned}$$ By substituting Eqs. (\[3.1\]) and (\[1.3\]), then, Eq. (\[3.2\]) becomes $$\begin{aligned}
m^{2}\phi&=&-\frac{\partial^{2}\phi}{\partial t^{2}}+i\frac{2f}{\rho}\frac{\partial\phi}{\partial t}+\frac{f^{2}}{\rho^{2}}\,\phi+\frac{\partial^{2}\phi}{\partial\rho^{2}}+\frac{1}{\rho}\frac{\partial\phi}{\partial\rho}+\frac{1}{\rho^{2}}\frac{\partial^{2}\phi}{\partial\varphi^{2}}\nonumber\\
[-2mm]\label{3.3}\\[-2mm]
&+&m\omega\,\phi-m^{2}\omega^{2}\rho^{2}\,\phi-2m\nu\,\rho\,\phi-\nu^{2}\,\rho^{2}\,\phi.\nonumber\end{aligned}$$
By following the steps from Eq. (\[1.5\]) to Eq. (\[7\]), we obtain the radial equation $$\begin{aligned}
\frac{d^{2}R}{d\rho^{2}}+\frac{1}{\rho}\,\frac{dR}{d\rho}-\frac{\gamma^{2}}{\rho^{2}}\,R+\frac{2f\mathcal{E}}{\rho}\,R-2m\nu\,\rho\,R-\theta^{2}R+\beta\,R=0,
\label{3.4}\end{aligned}$$ where the parameters $\beta$ and $\gamma$ are defined in Eq. (\[7\]) and the parameter $\theta$ is defined as $$\begin{aligned}
\theta^{2}=m^{2}\omega^{2}+\nu^{2}.
\label{3.5}\end{aligned}$$
Let us perform a change of variables given by $\xi=\sqrt{\theta}\,\rho$, then, the radial equation (\[3.4\]) becomes $$\begin{aligned}
\frac{d^{2}R}{d\xi^{2}}+\frac{1}{\xi}\,\frac{dR}{d\xi}-\frac{\gamma^{2}}{\rho^{2}}\,R+\frac{\tau}{\xi}\,R-\mu\,\xi\,R-\xi^{2}\,R+\frac{\beta}{\theta}\,R=0,
\label{3.6}\end{aligned}$$ where we have defined the parameters $$\begin{aligned}
\tau=\frac{2f\mathcal{E}}{\sqrt{\theta}};\,\,\,\,\,\mu=\frac{2m\nu}{\theta^{3/2}}.
\label{3.7}\end{aligned}$$
By analysing the asymptotic behaviour of the possible solutions to Eq. (\[3.6\]) as in the previous section, we have that we can write the function $R\left(\xi\right)$ in the form: $$\begin{aligned}
R\left(\xi\right)=e^{-\xi^{2}/2}\,e^{-\mu\xi/2}\,\xi^{\left|\gamma\right|}\,\bar{H}\left(\xi\right),
\label{3.8}\end{aligned}$$ where $\bar{H}\left(\xi\right)$ is an unknown function. After substituting Eq. (\[3.8\]) into Eq. (\[3.6\]), we obtain the following equation for the function $\bar{H}\left(\xi\right)$: $$\begin{aligned}
\frac{d^{2}\bar{H}}{d\xi^{2}}+\left[\frac{\left(2\left|\gamma\right|+1\right)}{\xi}-\mu-2\xi\right]\frac{d\bar{H}}{d\xi}+\left[\sigma+\frac{\vartheta}{\xi}\right]\bar{H}=0.
\label{3.9}\end{aligned}$$ where we have established that $$\begin{aligned}
\sigma=\frac{\beta}{\theta}+\frac{\mu^{2}}{4}-2-2\left|\gamma\right|;\,\,\,\,\vartheta=\frac{\mu}{2}\left(2\left|\gamma\right|+1\right)+\tau.
\label{3.10}\end{aligned}$$
Observer that Eq. (\[3.9\]) is also a biconfluent Heun equation [@heun] and the function $\bar{H}\left(\xi\right)=H_{\mathrm{B}}\left(2\left|\gamma\right|,\,\mu,\,\frac{\beta}{\theta}+\frac{\mu^{2}}{4},\,-2\tau,\,\xi\right)$ is the biconfluent Heun function. By following the steps from Eq. (\[13\]) to Eq. (\[16\]), we obtain the following recurrence relation: $$\begin{aligned}
a_{j+2}=\frac{\mu\left(j+1\right)+\vartheta}{\left(k+2\right)\left(k+2+2\left|\gamma\right|\right)}\,a_{j+2}-\frac{\sigma-2j}{\left(k+2\right)\left(k+2+2\left|\gamma\right|\right)}\,a_{j}.
\label{3.11}\end{aligned}$$ By starting from $a_{0}=1$ as in the previous section, the coefficients $a_{1}$ and $a_{2}$ becomes $$\begin{aligned}
a_{1}&=&\frac{\mu}{2}+\frac{\tau}{\left(1+2\left|\gamma\right|\right)};\nonumber\\
[-2mm]\label{3.12}\\[-2mm]
a_{2}&=&\frac{\vartheta\left(\vartheta+\mu\right)}{2\left(1+2\left|\gamma\right|\right)\left(2+2\left|\gamma\right|\right)}-\frac{\sigma}{2\left(2+2\left|\gamma\right|\right)}.\nonumber\end{aligned}$$
Now, we have that the series (\[13\]) becomes a polynomial of degree $n$ by imposing that: $$\begin{aligned}
\sigma=2n;\,\,\,\,a_{n+1}=0;
\label{3.13}\end{aligned}$$ where $n=1,2,3,\ldots$. From the condition $\sigma=2n$, we obtain $$\begin{aligned}
\mathcal{E}_{n,\,l}^{2}=m^{2}-m\omega+2\theta\left[n+\left|\gamma\right|+1\right]-\frac{m^{2}\nu^{2}}{\theta^{2}}.
\label{3.14}\end{aligned}$$
On the other hand, by analysing the condition $a_{n+1}=0$ for the ground state ($n=1$) as in the previous section, we obtain a third degree algebraic equation given by $$\begin{aligned}
\theta_{1,\,l}^{3}-\frac{2f^{2}\mathcal{E}_{1,\,l}^{2}}{\left(2+2\left|\gamma\right|\right)}\,\theta_{1,\,l}^{2}-2m\nu\,f\,\mathcal{E}_{1,\,l}\,\theta_{1,\,l}-m^{2}\nu^{2}\,\frac{\left(1+2\left|\gamma\right|\right)\left(3+2\left|\gamma\right|\right)}{2\left(2+\left|\gamma\right|\right)}=0;
\label{3.15}\end{aligned}$$ where we have labelled $\theta_{n,\,l}=\sqrt{m^{2}\,\omega_{n,\,l}^{2}+\nu^{2}}$ in order to establish that we are considering the angular frequency of the Klein-Gordon oscillator to be the parameter that can be adjusted in order that the condition $a_{n+1}=0$ can be satisfied and a polynomial solution to $\bar{H}\left(\xi\right)$ can be achieved. Since Eq. (\[3.15\]) has at least one real solution [@eug], then, the expression for the energy level of the ground state $\mathcal{E}_{1,\,l}$ can be obtained from this real solution. However, we do not write it here because the real solution of Eq. (\[3.15\]) is very long. Observe that, for other energy levels, different equations for $\theta_{n,\,l}$ or $\omega_{n,\,l}$ can be obtained from the condition $a_{n+1}=0$.
Conclusions
===========
We have seen that the effects of the Coulomb potential on the Klein-Gordon oscillator is given by the modification of the spectrum of energy of the Klein-Gordon oscillator, where the ground state is defined by the quantum number $n=1$ instead of the quantum number $n=0$ as given in Ref. [@kgo] and the values of the angular frequency are restricted and depend on the quantum numbers $\left\{n,\,l\right\}$ of the system. The meaning of this restriction of the values of the angular frequency is that only a set of values allow us to obtain a polynomial solution to the biconfluent Heun series. As an example, we have calculated the angular frequency of the ground state $n=1$ and obtained the expression of the energy level of the ground state.
We have also analysed the influence of a linear scalar potential and the Coulomb potential on the Klein-Gordon oscillator. We have seen that the energy levels of the Klein-Gordon oscillator are modified, where the ground state is also determined by the quantum number $n=1$ instead of the quantum number $n=0$, and possible values of the angular frequency are determined by the quantum numbers $\left\{n,\,l\right\}$ of the system. In particular, we have shown that the possible values of the angular frequency of the Klein-Gordon oscillator associated with the ground state of the system are determined by a third degree algebraic equation.
Recently [@bf], we have investigated the the effects of the Coulomb-type potential introduced by a coupling with the mass term on the spectrum of energy of the Klein-Gordon oscillator. Despite the Klein-Gordon oscillator is under the influence of a scalar potential, these two couplings yield different results, for instance, for the angular frequency $\omega_{1,\,l}$ and the relativistic energy level $\mathcal{E}_{1,\,l}$ associated with the ground state of the Klein-Gordon oscillator. It is worth mentioning that the influence of the Klein-Gordon oscillator and the Coulomb potential on a scalar particle can be of interest in studies of the quark-antiquark interaction [@bah], the Kaluza-Klein theory [@furtado], the Casimir effect [@casimir; @mb1] and relativistic effects on condensed matters systems such as effects associated with linear topological defects in solids [@kleinert; @kat; @moraesG2], the Aharonov-Bohm effect for bound states [@pesk; @fur5; @ab7] and persistent currents [@by; @tan].
[99]{}
M. K. Bahar and F. Yasuk, Advances in High Energy Physics, vol. 2013, Article ID 814985, 6 pages, 2013. Doi:10.1155/2013/814985.
R. Kumar and F. Chand, Phys. Scr. [**85**]{}, 055008 (2012).
S. H. Dong, Int. J. Theor. Phys. [**39**]{}, 1119 (2000).
S. H. Dong, Int. J. Theor. Phys. [**40**]{}, 559 (2001).
C. Quigg and J. L. Rosner, Phys. Rep. [**56**]{}, 167 (1979).
M. Chaichian and R. Kögerler, Ann. Phys. (NY), [**124**]{}, 61 (1980).
S. M. Ikhdair, Int. J. Mod. Phys. C [**20**]{}, 1563 (2009).
L. Dekar, L. Chetouani and T. F. Hammann, Phys. Rev. A [**59**]{}, 107 (1999).
A. D. Alhaidari, Phys. Rev. A [**66**]{}, 042116 (2002).
A. D. Alhaidari, Phys. Lett. A [**322**]{}, 72 (2004).
L. Serra and E. Lipparini, Europhys. Lett. [**40**]{}, 667 (1997).
M. Moshinsky and A. Szczepaniak, J. Phys. A: Math. Gen. [**22**]{}, L817 (1989).
P. Rozmej and R. Arvieu, J. Phys. A [**32**]{}, 5367 (1999).
A. Bermudez, M. A. Martin-Delgado and E. Solano, Phys. Rev. A [**76**]{}, 041801(R) (2007).
A. Bermudez, M. A. Martin-Delgado and A. Luis, Phys. Rev. A [**77**]{}, 063815 (2008).
A. Bermudez, M. A. Martin-Delgado and E. Solano, Phys. Rev. Lett. [**99**]{}, 123602 (2007).
A. Bermudez, M. A. Martin-Delgado and A. Luis, Phys. Rev. A [**77**]{}, 033832 (2008).
S. Bruce and P. Minning, Nuovo Cimento A [**106**]{}, 711 (1993).
V. V. Dvoeglazov, Nuovo Cimento A [**107**]{}, 1413 (1994).
N. A. Rao and B. A. Kagali, Phys. Scr. [**77**]{}, 015003 (2008).
A. Boumali, A. Hafdallah and A. Toumi, Phys. Scr. [**84**]{}, 037001 (2011).
B. Mirza and M. Mohadesi, Commun. Theor. Phys. [**42**]{}, 664 (2004).
B. Mirza, R. Narimani and S. Zare, Commun. Theor. Phys. [**55**]{}, 405 (2011).
M.-L. Liang and R.-L. Yang, Int. J. Mod. Phys. A [**27**]{}, 1250047 (2012).
Y. Xiao, Z. Long and S. Cai, Int. J. Theor. Phys. [**50**]{}, 3105 (2011).
J.-Y. Cheng, Int. J. Theor. Phys. [**50**]{}, 228 (2011).
A. de Souza Dutra and C.-S. Jia, Phys. Lett. A [**352**]{}, 484 (2006).
W.-C. Qiang, R.-S. Zhou and Y. Gao, Phys. Let. A, [**371**]{}, 201 (2007).
A. S. de Castro, Phys. Lett. A [**338**]{}, 81 (2005).
A. D. Alhaidari, H. Bahlouli and A. Al-Hasan, Phys. Lett. A [**349**]{}, 87, (2006).
F. Domingues-Adame, Phys. Lett. A [**136**]{}, 175 (1989).
Y. Xu, S. He and C.-S. Jia, Phys. Scr. [**81**]{}, 045001 (2010).
Q. Wen-Chao, Chinese Phys. [**12**]{}, 1054 (2003).
H. Motavalli and A. R. Akbarieh, Mod. Phys. Lett. A [**25**]{}, 2523 (2010).
F. Yasuk, A. Durmus and I. Boztosun, J. Math. Phys [**47**]{}, 082302 (2006).
A. L. Cavalcanti de Oliveira and E. R. Bezerra de Mello, Class. Quantum Grav. [**23**]{}, 5249 (2006).
W. Greiner, *Relativistic Quantum Mechanics: Wave Equations, 3rd Edition* (Springer, Berlin, 2000).
H. G. Dosch, J. H. Jansen and V. F. Müller, Phys. Norv. [**5**]{}, 2 (1971).
G. Soff, B. Müller, J. Rafelski and W. Greiner, Z. Naturforsch. A [**28**]{}, 1389 (1973).
E. R. Figueiredo Medeiros and E. R. Bezerra de Mello, Eur. Phys. J. C [**72**]{}, 2051 (2012).
A. Verćin, Phys. Lett. B [**260**]{}, 120 (1991).
J. Myrhein, E. Halvorsen and A. Verćin, Phys. Lett. B [**278**]{}, 171 (1992).
A. Ronveaux, *Heun’s differential equations* (Oxford University Press, Oxford, 1995).
K. Bakke, Ann. Phys. (NY) [**341**]{}, 86 (2014).
G. B. Arfken and H. J. Weber, [*Mathematical Methods for Physicists, sixth edition*]{} (Elsevier Academic Press, New York, 2005).
C. Furtado, B. G. C. da Cunha, F. Moraes, E. R. Bezerra de Mello and V. B. Bezerra, Phys. Lett. A [**195**]{}, 90 (1994).
K. Bakke and H. Belich, Eur. Phys. J. Plus [**127**]{}, 102 (2012).
K. Bakke and H. Belich, Ann. Phys. (NY) [**333**]{}, 272 (2013).
K. Bakke and C. Furtado, Ann. Phys. (NY) [**355**]{}, 48 (2015).
C. L. Chrichfield, Phys. Rev. D [**12**]{}, 923 (1975).
C. L. Chrichfield, J. Math. Phys. [**17**]{}, 261 (1976).
R. S. Kaushal, Phys. Lett. B [**57**]{}, 354 (1975).
E. Eichten [*el al*]{}, Phys. Rev. D [**21**]{}, 203 (1980).
E. Eichten [*el al*]{}, Phys. Rev. Lett. [**34**]{}, 369 (1975).
H. Tezuka, J. Phys. A: Math. Gen. [**24**]{}, 5267 (1991).
E. J. Austin, Mol. Phys. [**40**]{}, 893 (1980).
E. R. Vrscay, Phys. Rev. A [**31**]{} 2054 (1985).
K. Killingbeck, Rep. Prog. Phys. [**40**]{}, 963 (1977).
K. Killingbeck, Phys. Lett. A [**65**]{} 87 (1978).
R. P. Saxena and V. S. Varma, J. Phys. A: Math. Gen. [**15**]{}, L149 (1982).
E. Castro and P. Martín, J. Phys. A: Math. Gen. [**33**]{}, 5321 (2000).
G. Plante and A. F. Antippa, J. Math. Phys. [**46**]{}, 062108 (2005).
J. H. Noble and U. D. Jentschura, Int. J. Mod. Phys. A [**30**]{}, 1550002 (2015).
M. L. Glasser and N. Shawagfeh, J. Math. Phys. [**25**]{}, 2533 (1984).
H. Tezuka, AIP Advances [**3**]{}, 082135 (2013).
J. F. Gunion and L. F. Li, Phys. Rev. D [**12**]{}, 3583 (1975).
P. Leal Ferreira, Phys. Rev. D [**38**]{}, 2648 (1988).
F. Domínguez-Adame adn M. A. González, Europhys. Lett. [**13**]{}, 193 (1990).
C. Furtado, F. Moraes and V. B. Bezerra, Phys. Rev. D [**59**]{}, 107504 (1999).
N. R. Khusnutdinov and M. Bordag, Phys. Rev. D [**59**]{}, 064017 (1999).
H. F. Mota and K. Bakke, Phys. Rev. D [**89**]{}, 027702 (2014).
H. Kleinert, [*Gauge fields in condensed matter, vol. 2*]{}, (World Scientific, Singapore, 1989).
M. O. Katanaev and I. V. Volovich. Ann. Phys. (NY) [**216**]{}, 1 (1992).
F. Moraes, Braz. J. Phys [**30**]{}, 304 (2000).
M. Peshkin and A. Tonomura, *The Aharonov-Bohm Effect* (Springer-Verlag, in: Lecture Notes in Physics, Vol. 340, Berlin, 1989).
C. Furtado, V. B. Bezerra and F. Moraes, Phys. Lett. A [**289**]{}, 160 (2001).
C. Furtado, C. A. de Lima Ribeiro and S. Azevedo, Phys. Lett. A [**296**]{}, 171 (2002).
N. Byers and C. N. Yang, Phys. Rev. Lett. [**7**]{}, 46 (1961).
W.-C. Tan and J. C. Inkson, Phys. Rev. B [**60**]{}, 5626 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Matthew R. Bate'
title: Recent Advances in Binary Star Formation Using SPH
---
Introduction
============
The Smoothed Particle Hydrodynamics (SPH) numerical method was introduced by Lucy [@Lucy77] and Gingold and Monaghan [@GinMon77]. Its first application was in the field of star formation, where it was used to study whether or not a rapidly-rotating polytrope could undergo fission to form a close binary system [@Lucy77; @GinMon78]. Since this initial application, SPH has been widely used in the study of star formation, for example [@GinMon81; @MiyHayNar84; @Lattanzioetal85; @Durisenetal86; @SigKla97; @BonBas91; @BatBonPri95; @Heller93; @NelPap93; @ArtLub94; @Larwoodetal96; @Nelsonetal98; @VanCam98; @Chapmanetal92].
SPH is has many attributes which make it particularly well suited to the study of star formation. SPH is Lagrangian and does not require a computational grid. Thus, it can efficiently follow problems with large density contrasts since computational effort is not wasted simulating the low-density regions. Also, recent SPH implementations [@Evrard88; @HerKat89; @Benz90; @Monaghan92] use spatially and temporally varying smoothing lengths so that the resolution increases automatically with increasing density; the complex multi-grid and adaptive-grid schemes that are used for finite difference methods are avoided.
In this proceedings, we review some of the recent advances in the study of binary star formation that have been made using SPH. In Section \[jeansmass\], we discuss the importance of always resolving the Jeans mass in numerical studies of self-gravitating gas. While this has been demonstrated using various types of hydrodynamic code, we concentrate specifically on the problems that can arise if this criterion is not obeyed with SPH. In Section \[collstellar\], we demonstrate that, for the first time, it is now possible to perform three-dimensional hydrodynamic calculations which follow the collapse of a molecular cloud core to stellar densities. These calculations are performed with SPH. Finally, in Section \[accretion\], we discuss how SPH has been used to study the evolution of a protobinary system as it accretes from an infalling gaseous envelope and how this work can lead to predictions of the properties of binary stars.
The Importance of Resolving the Jeans Mass {#jeansmass}
==========================================
It has recently been realised that it is important that the Jeans mass/length is always resolved during a hydrodynamical calculation. This has been demonstrated both with SPH [@BatBur97; @Whitworth98] and with grid-based codes [@Trueloveetal97; @Trueloveetal98; @Boss98]. If this criterion is not obeyed, artificial fragmentation can be induced, or fragmentation can be inhibited. Essentially this is because when the resolution length/mass approaches the Jeans length/mass, collapse is artificially delayed due to viscous forces, softening of gravitational forces, or a combination of both. A good example is the collapse of an isothermal filament [@Trueloveetal97]. Such a filament should collapse without limit to a filamentary singularity without fragmenting. However, if the collapse perpendicular to the major-axis is delayed, small density perturbations along the filament may have enough time to grow to non-linear amplitudes and fragments may form along the bar.
Bate & Burkert [@BatBur97] first demonstrated the need for the Jeans mass criterion using self-gravitating SPH calculations. With SPH, the problem can be understood by considering the gravitational and pressure forces between two particles within a marginally Jeans-stable clump of gas of radius $R\approx h$ (Figure \[bate.forces\]). SPH codes typically soften the gravitational forces between neighbouring particles, using either the Plummer force law or kernel softening [@GinMon77; @HerKat89; @Benz90]. The characteristic gravitational softening length, $\epsilon$, may or may not be equal to the SPH hydrodynamic smoothing length, $h$, depending on the specific implementation. If $\epsilon=h$, the ratio of the gravitational and pressure forces between two particles is approximately constant for particle with separations $\simless h$. Thus, a Jeans-unstable clump of gas will collapse, while a Jeans-stable clump will be supported. However, although the ratio of the gravitational and pressure forces is approximately independent of the separation of the particles, the magnitude of the gravitational force decreases with separation due to the softening. Thus, while a Jeans-unstable clump of gas with a size much larger than the resolution length, $h$, will collapse at the correct rate, the collapse of clumps with a size $\approx h$ will be delayed. This is demonstrated in Figure \[bate.massden\].
Unfortunately, the effect is not always limited to a simple delay of the collapse. Given the right problem, artificial fragmentation can be induced (such as with the filament described above), or inhibited. In Figure \[bate.80000\] we show how the collapse of a particular molecular cloud core with an initial $m=2$ density perturbation results in a binary protostellar system with a bar of gas between them. This calculation was performed with $8\times 10^4$ particles, enough to resolved the Jeans mass/length until after the binary had formed. However, performing the same calculation with $1\times 10^4$ particles gives a different result: a single, dense bar of gas without the binary. In this calculation, the Jeans mass becomes unresolved before $t=1.20$ and the collapse of each of the two over-dense regions resulting from the original $m=2$ density perturbation is delayed. The collapse of the larger-scale elongated region, however, continues, leading to the formation of a bar rather than a binary.
It is important to note that, with SPH, the effect of not resolving a Jeans mass also depends on the way the gravitational softening and hydrodynamic smoothing are implemented! If $\epsilon < h$, the ratio of the gravitational to pressure forces between two particles increases with decreasing separation (Figure \[bate.forces\]). This may lead to an instability in which a group of particles within a Jeans-stable clump collapse artificially. As demonstrated in Figure \[bate.10000eh\], this can lead to artificial fragmentation. Alternately, if $\epsilon > h$, the pressure forces between particles within a Jeans-unstable clump may exceed the gravitational forces and the clump will be artificially supported against collapse.
Clearly, the best SPH implementation is one where $\epsilon=h$ always. In this case, collapse of Jeans-unstable clumps with a size similar to that of the resolution length will still be delayed, but the possibilities of artificial collapse within Jeans-stable clumps or the supporting of Jeans-unstable clumps against collapse are eliminated. Then, in order to avoid the collapse of Jeans-unstable clumps being delayed significantly, enough particles should be used so that the Jeans length/mass is always resolved.
How many particles are necessary to ensure that the Jeans length/mass is resolved? With SPH, the spatial resolution is given by the smoothing length which is usually variable in time and space. The smoothing lengths are set by ensuring that each particle contains a certain number of neighbours, $N_{\rm neigh}$, or equivalently a fixed mass, within two smoothing lengths. Thus, in contrast to a grid-based code which has spatially-limited resolution, SPH has mass-limited resolution which [*automatically*]{} gives greater spatial resolution in regions of higher density. Therefore, with SPH, it is necessary to ensure that the minimum resolvable mass is always less than the Jeans mass. In practice, Bate & Burkert found that a Jeans mass should always be represented by at least $\approx 2 N_{\rm neigh}$ particles.
Collapse of a Molecular Cloud to Stellar Densities {#collstellar}
==================================================
The mass-limited resolution of SPH is ideal for studying the collapse and fragmentation of molecular cloud cores because there is a minimum Jeans mass in the problem (Figure \[bate.tmr\]). By contrast, there is no minimum Jeans length. This is a problem for grid-based codes which must resort to nested or adaptive grids [@BurBod93; @Trueloveetal97; @Trueloveetal98]. With SPH, if the number of particles used is sufficient to resolve the minimum Jeans mass, a calculation can be followed to arbitrary densities with the required spatial resolution given automatically with increasing density.
Recently, this ability of SPH was used to perform the first three-dimensional calculations ever to follow the collapse of a molecular cloud core to stellar densities [@Bate98]. The calculations followed the collapse of an initially uniform-density molecular cloud core of mass $M=1 {\rm \ M}_\odot$ and radius $R=7 \times 10^{16} {\rm \ cm}$.
The minimum resolvable mass in the SPH code was $\approx 2 N_{\rm neigh} = 100$ particles. Thus, to enable the minimum Jeans mass during the calculation ($\approx 4 \times 10^{-4}\ {\rm M}_{\odot}$) to be resolved, the calculation used $3 \times 10^5$ equal-mass particles.
The code did not include radiative transfer. Instead, to model the behaviour of the gas during the different phases of collapse, a piece-wise polytropic equation of state, $P=K \rho^{\gamma}$, was used, where $P$ is the pressure, $\rho$ is the density, $K$ gives the entropy of the gas, and the ratio of specific heats, $\gamma$, was varied as $$\label{polytropic}
\gamma = \cases {\begin{array}{lll}
1 & & \rho \leq 1.0 \times 10^{-13} \cr
7/5 & \ 1.0 \times 10^{-13} < \hspace{-6pt} & \rho \leq 5.7 \times 10^{-8}
\cr
1.15 & \ 5.7 \times 10^{-8\ } < \hspace{-6pt} & \rho < 1.0 \times 10^{-3} \cr
5/3 & & \rho > 1.0 \times 10^{-3} \cr
\end{array}}$$ where the densities are in ${\rm g\ cm}^{-3}$ (see Figure \[bate.tmr\]). The values of $\gamma$ and the transition densities were derived from Tohline [@Tohline82]. The variable value of $\gamma$ mimics the following behaviour of the gas. The collapse is isothermal ($\gamma=1$) until the gas becomes optically thick to infrared radiation at $\rho \approx 10^{-13} {\rm \ g \ cm}^{-3}$, beyond which $\gamma=7/5$ (appropriate for a diatomic gas). When the gas reaches a temperature of $\approx$ 2000 K ($\rho = 5.7 \times 10^{-8} {\rm \ g \ cm}^{-3}$), molecular hydrogen begins to dissociate and the temperature only slowly increases with density. In this phase $\gamma = 1.15$ is used to model [*both*]{} the decreasing mean molecular weight and the slow increase of temperature with density, the latter of which has an effective $\gamma \approx 1.10$. Finally, when the gas has fully dissociated ($\rho \approx 10^{-3} {\rm \ g \ cm}^{-3}$), the gas is monatomic and $\gamma=5/3$. The value of $K$ is defined such that when the gas is isothermal, $K = c_{\rm s}^2$ with $c_{\rm s} = 2.0 \times 10^4 {\rm \ cm\ s^{-1}}$, and when $\gamma$ changes the pressure is continuous.
Collapse of an initially-static cloud
-------------------------------------
To test that the above equation of state captures the important elements of the gas’s behaviour, spherically-symmetric, one-dimensional (1-D), finite-difference calculations were performed of the collapse of an initially-static molecular cloud core with the above parameters and equation of state. The results are shown in Figure \[bate.static\] (solid line). These results are in good agreement with the results from 1-D calculations incorporating radiative transfer (e.g. Larson 1969; Winkler & Newman 1980a, b).
The three-dimensional (3-D) SPH code was also tested on the same problem to check that the SPH code can indeed accurately resolve the collapse down to stellar densities (Figure \[bate.static\], dotted line). There is excellent agreement between the results from the 1-D finite-difference code and those from the 3-D SPH code.
Collapse of a rotating cloud
----------------------------
Three-dimensional calculations are required if the molecular cloud core is rotating. In Figures \[bate.rotmassdens\] to \[bate.finalstate2\] we present results from the collapse of a cloud core which is identical to that in the previous section, but which is initially in solid-body rotation with angular frequency $\Omega=7.6 \times 10^{-14}$ rad s$^{-1}$. Thus, the ratio of rotational energy to the magnitude of the gravitational potential energy is $\beta=0.005$ (i.e. the cloud is rotating quite slowly).
The evolution of the calculation is as follows (Figure \[bate.rotmassdens\]). The initial collapse is isothermal. When the density surpasses $10^{-13}$ g cm$^{-3}$, the gas in the center is assumed to become optically thick to infrared radiation and begins to heat ($t=1.009~t_{\rm ff}$). The heating stops the collapse at the center and the first hydrostatic core is formed ($t=1.015~t_{\rm ff}$) with maximum density $\approx 2 \times 10^{-11}$ g cm$^{-3}$, mass $\approx 0.01 {\rm \ M}_\odot$ ($\approx 3 \times 10^3$ particles), and radius $\approx 7$ AU. As the first core accretes, its maximum density only slowly increases at first. However, the first core is rapidly rotating, oblate, and has $\beta \approx 0.34 > 0.274$, making it dynamically unstable to the growth of non-axisymmetric perturbations [@Durisenetal86; @ImaDurPic99]. At $t \approx 1.023~t_{\rm ff}$, after about 3 rotations, the first core becomes violently bar-unstable and forms trailing spiral arms (Figure \[bate.bar\]). This leads to a rapid increase in maximum density (Figure \[bate.rotmassdens\]) as angular momentum is removed from the central regions of the first core (now a disc with spiral structure) by gravitational torques ($t=1.023-1.030~t_{\rm ff}$). An MPEG animation of this bar instability is provided on this CD-ROM (bate1.mpg). When the maximum temperature reaches 2000 K, molecular hydrogen begins to dissociate, resulting in a rapid second collapse to stellar densities ($t=1.030~t_{\rm ff}$). The collapse is again halted at a density of $\approx 0.007$ g cm$^{-3}$ with the formation of the second hydrostatic, or stellar, core. The initial mass and radius of the stellar core are $\approx 0.0015 {\rm \ M}_\odot$ ($\approx 5 \times 10^2$ particles) and $\approx 0.8 {\rm \ R}_\odot$, respectively. Finally, an inner circumstellar disc begins to form around the stellar object, within the region undergoing second collapse. The calculation is stopped when the stellar object has a mass of $\approx 0.004 {\rm \ M}_\odot$ ($\approx 1.2 \times 10^3$ particles), the inner circumstellar disc has extended out to $\approx 0.1$ AU, and the outer disc (the remnant of the first hydrostatic core) contains $\approx 0.08 {\rm \ M}_\odot$ ($\approx 2.4 \times 10^4$ particles) and extends out to $\approx 70$ AU. Note that the massive outer disc forms [*before*]{} the stellar core. This final state is illustrated in Figures \[bate.finalstate\] and \[bate.finalstate2\]. If the calculation was evolved further, the inner circumstellar disc would continue to grow in radius as gas with higher angular momentum fell in. Eventually, the inner disc would meet the outer disc with only a small molecular dissociation region between the two.
Close binary stellar systems
----------------------------
The ability to perform three-dimensional calculations which follow the collapse of molecular cloud cores to stellar densities allows us to study the formation of close ($\simless 1$ AU) binary stellar systems. Currently, there is no accepted mechanism for forming close binaries; the proposal that close binary systems form via the fission of a rapidly-rotating protostellar object has been discredited by studies of rapidly-rotating polytropes [@Durisenetal86]. Although fission itself appears not to operate, it is possible that fragmentation can still occur due to the growth of non-axisymmetric perturbations in rotationally-supported objects. Only two studies have looked at this possibility in detail [@Boss87; @BonBat94b], and the latter of these finds that fragmentation of a massive circumstellar disc on scales ($\simless 1$ AU) may be possible. However, in both these studies, only the region inside the first hydrostatic core was modelled and the initial conditions were chosen somewhat artificially. The ability to perform three-dimensional calculations which follow the collapse of molecular cloud cores to stellar densities now allows us to study the formation of close binaries from the collapse of larger-scale ($\approx 10000$ AU) molecular cloud cores.
The evolution of an accreting protobinary system {#accretion}
================================================
The favoured mechanism for the formation of most binary stellar systems is the fragmentation of a collapsing molecular cloud core. Fragmentation has been studied numerically for $\approx 20$ years. These calculations appear to show that it is possible to form binaries with similar properties to those that are observed via fragmentation. However, they have not allowed us to predict the fundamental properties of stellar systems such as the fraction of stellar systems which are binary or the properties of binary systems (e.g. the distributions of mass ratios, separations, and eccentricities and the properties of discs in pre-main-sequence systems).
There are two primary reasons for this lack of predictive power. First, the results of fragmentation calculations depend sensitively on the initial conditions, which are poorly constrained. The second problem is that of accretion. Fragmentation calculations are typically stopped soon after the fragmentation occurs, when the binary or multiple protostellar system contains only a small fraction of the total mass of the original cloud [@Boss86; @BonBat94b]. However, because much of the gas contained in the original cloud still has to fall on to the system and be accreted, the final properties of the stellar system are unknown. Following the calculation significantly beyond the point at which fragmentation occurs is extremely computationally intensive. Thus, it is impossible to perform the number of calculations that are required to predict the statistical properties of binary stellar systems – even if we knew the distribution of the initial conditions. On the other hand, if we can overcome this second difficulty, we can make theoretical predictions about the properties of binary stars and, by comparing these predictions to the observed properties of binary systems, we may be able to better constrain the initial conditions for star formation.
The Effects of Accretion on a Protobinary System
------------------------------------------------
Using SPH, Bate & Bonnell [@BatBon97] studied and quantified how the properties of a binary system are affected by the accretion of a small amount of gas from an infalling gaseous envelope. They found that the effects depend primarily on the specific angular momentum of the gas and the binary’s mass ratio (see also [@Artymowicz83; @Bate97]). Generally, accretion of gas with low specific angular momentum decreases the mass ratio and separation of the binary, while accretion of gas with high specific angular momentum increases the separation and drives the mass ratio toward unity. From these results, they predicted that closer binaries should have mass ratios that are biased toward equal masses compared to wider systems, since the gas which falls on to a closer system is likely to have more specific angular momentum, relative to the binary, than for a wider system.
They also studied the process of disc formation around an accreting protobinary system and found that for each protostar, a circumstellar disc was only formed if the specific angular momentum of the infalling gas was greater than the specific orbital angular momentum of that protostar about the centre of mass of the binary (Figure \[bate.accretion\]). This is because, to be capture by one of the protostars, the gas much achieve the same specific orbital angular momentum as that of the protostar. If the gas has more specific angular momentum initially, some of its angular momentum goes into forming a disc around the protostar. However, if it has less specific angular momentum initially, there is no excess angular momentum to form a circumstellar disc, and it must gain angular momentum even to be captured by the protostar. In this case, the infalling gas gains angular momentum as it falls on to the protostar in a Bondi-Hoyle-type accretion stream. In practice, this means that a circumstellar disc is almost always formed around the primary, but the secondary does not have a circumstellar disc unless the infalling gas has more specific angular momentum that some critical value. In a similar way, the formation of a circumbinary disc only begins when the specific angular momentum of the infalling gas is great enough for the gas to form a circular orbit at a radius greater than that of the secondary from the centre of mass of the binary.
Development of a Protobinary Evolution Code
-------------------------------------------
Using the quantitative results of Bate & Bonnell [@BatBon97], Bate [@Bate2000] developed a protobinary evolution (PBE) code which follows the evolution of a protobinary system as it accretes from its initial to its final mass, but does so in far less time than would be required for a full hydrodynamic calculation.
This code is based on the following model for the formation of binary stellar systems (Figure \[bate.model\]). The model begins with a molecular cloud core of known initial density and angular momentum profile. It is assumed that this cloud begins to collapse and that a ‘seed’ binary system is formed at the centre, presumably via some sort of fragmentation. The ‘seed’ binary has mass ratio $q \leq 1$, separation $a$, and is assumed to have a circular orbit. It initial consists of only a small fraction of the total mass of the core and is assumed to have formed from the gas that was originally contained within a sphere of radius $r$, at the centre of the initial cloud (Figure \[bate.model\]). For the results presented in this proceedings, the separation of the ‘seed’ binary is set by assuming that the angular momentum of the gas from which the binary forms is equal to the orbital angular momentum of the binary.
Subsequently, the binary accretes the remainder of the initial cloud (which falls on to the binary) and the binary’s properties evolve due to the accretion. This evolution is calculated by taking a thin shell of gas of thickness d$r$ (Figure \[bate.model\]), surrounding the sphere from which the binary was formed, dividing the shell into small elements of gas, and calculating the effect that each element of gas has on the protobinary when it is accreted (using the results of Bate & Bonnell [@BatBon97]). The binary’s parameters (masses and separation) are updated, and the next shell of gas is considered until the whole cloud is accreted on to the binary. The amount of gas which settles into a circumbinary disc is also recorded. In this way, the code calculates the evolution of the binary from its initial to its final state when all of the original cloud’s gas is contained either in one of the two stars or their surrounding discs.
Testing the Protobinary Evolution Code
--------------------------------------
To test how accurately the PBE code describes the evolution of a ‘seed’ binary as it accretes from its initial to its final mass, the PBE results were compared to those from full SPH calculations. Two test cases were performed. The first followed the formation of a binary system from the collapse of an initially uniform-density, spherical molecular cloud core in solid-body rotation. The ‘seed’ binary was assumed to have a mass ratio of $q=0.6$ and a mass of $1/10$ the initial cloud mass. The second test case was similar, except that the progenitor cloud was centrally-condensed with a 1/r-density distribution and the cloud had a total mass of only 5 times the ‘seed’ binary’s mass. A full discussion of the test cases is given by [@Bate2000].
### Test Case 1
The evolution of the mass ratio, separation and amount of gas in the circumbinary disc are given for the PBE code and for a full SPH calculation in Figure \[bate.testcase1\]. The curves are given as functions of the amount of gas that has fallen on to the binary, $M_{\rm acc}$, relative to the binary’s initial mass. In addition, an MPEG animation of the SPH calculation is included on this CD-ROM as bate2.mpg. The CPU time required to evolve the SPH calculation until the entire cloud falls on to the binary in is prohibitively long, which, after all, is the reason that the PBE code was developed in the first place. It takes $\approx 60$ orbits for the binary to increase its mass by a factor of 6 (i.e. $\approx 60$% of the total cloud was accreted). The SPH calculation took $\approx 5$ months on a 170 MHz Sun Ultra workstation with a GRAvity-PipE (GRAPE) board used to calculate the gravitational forces and neighbouring SPH particles. The evolution with the PBE code took a few seconds!
Although the SPH calculation did not run to completion, we can compare the evolution as the binary’s mass increases by a factor of 6 (Figure \[bate.testcase1\]). Generally, there is good agreement between the PBE and SPH codes. The mass ratio is predicted to within 5% over the entire evolution and the separation to within 15%. In fact, as discussed in [@Bate2000], the small differences between the PBE and SPH results reflect unphysical treatment of the circumstellar discs by the SPH code rather than a problem with the PBE code. For example, the slower rate of increase of the mass ratio initially is due to the circumsecondary disc not being resolved correctly in the SPH calculation, and the larger separation when $M_{\rm acc}\simgreat 3$ is due to unphysically-rapid viscous evolution of the circumstellar discs which transfers angular momentum into the binary’s orbit too quickly. The greatest difference between the PBE and SPH results is that the PBE code predicts that a circumbinary disc should be formed around the binary whereas no circumbinary disc is formed in the SPH calculation. This is due to the larger separation of the binary when $M_{\rm acc}\simgreat 3$ and the large shear viscosity in the SPH calculation.
### Test Case 2
Unlike test case 1, the PBE code predicts that a massive circumbinary disc should be produced very early in the evolution of test case 2. Thus, test case 2 provides a better test of how well the PBE code predicts the formation of a circumbinary disc and its evolution. We note that, although the PBE code records the amount of gas which settles into a circumbinary disc, it does not attempt to take account of the interaction between the binary and the circumbinary disc. In reality, this interaction is expected to result in the transfer of angular momentum from the orbit of the binary into the gas of the circumbinary disc and, hence, in a smaller separation. Furthermore, if the separation decreases, more of the infalling gas would be expected to settle into the circumbinary disc and, for the same increase in the binary’s mass, the mass ratio should increase more rapidly because the gas has a greater specific angular momentum relative to that of the binary. Thus, if a massive circumbinary disc is formed, the PBE code is expected to over-estimate the binary’s separation, under-estimate the mass in the circumbinary disc, and slightly under-estimate the mass-ratio of the binary.
The evolution of the mass ratio, separation and amount of gas in the circumbinary disc are given for test case 2 in Figure \[bate.testcase2\]. An MPEG animation of the SPH calculation is included on this CD-ROM as bate3.mpg. To avoid the problems that occurred due to the large shear viscosity in the SPH calculation for test case 1, the SPH calculation here uses a formulation with less shear viscosity (see [@Bate2000]). As with test case 1, due to the computational cost, the SPH calculations were stopped before all of the gas had fallen on to the binary. The SPH calculation took $\approx 4$ months on a 300 MHz Sun Ultra workstation (using a binary tree, not a GRAPE board). During the evolution, the binary performed $\approx 40$ orbits and $\approx 60$% of the total mass was accreted by the binary or settled into a circumbinary disc.
The agreement for the evolution of the mass ratio is even better than it was with test case 1 with differences between the PBE and SPH results of $\simless 3$%. The separation follows the prediction of the PBE code to better than $3$% until the circumbinary disc begins to form. Once the circumbinary disc attains approximately 5% of the binary’s mass, however, the separation is always smaller than predicted by the PBE code. As described above, this is expected because the PBE code neglects the separation-decreasing effect of the interaction between the binary and the circumbinary disc. This also explains why the PBE code slightly under-estimates the binary’s mass ratio. However, it is pleasing to see that even neglecting the interaction between the binary and the circumbinary disc, the PBE code still predicts the mass of the circumbinary disc to within a factor of 2 of that given by the SPH code during the entire evolution.
The Evolution of Accreting Protobinary Systems
----------------------------------------------
As we have seen, the PBE code gives a relatively accurate description of the evolution of an accreting protobinary, but does so $\sim 10^6$ times faster than a full hydrodynamic SPH calculation. This allows us to perform many calculations to study how the evolution of a binary as it accretes to its final mass depends on its initial mass ratio and on the properties of the molecular cloud core from which it formed.
As examples, we give the evolution of ‘seed’ binaries that form from two types of molecular cloud core. A greater range of molecular cloud cores is considered in [@Bate2000]. Figure \[bate.uniform\] presents the evolution of binaries formed from molecular cloud cores which had uniform-density and were in solid-body rotation before they began to collapse dynamically. In Figure \[bate.1r\], the molecular cloud cores had radial density profiles of $\rho \propto 1/r$ with solid-body rotation, initially. Evolutionary curves are provided for ‘seed’ binaries with initial mass ratios ranging from $q=0.1-1.0$.
In all cases, the long-term evolution is towards a mass ratio of unity, since the material that falls in later has higher specific angular momentum relative to that of the binary. Thus, the more the binary accretes relative to its initial mass, the stronger the tendency is for the mass ratio to be driven to unity. Similarly, the more the binary accretes relative to its initial mass, the more likely it is to be surrounded by a circumbinary disc. Note that, in the previous sections, we found that when a massive circumbinary disc is formed, the PBE code tends to over-estimate the separation, and under-estimate the mass of the circumbinary disc and the binary’s mass ratio. Thus, if anything, the evolutionary curves in Figures \[bate.uniform\] and \[bate.1r\] tend to under-estimate the binary’s mass ratio and the mass of the circumbinary disc.
The Properties of Binary Stars
------------------------------
The aim of developing the PBE code was to make it possible to predict some of the properties of binary stars and, by comparing these to the observed properties of binary systems, to constrain the initial conditions for binary star formation.
In order to obtain predictions about the properties of binaries we note that, generally, the initial mass of a ‘seed’ binary is smaller for those binaries with smaller separations. This relationship between a ‘seed’ binary’s mass and its separation is observed from fragmentation calculations [@Boss86; @BonBat94b] and is easily understood from a Jeans-mass argument [@Bate2000]. In order for fragmentation to occur, the Jeans length at the time of fragmentation must be less than or approximately equal to the separation of the binary which is formed. However, for a constant temperature, the Jeans mass depends linearly on the Jeans length. Thus, the smaller the separation of the ‘seed’ binary, the smaller its initial mass. Generally, ‘seed’ binaries with separations $\simless 10$ AU are expected to have masses $\approx 0.01 {\rm M}_{\odot}$, while for larger separations, the ‘seed’ mass is expected to increase approximately linearly (i.e. ‘seed’ binaries with separations of 100-1000 AU should have initial masses of $\approx 0.1-1.0 {\rm M}_{\odot}$).
This dependence of the initial mass on the separation means that to form binaries with the same final total mass, the closer systems need to accrete more material, relative to their initial mass. Therefore, from the evolutionary curves of Figures \[bate.uniform\] and \[bate.1r\], closer systems are more likely to have equal-mass components than wider systems.
This prediction is supported by surveys of main-sequence G-dwarf stellar systems. Duquennoy & Mayor [@DuqMay91] found that the mass-ratio distribution, averaged over binaries with all separations, increases toward small mass ratios. However, there is mounting evidence that the mass-ratio distributions differ between short and long-period systems with the distribution for close binary systems ($P <
3000$ days; $a \simless 5$ AU) consistent with a uniform distribution [@Mazehetal92; @HalMayUdr98]. Thus, relative to wide systems, the close systems are biased toward mass ratios of unity.
The fraction by which the mass of a ‘seed’ binary must be increased in order for its mass ratio to approach unity depends on the conditions in the molecular cloud core. Generally, the less centrally-condensed a core is, the easier it is to form a binary system with a low mass ratio (c.f. Figures \[bate.uniform\] and \[bate.1r\]). We can use this dependence of the evolutionary curves on the type of molecular cloud core to attempt to constrain the initial conditions for binary star formation.
Duquennoy & Mayor [@DuqMay91] found that binaries containing G-dwarfs with separations $\simgreat 30$ AU generally have unequal masses (typically $q\approx 0.3$). Such binaries are likely to have accreted from a few to ten times there initial mass. For uniform-density cores (Figure \[bate.uniform\]), the observed mass-ratio distribution can easily be obtained. Cores with $\rho \propto 1/r$ result in higher mass ratios than uniform-density cores, but it is still possible to envisage a spectrum of ‘seed’ mass ratios which gives a final mass-ratio distribution which is consistent with the observations of wide binaries.
However, close binaries ($\simless 5$ AU) have initial masses of $\approx 0.01 {\rm M}_{\odot}$. Thus, they are expected to have to accrete up to 100 times their initial mass from the infalling gaseous envelope before systems with G-dwarf primaries are obtained, yet the observed mass-ratio distribution is approximately flat (i.e. approximately 1/2 the binaries have $q<0.5$). It is effectively impossible for cores in solid-body rotation to produce such a mass ratio distribution if they are significantly centrally-condensed (Figure \[bate.1r\]). Even with uniform-density cores most of the ‘seed’ binaries would need to have mass ratios $q<0.1$, which is unlikely.
However, the PBE code only evolves circular binaries whereas most binaries have significant eccentricity [@DuqMay91]. For the same semi-major axis, eccentric binaries have less angular momentum than circular binaries meaning that the clouds from which they formed may be rotating more slowly and, thus, the gas in the envelope may have less angular momentum. This would result in slower evolution toward equal masses for eccentric binaries. Taking the effects of eccentricity into account, it is quite possible that the observed binary mass ratios could be produced by the collapse of molecular cloud cores with radial density profiles less centrally-condensed than $\rho
\propto 1/r$. However, even accounting for eccentric binaries, it seems virtually impossible that the observed G-dwarf binary systems could have been formed from molecular cloud cores with density profiles that were more centrally-condensed than $\rho \propto 1/r$.
The above conclusion that closer binaries should have mass ratios that are biased toward unity compared to wider systems with the same total mass is just one of many predictions that may be derived using the PBE code (see [@Bate2000]). Others include: closer binaries are more likely to have circumbinary discs than wider binaries; brown dwarf companions to solar-type stars should be very rare at separations $\simless 5$ AU, but their frequency should increase at larger separations.
Conclusions
===========
The Lagrangian nature of SPH and its inherent ability to provide finer spatial resolution in regions of higher density make it a powerful tool which is ideally suited for studying star formation.
Recent advances in the study of binary star formation that have been made using SPH include: the realisation that it is essential that the Jeans mass is always resolved in numerical studies of self-gravitating gas; the ability to perform three-dimensional hydrodynamic calculations which follow the collapse of a molecular cloud core to stellar densities; the study of the effects of accretion on a protobinary systems and the development of a code which enables the evolution of accreting binary systems to be followed $\sim 10^6$ times faster than a full hydrodynamic calculation. The latter of these developments has resulted in the first firm predictions of the properties of binary stars for one particular model of binary star formation.
Artymowicz P., 1983, Acta Astronomica, 33, 223
Artymowicz P., Lubow S. H., 1994, ApJ, 421, 651
Bate M. R., 1997, MNRAS, 285, 16
Bate M. R., 1998, ApJ, 508, L95
Bate M. R., 2000, MNRAS, submitted
Bate M. R., Bonnell I. A., 1997, MNRAS, 285, 33
Bate M. R., Bonnell I. A., Price N. M., 1995, MNRAS, 277, 362
Bate M. R., Burkert A., 1997, MNRAS, 288, 1060
Benz W., 1990, in Buchler J. R., ed., The Numerical Modeling of Nonlinear Stellar Pulsations: Problems and Prospects. Kluwer, Dordrecht, p. 269
Bonnell I. A., Bastien P., 1991, ApJ, 374, 610
Bonnell I. A., Bate M. R., 1994, MNRAS, 271, 999
Boss A. P., 1986, ApJS, 62, 519
Boss A. P., 1987, ApJ, 319, 149
Boss A. P., 1998, ApJ, 501, 77
Burkert A., Bodenheimer P., 1993, MNRAS, 264, 798
Chapman S., Pongracic H., Disney M., Nelson A., Turner J., Whitworth A., 1992, Nature, 359, 207
Durisen R. H., Gingold R.A., Tohline J.E., Boss A.P., 1986, ApJ, 305, 281
Duquennoy A., Mayor M., 1991, A&A, 248, 485
Evrard A. E., 1988, MNRAS, 235, 911
Lucy L., 1977, AJ, 82, 1013
Gingold R. A., Monaghan J. J., 1977, MNRAS, 181, 375
Gingold R. A., Monaghan J. J., 1978, MNRAS, 184, 481
Gingold R. A., Monaghan J. J., 1981, MNRAS, 197, 461
Halbwachs J. L., Mayor M., Udry S., 1998, in Rebolo R., Martin E. L., Zapatero Osorio M.R., eds., Brown Dwarfs and Extrasolar Planets (ASP Conf. Ser. 134). Brigham Young University, Provo, p. 308
Heller, C. H., 1993, ApJ, 408, 337
Hernquist L., Katz N., 1989, ApJS, 70, 419
Imamura J. N., Durisen R. H., Pickett B. K., 1999, ApJ, in press
Larwood J. D., Nelson R. P., Papaloizou J. C. B., Terquem C., 1996, MNRAS, 282, 597
Lattanzio J. C., Monaghan J. J., Pongracic H., Schwarz M. P., 1985, MNRAS, 215, 125
Mazeh T., Goldberg D., Duquennoy A., Mayor M., 1992, ApJ, 401, 265
Miyama S. M., Hayashi C., Narita S., 1984, ApJ, 279, 621
Monaghan J. J., 1992, ARA&A, 30, 543
Nelson A. F., Benz W., Adams F. C., Arnett D., ApJ, 502, 342
Nelson R., Papaloizou J. C., 1993, MNRAS, 265, 905
Sigalotti L. Di G., Klapp J., A&A, 319, 547
Tohline, J. E., 1982, Fund. Cos. Phys., 8, 1
Truelove J. K., Klein R. I., McKee C. F., Holliman J. H. II, Howell L. H., Greenough J. A., 1997, ApJ, 489, L179
Truelove J. K., Klein R. I., McKee C. F., Holliman J. H. II, Howell L. H. Greenough J. A., & Woods D. T., 1998, ApJ, 495, 821
Vanhala, H. A. T., Cameron A. G. W., 1998, ApJ, 508, 291
Whitworth A. P., 1998, MNRAS, 296, 442
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we study a security problem of protecting secrets with multiple protections and minimum costs. The target system is modeled as a discrete-event system (DES) in which a few states are secrets, and there are multiple subsets of protectable events with different cost levels. We formulate the problem as to ensure that every string that reaches a secret state (from the initial state) contains a specified number of protectable events and the highest cost level of these events is minimum. We first provide a necessary and sufficient condition under which this security problem is solvable, and then propose an algorithm to solve the problem based on the supervisory control theory of DES. The resulting solution is a protection policy which specifies at each state which events to protect and the highest cost level of protecting these events is minimum. Finally, we demonstrate the effectiveness of our solution with a network security example.'
author:
- 'Shoma Matsui and Kai Cai [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'reference.bib'
title: '**Secret Securing with Multiple Protections and Minimum Costs**'
---
[^1]: S. Matsui and K. Cai are with the Department of Electrical and Information Engineering at Osaka City University, Osaka, Japan. S. Matsui: `[email protected]`, K. Cai: `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the first scientific images obtained with a deformable secondary mirror adaptive optics system. We utilized the 6.5m MMT AO system to produce high-resolution (FWHM=0.07$\arcsec$) near infrared (1.6$\mu m$) images of the young ($\sim 1$ Myr) Orion Trapezium $\theta^{1}$ Ori cluster members. A combination of high spatial resolution and high signal to noise allowed the positions of these stars to be measured to within $\sim 0.003\arcsec$ accuracies. We also present slightly lower resolution (FWHM$\sim$0.085$\arcsec$) images from Gemini with the Hokupa’a AO system as well. Including previous speckle data [@wei99], we analyze a six year baseline of high-resolution observations of this cluster. Over this baseline we are sensitive to relative proper motions of only $\sim0.002\arcsec$/yr (4.2 km/s at 450 pc). At such sensitivities we detect orbital motion in the very tight $\theta^{1}$ Ori $B_{2}B_{3}$ (52 AU separation) and $\theta^{1}$ Ori $A_{1}A_{2}$ (94 AU separation) systems. The relative velocity in the $\theta^{1}$ Ori $B_{2}B_{3}$ system is $4.2\pm2.1$ km/s. We observe $16.5\pm5.7$ km/s of relative motion in the $\theta^{1}$ Ori $A_{1}A_{2}$ system. These velocities are consistent with those independently observed by [@sch03] with speckle interferometry, giving us confidence that these very small ($\sim 0.002\arcsec$/yr) orbital motions are real. All five members of the $\theta^{1}$ Ori $B$ system appear likely gravitationally bound ($B_{2}B_{3}$ is moving at $\sim1.4$ km/s in the plane of the sky w.r.t. $B_{1}$ where $V_{esc} \sim 6$ km/s for the B group). The very lowest mass member of the $\theta^{1}$ Ori $B$ system ($B_{4}$) has $K^\prime \sim 11.66$ and an estimated mass of $\sim 0.2 M_{\sun}$. There was very little motion ($4\pm 15$ km/s) detected of $B_4$ w.r.t $B_1$ or $B_2$, hence $B_4$ is possibly part of the $\theta^{1}$ Ori $B$ group. We suspect that if this very low mass member is physically associated it most likely is in an unstable (non-hierarchical) orbital position and will soon be ejected from the group. The $\theta^{1}$ Ori $B$ system appears to be a good example of a star formation “mini-cluster” which may eject the lowest mass members of the cluster in the near future. This “ejection” process could play a major role in the formation of low mass stars and brown dwarfs.'
author:
- 'Laird M. Close$^1$, Francois Wildi$^1$, Michael Lloyd-Hart$^1$, Guido Brusa$^{12}$, Don Fisher$^1$, Doug Miller$^1$, Armando Riccardi$^2$, Piero Salinari$^2$, Donald W. McCarthy$^1$, Roger Angel$^1$, Rich Allen$^1$, H.M. Martin$^1$, Richard G. Sosa$^1$, Manny Montoya$^1$, Matt Rademacher$^1$, Mario Rascon$^1$, Dylan Curley$^1$, Nick Siegler$^1$, Wolfgang J. Duschl$^3$'
title: 'High Resolution Images of Orbital Motion in the Trapezium Cluster: First Scientific Results from the MMT Deformable Secondary Mirror Adaptive Optics System'
---
Introduction
============
The detailed formation of stars is still a poorly understood process. In particular, the formation of the lowest mass stars and brown dwarfs is uncertain. Detailed 3D simulations of star formation by [@bat02] suggest that stellar embryos form into “mini-clusters” which dynamically decay “ejecting” the lowest mass members. Such theories can explain why there are far more field brown dwarfs (BD) compared to BD companions of solar type stars [@mcc03] or early M stars [@hin02]. Moreover, these theories which invoke some sort of dynamical decay [@dur01] or ejection [@rep02] suggest that there should be no wide ($>20$ AU) very low mass (VLM; $M_{tot}<0.185
M_{\sun}$) binary systems observed. Indeed, the AO surveys of [@clo03a] and the HST surveys of [@rei01a; @bur03; @bou03; @giz03] have not discovered any wide ($>16$ AU) VLM systems of the 34 systems known to date. As well, the dynamical biasing towards the ejection of the lowest mass members naturally suggests that the frequency of VLM binaries should be much less ($\la 5\%$ for $M_{tot}\sim 0.16
M_{\sun}$) than for more massive binaries ($\sim 60\%$ for $M_{tot}
\sim 1 M_{\sun}$). Indeed, observations suggest that the binarity of VLM systems with $M_{tot} \la 0.185 M_{\sun}$ is $10-15\%$ [@clo03a; @bur03] which, although higher than predicted is still lower than that of the $\sim 60\%$ of G star binaries [@duq91].
Despite the success of these decay or ejection scenarios in predicting the observed properties of binary stars, it is still not clear that “mini-clusters” even exist in the early stages of star formation. To better understand whether such “mini-clusters” do exist we have examined the closest major OB star formation cluster for signs of such mini-clusters. Here we focus on the $\theta^1$ Ori stars in the Trapezium cluster. Trying to determine if some of the tight star groups in the Trapezium cluster are gravitationally bound is a first step to determining if bound “mini-clusters” exist. In particular, we will examine the case of the $\theta ^1$ Ori B and A groups.
The Trapezium OB stars ($\theta^1$ Ori A, B, C, D, and E) consists of the most massive OB stars located at the center of the Orion Nebula star formation cluster (for a review see [@gen89]). Due to the luminous nature of these stars they have been the target of several high-resolution imaging studies. Utilizing only tip-tilt compensation [@mcc94] mapped the region at $K^\prime$ from the 3.5-m Calar Alto telescope. They noted that $\theta^1$ Ori B was really composed of 2 components ($B_1$ & $B_2$) about $\sim 1\arcsec$ apart. Higher $\sim
0.15\arcsec$ resolutions were obtained from the same telescope by [@pet98] with speckle holographic observations. At these higher resolutions [@pet98] discovered that $\theta^1$ Ori $B_2$ was really itself a $0.1\arcsec$ system ($B_2$ & $B_3$) and that $\theta^1$ Ori A was really a $\sim0.2\arcsec$ binary ($A_1$ & $A_2$). A large AO survey of the inner 6 square arcminutes was carried out by [@sim99], who discovered a very faint (100 times fainter than $B_1$) object ($B_4$) located just $0.6\arcsec$ between $B_1$ and $B_2$. Moreover, a spectroscopic survey [@abt91] showed that $B_1$ was really an eclipsing spectroscopic binary ($B_1$ & $B_5$; sep. 0.13 AU; period 6.47 days). As well, $\theta^1$ Ori $A_1$ was also found to be a spectroscopic binary ($A_1$ & $A_3$; sep. 1 AU; [@bos89] ). [@wei99] carried out bispectrum speckle interferometric observations at the larger Russian SAO 6-m telescope (2 runs in 1997 and 1998). These observations showed $\theta^1$ Ori C was a very tight 0.033$\arcsec$ binary. These observations also provided the first set of accurate relative positions for these stars. [@sch03] has continued to monitor this cluster of stars and has independently detected an orbital motion (of $\Delta PA \sim
6^{\circ}$ for $\theta^{1}$ Ori $A_2$ around $A_1$ and a $\Delta PA$ of $\sim 8^{\circ}$ for $\theta^{1}$ Ori $B_3$ around $B_2$ over a 5.5 yr baseline). They conclude that this is real orbital motion. We present additional recent AO observations of these binaries as an independent check to confirm that these motions are indeed real.
We first utilized the Gemini telescope (with the Hokupa’a AO system) and then observed $\theta^1$ Ori B during commissioning of the world’s first secondary deformable mirror at the 6.5-m MMT telescope. In this paper we outline how the observations were carried out, and how the stellar positions were measured. We fit the observed positions to calculate velocities (or upper limits) for the $\theta^1$ Ori B & A stars. In agreement with [@sch03], we find that there is good evidence that the $\theta^1$ Ori B group may be a bound “mini-cluster” and that the $\theta^1$ Ori A group is also likely gravitationally bound.
OBSERVATIONS
============
We have utilized the University of Arizona adaptive secondary AO system to obtain the most recent high resolution images of the young stars in the Trapezium cluster (the $\theta^{1}$ Ori group).
The World’s First Adaptive Secondary AO System Scientific Results
-----------------------------------------------------------------
The 6.5 m MMT telescope has a unique adaptive optics system. To reduce the aberrations caused by atmospheric turbulence all AO systems have a deformable mirror which is updated in shape at $\sim 500$ Hz. Until now all adaptive optics systems have located this deformable mirror (DM) at a re-imaged pupil (effectively a compressed image of the primary mirror). To reimage the pupil onto a DM typically requires 6-8 warm additional optical surfaces which significantly increases the thermal background and decreases the optical throughput of the system [@llo00]. However, the MMT utilizes a completely new type of DM. This DM is both the secondary mirror of the telescope and the DM of the AO system. In this manner there are no additional optics required in front of the science camera. Hence the emissivity is lower and the possibility of thermal IR AO imaging [@clo03b; @bil03] becomes a reality.
The DM consists of 336 voice coil actuators that push on 336 small magnets glued to the backsurface of a thin (2.0 mm thick) 642 mm aspheric ULE glass “shell” (for a detailed review of the secondary mirror see [@bru03a; @bru03b]). We have complete positional control of the surface of this reflective shell by use of a capacitive sensor feedback loop. This positional feedback loop allows one to position an actuator of the shell to within 4 nm rms (total surface errors amount to only 40 nm rms over the whole secondary). The AO system samples at 550 Hz using 108 active subapertures. For a detailed review of the MMT AO system see [@wil03; @wil03b] and references within.
MMT AO Observations
-------------------
During our second engineering run we observed the $\theta^{1}$ Ori B group on the night of Jan 20, 2003 (UT). The AO system corrected the lowest 52 system modes and was updated at 550 Hz. The closed loop bandwidth was estimated at 30 Hz 0 dB. Without AO correction our images had FWHM=$0.6\arcsec$, after AO correction our 23 second images had improved to FWHM=$0.070\arcsec$ (close to the diffraction limit of $0.056\arcsec$ in the H-band). A detailed analysis suggested that during our engineering run a 40 Hz vibration in the MMT telescope increased our FWHM by $\sim 0.015\arcsec$ and decreased our Strehl by a factor of two. We are in the process of identifying and decreasing the effect of this 40 Hz vibration. In any case, as Figure \[fig2\] clearly shows, there is a large improvement in image quality (the Strehl increases by 20 times) with the adaptive secondary AO system.
### The Indigo Near-IR Video Camera
Since these observations were carried out during the engineering run we utilized a commercially available 320x256 InGaAs 0.9-1.68 $\mu m$ “Merlin-NIR” video camera. Although this commercial camera (produced by the Indigo company) is not nearly as sensitive as our facility AO camera (AIRES; [@mcc98]) it still provides excellent dynamical information about the performance of the AO system on bright objects (it will be replaced by the ARIES camera in the fall of 2003). Here we use it as a simple NIR (H band) science camera.
The Indigo camera was fed by a relay lens that converted the f/15 AO corrected beam to a f/39 beam yielding $0.0242\pm0.0020\arcsec$ per $30\mu m$ pixel (providing a $7.7\times 6.2\arcsec$ FOV). Astrometric standards ADS 8939 and ADS 7158 were observed to calibrate this platescale and error (see Figures \[fig2a\] & \[fig2b\]). It was found that the direction of north was slightly ($0.113^{\circ}$) east of Indigo’s Y axis (when the parallactic angle was zero (transit) and one is looking towards the south). During this commissioning run we did not observe with the MMT Cassegrain derotator tracking field rotation, hence all images must be rotated by the appropriate parallactic angle (plus $0.113^{\circ}$) to have north up and east to the left on the Indigo camera.
The camera was mounted under a high optical quality dichroic which sent the visible light (0.5-1 $\mu m$) to the 108 subaperture shack-Hartmann wavefront sensor (WFS). The infrared light ($\lambda >
1 \mu m$) was transmitted to the Indigo camera. The camera had a standard H band filter ($1.6\mu m$) mounted 3 inches from focus in a light-tight barrel.
To maximize the sensitivity of the Indigo camera we carried out a standard “2-point” calibration on a both a dark (cold) flat field source and on a bright (hot) source to scale the automatic gain control/dynamic range of the camera’s electronics. This appeared to yield images that were auto flat fielded to a few percent in accuracy when the counts were between the linear range defined by the dark and bright calibration flats. The camera was remotely controlled via a serial port. Digital (16 bit) data were streamed to the control PC’s hard drive. Data could be acquired as fast as 50 frames per second (although data in this paper was acquired at 15 frames/sec to sample longer periods on the sky). Integration times can range from 1-16000 $\mu s$. The lack of a longer integration time (since the camera is primarily intended for commercial high-background, high-bandwidth applications) leads to most sources being read-noise limited. However, we found that point sources of $H\sim11$ could be detected in 3 s of total exposure (200 16 ms frames) with AO correction at the MMT. Although insensitive by most astronomical standards the Indigo camera is able to capture temporal events of durations as short as 1 $\mu s$. In this paper we will focus on the ability of the Indigo camera to produce high resolution ($0.07\arcsec$) images of the $\theta^{1}$ Ori B group.
### Reducing the Indigo MMT AO Data
For the $\theta^{1}$ Ori B group we obtained 7 series of 200x16 ms data cubes with the Indigo camera. The data from each cube was simply averaged together to produce 7 individual 3.2 second exposures. A similarly reduced cube of “sky” images was subtracted from each data set. These 7 sky-subtracted exposures were then rotated (in IRAF) by the current parallactic angle (plus the $0.113^{\circ}$ offset) so north was aligned with the Y axis, and east is the negative X axis. Then each of the 7 images were cross-correlated and aligned with a cubic spline interpolator. Then the final stack of images were median combined to produce the final image. The final image is displayed in Figure \[fig3\].
Hokupa’a/Gemini Images of the Trapezium
---------------------------------------
In addition to our excellent MMT images of the $\theta^{1}$ Ori B group we also have an epoch of $K^{\prime}$ images of the central $30\arcsec$ of the cluster. These Hokupa’a/Gemini [@gra98; @clo98a] AO images were taken September 19, 2001. We acquired a series of 10 short (1 s) images and dithered the telescope in a 10x10$\arcsec$ box while AO guiding off $\theta^{1}$ Ori B itself (as in the case of the MMT AO observations). We utilized the QUIRC IR camera [@hod96] with a calibrated platescale of $0.0199\pm0.0002\arcsec$/pix [@pot02a].
### Reducing the Gemini data
We have developed an AO data reduction pipeline in the IRAF language which maximizes sensitivity and image resolution. This pipeline is standard IR AO data reduction and is described in detail in [@clo02a; @clo02b].
The pipeline cross-correlates and aligns each image, then rotates each image so north is up (to an accuracy of $\pm0.3$ degrees) and east is to the left, then median combines the data with an average sigma clip rejection at the $\pm2.5 \sigma$ level. By use of a cubic-spline interpolator the script preserves image resolution to the $<0.02$ pixel level. Next the custom IRAF script produces two final output images, one that combines all the images taken (see Figure \[fig4\]) and another where only the sharpest 50% of the images are combined (this high-Strehl image was very similar to that shown in Figure \[fig4\], just a bit noisier – and so was not further analyzed).
The final image (see Figures \[fig4\] and \[fig5\]) has FWHM=$0.085\arcsec$ which is just slightly worse than the MMT data. Even though Gemini is a larger telescope (8.2-m), Hokupa’a’s fitting error (36 elements over 50 meters$^2$) is worse than that of the MMT (52 modes over 33 meters$^2$), hence higher resolution images can result from the smaller of the two telescopes (Gemini has a diffraction-limit of $0.056\arcsec$ at $K^{\prime}$ similar to that of the MMT at H). However, Hokupa’a’s curvature WFS could guide on much fainter (R$\sim17$) guide stars [@clo02a; @clo02b; @sie02a].
Reductions
==========
In Table \[tbl-1\] we present the analysis of our MMT and Gemini images in Figures \[fig3\] and \[fig4\]. The photometry was based on DAOPHOT’s PSF fitting photometry task ALLSTARS [@ste87]. The PSF used was $\theta^{1}$ Ori $B_{1}$ itself. Since all the members of the $\theta^{1}$ Ori $B$ group are located within $1\arcsec$ of $\theta^{1}$ Ori $B_{1}$ the PSF fit is excellent (there is no detectable change in PSF morphology due to anisoplanatic effects inside the $\theta^1$ Ori B group [@dio00]).
Since the PSF model was so accurate and the data had such high signal to noise (and high resolution) it was possible for DAOPHOT to measure relative positions to within $0.003\arcsec$. We estimate this error based on the scatter of the $\theta^{1}$ Ori $B_{1}B_{2}$ separation (which should be very close to a constant since the $B_{1}B_{2}$ system has an orbital period of $\sim 2000$ yr). The lack of any motion between $B_1$ and $B_2$ is also confirmed by [@sch03]. Our data is summarized in Table \[tbl-1\]. Linear (weighted) fits to the data in Table \[tbl-1\] (Figures \[fig6\] to \[fig11\]) yield the velocities shown in Table \[tbl-1\]. The overall error in the relative proper motions observed is $\sim 0.002\arcsec$/yr in proper motion ($\sim 4$ km/s).
ANALYSIS
========
With these accuracies it is now possible to determine whether these stars in the $\theta^{1}$ Ori $B$ group are bound together, or merely chance projections in this very crowded region. As can be seen from Table \[tbl-1\] and Figures \[fig6\] – \[fig11\] there is very little relative motion between any of the members of the $\theta^{1}$ Ori $B$ group. Therefore it is possible that the group is physically bound together.
If we adopt the masses of each star from the [@sie97; @ber96] tracks fit by [@wei99] we find masses of: $B_1\sim7M_{\sun} $; $B_2\sim3M_{\sun} $; $B_3\sim2.5M_{\sun} $; $B_4\sim0.2M_{\sun} $; $B_5\sim 7 M_{\sun}$; $A_1\sim20M_{\sun} $; $A_2\sim 4M_{\sun} $; and $A_3\sim 2.6 M_{\sun}$. Based on these masses (which are similar to those adopted by [@sch03]) we can comment on whether the observed motions are less than the escape velocities expected for simple face-on circular orbits.
Our combination of high spatial resolution and high signal to noise yields an error in the proper motions of only $\sim 0.002\arcsec$/yr according to the scatter in the $B_1B_2$ and $B_1B_3$ systems (see Table \[tbl-1\]). We have observed orbital motion in the very tight $\theta^{1}$ Ori $B_{2}B_{3}$ (see Figure \[fig9\]) and $\theta^{1}$ Ori $A_{1}A_{2}$ (see Figure \[fig11\]) systems, with 52 and 94 AU separations; respectively.
Is the $\theta^1$ Ori $B_2B_3$ System Physical?
-----------------------------------------------
The relative velocity in the $\theta^{1}$ Ori $B_{2}B_{3}$ system (in the plane of the sky) is $\sim 4.2\pm2.1$ km/s (mainly in the azimuthal direction; see Figure \[fig9\]). This is a reasonable $V_{tan}$ since an orbital velocity of $\sim 6.7$ km/s is expected from a face-on circular orbit from a $\sim5.5 M_{\sun}$ binary system like $\theta^{1}$ Ori $B_{2}B_{3}$ with a 52 AU projected separation. It is worth noting that this velocity is also greater than the $\sim 3$ km/s [@hil98] dispersion velocity of the cluster. Hence it is most likely that these two $K^\prime =7.6$ and $K^\prime
=8.6$ stars (separated by just 0.116$\arcsec$) are indeed in orbit around each other. Moreover, there are only 10 stars known to have $K^\prime < 8.6$ in the inner $30\times 30\arcsec$ (see Figure \[fig5\]), we can estimate that the chances of finding two bright ($K^\prime < 8.6$) stars within $0.116\arcsec$ is a small $<10^{-4}$ probability.
Our observed velocity of $0.93\pm0.49^{\circ}$/yr is consistent (in both direction and magnitude) with the $1.4^{\circ}$/yr observed by [@sch03]. This suggests that the AO and speckle datasets are both detecting real motion. Moreover, since this motion is primarily azimuthal strongly suggests an orbital arc of $B_3$ orbiting $B_2$.
Is the $\theta^1$ Ori $A_1A_2$ System Physical?
-----------------------------------------------
We observe $\sim 16.5\pm5.7$ km/s of relative motion in the $\theta^{1}$ Ori $A_{1}A_{2}$ system (mainly in the azimuthal direction; see Figure \[fig11\]). This is higher than the average dispersion velocity of $\sim 3$ km/s but still close to an estimated periastron velocity of the $\sim 20 M_{\sun}$ $A_{1}A_{2}$ system (projected separation of 94 AU). Hence it is highly likely that these two $K^\prime =6.0$ and $K^\prime =7.6$ stars (separated by just 0.21$\arcsec$) are indeed in orbit around each other. In addition, there are only 8 stars known to have $K^\prime < 7.6$ in the inner $30\times 30\arcsec$ (see Figure \[fig5\]), we can estimate that the chances of finding two bright ($K^\prime < 8.6$) stars within $0.21\arcsec$ is a small $<4\times 10^{-4}$ probability.
Our observed velocity of $16.5\pm5.7$ km/s is consistent (in both direction and magnitude) with the $\sim 10.3$ km/s observed by [@sch03]. This again suggests that the AO and speckle datasets are both detecting real motion of $A_2$ orbiting $A_1$.
Is the $\theta^1$ Ori B Group Stable?
-------------------------------------
The pair $B_1B_5$ is moving at $\sim 1.4 \pm 4.4$km/s in the plane of the sky w.r.t. to the pair $B_2B_3$ where the escape velocity $V_{\mathrm esc} \sim 6$km/s for this system. Hence these pairs are very likely gravitationally bound together. However, radial velocity measurements will be required to be absolutely sure that these 2 pairs are bound together.
### Is the Orbit of $\theta^1$ Ori $B_4$ Stable?
The situation is somewhat different for the faintest component of the group, $B_4$. It has $K = 11.66$ mag which according to Hillenbrand & Carpenter (2000) suggests a mass of only $\sim 0.2\,{\mathrm
M}_\odot$. Since there are only 20 stars known to have $K < 11.66$ in the inner $30 \time 30^{\prime\prime}$ (see Fig. 6), we can estimate that the chances of finding a $K < 11.66$ star within $0.6^{\prime\prime}$ of $B_1$ is a small $< 8 \times 10^{-3}$ probability. Our two AO measurements (and the one speckle detection of [@sch03]) did not detect a significant velocity of $B_4$ w.r.t. $B_1$ ($4\pm15$ km/s; see Figures \[figB4\_sep\] & \[figB4\_pa\]). Together with the escape velocity of $\sim 6$ km/s, this points towards $B_4$ being also a gravitationally bound member of the $\theta^1$ Ori B group.
On the other hand, its mass and its location w.r.t. to the other four groups members makes it highly unlikely that $B_4$ is on a stable orbit within the group. To reconcile these conflicting observations, one may think of [*(a)*]{} $B_4$’s projected distances from the other B group members being considerably smaller than the true distance thus making a stable orbit much more likely, or [*(b)*]{} $B_4$’s current motion pointing almost exactly along our line of sight, allowing for a higher true velocity, or [*(c)*]{} $B_4$ being a chance projection of an object not related to the other four members of the B group. Without additional astrometric data, we cannot yet decide which of these three possibilities is the most likely.
### Is the orbit of $B_3$ around $B_2$ and of $B_5$ around $B_1$ stable longterm?
$B_{1}B_{5}$, and $B_{2}B_{3}$ are two binaries with projected separations of 0.13 AU ($B_1B_5$) and 52 AU ($B_2B_3$); respectively. The two pairs are separated by a projected distance of 415 AU. The distance $D_{B_1B_5} \sim 3 \times 10^{-4} \times D_{B_1B_5 B_2B_3}$ and thus the $B_1B_5$ system is stable. Much more interesting is the case of $B_2B_3$. Their projected distance is not very small compared to their projected distance (D) from the $B_1B_5$ pair:$ D_{B_2B_3} \sim 0.12 \times D_{B_1B_5 B_2B_3}$. Thus the stability of the $B_2B_3$ orbit needs a more detailed analysis since it is possible that $B_3$ may be ejected in the future.
[@egg95] have given an empirical criterion for the long-term stability of the orbits of hierarchical triple systems, based on the results of their extensive model calculations [@kis94; @kis94a; @egg95]. Their analytic stability criterion is good to about $\pm 20\%$, and is meant to indicate stability for another $10^2$ orbits. Given the uncertainties of the masses of the members of the B group, this accuracy is sufficient for our present discussion.
The orbital period of the two binaries w.r.t. each other is $P_{(15)/(23)} \sim 1920\,$yrs, while the orbital period of $B_3$ w.r.t $B_2$ amounts to $P_{2/3} \sim 160\,$yrs. For the calculation of both periods, we have assumed the masses as given above, and circular orbits in the plane of the sky. This leads to a period ratio $X =
P_{(15)/(23)}/P_{2/3} \sim 12$. Eggelton & Kiseleva’s stability criterion requires $X \ge X_{\mathrm crit} = 10.08$ for the masses in the B group. This means that within the accuracy limits of our investigation, the binary $B_2B_3$ is just at the limit of stability. The stability criterion depends also on the orbits’ eccentricities. In our case, already mild eccentricities of the order of $e \sim 0.1$ (as can be expected to develop in hierarchical triple systems; see, e.g., Georgakarakos 2002), make the B group unstable. While we cannot decide yet whether the pair $B_2B_3$ orbit each other in a stable way, it is safe to say that that the “triple” $B_{1}B_{5}$, $B_2$, and $B_3$ is not a simple, stable hierarchical triple system.
The $\theta^1$ Ori B system seems to be a good example of a highly dynamic star formation “mini-cluster” which is possibly in the process of ejecting the lowest-mass member through dynamical decay [@dur01], and breaking up the gravitational binding of the widest of the close binaries (the $B_{2}B_{3}$ system). The “ejection” of the lowest-mass member of a formation “mini-cluster” could play a major role in the formation of low mass stars and brown dwarfs [@rei01a; @bat02; @dur01; @clo03a]. The breaking up of binaries, of course, modifies the binary fraction of main sequence stars considerably as well.
Future observations
===================
In our opinion it is most likely that these $\theta ^1 $ Ori A & B group stars are bound. We caution, however, that the motion of each of these stars could currently be fit equally well by linear motion (not orbital arcs). Future high resolution observations are required to see if these stars follow true orbital arcs around each other proving that they are interacting. In particular, future observations of the $\theta^1$ Ori $B_4$ positions would help reduce the scatter in the velocity data and indicate if it is indeed part of the $\theta^1$ Ori B group.
Future observations should also try to determine the radial velocities of these stars. Once radial velocities are known one can calculate unambiguously if these systems are bound. Such observations will require both very high spatial and spectral resolutions. This might be possible with such future instruments like the future AO fed ARIES instrument.
These MMT observations were possible due to the hard work of the entire Center for Astronomical Adaptive Optics (CAAO) staff at the University of Arizona. In particular, we would like to thank Tom McMahon, Kim Chapman, Doris Tucker, and Sherry Weber for their endless support of this project. We thank the anonymous referee for helpful comments that produced a better paper. The Indigo H band filter holder was installed by graduate student Melanie Freed. The adaptive secondary mirror is a joint project of University of Arizona and the Italian National Institute of Astrophysics - Arcetri Observatory. We would also like thank the whole MMT staff for their excellent support and flexibility during our commissioning run at the telescope.
The Hokupa’a AO observations were supported by the University of Hawaii AO group. (D. Potter, O. Guyon, & P. Baudoz). Support for Hokupa’a comes from the National Science Foundation. These results were based, in part, on observations obtained at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the Particle Physics and Astronomy Research Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), CNPq (Brazil) and CONICET (Argentina).
The secondary mirror development could not have been possible without the support of the Air Force Office of Scientific Research under grant AFOSR F49620-00-1-0294. LMC acknowledges support from NASA Origins grant NAG5-12086 and NSF SAA grant AST0206351.
Abt H.A., Wang R., Cardona O., 1991, , 367, 155 Bate, M.R., Bonnell, I.A., Bromm, V. 2002, , 332, L65 Bernasconi P.A., & Maeder A. 1996, A&A, 307, 829 Biller, B., et al. 2003, , submitted. Brusa, G., et al. 2003a, Proc. SPIE 4839, 691. Brusa, G., et al. 2003b, Proc. SPIE in prep. Burgasser, A. et al. 2003a, , in press Burrows, A., Hubbard, W. B., Lunine, J. I., Marley, M. S., Saumon, D. 2000, Protostars and Planets IV (Tucson: University of Arizona Press, eds Mannings, V., Boss, A.P., Russell, S. S.), p. 1339 Bouy, H., Brandner W., Martín, E., Delfosse, X., Allard, F., & Basri, G. 2003, , in press Bossi M., Gaspani A., Scardia M., Tadini M., 1989, A&A, 222, 117 Chabrier, G., Baraffe, I., Allard, F., & Hauschildt, P. 2000, , 542, 464 Close, L.M., Roddier, F.J., Roddier, C.A., Graves, J.E., Northcott, M.J., Potter, D. 1998, Proc. SPIE Vol. 3353, p. 406-416. Adaptive Optical System Technologies, D. Bonaccini, R.K. Tyson, Eds Close, L. M. 2000, Proc. SPIE Vol. 4007, p758-772. Adaptive Optical Systems Technology, P.L. Wizinowich, Ed. Close, L.M. et. al. 2002a, , 566, 1095. Close, L.M. et. al. 2002b, , 567, L53. Close, L.M. 2002, Proc. SPIE Vol. 4834-12 Research Prospects on Large 6.5-10m Telescopes. Aug 2002, Kona. in press Close, L.M., Siegler, N., Freed, M., Biller, B. 2003a , 587, 407 Close, L.M. et al. 2003b , submitted Diolaiti, E., Bendinelli, O. Bonaccini, D.; Close, L Currie, D. Parmeggiani, G. 2000, A&AS 147, 335 Durisen, R.H., Sterzik, M.F., & Pickett, B.K. 2001, A&A, 371, 952 Duquennoy, A., Mayor, M. 1991, , 248, 485 Eggelton P., Kiseleva L., 1995, , 455, 640 Fischer, D. A., Marcy, G. W. 1992, , 396, 178 Freed, M., Close, L.M., & Siegler, N. 2003, , 584, 453 Genzel R., Stutzki J., 1989, ARA&A 27, 41 Graves, J.E., Northcott, M.J., Roddier, F.J., Roddier, C.A., Close, L.M. 1988, Proc. SPIE Vol. 3353, p. 34-43. Adaptive Optical System Technologies, D. Bonaccini, R.K. Tyson, Eds. Gizis, J.E. et al. 2003, , in press Hillenbrand L.A., & Carpenter J. 2000 , 540, 236 Hillenbrand L.A., & Hartmann L.W. 1998 , 492, 540 Hinz J.L., McCarthy D.W., Simons, D.A., Henry T.J., Kirkpatrick J.D., McGuire P.C. 2002, , 123, 2027 Hodapp, K.-W., Hora, J. L., Hall, D. N. B., Cowie, L. L., Metzger, M., Irwin, E., Vural, K., Kozlowski, L. J., Cabelli, S. A., Chen, C. Y., Cooper, D. E., Bostrup, G. L., Bailey, R. B., Kleinhans, W. E. 1996, New Astronomy, 1, 177 Kiseleva L.G., Eggelton P.P., Anosova J.P., 1994, MNRAS 267, 161 Kiseleva L.G., Eggelton P.P., Orlov V.V., 1994, MNRAS 270, 936 Lloyd-Hart M. 2000, PASP 112, 264 McCaughrean M.J, & Stauffer J.R., 1994, , 108, 1382 McCarthy C. et al. IAU conf. “Brown Dwarfs” Kona in press McCarthy D.W. et al. 1998, Proc. SPIE 3354 750 McDonald, J. M., & Clarke, C. J. 1993, , 262, 800 Petr M.G., Du Foresto V., Beckwith S.V.W., Richichi A., McCaughrean M.J. 1998, , 500, 825 Potter, D. et al. 2002a , 567, 113 Reid, I. N., Gizis, J.E., Kirkpatrick, J.D., Koerner, D. W. 2001a, , 121, 489 Reid, I. N., Burgasser, A. J., Cruz, K. L., Kirkpatrick, J. D., Gizis, J. E. 2001b, , 121, 1710 Reipurth, B. & Clarke, C. 2001, , 122, 432 Schertl, D., Balega, Y.Y., Preibisch, Th., & Weigelt, G. 2003, A&A, 402, 267 Siegler, N., Close, L.M., Freed, M. 2002, SPIE Kona conf. proc., in press Siegler, N., Close, L.M., Mamajek, E., Freed, M. 2003, , in press Siess L., Forestini M., Dougados C., 1997, A&A 324, 556 Simon, M., Close, L.M., & Beck, T. 1999, , 117, 1375 Sterzik, M. F., & Durisen, R. H. 1998 A&A, 339, 95 Stetson, P. B. 1987, , 99, 191 Wainscoat R. J., & Cowie, L.L. 1992, , 103, 332. Weigelt G., Balega, Y., Preibisch T., Schertl D., Scholler M., Zinnecker H. 1999, A&A, 347, L15 Wildi F., Brusa G., Riccardi A., Lloyd-Hart M., Martin H.M., L.M. Close proc. SPIE 4839, 155 Wildi F. et al. 2003b JOSA, in prep.
[lllllllll]{}
$B_{1}B_{2}$&$2.30\pm0.15$&&$0.942\pm0.020\arcsec$&&$254.9\pm1.0$&&SAO&10/14/97\
&&$1.31\pm0.10$&$0.942\pm0.020\arcsec$&&$254.4\pm1.0$&&SAO&11/03/98\
&&$2.07\pm0.05$&$0.9388\pm0.0040\arcsec$&&$255.1\pm1.0$&&GEMINI&09/19/01\
&$2.24\pm0.05$&&$0.9375\pm0.0030\arcsec$&&$255.1\pm1.0$&&MMT&01/20/03\
&&&&-0.0006$\pm0.0019\arcsec$/yr&&0.07$\pm0.25^\circ$/yr&&\
&&&&&&&&\
$B_{2}B_{3}$&$1.00\pm0.11$&&$0.114\pm0.05\arcsec$&&$204.3\pm4.0$&&SAO&10/14/97\
&&$1.24\pm0.20$&$0.117\pm0.005\arcsec$&&$205.7\pm4.0$&&SAO&11/03/98\
&&$1.04\pm0.05$&$0.1166\pm0.0040\arcsec$&&$207.8\pm1.0$&&GEMINI&09/19/01\
&$0.85\pm0.05$&&$0.1182\pm0.0030\arcsec$&&$209.7\pm1.0$&&MMT&01/20/03\
&&&&$0.0006\pm0.0010\arcsec$/yr &&0.93$\pm0.49^\circ$/yr &&\
&&&&&&&&\
$B_{1}B_{4}$&&$5.05\pm0.8$&$0.609\pm0.008\arcsec$&&$298.0\pm2.0$&&SAO&02/07/01\
&&$5.01\pm0.10$&$0.6126\pm0.0040\arcsec$&&$298.2\pm1.0$&&GEMINI&09/19/01\
&$4.98\pm0.10$&&$0.6090\pm0.0050\arcsec$&&$298.4\pm1.0$&&MMT&01/20/03\
&&&&$-0.0017\pm0.0033\arcsec$/yr &&0.18$\pm0.95^\circ$/yr &&\
&&&&&&&&\
$A_{1}A_{2}$&$1.51\pm0.15$&$1.38\pm0.10$&$0.208\pm0.030\arcsec$&&$343.5\pm5.0$&&Calar Alto&11/15/94\
&&$1.51\pm0.05$&$0.2215\pm0.005\arcsec$&&$353.8\pm2.0$&&SAO&11/03/98\
&&$1.62\pm0.05$&$0.2051\pm0.0030\arcsec$&&$356.9\pm1.0$&&GEMINI&09/19/01\
&&&&$-0.0064\pm0.0027\arcsec$/yr &&2.13$\pm0.73^\circ$/yr &&\
![ A typical example of how the the Adaptive Optics (AO) system can make very sharp images. With AO “OFF” $\theta^1$ Ori B appears to be just 2 stars. With AO turned “ON” it is clearly a tight group of 4 visual stars. Note how with AO correction the peak intensity increases by 20 times and the resolution becomes ten times better.[]{data-label="fig2"}](f1_small.eps){width="\columnwidth"}
![ An H band MMT AO image of the astrometric binary ADS 8939 (WDS 13329+3454; STT 269AB). The well known orbit (WDS Grade level 2) of this binary star predicted a separation of $0.265\arcsec$ and a PA of $218.237^{\circ}$ for UT Jan 19, 2003 (the night of this observation). For these values we derived that the Indigo camera had a platescale of $0.0242\arcsec$/pixel. This 10 second integration had a mid-point time of UT 12:21:30, hence the parallactic angle during this exposure was $-107.6^{\circ}$. Rotating the image by $-107.6^{\circ}$ (clockwise) resulted in a measured PA of $218.35^{\circ}$ which indicates North is $0.113^{\circ}$ east of the Indigo’s Y axis. Linear color scale. North is up and east is left.[]{data-label="fig2a"}](f2_small.eps){width="\columnwidth"}
![ An H band image of the astrometric binary ADS 7158 (WDS 09036+4709; A 1585). The well known orbit (WDS Grade level 2) of this binary star predicted a separation of $0.111\arcsec$ and a PA of $312.764^{\circ}$ for UT Jan 20, 2003 (the night of this observation). We utilized these values to check the $0.0242\arcsec$/pixel platescale and orientation (north being $0.113^{\circ}$ east of the Indigo’s Y axis) that were obtained from the ADS 8939 observations for the Indigo camera (see Figure \[fig2a\]). The above 10 second integration had a mid point time of UT 8:18:50, hence the parallactic angle during this exposure was $-171.0^{\circ}$. Rotating the image by $-171.0^{\circ}$ (and correcting for the $0.113^{\circ}$ misalignment of the Y axis) resulted in a measured PA of $312.146^{\circ}$ which which is incorrect by $0.62^{\circ}$. Hence we conservatively estimate our PA is calibrated to with $\pm1^{\circ}$. The separation of ADS 7158 is 4.677 pixels suggesting a platescale of $0.0241\arcsec$/pixel. Hence we estimate a conservative $\pm0.002\arcsec$ error in the Indigo platescale of $0.0242\arcsec$/pixel. Logarithmic color scale, note the Airy rings around each component. North is up and east is left.[]{data-label="fig2b"}](f3_small.eps){width="\columnwidth"}
![ Detail of the $\theta^{1}$ Ori B group as imaged at $0.077\arcsec$ resolution (in the H band) with the MMT AO system and the Indigo IR camera. Logarithmic color scale. North is up and east is left. Note that the object “$B_1$” is really an eclipsing spectroscopic binary ($B_1B_5$); where the unseen companion $B_5$ orbits $B_1$ every 6.47 days [@abt91].[]{data-label="fig3"}](f4_small.eps){width="\columnwidth"}
![ The Gemini/Hokupa’a images of the $\theta^{1}$ Ori B group in the $K^\prime$ band. Resolution $0.085\arcsec$. Logarithmic color scale. North is up and east is left. []{data-label="fig4"}](f5_small.eps){width="\columnwidth"}
![ The upper part of the $\theta^{1}$ Ori cluster as imaged over $30\times 30\arcsec$ FOV at Gemini with the Hokupa’a AO system. Logarithmic color scale. North is up and east is left. Note that the object “$A_1$” is really a spectroscopic binary ($A_1A_3$); where the unseen companion $A_3$ is separated from $A_1$ by 1 AU [@bos89] []{data-label="fig5"}](f6_small.eps){width="\columnwidth"}
![ The separation between $\theta^{1}$ Ori $B_1$ and $B_2$. Note how over five years of observation there has been little significant relative proper motion observed (-0.0006$\pm0.0019\arcsec$/yr; which is insignificantly different from a constant). If the group is gravitationally bound the separation should be roughly constant over five years. The observed rms scatter from a constant value is indeed a mere $\pm0.0019\arcsec$, suggesting the whole $\theta^{1}$ Ori B group is likely physically bound together. The first 2 data points are speckle observations from the 6-m SAO telescope [@wei99], the next point is from our Gemini/Hokupa’a observations and the last data point is from the MMT AO observations. []{data-label="fig6"}](f7.eps){width="\columnwidth"}
![ The position angle between $\theta^{1}$ Ori $B_1$ and $B_2$. Note how over five years of observation there has been no significant relative proper motion observed (0.07$\pm0.25^\circ$/yr which is insignificantly different from a constant). The error from a constant value is a mere $\pm0.3^\circ$.The first 2 data points are speckle observations from the 6-m SAO telescope [@wei99], the next point is from our Gemini/Hokupa’a observations and the last data point is from the MMT AO observations. []{data-label="fig7"}](f8.eps){width="\columnwidth"}
![ The separation between $\theta^{1}$ Ori $B_2$ and $B_3$. Note the lack of any significant relative motion ($0.0006\pm0.0010\arcsec$/yr). The rms scatter from a constant value is only $0.001\arcsec$. There appears to very little change in the separation of the $B_2B_3$ system. The first 2 data points are speckle observations from the 6-m SAO telescope [@wei99], the next point is from our Gemini/Hokupa’a observations and the last data point is from the MMT AO observations. []{data-label="fig8"}](f9.eps){width="\columnwidth"}
![ The position angle of $\theta^{1}$ Ori $B_2$ and $B_3$. Here we observe what may be real orbital motion of $B_3$ moving counter-clockwise (at 0.93$\pm0.49^\circ$/yr; correlation significant at the 99.2% level) around $B_2$. This small amount of motion is consistant with the $B_2B_3$ system being bound. The first 2 data points are speckle observations from the 6-m SAO telescope [@wei99], the next point is from our Gemini/Hokupa’a observations and the last data point is from the MMT AO observations. []{data-label="fig9"}](f10.eps){width="\columnwidth"}
![ The separation between $\theta^{1}$ Ori $A_1$ and $A_2$. There is a small negative changes in the orbital separation ($-0.0064\pm0.0027\arcsec$/yr) as $A_2$ moves towards $A_1$. The first data point is from speckle observations at the 3.5-m Calar Alto telescope [@pet98], the next point is from a speckle observation from the 6-m SAO telescope [@wei99], the last point is from our Gemini/Hokupa’a observations. []{data-label="fig10"}](f11.eps){width="\columnwidth"}
![ The position angle of $\theta^{1}$ Ori $A_1$ and $A_2$. There does appear to be significant changes in the position angle as $A_2$ moves counter clockwise (at 2.13$\pm0.73^\circ$/yr) around $A_1$. This relatively small motion is consistent with the $A_1A_2$ system being bound. The first data point is from speckle observations at the 3.5-m Calar Alto telescope [@pet98], the next point is from a speckle observation from the 6-m SAO telescope [@wei99], the last point is from our Gemini/Hokupa’a observations. []{data-label="fig11"}](f12.eps){width="\columnwidth"}
![ The separation between $\theta^{1}$ Ori $B_1$ and $B_4$. Note how over three years of observation there has been little significant relative proper motion observed (-0.0017$\pm0.0033\arcsec$/yr; which is insignificantly different from a constant). If the low mass star $B_4$ is gravitationally bound to the B group the $B_1B_4$ separation should be roughly constant over these three years. The observed rms scatter from a constant value is indeed a mere $\pm0.0019\arcsec$, suggesting the whole $\theta^{1}$ Ori B group is likely physically bound together. The first data point is an speckle observation from the 6-m SAO telescope [@sch03], the next point is from our Gemini/Hokupa’a observations and the last data point is from the MMT AO observations. []{data-label="figB4_sep"}](f13.eps){width="\columnwidth"}
![ The position angle between $\theta^{1}$ Ori $B_1$ and $B_4$. Note how over three years of observation there has been no significant relative proper motion observed (0.18$\pm0.9^\circ$/yr which is insignificantly different from a constant). The error from a constant value is a mere $\pm0.3^\circ$.The first data point is a speckle observation from the 6-m SAO telescope [@sch03], the next point is from our Gemini/Hokupa’a observations and the last data point is from the MMT AO observations. []{data-label="figB4_pa"}](f14.eps){width="\columnwidth"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Data mining is about obtaining new knowledge from existing datasets. However, the data in the existing datasets can be scattered, noisy, and even incomplete. Although lots of effort is spent on developing or fine-tuning data mining models to make them more robust to the noise of the input data, their qualities still strongly depend on the quality of it. The article starts with an overview of the data mining pipeline, where the procedures in a data mining task are briefly introduced. Then an overview of the data preprocessing techniques which are categorized as the data cleaning, data transformation and data preprocessing is given. Detailed preprocessing methods, as well as their influenced on the data mining models, are covered in this article.'
author:
- '\'
bibliography:
- 'bibliography.bib'
title: 'Preprocessing Methods and Pipelines of Data Mining: An Overview'
---
Data Mining, Data Preprocessing, Data Mining Pipeline
Introduction
============
mining is a knowledge obtaining process: it gets data from various data sources and finally transforms the data into knowledge, thus provides insight to its application field. Data mining pipeline is a typical example of the end-to-end data mining system: they are an integration of all data mining procedures and deliver the knowledge directly from data source to human.
The purpose of data preprocessing is making the data easier for data mining models to tackle. The quality of data can have a significant influence on data mining models. It is considered that the data and features have already set the upper bound of the knowledge that can be obtained, and the data mining models are just about approximating the upper bound. Various preprocessing techniques are invented to make the data meet the input requirements of the model, improve the relevance of the prediction target, and make the optimization step of the model easier.
It is common that raw data obtained from the natural world is badly shaped. The problems include the appearance of missing values (e.g., a patient did not go through all the tests), duplications (e.g., annual income and monthly income), outlier values (e.g., age is -1) as well as contradictions (e.g., gender is male and is pregnant) in the dataset. Although the existing preprocessing techniques would not guarantee to solve all these problems, they could at least correct some of them and improve the performance of the models.
The data type and distribution of data are usually transformed before being sent to data mining models. The purpose of data transformation includes making the data meets the input requirement of the models, removing the noise of data, and making the distribution of data more suitable for applying optimization algorithms in the model training step.
The input for data mining models can be huge: they may have too many dimensions or of massive amount, which would make it difficult for the data mining model to train or cause troubles while transferring and storing the data. Data reduction techniques can reduce the problem by applying reduction on dimensions (known as dimensional reduction) or amounts of data (known as instance selection and sampling).
To implement preprocessing to data, Python and R are among the most popular tools. With bulks of packages such as scikit-learn[@scikit-learn] and PreProcess[@preprocess], most of the preprocessing algorithms covered in this paper can be implemented even without consideration of its details.
In the following section, the data mining pipeline and the primary procedures in the data mining pipeline will be introduced. From Section 2 on, we will focus on the steps in the data preprocessing work: Section 3 will introduce the techniques used in data cleaning, while Section 4 will cover the data transformation techniques. In the last section, data reduction techniques will be discussed.
Data Mining Pipeline
====================
The data mining pipeline is an integration of all procedures in a data mining task. While most of the data already exists in a data base, a data warehouse, or other types of data source[@han2011data], various steps should be taken in order to make them easier for a human to understand. An illustration of the data mining pipeline is given as in [ \[fig:datamining\]]{}. Generally speaking, the key procedures include *obtaining*, *scrubbing*, *exploring*, *modelling*, *interpreting*. These procedures are known as “OSEMN”[^1]. However, note that the pipeline is not a linear process in the real world, but a successive and long-lasting task. Methods in scrubbing and modeling procedures have to be tested and refined, the obtaining procedure may have to be adapted for different kinds of data sources, and the visualization and interpretation of the data may have to be adjusted for their audience, thus meet the audience’s demand. In the rest of this section, details of these procedures will be discussed.
Obtaining
---------
Obtaining the data is the most fundamental step in data mining since it is the data itself that decides what knowledge it may contain. Data base and data warehouse are among the primary sources of data, where the structured data can be fetched with query languages, usually SQL. The data warehouse is specially designed for organizing, understand, and making use of the data[@han2011data]: they are usually separated system from the operational database, having time-variant structure as well as structures that makes the subsequent analysis work easy, and most importantly, nonvolatile.
The obtained data can be archived as files and directly used for the subsequent procedures. They may also be reformatted and stored in a data base or a data warehouse, prepared for the data mining tasks in the future.
In the past, as well as in most common cases now, we regard the data obtaining step as the process for obtaining a dataset, regardless of how we obtained them. However, nowadays, new data is generating at an extreme speed: there are tons of data being created every second of every day. Some services, such as the public opinion monitoring and the recommendation system, do need the newly generated data: they have strong demand for being on time. In these circumstances, the concept of “stream” is comparatively more important than a dataset. A stream is a real-time representation of data. Under this concept, models and algorithms that can run online are developed[@gaber2005mining]. For the stream mining tasks, the goal of data obtaining is no longer obtaining a dataset, but a real-time input source.
Scrubbing
---------
Scrubbing is about the cleaning and preprocessing of the data, aiming to make the data have a unified format and easy to be modeled. As for the detailed concepts and techniques in data scrubbing, the readers could refer to the following sections of this paper, since most of them will be covered in the overview of data preprocessing.
Exploring
---------
Before modeling the data, people may want to get to know the underlying distribution of the data, the correlation between variables, and their correlation with the labels. Assumptions can be made in this step. For instance, people may assume smoking is highly correlated with lung cancer. These assumptions are important because they would provide indications to other procedures in a data mining task, including helping to choose a suitable model, and helping to justify your work in the interpreting of data.
Tools for exploring the data and verifying the assumptions are usually statistical analysis and data visualization. The statistical analysis gives us theoretical probabilities, known as significance level, of our assumption being incorrect, while data visualization tools, such as ggplot[@ggplot] and D3[@bostock2011d3], give us the impressions about the distribution of the data, help people to verify their assumption conceptually. Also, new patterns that are ignored in the assumption step might be found in the visualization step.
Modeling
--------
With underlying patterns existed in the data source, modeling makes it possible to represent the pattern explicitly with the data mining models. For a data mining task, modeling would usually split the data into the training set and test set thus could score the accuracy of the model on a relatively “new” dataset. If the model contains hyperparameters, such as the parameter $k$ in a K-Nearest Neighbor (KNN) model, a cross validation set will be created for obtaining the best set of the hyperparameter.
For most of the data mining models, loss functions are defined. Generally, a loss function will have lower value if the model performs well. Besides, it usually has special features such as convexity, which makes the gradient-based optimizing algorithm performs better. With trainable parameters, a model’s training step is about adjusting its parameter so that it gains lower loss on its training data. The specific definition of loss function depends on the model itself and the task. Mean squared error ($\sum_{i=1}^n (\hat{y}_i - y_i^2$) for regression task and cross entropy ($-\sum_{i=1}^n [y_i\log \hat{y_i} + (1 - y_i)\log (1 - \hat{y_i})]$) for classification task are frequently accepted loss functions.
There are too many kinds of data mining models existed; their tasks include clustering, classification, and regression. The complexity of the models also varies: simple models such as linear regression only have a few parameters, a small amount of data will make the training step converge, while complex models such as AlexNet have millions of parameters[@krizhevsky2012imagenet], and their training also require huge dataset. However, complex does not mean better: the model should be decided according to the predicting target of the task, the dataset size, the data type, etc. Sometimes, it is necessary to run different models over one dataset and find the most suitable data mining model.
Interpreting
------------
While previous steps are generally pure science, the interpreting step is more humanistic. Knowledge can be extracted from the input data, but it takes extra effort to convince people to accept this knowledge. Although complicated statistics and models make the work looks more professional, for laymen, graphs, tables and well explained accuracy make it easier to understand and accept. Besides, social skills such as story telling and emotion quotient are also important: this procedure is about the human, not the data.
Data Cleaning
=============
The data obtained from the natural world is usually badly shaped. Some of the problems, such as outlier, may affect the data mining model and produce a biased result. For instance, the outlier may affect the K-Means clustering algorithm by substantially distort the distribution of data[@velmurugan2010computational]. Other problems, if not handled, will make it impossible for the data to be analyzed by models, such as the Not a Number (nan) values in a data vector. Data cleaning techniques, including missing value handling and outlier detection, were issued to tackle the problem. They make the gathered data suitable as the input of the model.
Missing Values Handling
-----------------------
Missing values is a typical kind of data incompleteness of dataset. Most of the data mining models would not tolerate the missing values of its input data: these values can not be used for comparison, not available for categorizing, and can not be operated with arithmetic algorithms. Thus, it is necessary to handle the missing values before pushing the dataset to data mining models. The easiest way to deal with missing value is to drop the entire sample. This method is effective if the proportion of missing value in a dataset is not significant, however, if the number of missing values is not suitable for ignoring, or the percentage of missing value for each attribute is different[@han2011data], dropping the samples with missing value would reduce the amount of dataset dramatically, the information contained in the dropped samples is not made use of. Another way to deal with missing values is by filling them, and there are varies methods for finding the suitable value to fill the missing value, some of them are listed as follows.
### Use special value to represent missing
Sometimes the missing value itself has some meaning. For instance, in a patient’s medical report, the missing value for uric acid means the patient did not go through the renal function test. Thus, using a certain value such as -1 makes sense, for they can be operated like normal values while having special meaning in the dataset.
### Use attribute statistics to fill
Statistics such as mean, median, or mode can be obtained from non-missing values in the missing value’s corresponding attribute. It is said that for a skewed dataset, the median would be a better choice[@han2011data]. However, this technique does not take the sample’s other non-missing attributes into account.
### Predicting the value with known attributes
If we assume, there exists a correlation between attributes, filling the missing value can be modeled as a prediction problem: predicting the value with the non-missing attributes with other samples as training data. The predicting methods includes regression algorithms, decision trees[@han2011data], and K-Means[@patil2010missing].
### Assigning all possible values
For categorical attributes, given an example $E$ with $m$ possible values for its missing value, then $E$ can be replaced with $m$ new examples $E_1, E_2, \dots, E_m$. This missing value filling technique assumes the missing attribute does not matter for the example. Thus the value can be anyone in its domain[@grzymala1991unknown].
Comparison of different missing value filling techniques is done in [@grzymala2000comparison]. In the work, different missing value filling techniques are tested on ten datasets for running simple and extended classification methods. The conclusion shows C4.5 decision tree method performs the best, ignoring the samples with missing values and assigning all possible values also performed well, and filling the values with mode perform the worst. However, the performance of missing value filling techniques may differ because of the feature of the dataset. As a result, most of the techniques are worth trying for a data mining task.
Outlier Detection
-----------------
Outlier refers to the data sample that has a massive distance to most of the other samples. Although the rare case does not necessarily mean wrong (e.g., age = 150), most outliers are caused by measurement error or wrong recording, thus ignoring a rarely appearing case would not harm a lot. Although some of the models are robust against outliers, outlier detection is still recommended in data preprocessing work.
Statistics-based outlier detection algorithms are among the most commonly used algorithms, which assume an underlying distribution of the data[@kantardzic2011data] and regard the data examples which corresponding probability density lower than a certain threshold as the outliers. As the underlying distribution is unknown for most cases, the normal distribution is a good substitute, and its parameter could be estimated by the mean value and standard deviation of the data. The Mahalanobis distance[@kantardzic2011data], as in , is a scale-irrelevant distance between two data samples. The outlier can be decided by comparing the Mahalanobis distance between each sample and the mean value of all samples. Box-plot, as another kind of statistics-based outlier detection technique, can give the graphical representation of outlier by plotting the lower quartile and upper quartile along with the median[@walfish2006review].
$$\label{eq:mahalanobis}
D_M(x, y) = \sqrt{(x - y)^T\Sigma^{-1}(x - y)}$$
Without making any assumption of the distribution of the data, the distance-based outlier detection algorithm can detect the outlier by analyzing the distance between every two samples, thus determine the outliers. Simple distance-based outlier detection algorithms are not suitable for a large dataset, since for $n$ samples with $m$ dimensions, their complexity is usually $O(n^2m)$[@kantardzic2011data], and each computation requires scanning all the samples. However, an extended cell-based outlier detection algorithm is developed in[@knorr2000distance], which guarantee linear complexity over the dataset volume and no more than three dataset scans. The experiment shows this algorithm is the best for the dataset with dimension less than 4.
Sometimes, with consideration of temporal and spatial locality, an outlier may not be a separate point, but a small cluster. Cluster-based outlier detection algorithms consider clusters with small size as outlier clusters and clean the dataset by removing the whole cluster[@ben2005outlier][@duan2009cluster].
Data Transformation
===================
The representation of data in different attributes varies: some are categorical, while some are numerical. For categorical values, they can be nominal, binary or ordinal[@han2011data], and for numerical data, they can also have different statistical features including mean values and standard deviations. However, not all kinds of data meet the requirement of data mining models. Also, the difference among data attributes may bring troubles for the subsequent optimization work of data mining models. Data transformation is about modifying the representation of data so that they are qualified to be the input for data mining models, as well as making the optimization algorithm of the data mining model easier to take effect.
Numeralization
--------------
Categorical values widely exists in the natural world, some of the operations, such as calculating the entropy between groups, can be done directly over categorical data. However, most operations are not applicable to categorical data. Thus, categorical data is supposed to be encoded into numerical data, making it meets the requirements of the models. The following encoding techniques are adopted for numeralization.
- *One-Hot encoding*: Regard each possible value of the categorical data as a single dimension, and use $1$ for the dimension which the sample belongs to the category, otherwise $0$.
- *Sequential encoding*: For each possible value of the categorical data, assign it with a unique and numerical index. This is implemented as a kind of word encoding, as in[@angelidis2018multiple].
- *Customized encoding*: Customized encoding is based on rules designed for a certain task. For instance, word2vec[@mikolov2013distributed] is an encoding that can turn a word into a $300$ dimensional vector, with consideration of the word’s meaning.
Generally, one-hot encoding is suitable for categorical data with fewer possible values; if there are too many possible values, such as English words, the encoded dataset would be huge and sparse. Sequential encoding would not produce huge output, but the encoded data is not as easy to separate as one-hot encoded data. Customized encoding, if carefully designed, usually perform well over a certain kind of task, but for other tasks, the encoding should be redesigned, and its design can take lots of efforts.
Discretization
--------------
Discretization of data is applied sometimes to meet the requirement of input of models, such as Naive Bayes which require its input to be countable[@rish2001empirical]. Also, it can smooth the noise. Discretization of data does not necessarily make the data categorical, but make the continuous values countable. The discretization of data can be achieved with unsupervised learning methods such as putting data into equal-width or equal-frequency slots, known as *binning*, or clustering. Some supervised learning methods such as *decision tree* can also be used for discretization of data[@han2011data].
Normalization
-------------
Since different attribute usually adopts a different unit system, their mean value and standard deviation are usually not identical. However, the numerical difference would make some of the attributes look more “important”, while others are not[@han2011data]. This impression could cause trouble for some models; one of the typical ones is KNN: larger value would strongly affect the distance comparison, making the model mainly consider attributes tend to have larger numerical values. Besides, for neuron network models, the different unit system will also have a negative influence on gradient descent optimizing methods, forcing it to adopt a smaller learning rate. To tackle the problem mentioned above varies of normalization methods are issued, some of them are listed as follows.
### Min-max normalization
Min-max normalization is used for mapping the attribute from its range $[lb, ub]$ to another range $[lb_{new}, ub_{new}]$; the target range is usually $[0, 1]$ or $[-1, 1]$ [@garcia2015data]. For a sample with value $v$, the normalized value $v'$ is given as in .
$$\label{eq:minmax}
v' = \frac{v - lb}{ub - lb} (ub_{new} - lb_{new}) + lb_{new}$$
### Z-score normalization
If the underlying range of an attribute is unknown or outlier exists, min-max normalization is not feasible or could be strongly affected[@garcia2015data]. Another normalization approach is to transform the data so that it would have $0$ as mean and $1$ as standard deviation. Given the mean $\mu$ and standard deviation $\sigma$ of the attribute, the transformation is represented as in .
$$\label{eq:zcore}
v' = \frac{v - \mu}{\sigma}$$
Note that if $\mu$ and $\sigma$ are unknown, they can be substituted with the sample mean and standard deviation.
### Decimal scaling normalization
An easier way to implement the normalization is to shift the floating point of the data so that each value in an attribute would have an absolute value less than $1$, the transformation is given as in .
$$\label{eq:decimal_scaling}
v' = \frac{v}{10^j}$$
For some cases, different attributes have an identical or similar unit system, such as the preprocessing of RGB-colored imaged. In these cases, normalization is not necessary. However, if this is not guaranteed, normalization is still recommended for all data mining tasks.
Numerical Transformations
-------------------------
The transformation over dataset can help to obtain additional attributes. These features obtained by transformation could be unimportant for some data mining models, such as neural network, which have superior fitting potential. However, for relatively simpler models with fewer parameters, linear regression, for example, the transformed features do help the model to get better performance(as in [ \[fig:boxcox\]]{}), for they could provide additional indication of the relationship between attributes. The transformation would also be essential for scientific discoveries and machine controls[@lin2002attribute].
Generally, given attributes set $\{a_1, a_2, \dots, a_p\}$, the numerical transformation can be represented as in . Theoretically, $f$ can be any function, however, since the input data is finite, $f$ can take polynomial forms[@lin2002attribute].
$$\label{eq:general_trans}
x' = f(a_1, a_2, \dots, a_p)$$
The commonly used representation of $f$ includes *polynomial based transformation*, *approximation based transformation*, *rank transformation* and *box-cox transformation*[@garcia2015data]. The parameters in the transformation formula could be obtained by subjective definition (for situations where people know the relationship between attributes and labels well), by brute search[@garcia2015data] or by applying maximum likelihood method.
Data Reduction
==============
The amount of data in a data warehouse or a dataset can be huge, causing difficulties for data storage and processing when working on a data mining task, while not every model needs a huge amount of data to train. On the other hand, although the data may have lots of attributes, there could be unrelated features as well as the interdependence between features[@kotsiantis2006data]. Data reduction is the technique that helps to reduce the amount or dimension, or both, of a dataset, thus making the model’s learning process more efficient as well as helping the model to obtain better performance, including preventing overfitting problem, and fix the skewed data distribution.
Dimensional reduction
---------------------
The dimension reduction technique is about reducing the dimensionality of data samples, thus reduce the total size of the data. As the number of attributes is reduced for a sample, there is less information contained in it. A good dimensional reduction algorithm will keep more general information: this could make it more difficult for models to become overfitted.
Some dimensional reduction techniques pose a dimensional reduction transformation over a dataset, generating new data samples, which have a fewer number of attributes than before. The transformations have different criteria. Principal component analysis, known as PCA, could reduce the dimension of data while keeping the maximum variance of data[@jolliffe2011principal]. This is achieved by multiplying matrix $A = (a_1, a_2, \dots, a_p)^T$ to the dataset $X$ and keep the top $k$ dimensions ($a_i$ stands for the normalized eigenvector corresponds to the $i^{th}$ greatest eigenvalue of the covariance matrix of the dataset). By contrast, linear discriminant analysis (LDA), is meant to maximize the component axes for class-separation. The implementation of LDA is similar to PCA; the only difference is it replaces the covariance matrix with the scatter matrix of samples. A graphical illustration of the difference between PCA and LDA is given as [ \[fig:pcalda\]]{}. In comparatively more situations, LDA outperforms PCA. PCA may outperform LDA when the data amount is small, or the data is nonuniformly sampled[@martinez2001pca]. Other dimensional reduction algorithms include factor analysis (assuming a lower dimensional underlying distribution), projection pursuit(measuring the aspect of non-Gaussianity)[@fodor2002survey], and wavelet transform[@han2011data].
Feature selection is another dimensional reduction technique: it is about removing irrelevant or correlated attributes from the dataset while keeping the other relatively independent attributes untouched. Feature selection is more than simply selecting the feature that has greater relevance with the variable to predict, the relationship between attributes are also supposed to be taken into consideration: the goal is to find a sufficiently good subset of features to predict[@kotsiantis2006data]. Feature selection methods can be divided into three types, as follows[@guyon2003introduction].
- *filter*: Directly select the feature based on attribute level criteria, including information gain, correlation score, or chi-square test. The filter method does not take the data mining model into consideration.
- *wrapper*: Use techniques to search through the potential subsets, according to their performance on the data mining model. Greedy strategies, including forward selection and backward elimination[@guyon2003introduction], are issued in order to reduce the time consumption.
- *embedded*: Embed the feature selection into the data mining model. Usually, the weight over different attributes would act as feature selection. A typical example is the regularization term of the loss function in linear regression, known as Lasso regression (for L1 regularization) and ridge regression (for L2 regularization).
Instance selection and sampling
-------------------------------
Both instance selection and sampling are about achieving the reduction of data by reducing the amount of data, seeking the chance to train the model with minimum performance loss[@garcia2015data], while based on different criteria for selecting (or dropping) instance.
Most instance selection algorithm is based on fine-tuning classification models. To help the model make better decision, *condensation algorithm* and *edition algorithm* are issued[@garcia2015data]. The condensation removes the samples lie in the relative center area of the class, assuming they do not contribute much in classification. *Condensed nearest neighbor*[@hart1968condensed], for instance, select instance by adding all the samples that cause a mistake to a K-Nearest Neighbor classifier. Edition algorithm removes the samples close to boundary, hoping to give the classifier a smoother decision boundary. Related algorithms include a clustering-based algorithm to select the center of clusters[@lumini2006clustering],
Compared with instance selection methods, sampling is a faster and easier way to reduce the number of instances, since almost no complex selection algorithm is required for sampling methods: they only focus on reducing the amount of the data samples. The easiest sampling technique is *random sampling*, which collect a certain amount or portion of samples from the dataset randomly. For skewed datasets, *stratified sampling*[@kotsiantis2006data] is more adapted, since it takes the appearance frequency of labels from different classes into account and assigns a different probability of data with different labels being chosen, thus makes the sampled dataset more balanced.
Summary and Outlook
===================
Data mining, as a technique to discover additional information from a dataset, can be integrated as a pipeline, in which obtaining, scrubbing, exploring, modeling, and interpreting are the key steps. The purpose of the data mining pipeline is to tackle realistic problems, including reviewing the past and predicting the future. The specific technique used in each step should be selected with care to give the best performance to the pipeline.
The success of a data mining model depends on the proper data preprocessing work. The unpreprocessed data can be of unsuitable format for model input, causing instability for the optimization algorithm of the model, having a great impact on the model’s performance because of its noise and outliers, and causing performance problems on the model’s training process. With careful selection of preprocessing steps, these problems can be reduced or avoided.
Data type transformation techniques as well as missing value handling techniques makes it possible for models for processing different types of data. By applying normalization, the unit system of different attributes would be more unified, reducing the probability of an optimization algorithm to miss the global minimum. For simpler models, numerical transformation can provide richer features to the model, thus enhance the model’s ability to discover more underlying relationships between features and labels. For the overfitting problems of the model, dimensional reduction techniques help model find the more general information about samples instead of the too detailed features by reducing the dimension of feature, thus remove some unimportant information. And for the performance of the model training, both dimensional reduction techniques and instance selection techniques would improve the training performance by reducing the total amount of data.
[^1]: http://www.dataists.com/2010/09/a-taxonomy-of-data-science/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The local reference frame (LRF) acts as a critical role in 3D local shape description and matching. However, most of existing LRFs are hand-crafted and suffer from limited repeatability and robustness. This paper presents the first attempt to learn an LRF via a Siamese network that needs weak supervision only. In particular, we argue that each neighboring point in the local surface gives a unique contribution to LRF construction and measure such contributions via learned weights. Extensive analysis and comparative experiments on three public datasets addressing different application scenarios have demonstrated that LRF-Net is more repeatable and robust than several state-of-the-art LRF methods (LRF-Net is only trained on one dataset). In addition, LRF-Net can significantly boost the local shape description and 6-DoF pose estimation performance when matching 3D point clouds.'
author:
- Angfan Zhu
- Jiaqi Yang
- Chen Zhao
- Ke Xian
- 'Zhiguo Cao[^1] Xin Li'
bibliography:
- 'ijcai20.bib'
title: 'LRF-Net: Learning Local Reference Frames for 3D Local Shape Description and Matching'
---
Introduction
============
The local reference frame (LRF) is a canonical coordinate system established in the 3D local surface, which is a useful geometric cue for 3D point clouds. LRF possesses two intriguing traits. One is that rotation invariance can be achieved via LRF if the local surface is transformed with respect to the LRF [@rops2013l]. The other is that useful geometric information can be mined with LRF [@petrelli2011repeatability]. These make LRF popular in many geometric relevant tasks, especially for local shape description and six-degree-of-free (6-DoF) pose estimation.
For local shape description, two corresponding local surfaces can be converted into the same pose and full 3D geometric information can be employed, which is beneficial to improving the performance of local descriptors. Some hand-crafted local shape descriptors, e.g., signature of histograms of orientations (SHOT) [@shot2010] and signature of rotational projection statistics (RoPS) [@rops2013l], estimate an LRF from the local surface and then translate local geometric information with respect to the estimated LRF into distinctive and rotation-invariant feature representations. Some learned local descriptors, e.g., [@gojcic2019perfect] and [@spezialetti2019learning], leverage LRFs to overcome the limitation of geometric deep learning networks of being sensitive to rotations. Therefore, LRF is critical for both traditional and learned local shape descriptors. For 6-DoF pose estimation, an LRF can significantly improves its efficiency. Traditional 6-DoF pose estimation is usually performed via RANSAC [@derpanis2010overview], which randomly selects inlier correspondences from an initial correspondence pool to for pose prediction. Such random sampling method is neither reliable nor computational efficient [@deng20193d]. By contrast, we can directly predict an initial pose via two corresponding LRFs, reducing the computational complexity from $O(n^3)$ to $O(n)$.
The desirable properties for LRF are twofold [@shot2010]. The first one is the invariance to rigid transformation (e.g., translations and rotations). The second one is the robustness to common disturbances (e.g., noise, clutter, occlusion and varying mesh resolutions). To achieve these goals, many LRF methods have been proposed in the past decade and they can be categorized into two classes [@yang2018toward]: covariance analysis (CA) [@mian2010repeatability; @shot2010] or point spatial distributions (PSD)-based [@petrelli2011repeatability; @petrelli2012repeatable; @yang2017toldi]. CA-based LRFs are based on the computation of eigenvectors of a covariance matrix calculated either for the points or triangles in the local surface. PSD-based LRFs usually calculate estimate axes successively, where the main efforts are put on the determination of the $x$-axis [@yang2018toward]. However, most CA-based LRFs still suffer from sign ambiguity, and PSD-based LRFs show limited robustness to high levels of noise and variations of mesh resolution [@petrelli2012repeatable]. Methods in both classes usually apply a weighted strategy to improve their repeatability performance. However, their weights are determined heuristically, and the repeatability performance in challenging 3D matching cases cannot be guaranteed.
Motivated by existing considerations, we propose a learned approach toward LRF estimation (named LRF-Net), which considers the contribution of all neighboring points (Fig. \[fig\_fig1\]). Our key insight is that each neighboring point in the local surface gives a unique contribution to LRF construction, which can be quantitatively represented by assigning weights to these points. Given a local surface centered at a keypoint, we first resort to the normal of the keypoint computed within a subset of the radius neighbors for the calculation of its $z$-axis. Its repeatability has been confirmed in [@petrelli2011repeatability]. Compared with $z$-axis, estimating the $x$-axis is more challenging, due to noise, clutter, and occlusion. By collecting angle and distance attributes within a local neighborhood, we can formulate the estimation of $x$-axis as a weighted prediction problem with respect to these geometric attributes. Unlike previous CA-based and PSD-based approaches, such learned strategy of determining weights is shown to be invariant to rigid transformation and robust to noise, clutter, occlusion and varying mesh resolutions. Our network can be trained in a weakly supervised manner. Specifically, it needs the corresponding relationships between local patches only, instead of ground-truth LRFs and/or exact pose variation information between patches. We have conducted a set of experiments on three public datasets to comprehensively evaluate the proposed LRF-Net. Extensive analysis and comparative experiments on three public datasets addressing different application scenarios have demonstrated that LRF-Net is more repeatable and robust than several state-of-the-art LRF methods (LRF-Net is only trained on one dataset). In addition, LRF-Net can significantly boost the local shape description and 6-DoF pose estimation performance when matching 3D point clouds. The major contributions of this paper are summarized as follows:
- LRF-Net, based on a Siamese network that needs weak supervision only, is proposed that achieves the state-of-the-art repeatability performance under the impacts of noise, varying mesh resolutions, clutter and occlusion. To the best of our knowledge, we are the first to concentrate on designing LRF for local surfaces with deep learning.
- LRF-Net can significantly boost the performance of local shape description and 6-DoF pose estimation.
The rest of this paper is organized as follows. Section 2 presents a detailed description of our proposed LRF-Net. Section 3 presents the experimental evaluation of LRF-Net on three public datasets with comparisons with several state-of-the-art methods. Several concluding remarks are drawn in Section 4.
Method
======
This section represents the details of our proposed LRF-Net for 3D local surface. We first introduce the technique approach for calculating the three axes for an LRF and then describes a weakly supervised approach for training LRF-Net.
A Learned LRF Proposal
----------------------
The whole architecture of LRF-Net in shown in Fig. \[fig\_LRFNet\](a). LRF-Net predicts the direction of three axes successively. For a local surface, we first estimate its $z$-axis via its normal vector computed over a small subset of the local point set. Then, unique weights are learned for each point in the local surface. The $x$-axis is calculated by integrating projection vectors with learned weights using a vector-sum operation. At last, the $y$-axis is calculated by the cross-product operation between $z$-axis and $x$-axis.\
\
**LRF definition:** Given a local surface $\bf{Q}$ centered at keypoint $\bf{p}$, the LRF at $\bf{p}$ (denoted by $\bf{L_p}$) can be represented as : $$\bf{L_p}=[\bf{x(p)}, \bf{z(p)}\times \bf{x(p)}, \bf{z(p)}],\label{eq_lrf}$$ where $\bf{x(p)}$, $\bf{y(p)}$, and $\bf{z(p)}$ denote the $x$-axis, $y$-axis, and $z$-axis of $\bf{L_p}$, respectively. As three axes are orthogonal, the estimation of LRF therefore contains two parts: estimation of the $z$-axis and the $x$-axis.
A naive way to learn an LRF for the local surface is to train a network that directly regresses the axes. The premise is that ground-truth LRFs are labeled for local surfaces. Unfortunately, the network trained in this manner meets two difficulties. The first one is that the definition of ground-truth LRFs for local surfaces remain an open issue in the community [@yang2018toward]. The second one, which is more important, is that the orthogonality of three axes cannot be guaranteed. We suggest estimating $z$-axis and $x$-axis independently.\
\
**z-axis:** As for $z$-axis, we take the normal of the keypoint as the $z$-axis., which has been confirmed [@petrelli2011repeatability] to be quite repeatable. To resist the impact fo clutter and occlusion, we collect a small subset of the local surface to calculate the normal. For more details, readers are referred to [@yang2017toldi].\
\
**x-axis:** Once the $z$-axis is determined, the remaining task is to compute the $x$-axis. Compared with $z$-axis, $x$-axis is more challenging due to noise, clutter, and occlusion [@yang2018toward]. We argue that each neighboring point in the local surface gives a unique contribution to LRF construction. Hence, we predict a weight for each neighboring point and leverage all neighboring points with learned weights for $x$-axis prediction. The main steps are as follows.
First, to make the estimate LRF invariant to rigid transformation, our network consumes with invariant geometric attributes, rather than point coordinates. In particular, two attributes, i.e., relative distance $a_{dist}$ and surface variation angle $a_{angle}$ are used in LRF-Net as illustrated in Fig. \[fig\_LRFNet\](b). For a neighbor $\bf{q}_i$ of $\bf{p}$, the two attributes of $\bf{q}_i$ are computed as: $$\begin{cases}
a_{dist}^i = \left\|\bf{pq}_{i}\right\|/r\\
a_{angle}^i = \cos(\bf{z(p)}, \bf{pq}_{i})
\end{cases},
\label{eq_attrib}$$ where $ \left\|\cdot\right\| $ is the $L_2$ norm and $r$ represents the support radius of the local surface. The range of $a_{angle}$ and $a_{dist}$ are $[-1, 1]$ and $[0, 1]$, respectively. Thus, every radius neighboring point represented by two attributes that will be encoded to a weight value via LRF-Net later. The employed two attributes in LRF-Net have two merits at least. First, the unique spatial information of a radius neighboring point in the local surface can be well represented, as shown in Fig. \[fig\_insight\]. Both attributes are complementary to each other. Second, the two attributes are calculated with respect to the keypoint, which are rotation invariant. It makes the learned weights rotation invariant as well.
Second, with geometric attributes being the input, we use a U-Net with multilayer perceptions (MLP) layers only to predict weights for neighboring points. The details of the network are illustrated in Fig. \[fig\_net\]. The network is very simple, however, is sufficient to predict stable and informative weights for neighboring points (as will be verified in the experiments).
Third, because $x$-axis is orthogonal to $z$-axis, we project each neighbor $\bf{q}_{i}$ on the tangent plane $\bf S$ of the $z$-axis and compute a projection vector for $\bf{q}_{i}$ as: $${\bf{v}}_{i}={\bf{pq}}_{i}-({\bf{pq}}_{i}\cdot {\bf{z(p)}})\cdot \bf{z(p)}.
\label{eq_vector_x}$$ We integrate all weighted projection vectors in a weighted vector-sum manner: $${\bf{x(p)}}=\sum_{i=1}^{n}w_i{\bf{v}}_{i}/\left\|\sum_{i=1}^{n}w_i{\bf{v}}_{i}\right\|,
\label{eq_axis_x}$$ where $n$ denotes the total number of radius neighbors of keypoint $\bf{p}$ and $w_{i}$ is a learned weight by LRF-Net. Another way for determining the $x$-axis, based on these weights, is choosing the vector with the maximum weight, as in many PSD-based LRFs [@petrelli2011repeatability; @petrelli2012repeatable]. However, it fails to leverage all neighboring information and we will shown that it is inferior to the vector-sum operation in the experiments.\
\
**y-axis:** Based on the calculated $z$-axis and $x$-axis, the $y$-axis can be computed by the cross-product between them.
Weakly Supervised Training Scheme
---------------------------------
Our training data are constituted by a series of corresponding local surface patches. The corresponding relationship is obtained based on the ground-truth rigid transformation of two whole point clouds. In particular, LRF-Net needs the corresponding relationships between local surface patches only, rather than ground-truth LRFs and/or exact pose variation information between patches. Therefore, our network can be trained in a weakly supervised manner.
We train our LRF-Net with two streams in a Siamese fashion where each stream independently predicts an LRF for a local surface. Specifically, two streams take the local surfaces of keypoints $\bf{p}_{m}$ and $\bf{p}_{s}$ as inputs, respectively. Here, $\bf{p}_{m}$ and $\bf{p}_{s}$ are two corresponding keypoints sampled from the model and scene point cloud. Both streams share the same architecture and underlying weights. We use the predicted LRFs ${\bf L}_m$ and ${\bf L}_s$ by two stream to transform the local surfaces ${\bf Q}_m$ and ${\bf Q}_s$ to the coordinate system of the two LRFs. Then, we calculate the Chamfer Distance [@deng2018ppf] between two transformed local surfaces as the loss function to train LRF-Net: $$Loss = d_{cham}({\bf L}_m\cdot{\bf Q}_m, {\bf L}_s\cdot{\bf Q}_s),
\label{eq_loss}$$ where $$\begin{split}
& d_{cham}(X,\hat{X})=\\
& \min \left\{ \frac{1}{|X|}\sum_{x\in X}\min_{\hat{x}\in \hat{X}}||x - \hat{x}||, \frac{1}{|\hat{X}|}\sum_{\hat{x}\in \hat{X}}\min_{x\in X}||x-\hat{x}|| \right\}.
\label{eq_cham}
\end{split}$$
Remarkably, our opinion is that it is difficult to define a “good” LRF for a single local surface. For 3D shape matching, LRFs that can align the poses of two local surface patches are judged as repeatable. This motivates us to consider two local patches simultaneously and employ the Chamfer Distance to train the network.
Experiments
===========
In this section, we first evaluate the repeatability performance of our LRF-Net on three standard datasets, including the Bologna retrieval (BR) dataset [@tombari2013performance], the UWA 3D modeling (UWA3M) dataset [@mian2006novel], and the UWA object recognition (UWAOR) dataset [@mian2006three], together with a comparison with other state-of-the-art LRFs. Second, we apply our LRF-Net perform local shape description and 6-DoF pose estimation to verify the practicability of our method. Third, analysis experiments are conducted to improve the explainability of the proposed LRF-Net.
Experimental Setup
------------------
The details of our experiments including the description of datasets and the illustration for all compared methods are introduced before evaluation. The experiments were conducted on a Windows Server with an Intel Xeon E5-2640 2.39 GHz CPU and 96 GB of RAM. We train our LRF-Net using a batch size of 512 local surface pairs and leverage the ADAM optimizer with an initial learning rate of 1e-4, which decays $5\%$ every epoch. Each sampled local surface contains 256 points. The max epoch count is set to 20.
### Datasets
Our experimental datasets includes three standard datasets with different application scenarios. The variety among these public 3D datasets definitely helps us to evaluate the performance of our method in a comprehensive manner. The main properties of these datasets are summarized in Table \[tabel\_dataset\].
These dataset are also injected with five levels of Gaussian noise (i.e., from 0.1 mr to 0.5 mr Gaussian noise) and four levels of mesh decimation (i.e., $\frac{1}{2}$, $\frac{1}{4}$, $\frac{1}{8}$ and $\frac{1}{16}$ of original mesh resolution). Here, the unit mr denotes mesh resolution. Remarkably, *the noise-free BR dataset is used to train our LRF-Net, the rest noisy data in the BR dataset and data in the UWA3M dataset and the UWAOR dataset are used for testing*.
[c|ccc]{} **Dateset** & BR & UWA3M & UWAOR\
**Scenario** & Retrieval & Registration & Recogntion\
**Challenge** &
----------
Gaussian
noise
----------
: Experimental datasets and inherited properties[]{data-label="tabel_dataset"}
&
----------------
holes, missing
region, and
self-occlusion
----------------
: Experimental datasets and inherited properties[]{data-label="tabel_dataset"}
&
---------------
clutter
and occlusion
---------------
: Experimental datasets and inherited properties[]{data-label="tabel_dataset"}
\
**\# Models** & 6 & 4 & 5\
**\# Scenes** & 18 & 75 & 50\
& 18 & 75 & 188\
### Compared Methods
We compare our LRF-Net with several existing LRF methods for a through evaluation. Specifically, the compared methods are proposed by Mian et al. [@mian2010repeatability], Tombari et al. [@shot2010], Petrelli et al. [@petrelli2012repeatable], Guo et al. [@rops2013l] and Yang et al. [@yang2017toldi], respectively. We dub them as *Mian*, *Tombari*, *Petrelli*, *Guo*, and *Yang*, respectively. To compare fairly, we keep the support radius of all the LRFs as 15 mr. The properties of these LRFs are shown in Table \[tabel\_param\].
To evaluate the local shape description performance of our method, we replace the LRF in four LRF-based descriptors (i.e., snapshots [@malassiotis2007snapshots], SHOT [@shot2010], RoPS [@rops2013l] and TOLDI [@yang2017toldi]) and assess the performance variations. To measure the 6-DoF pose estimation performance of our method, we adapt LRF-Net to the RANSAC pipeline and compare with the original RANSAC [@fischler1981random].
**Method** *Mian* *Tombari* *Guo* *Petrelli* *Yang* *Ours*
--------------- -------- ----------- ------- ------------ -------- --------
**Category** CA CA CA PSD PSD PSD
**Date type** P P M P P P
**Weight** $-$ H H H H L
: Properties of six LRF methods. H and L respectively represent hand-crafted and learned methods for point weight calculation; P and M respectively denote point cloud and mesh.[]{data-label="tabel_param"}
Performance Evaluation of LRF-Net
---------------------------------
### Repeatability Performance
We evaluate the repeatability of all LRFs via the popular $MeanCos$ [@shot2010] metric, which measures overall angular error between two LRFs. The repeatability results of evaluated LRFs are shown in Fig. \[fig\_meancos\] and Fig. \[fig\_robustness\]. Several observations can be made from these figures.
First, as witnessed by Fig. \[fig\_meancos\], our LRF together with *Tombari*, *Petrelli*, and *Yang* achieve decent performance on the BR dataset. On the UWA3M and UWAOR datasets, our LRF-Net achieves the best performance. Second, as shown in Fig. \[fig\_robustness\](a), LRF-Net and *Tombari* achieve a comparably stable performance on the BR dataset with respect to different levels of Gaussian noise. Fig. \[fig\_robustness\](b) and Fig. \[fig\_robustness\](c) indicate that LRF-Net achieves the best performance under all levels of Gaussian noise on the UWA3M and UWAOR datasets, surpassing the others by a very significant gap. Note that UWA3M and UWAOR datasets also include nuisances such as clutter, self-occlusion, and occlusion. Third, results in Fig. \[fig\_robustness\](d)-(f) suggest that LRF-Net is the best competitor with $\frac{1}{2}$, $\frac{1}{4}$, and $\frac{1}{8}$ mesh decimation on all datasets.
These results clearly demonstrate the strong robustness of our LRF-Net with respect to Gaussian noise, mesh decimation, clutter, and occlusion. The reasons are at least twofold. One is that all points are leveraged to generate the critical $x$-axis, which guarantees the robustness to Gaussian noise and low level mesh decimation. The other is that a LRF-Net can learn stable and informative weights for neighboring points. It can improve the robustness of LRF-Net to common nuisances.
### Local Shape Description Performance
We further evaluate our LRF-Net by replacing the LRFs in four LRF-based descriptors (i.e., snapshots, SHOT, RoPS, and TOLDI) with our LRF-Net. Then we compare their descriptor matching performance measured via recall vs. 1-precision curve (RPC) [@guo2016comprehensive; @shot2010]. Notably, the original LRF methods employed by snapshots, SHOT, RoPS, and TOLDI are *Mian*, *Tombari*, *Guo* *Yang*, respectively. We conduct this experiment on the original BR, UWA3M, and UWAOR datasets. Fig. \[fig\_RPC\] reports the RPC results of the all tested descriptors.
As witnessed by the figure, all LRF-based descriptors equipped with our LRF-Net outperform their original versions. Specifically, snapshots achieves a dramatic performance improvement with our LRF-Net on the BR dataset; the performance of SHOT also climbs significantly on the UWA3M and UWAOR datasets with the help of the proposed LRF-Net. Therefore, we can draw a conclusion that LRF plays an important role in local shape description, where a repeatable LRF can effectively improve the description performance of an LRF-based descriptor without changing its feature representation. It also indicates that the proposed LRF-Net can bring positive impacts on a number of existing local shape descriptors.
### 6-DoF Pose Estimation Performance
A general 6-DoF pose estimation process with local descriptors is achieved by correspondence generation and pose estimation from correspondences with potential outliers [@derpanis2010overview]. RANSAC is arguablly the de facto 6-DoF pose estimator in many applications. However, a key limitation of RANSAC is that the computational complexity of RANSAC is $O(n^3)$ and estimating a reasonable pose requires a huge number of iterations. With LRFs, a single correspondence is able to generate a 6-DoF pose, decreasing the computational complexity from $O(n^3)$ to $O(n)$. Therefore, we apply LRF-Net to 6-DoF pose estimation, following a RANSAC-fashion pipeline. The difference is that we sample one correspondence per iteration. Two criteria, i.e., the rotation error $err_r$ between our predicted rotation $R$ and the ground-truth one $R_{GT}$, and the translation error $err_t$ between the predicted translation vector $T$ and the ground truth one $T_{GT}$ [@mian2006novel], are employed for evaluating the performance of 6-DoF pose estimation.
The initial feature correspondence set is generated by first matching TOLDI (equipped with our LRF-Net) descriptors and keeping 100 correspondences with the highest similarity scores. 100 and 1000 iterations are assigned to our method and RANSAC. The average rotation errors and translation errors of the two estimators on three experimental datasets are shown in Table \[tabel\_pose\].
-- -- ------------------------------------- -- --
**BR & **UWA3M & **UWAOR\
& $err_t$ & 0.000 & 7.929 & 9.513\
& $err_r$ & 0.0298 & 0.696 & 0.769\
& $err_t$ & 0.000 & 6.088 & 4.392\
& $err_r$ & 0.0239 & 0.608 & 0.405\
******
-- -- ------------------------------------- -- --
: 6-DoF Pose estimation performance on three experimental datasets.[]{data-label="tabel_pose"}
Two salient observations can be made from the table. First, both RANSAC and our method manage to achieve accurate pose estimation results on the BR dataset that contains point cloud pairs with large overlapping ratios. However, our method only needs $\frac{1}{10}$ of the iterations required for RANSAC. Second, on more challenging datasets, i.e., UWA3M and UWAOR, our method significantly outperforms RANSAC. This demonstrates that LRF-Net can improve the accuracy and efficiency of RANSAC for 6-DoF pose estimation simultaneously.
Analysis Experiments
--------------------
### Verifying the Rationality of LRF-Net
To verify the rationality of the main technique components of our LRF-Net, we conduct the following experiments. First, in order to verify the choice of weighted vector-sum operation for $x$-axis calculation, we test the approach using the vector with the maximum weight as the $x$-axis (dubbed “Max”). Second, to demonstrate that the axes of LRF is not suitable to be directly regressed, we compare our method with the one regressing $x$-axis via a network shown in the left of Fig. \[fig\_ablation\] (dubbed “DR”). The results are shown in the right of Fig. \[fig\_ablation\].
Clearly, LRF-Net achieves the best performance among tested methods. It verifies that learning weights rather than directly learning axes is more reasonable. In addition, vector-sum is more appropriate for integrating projection vectors with learned weights for LRF-Net.
### Visualization
Fig. \[fig\_weight\] visualizes the learned weights by our LRF-Net for several sample local surfaces, which presents two interesting findings. First, closer points do not seem to have greater contributions. It is a common assumption for many existing CA- and PSD-based LRF methods, including *Tombari*, *Guo*, and *Yang*, that closer points should have greater weights. However, they are inferior to our LRF-Net in terms of repeatability performance. Second, $x$-axis estimation is generally determined by a particular area, rather than a single salient point as employed by many PSD-based methods, e.g., *Petrelli*. These visualization results also demonstrate our opinion that each neighboring point in the local surface gives a unique contribution to LRF construction.
Conclusion
==========
In this paper, we have proposed LRF-Net, a learned LRF for 3D local surface that is repeatable and robust to a number of nuisances. LRF-Net assumes that each neighboring point in the local surface gives a unique contribution to LRF construction and measure such contributions via learned weights. Experiments showed that our LRF-Net outperforms many state-of-the-art LRF methods on datasets addressing different application scenarios. In addition, LRF-Net can significantly boost the local shape description and 6-DoF pose estimation performance. In the future, we expect further improving the LRF-Net by considering RGB cues and multi-scale geometric information.
[^1]: Contact Author
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop a simple and accurate method to solve fractional variational and fractional optimal control problems with dependence on Caputo and Riemann–Liouville operators. Using known formulas for computing fractional derivatives of polynomials, we rewrite the fractional functional dynamical optimization problem as a classical static optimization problem. The method for classical optimal control problems is called Ritz’s method. Examples show that the proposed approach is more accurate than recent methods available in the literature.'
author:
- Salman Jahanshahi
- 'Delfim F. M. Torres'
date: 'Submitted: 31-Dec-2014 / Revised: 21-Oct-2015 and 17-Jan-2016 / Accepted: 24-Jan-2016.'
title: A Simple Accurate Method for Solving Fractional Variational and Optimal Control Problems
---
Introduction {#sec1}
============
In view of its history, fractional (non-integer order) calculus is as old as the classical calculus [@MR2218073; @MR1347689; @MR3181071]. Roughly speaking, there is just one definition of fractional integral operator, which is the Riemann–Liouville fractional integral. There are, however, several definitions of fractional differentiation, e.g., Caputo, Riemann–Liouville, Hadamard, and Wily differentiation [@MR3224387; @MR2960307]. Each type of fractional derivative has its own properties, which make richer the area of fractional calculus and enlarge its range of applicability [@MR2768178; @MR3188372].
There are two recent research areas where fractional operators have a particularly important role: the fractional calculus of variations and the fractional theory of optimal control [@sal34; @book:adv:FCV; @MR2984893]. A fractional variational problem is a dynamic optimization problem, in which the objective functional, as well as its constraints, depends on derivatives or integrals of fractional order, e.g., Caputo, Riemann–Liouville or Hadamard fractional operators. This is a generalization of the classical theory, where derivatives and integrals can only appear in integer orders. If at least one non-integer (fractional) term exists in its formulation, then the problem is said to be a fractional variational problem or a fractional optimal control problem. The theory of the fractional calculus of variations was introduced by Riewe in 1996, to deal with nonconservative systems in mechanics [@sal1; @sal2]. This subject has many applications in physics and engineering, and provides more accurate models of physical phenomena. For this reason, it is now under strong development: see [@sal5; @sal3; @sal4; @MR3103208; @MR3162654; @MR3200762] and the references therein. For a survey, see [@MR3221831].
There are two main approaches to solve problems of the fractional calculus of variations or optimal control. One involves solving fractional Euler–Lagrange equations or fractional Pontryagin-type conditions, which is the indirect approach; the other involves addressing directly the problem, without involving necessary optimality conditions, which is the direct approach. The emphasis in the literature has been put on indirect methods [@book:adv:FCV; @MR2984893; @sal6; @sal7]. For direct methods, see [@sal34; @sal3]. Furthermore, Almeida et al. developed direct numerical methods based on the idea of writing the fractional operators in power series, and then approximating the fractional problems with classical ones [@MyID:294; @sal8]. In this paper, we use a different approach to solve fractional variational problems and fractional optimal control problems, based on Ritz’s direct method. The idea is to restrict admissible functions to linear combinations of a set of known basis functions. We choose basis functions in such a way that the approximated function satisfies the given boundary conditions. Using the approximated function and its derivatives whenever needed, we transform the functional into a multivariate function of unknown coefficients. Recently, Dehghan et al. in [@sal9] used the Rayleigh–Ritz method, based on Jacobi polynomials, to solve fractional optimal control problems. Several illustrative examples show that our results are more accurate and more useful than the ones introduced in [@sal8; @sal9].
The paper is organized as follows. In Section \[sec2\], we present some necessary preliminaries on fractional calculus. In Section \[sec3\], we investigate Ritz’s method for solving three kinds of fractional variational problems. In Section \[sec4\], we solve five examples of fractional variational problems and three examples of fractional optimal control problems, comparing our results with previous methods available in the literature. The main conclusions are given in Section \[sec:conc\].
Preliminaries and Notations About Fractional Calculus {#sec2}
=====================================================
Riemann–Liouville fractional integrals are a generalization of the $n$ fold integral, $n\in \mathbb{N}$, to real value numbers. Using the usual notation in the theory of fractional calculus, we define the Riemann–Liouville fractional integrals as follows.
The left and right Riemann–Liouville fractional integrals of order $\alpha>0$ of a given function $f$ are defined by $$_aI_x^\alpha f(x):=\frac{1}{\Gamma(\alpha)}
\int_a^x(x-t)^{\alpha-1}f(t)dt$$ and $$_xI_b^{\alpha}f(x):=\frac{1}{\Gamma(\alpha)}\int_x^b(t-x)^{\alpha-1}f(t)dt,$$ respectively, where $\Gamma$ is Euler’s gamma function, that is, $$\Gamma(x):=\int_0^\infty t^{x-1}e^{-t}dt,$$ and $a<x<b$.
The left Riemann–Liouville fractional operator has the following properties: $${_aI_x^\alpha} {_aI_x^\beta}
= {_aI_x}^{\alpha+\beta},
\quad {_aI_x^\alpha} {_aI_x^\beta}
= {_aI_x^\beta} {_aI_x^\alpha},$$ $$\label{eq11}
_aI_x^\alpha (x-a)^n=\frac{\Gamma(n+1)}{\Gamma(\alpha+n+1)}(x-a)^{\alpha+n}, \quad x>a,$$ where $\alpha,\beta\geq 0$ and $n\in \mathbb{N}_0=\{0,1,\ldots\}$. Similar relations hold for the right Riemann–Liouville fractional operator. Now, using the definition of fractional integral, we define two kinds of fractional derivatives.
Let $\alpha>0$ with $n-1<\alpha\leq n$, $n\in\mathbb{N}$. The left and right Riemann–Liouville fractional derivatives of order $\alpha>0$ of a given function $f$ are defined by $$_aD_x^{\alpha}f(x):=D^n(_aI_x^{n-\alpha})f(x)
=\frac{1}{\Gamma(n-\alpha)}
\frac{d^n}{dx^n}\int_a^x(x-t)^{n-\alpha-1}f(t)dt$$ and $$_xD_b^\alpha f(x):=(-1)^nD^n(_xI_b^{n-\alpha})f(x)
=\frac{(-1)^n}{\Gamma(n-\alpha)}
\frac{d^n}{dx^n}\int_x^b(t-x)^{n-\alpha-1}f(t)dt,$$ respectively.
The following relations hold for the left Riemann–Liouville fractional derivative: $${_aD_x^\alpha} {_aD_x^\beta}
= {_aD_x}^{\alpha+\beta},
\quad {_aD_x^\alpha} {_aD_x^\beta}
= {_aD_x^\beta} {_aD_x^\alpha},$$ $$_aD_x^\alpha k= \frac{k (x-a)^\alpha}{\Gamma(1-\alpha)},$$ $$\label{eq12}
_aD_x^\alpha (x-a)^n=\frac{\Gamma(n+1)}{\Gamma(n-\alpha+1)}(x-a)^{n-\alpha}, \quad x>a,$$ where $\alpha,\beta\geq 0$, $k$ is a constant and $n\in \mathbb{N}_0$. Similar relations hold for the right Riemann–Liouville fractional derivative.
Another type of fractional derivative, which also uses the Riemann–Liouville fractional integral in its definition, was proposed by Caputo [@sal16].
Let $\alpha>0$ with $n-1<\alpha\leq n$, $n\in\mathbb{N}$. The left and right Caputo fractional derivatives of order $\alpha>0$ of a given function $f$ are defined by $$_a^CD_x^\alpha f(x):=(_aI_x^{n-\alpha})D^n f(x)
=\frac{1}{\Gamma(n-\alpha)}\int_a^x(x-t)^{n-\alpha-1}f^{(n)}(t)dt$$ and $$_x^CD_b^\alpha f(x):=(-1)^n(_xI_b^{n-\alpha})D^n f(x)
=\frac{(-1)^n}{\Gamma(n-\alpha)}\int_x^b(t-x)^{n-\alpha-1}f^{(n)}(t)dt,$$ respectively.
The following relations hold for the left Caputo fractional derivative: $${_a^CD_x^\alpha} \, {_a^CD_x^\beta}
= {{_a^CD_x^{\alpha+\beta}}},
\quad {_a^CD_x^\alpha} \, {_a^CD_x^\beta}
= {_a^CD_x^\beta} \, {_a^CD_x^\alpha},$$ $$_a^CD_x^\alpha k = 0,$$ $$\label{eq13}
_a^CD_x^\alpha (x-a)^n=\left\{
\begin{array}{ll}
\frac{\Gamma(n+1)}{\Gamma(n-\alpha+1)}(x-a)^{n-\alpha},
& n\geq [\alpha],\\
0, & n< [\alpha],
\end{array} \right.$$ where $\alpha,\beta\geq 0$, $k$ is a constant, $x>a$, $n\in \mathbb{N}_0$, and $[\cdot]$ is the ceiling function, that is, $[\alpha]$ is the smallest integer greater than or equal to $\alpha$. Similar relations hold for the right Caputo fractional derivative. Moreover, the following relations between Caputo and Riemann–Liouville fractional derivatives hold: $$_a^CD_x^\alpha f(x)={_aD_x^\alpha} f(x)
-\sum_{k=0}^{n-1}\frac{f^{(k)}(a)}{\Gamma(k-\alpha+1)}(x-a)^{k-\alpha}$$ and $$_x^CD_b^\alpha f(x)={_xD_b^\alpha} f(x)
-\sum_{k=0}^{n-1}\frac{f^{(k)}(b)}{\Gamma(k-\alpha+1)}(b-x)^{k-\alpha}.$$ Therefore, if $f\in C^n[a,b]$ and $f^{(k)}(a)=0$, $k=0,1,\ldots,n-1$, then $_a^CD_x^\alpha f= {_aD_x^\alpha} f$; if $f^{(k)}(b)=0$, $k=0,1,\ldots,n-1$, then $_x^CD_b^\alpha f={_xD_b^\alpha} f$.
Main Results {#sec3}
============
Consider the following fractional variational problem: find $y\in C^n[a,b]$ in such a way to minimize or maximize the functional $$\label{eq3.1}
J\{y\}=\int_a^b L(t,y(t),D^\alpha y(t))dt, \quad n-1\leq \alpha<n,$$ subject to boundary conditions $$\label{eq3.2}
y^{(k)}(a)=u_{k,a},\quad y^{(k)}(b)=u_{k,b}, \quad k=0,1,\ldots,n-1,$$ where $L$ is the Lagrangian, assumed to be continuous with respect to all its arguments, $D^\alpha$ is a fractional operator (left or right Riemann–Liouville fractional integral or derivative or left or right Caputo fractional derivative), and $u_{k,a}$ and $u_{k,b}$, $k=0,1,\ldots,n-1$, are given constants. To solve problem – with our method, we need to recall the following classical theorems from approximation theory.
\[thm:sw\] Let $K$ be a compact metric space and $A\subset C(K, R)$ a unital sub-algebra, which separates points of $K$. Then, $A$ is dense in $C(K, R)$.
\[thm:wa\] If $f \in C([a, b], \mathbb{R})$, then there is a sequence of polynomials $P_n(x)$ that converges uniformly to $f(x)$ on $[a, b]$.
\[theorem3\] Let $P_N(x):=\sum_{i=0}^N c_i (x-a)^i$ be a polynomial. Then, $$\label{eq14}
_aI_x^\alpha P_N(x)=\sum_{i=0}^N c_i
\frac{\Gamma(i+1)}{\Gamma(\alpha+i+1)}(x-a)^{\alpha+i},$$ $$\label{eq15}
_aD_x^\alpha P_N(x)=\sum_{i=0}^N c_i
\frac{\Gamma(i+1)}{\Gamma(i-\alpha+1)}(x-a)^{i-\alpha},$$ $$\label{eq16}
_a^CD_x^\alpha P_N(x)
=\sum_{i = [\alpha]}^{N}\frac{\Gamma(i+1)}{\Gamma(i-\alpha+1)}(x-a)^{i-\alpha},$$ for $x>a$.
Follows from relations , and and the linearity property of the fractional operators. $\qed$
For solving fractional variational problems involving right fractional operators, we use the following theorem.
\[theorem4\] Let $P_N(x):=\sum_{i=0}^{N} c_i (b-x)^i$ be a polynomial. Then, $$\label{eq30}
_xI_b^\alpha P_N(x)=\sum_{i=0}^N c_i
\frac{\Gamma(i+1)}{\Gamma(\alpha+i+1)}(b-x)^{\alpha+i},$$ $$\label{eq31}
_xD_b^\alpha P_N(x)=\sum_{i=0}^N c_i
\frac{\Gamma(i+1)}{\Gamma(i-\alpha+1)}(b-x)^{i-\alpha},$$ $$\label{eq32}
_x^CD_b^\alpha P_N(x)
=\sum_{i=[\alpha]}^{N}\frac{\Gamma(i+1)}{\Gamma(i-\alpha+1)}(b-x)^{i-\alpha},$$ for $x<b$.
Similar to the proof of Theorem \[theorem3\]. $\qed$
Without any loss of generality, from now on we consider $a = 0$, $b = 1$ and $t\in [0, 1]$ in the fractional variational problem –.
Fractional Variational Problems Involving Left Operators {#sec:MR:FVP:LO}
--------------------------------------------------------
Consider the fractional variational functional involving a left fractional operator, subject to boundary conditions . To find the function $y(x)$ that solves problem –, we put $$\label{eq:yN:L}
y(x)\simeq y_N(x)=\sum_{i=0}^N c_i x^i.$$ Then, by substituting and relations – into , we obtain $$\label{eq18}
J[y]=J[c_0,c_1,\ldots ,c_N]=\int_0^1 L(x,y_N(x),D^\alpha y_N(x))dx$$ subject to boundary conditions $$\label{eq19}
y_N^{(k)}(0)=u_{k,0}, \quad y_N^{(k)}(1)=u_{k,1},
\quad k=0,1,\ldots, n-1,$$ which is an algebraic function of unknowns $c_i$, $i=0,1,\ldots, N$. To optimize the algebraic function $J$, we act as follows. We should find $c_0,c_1,\ldots,c_N$, such that $y_N$ satisfies boundary conditions . This means that the following relations must be satisfied: $$\label{eq20}
y_N^{(k)}(0)=u_{k,0}, \quad y_N^{(k)}(1)=u_{k,1},
\quad k=0,1,\ldots,n-1.$$ Using the well known Hermite interpolation and the relations , we calculate $c_0,c_1,\ldots,c_n$. For obtaining the values of $c_{n+1},c_{n+2},\ldots,c_N$, firstly we use the $N+1$ point Gauss–Legendre quadrature rule, which is exact for every polynomial of degree up to $2N+2$. Secondly, we calculate the exact value of the integral in the right-hand side of . Then, according to differential calculus, we must solve the following system of equations: $$\label{eq21}
\frac{\partial J}{\partial c_j}=0,
\quad j=n+1,n+2,\ldots,N,$$ which, depending on the form of $L$, is a linear or nonlinear system of equations. Furthermore, we choose the value of $N$ such that $|y_{N+1}(x)-y_N(x)|\cong 0(x)$, where $0(x)$ is the null polynomial.
Fractional Variational Problems Involving Right Operators
---------------------------------------------------------
Now consider the fractional variational functional involving a right fractional operator, subject to boundary conditions . To find function $y(x)$ that solves problem –, we put $$\label{eq:yN:R}
y(x)\simeq y_N(x)=\sum_{i=0}^N c_i (1-x)^i.$$ Then, substituting and relations – into , we obtain $$\label{eq:alg:R}
J[y]=J[c_0,c_1,\ldots ,c_N]=\int_0^1 L(x,y_N(x),D^\alpha y_N(x))dx$$ subject to boundary conditions $$y_N^{(k)}(0)=u_{k,0}, \quad y_N^{(k)}(1)=u_{k,1}, \quad k=0,1,\ldots, n-1,$$ which is an algebraic function of unknowns $c_i$, $i=0,1,\ldots, N$. To optimize the algebraic function , we act as explained in Section \[sec:MR:FVP:LO\].
Fractional Optimal Control Problems
-----------------------------------
A fractional optimal control problem requires finding a control function $u(t)$ and the corresponding state trajectory $x(t)$, that minimizes (or maximizes) a given functional $$\label{eq9}
J\{x,u\}=\int_a^b L\left(t,x(t),u(t)\right) dt$$ subject to a fractional dynamical control system $$\label{eq7}
D^\alpha x(t)=f(t,x(t),u(t))$$ and boundary conditions $$\label{eq:foc:bc}
x^{(k)}(a)=u_{k,a},\quad x^{(k)}(b)=u_{k,b},
\quad k=0,1,\ldots,n-1,$$ where $D^\alpha$ is a fractional operator, $\alpha$ is a positive real number, and $f$ and $L$ are two known functions. For more details, see [@MyID:294; @sal14; @MR2433010; @MR2386201; @sal15] and the references therein.
Here we restrict our attention to those fractional optimal control problems –, for which one can solve with respect to $u$ and write $$\label{eq8}
u(t)=g\left(t,x(t),D^\alpha x(t)\right).$$ Then, by substituting into , we obtain the following fractional variational problem (FVP): find function $x(t)$ that extremizes the functional $$\label{3q10}
J\{x\}=\int_a^b L\left(t,x(t),g(t,x(t),D^\alpha x(t))\right) dt$$ subject to boundary conditions $$\label{3q10:bc}
x^{(k)}(a)=u_{k,a},\quad x^{(k)}(b)=u_{k,b},
\quad k=0,1,\ldots,n-1.$$ We solve the FVP – by using the method illustrated in previous sections, and then we find control function $u(t)$ by using .
Similarly to classical Ritz’s method, the trial functions are selected to meet boundary conditions (and any other constraints). The exact solutions are not known; and the trial functions are parametrized by adjustable coefficients, which are varied to find the best approximation for the basis functions used. The choice of the basis functions depends on the solution space. Since our problems involve finding $C^n$ solutions, we choose the basis functions to be polynomials. If the solution space is another one, like the $L_p$ space or the space of harmonic functions, then we should choose some basis functions like Fourier or wavelet basis.
For choosing the number $N$ one needs to decide on the required accuracy. One can stop when the difference between two consecutive approximations $y_N$ and $y_{N-1}$ or their respective functional values is smaller than a desired tolerance, that is, when $\left\|y_N - y_{N-1}\right\| \leq \varepsilon$ or when $\left|J[y_N] - J[y_{N-1}]\right| \leq \varepsilon$ for some given $\varepsilon$. Note that for a minimization (maximization) problem, the functional $J$ is always a non-increasing (non-decreasing) function of $N$: $J[y_N] \leq J[y_{N-1}]$ ($J[y_N] \geq J[y_{N-1}]$). Moreover, if $y$ is the (unknown) exact solution of the fractional variational problem at hand, then $y_N$ converges uniformly to $y$ (Theorems \[thm:sw\] and \[thm:wa\]).
Throughout the paper, $N$ is greater than $n$.
In the next section, we solve five fractional problems of the calculus of variations and three fractional optimal control problems. Six of the eight problems involve left fractional operators, while the other two involve right operators.
Illustrative Examples {#sec4}
=====================
We begin by solving two problems of the calculus of variations involving a left Riemann–Liouville fractional derivative. These problems were recently investigated by Almeida et al. in [@sal34]. We also solve three examples involving left and right Caputo fractional derivatives, which were recently considered by Dehghan et al. in [@sal9], and we compare the results. Finally, we solve three fractional optimal control problems that were investigated before in [@MyID:294; @sal15; @sal17]. For all examples, the error between the exact solution $y$ and the approximate solution $y_N$, found using our method, is computed as follows: $$\label{error}
Error\{y,y_N\}=\int_0^1(y(x)-y_N(x))^2 dx.$$ The results show that our method is very simple but very accurate, providing better results than those found in the literature.
[@sal34; @sal8] \[example3.1\] Let $\alpha \in ]0, 1[$. Consider the following FVP: $$\label{eq3.10}
\text{minimize} \quad J\{y\}
=\int_0^1 \left(_0D_x^\alpha y(x)-(y')^2(x)\right) dx,$$ subject to $$\label{eq3.10:bc}
y(0)=0, \quad y(1)=1.$$ The exact solution to problem – is $$y(x)=-\frac{1}{2\Gamma(3-\alpha)}(1-x)^{2-\alpha}
+\left(1-\frac{1}{2\Gamma(3-\alpha)}\right)x
+\frac{1}{2\Gamma(3-\alpha)}$$ [@sal34; @sal8]. With the given boundary conditions , we consider $$y_N(x)=\left(1-\sum_{i=2}^N c_i\right) x
+\sum_{i=2}^N c_i x^i.$$ Using our method, we calculate the $c_i$’s by solving the system of equations . For $\alpha=0.5$ and $N=2$, we have $$y(x)=0.376126 - 0.376126 (1 - x)^{1.5} + 0.623874 x,$$ $$y_2(x)=1.22568 x - 0.225676 x^2.$$ Figure \[Fig:1\] plots the result for $\alpha=0.5$ and $N=2$. The errors computed with for different values of $\alpha$ and $N$ are presented in Table \[Table3.3\].
![[Exact solution to problem – of Example \[example3.1\] with $\alpha=0.5$ (the solid line) versus our approximate solution with $N=2$ (the dashed line).]{}[]{data-label="Fig:1"}](Figure1.eps)
$\alpha$ $N=3$ $N=4$ $N=5$ $N=6$
---------- -------------------- -------------------- -------------------- --------------------- -- --
$0.25$ $1.7\times10^{-7}$ $1.9\times10^{-8}$ $3.6\times10^{-9}$ $9.1\times10^{-10}$
$0.5$ $9.7\times10^{-6}$ $1.5\times10^{-7}$ $3.5\times10^{-8}$ $1.1\times10^{-8}$
$0.75$ $1.6\times10^{-6}$ $3.4\times10^{-7}$ $9.9\times10^{-8}$ $3.5\times10^{-8}$
: [Errors obtained for problem – of Example \[example3.1\] when computed with for $\alpha=0.25,0.5,0.75$ and $N=3,4,5,6$.]{}[]{data-label="Table3.3"}
Our approximate results are better than the ones presented in [@sal34; @sal8].
[@sal34] \[example3.2\] Let $\alpha=0.5$ and consider the minimization problem $$\label{eq:J:ex3.2}
\text{minimize} \quad J\{y\}=\int_0^1 \left(_0D_x^\alpha y(x)-\frac{16\Gamma(6)}{\Gamma(5.5)}x^{4.5}
+\frac{20\Gamma(4)}{\Gamma(3.5)}x^{2.5}-\frac{5}{\Gamma(2.5)}x^{0.5}\right)^4 dx,$$ subject to the boundary conditions $$\label{eq:J:ex3.2:bc}
y(0)=0, \quad y(1)=1.$$ The minimizer to the FVP – is given by $y(x)=16x^5-20x^3+5x$. For $N=5$, we obtain $$y_5(x)=5x + (2.3995\times 10^{-10}) x^2 {- 20} x^3 + (7.70126\times 10^{-10})x^4 + 16x^5.$$ Figure \[Fig:3\] shows that our results are more accurate than the results presented in [@sal34].
\
Below we solve three problems of the calculus of variations, which were recently solved by Dehghan et al. in [@sal9]. Our results show that our method is also more accurate than the method introduced in [@sal9].
[@sal9] \[example3.3\] Consider the following FVP: $$\label{eq:J:ex3.3}
\text{minimize}\quad J\{y\}=\frac{1}{2}\int_0^1(_0^CD_x^\alpha y(x)-f(x))^2dx,
\quad 0<\alpha <1,$$ where $f(x)$ is given by $$f(x)=\frac{\Gamma(\beta +1)}{\Gamma(1+\beta -\alpha)}x^{\beta -\alpha},$$ subject to the boundary conditions $$\label{eq:J:ex3.3:bc}
y(0)=0, \quad y(1)=1.$$ The exact solution to this problem is $y(x)=x^\beta$ [@sal9]. In Figure \[Fig:4\], the exact solution for $\alpha=0.5$ and $\beta=2.5$ versus our numerical solution for $N=3$ is plotted. Note that for $\beta=k$, $k \in \mathbb{N}$, the error is equal to zero when $N\geq k$.
![[Exact solution to problem – of Example \[example3.3\] with $\alpha=0.5$ and $\beta=2.5$ (the solid line) versus our approximate solution with $N=3$ (the dashed line).]{}[]{data-label="Fig:4"}](Figure4.eps)
[@sal9] \[3x\] Consider the following FVP, depending on a right Caputo fractional derivative: $$\label{eq:ex:3x}
\text{minimize}\quad J\{y\}=\frac{1}{2}\int_0^1(_x^CD_1^\alpha y(x)-1)^2dx,
\quad 0<\alpha <1,$$ subject to boundary conditions $$\label{eq:ex:3x:bc}
y(0)=1+\frac{1}{\Gamma(1+\alpha)}, \quad y(1)=1.$$ The exact solution to – is given by $$y(x)=1+\frac{(1-x)^\alpha}{\Gamma(1+\alpha)}.$$ An approximate solution obtained with our method is shown in Figure \[Fig:5\]. See Table \[Table3.4\] for the error of our approximations.
![[Exact solution to problem – of Example \[3x\] with $\alpha=0.75$ (the solid line) versus our approximate solution with $N=6$ (the dashed line).]{}[]{data-label="Fig:5"}](Figure5.eps)
$\alpha$ $N=2$ $N=3$ $N=4$ $N=5$ $N=6$
---------- -------------------- -------------------- -------------------- -------------------- -------------------- --
$0.75$ $5.7\times10^{-4}$ $1.2\times10^{-4}$ $3.6\times10^{-5}$ $1.6\times10^{-5}$ $7.2\times10^{-6}$
: [Errors obtained for problem – of Example \[3x\], computed with for $\alpha=0.75$ and $N=2,3,4,5,6$.]{}[]{data-label="Table3.4"}
[@sal9] \[example4\] As a second example involving a fractional right operator, consider the following FVP with $0<\alpha <1$: $$\label{eq:prb:ex4}
\text{minimize}\quad J\{y\}=\int_0^1\left(_x^CD_1^\alpha y(x)+y(x)-(1-x)^\beta
-\frac{\Gamma(\beta +1)(1-x)^{\beta -\alpha}}{\Gamma(\beta -\alpha +1)}\right)^2 dx,$$ subject to boundary conditions $$\label{eq:prb:ex4:bc}
y(0)=1, \quad y(1)=0.$$ In this case the exact solution is $y(x)=(1-x)^\beta$ [@sal9]. Figure \[fig6\] shows the exact solution for $\alpha=0.39$ and $\beta=3$ versus our numerical solution with $N=3$. The errors of our approximations, for different values of $(\alpha,\beta)$, $\beta\neq3$, are listed in Table \[Table3.5\]. As with Example \[example3.3\], the error for $\beta=k$, $k \in \mathbb{N}$, is equal to zero when we consider $N\geq k$.
![[Exact solution to problem – of Example \[example4\] with $\alpha=0.39$ and $\beta=3$ (the solid line) versus our approximate solution with $N=3$ (the dashed line).]{}[]{data-label="fig6"}](Figure6.eps)
$(\alpha,\beta)$ $N=3$ $N=4$ $N=5$
------------------ -------------------- -------------------- -------------------- -- -- --
$(0.39,1.5)$ $6.1\times10^{-6}$ $8.1\times10^{-7}$ $1.7\times10^{-8}$
$(0.39,2.5)$ $1.3\times10^{-6}$ $4.5\times10^{-8}$ $3.7\times10^{-9}$
$(0.59,1.5)$ $6.2\times10^{-6}$ $8.1\times10^{-7}$ $4.5\times10^{-9}$
$(0.59,2.5)$ $1.3\times10^{-6}$ $4.5\times10^{-8}$ $4.5\times10^{-9}$
: [Errors for problem – of Example \[example4\] when computed with for different values of $(\alpha,\beta)$ and $N=3,4,5$.]{}[]{data-label="Table3.5"}
We finish by applying our method to three fractional optimal control problems.
[@sal15; @sal17] \[exam5\] Consider the following fractional optimal control problem (FOCP): $$\label{eq:J:OC:ex5}
\text{minimize}\quad J\{x,u\}=\int_0^1(tu(t)-(\alpha +2)x(t))^2 dt,$$ subject to the dynamical fractional control system $$\label{eq:J:OC:ex5:cs}
x'(t)+{_0^CD_t^\alpha} x(t)=u(t)+t^2$$ and the boundary conditions $$\label{eq:J:OC:ex5:bc}
x(0)=0, \quad x(1)=\frac{2}{\Gamma(3+\alpha)}.$$ The exact solution is given by $$(x(t),u(t))=\left(\frac{2t^{\alpha+2}}{\Gamma(\alpha +3)},
\frac{2t^{\alpha+1}}{\Gamma(\alpha +2)}\right).$$ To solve this problem with our method, we take $$x_N(t)=\left(\frac{2}{\Gamma(3+\alpha)}
-\sum_{i=1}^N c_i\right)t+\sum_{i=1}^N c_i t^i.$$ The results for $\alpha=0.5$ and $N=3,4$ are plotted in Figure \[fig7\]. Table \[Table3.9\] shows the errors for $\alpha=0.5$ and $N=2,3,4$.
\
\
Errors $N=2$ $N=3$ $N=4$
------------------ -------------------- -------------------- -------------------- -- -- --
$Error\{x,x_N\}$ $1.2\times10^{-4}$ $7.1\times10^{-7}$ $2.8\times10^{-8}$
$Error\{u,u_N\}$ $7.1\times10^{-3}$ $9.6\times10^{-5}$ $7.9\times10^{-6}$
: [Errors for problem – of Example \[exam5\] when computed with for $\alpha=0.5$ and $N=2,3,4$.]{}[]{data-label="Table3.9"}
[@sal17] \[exam6\] Consider now the following FOCP: $$\label{eq:J:OC:ex6}
\text{minimize}\quad J\{x,u\}=\int_0^1(u(t)-x(t))^2 dt,$$ subject to the fractional dynamical control system $$\label{eq:J:OC:ex6:cs}
x'(t)+{_0^CD_t^\alpha} x(t)=u(t)-x(t)+\frac{6t^{\alpha +2}}{\Gamma(\alpha +3)}+t^3$$ and the boundary conditions $$\label{eq:J:OC:ex6:bc}
x(0)=0, \quad x(1)=\frac{6}{\Gamma(4+\alpha)}.$$ The exact solution to – is given by $$(x(t),u(t))=\left(\frac{6t^{\alpha+3}}{\Gamma(\alpha +4)},
\frac{6t^{\alpha+3}}{\Gamma(\alpha +4)}\right).$$ To solve problem – with our method, we take $$x_N(t)=\left(\frac{6}{\Gamma(4+\alpha)}
-\sum_{i=1}^N c_i\right)t+\sum_{i=1}^N c_i t^i.$$ The results for $\alpha=0.5$ and $N=3,4$ are plotted in Figure \[fig8\]. The errors for $\alpha=0.5$ and $N=2,3,4$ are shown in Table \[Table3.10\].
\
\
Errors $N=2$ $N=3$ $N=4$
------------------ --------------------- -------------------- -------------------- -- -- --
$Error\{x,x_N\}$ $7.3\times 10^{-4}$ $2.5\times10^{-6}$ $8.4\times10^{-9}$
$Error\{u,u_N\}$ $5.2\times 10^{-2}$ $3.9\times10^{-4}$ $2.2\times10^{-6}$
: [Errors for problem – of Example \[exam6\] when computed with for $\alpha=0.5$ and $N=2,3,4$.]{}[]{data-label="Table3.10"}
In [@MyID:294] the authors propose a direct method for solving fractional optimal control problems, which involves approximating the initial fractional order problem by a new one with integer order derivatives only. The latter problem is then discretized, by application of finite differences, and solved numerically. Our method is simpler and does not involve the solution of a nonlinear programming problem through AMPL and IPOPT. Moreover, as we see next, it provides a much better result when compared with the example given in [@MyID:294].
[@MyID:294] \[exam11\] Consider the FOCP of [@MyID:294]: $$\label{eq:exam11}
\mathrm{minimize}
\quad J\{x,u\}= \int_0^1 (u^2(t)- 4x(t))^2 dt,$$ subject to the control system $$\label{eq:exam12}
x'(t)+{_0^CD_t^{0.5}} x(t)=u(t)+ \frac{2}{\Gamma(2.5)} t^{1.5},
\quad t\in [0,1],$$ and the boundary conditions $$\label{eq:exam13}
x(0)=0, \quad \text{and} \quad x(1)=1.$$ The exact solution to – is given by $$(x(t),u(t))= (t^2,2t)$$ (see [@MyID:294]). To solve problem – with the method of this paper, we take $$x_N(t)= \left(1-\sum_{i=1}^N c_i\right)t+ \sum_{i=1}^N c_it^i.$$ In this case our method gives the exact solution $x(t) = t^2$ for $N=2$. This is in contrast with the results in [@MyID:294], for which the exact solution is not found. In fact, our method provides here better results with just 2 steps ($c_1 = 0$ and $c_2 = 1$) than the method introduced in [@MyID:294] with 100 steps.
According to relations and , accuracy depends on $N$. When the order $\alpha$ of the derivative changes, then the problem under consideration also changes. We were not able to find any general relation between $\alpha$ and the accuracy, that is, a general pattern on how $N$ changes with $\alpha$, for a fixed precision. This seems to depend on the particular situation at hand.
Conclusions {#sec:conc}
===========
We introduced a numerical method, based on Ritz’s direct method, to solve problems of the fractional calculus of variations and fractional optimal control. The idea of this approach is simple: by using some known basis functions, we construct an approximate solution, which is a linear combination of the basis functions, carrying out a finite-dimensional minimization among such linear combinations and approximating the exact solution of the fractional optimal control problem. From the simulation results, the proposed approach is surprisingly accurate, and the approximate solution is nearly a perfect match with the exact solution, which is superior to the results from previous numerical methods available in the literature.
This work is part of first author’s PhD project. It was partially supported by Islamic Azad University, Tehran, Iran; and CIDMA-FCT, Portugal, within project UID/MAT/04106/2013. Jahanshahi was also supported by a scholarship from the Ministry of Science, Research and Technology of the Islamic Republic of Iran, to visit the University of Aveiro, Portugal, and work with Professor Torres. The hospitality and the excellent working conditions at the University of Aveiro are here gratefully acknowledged. The authors are indebted to an anonymous referee for a careful reading of the original manuscript and for providing several suggestions, questions, and remarks. They are also grateful to the Editor-in-Chief, Professor Giannessi, and Ryan Loxton, for English improvements.
[99]{}
Kilbas, A.A., Srivastava, H.M., Trujillo, J.J.: Theory and applications of fractional differential equations. North-Holland Mathematics Studies, 204, Elsevier, Amsterdam (2006)
Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional integrals and derivatives. Translated from the 1987 Russian original. Gordon and Breach, Yverdon (1993)
Valério, D., Tenreiro Machado, J., Kiryakova, V.: Some pioneers of the applications of fractional calculus. Fract. Calc. Appl. Anal. 17:2, 552–578 (2014)
de Oliveira, E.C., Tenreiro Machado, J.A.: A review of definitions for fractional derivatives and integral. Math. Probl. Eng. 2014:238459, 6 pp (2014)
Ortigueira, M.D., Trujillo, J.J.: A unified approach to fractional derivatives. Commun. Nonlinear Sci. Numer. Simul. 17:12, 5151–5157 (2012)
Ortigueira, M.D.: Fractional calculus for scientists and engineers. Lecture Notes in Electrical Engineering, 84, Springer, Dordrecht (2011)
Tenreiro Machado, J.A., Baleanu, D., Chen, W., Sabatier, J.: New trends in fractional dynamics. J. Vib. Control 20:7, 963 (2014)
Almeida, R., Pooseh, S., Torres, D.F.M.: Computational methods in the fractional calculus of variations. Imp. Coll. Press, London (2015)
Malinowska, A.B., Odzijewicz, T., Torres, D.F.M.: Advanced methods in the fractional calculus of variations. Springer Briefs in Applied Sciences and Technology, Springer, Cham (2015)
Malinowska, A.B., Torres, D.F.M.: Introduction to the fractional calculus of variations. Imp. Coll. Press, London (2012)
Riewe, F.: Nonconservative Lagrangian and Hamiltonian mechanics. Phys. Rev. E (3) 53:2, 1890–1899 (1996)
Riewe, F.: Mechanics with fractional derivatives. Phys. Rev. E (3) 55:3, part B, 3581–3592 (1997)
Agrawal, O.P.: Fractional variational calculus in terms of Riesz fractional derivatives. J. Phys. A 40:24, 6287–6303 (2007)
Almeida, R., Torres, D.F.M.: Leitmann’s direct method for fractional optimization problems. Appl. Math. Comput. 217:3, 956–962 (2010) [arXiv:1003.3088]{}
Almeida, R., Torres, D.F.M.: Necessary and sufficient conditions for the fractional calculus of variations with Caputo derivatives. Commun. Nonlinear Sci. Numer. Simul. 16:3, 1490–1500 (2011) [arXiv:1007.2937]{}
Atanacković, T.M., Janev, M., Konjik, S., Pilipović, S., Zorica, D.: Expansion formula for fractional derivatives in variational problems. J. Math. Anal. Appl. 409:2, 911–924 (2014)
Baleanu, D., Garra, R., Petras, I.: A fractional variational approach to the fractional Basset-type equation. Rep. Math. Phys. 72:1, 57–64 (2013)
Bourdin, L., Odzijewicz, T., Torres, D.F.M.: Existence of minimizers for generalized Lagrangian functionals and a necessary optimality condition—application to fractional variational problems. Differential Integral Equations 27:7-8, 743–766 (2014) [arXiv:1403.3937]{}
Odzijewicz, T., Torres, D.F.M.: The generalized fractional calculus of variations. Southeast Asian Bull. Math. 38:1, 93–117 (2014) [arXiv:1401.7291]{}
Almeida, R., Khosravian-Arab, H., Shamsi, M.: A generalized fractional variational problem depending on indefinite integrals: Euler-Lagrange equation and numerical solution. J. Vib. Control 19:14, 2177–2186 (2013)
Blaszczyk, T., Ciesielski, M.: Numerical solution of fractional Sturm-Liouville equation in integral form. Fract. Calc. Appl. Anal. 17:2, 307–320 (2014)
Almeida, R., Torres, D.F.M.: A discrete method to solve fractional optimal control problems. Nonlinear Dynam. 80:4, 1811–1816 (2015) [arXiv:1403.5060]{}
Pooseh, S., Almeida, R., Torres, D.F.M.: Numerical approximations of fractional derivatives with applications. Asian J. Control 15:3, 698–712 (2013) [arXiv:1208.2588]{}
Dehghan, M., Hamedi, E.-A., Khosravian-Arab, H.: A numerical scheme for the solution of a class of fractional variational and optimal control problems using the modified Jacobi polynomials. J. Vib. Control, in press. DOI:10.1177/1077546314543727
Caputo, M.: Linear models of dissipation whose $Q$ is almost frequency independent. II, Fract. Calc. Appl. Anal. 11:1, 4–14 (2008)
Agrawal, O.P.: A general formulation and solution scheme for fractional optimal control problems. Nonlinear Dynam. 38:1-4, 323–337 (2004)
Frederico, G.S.F., Torres, D.F.M.: Fractional conservation laws in optimal control theory. Nonlinear Dynam. 53:3, 215–222 (2008) [arXiv:0711.0609]{}
Frederico, G.S.F., Torres, D.F.M.: Fractional optimal control in the sense of Caputo and the fractional Noether’s theorem. Int. Math. Forum 3:9-12, 479–493 (2008) [arXiv:0712.1844]{}
Pooseh, S., Almeida, R., Torres, D.F.M.: Fractional order optimal control problems with free terminal time. J. Ind. Manag. Optim. 10:2, 363–381 (2014) [arXiv:1302.1717]{}
Sweilam, N.H., Al-Ajami, T.M., Hoppe, R.H.W.: Numerical solution of some types of fractional optimal control problems. The Scientific World Journal 2013:306237, 9 pp (2013)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Dong Lai\
Center for Radiophysics and Space Research, Department of Astronomy\
Cornell University, Ithaca, NY 14853
---
\[1996/06/01\]
Neutron Star Kicks and Supernova Asymmetry
==========================================
[*Observations over the last decade have shown that neutron stars receive a large kick velocity (of order a few hundred to a thousand kms$^{-1}$) at birth. The physical origin of the kicks and the related supernova asymmetry is one of the central unsolved mysteries of supernova research. We review the physics of different kick mechanisms, including hydrodynamically driven, neutrino – magnetic field driven, and electromagnetically driven kicks. The viabilities of the different kick mechanisms are directly related to the other key parameters characterizing nascent neutron stars, such as the initial magnetic field and the initial spin. Recent observational constraints on kick mechanisms are also discussed.*]{}
Evidence for Neutron Star Kicks and Supernova Asymmetry
-------------------------------------------------------
It has long been recognized that neutron stars (NSs) have space velocities much greater than their progenitors’. A natural explanation for such high velocities is that supernova (SN) explosions are asymmetric, and provide kicks to the nascent NSs. In recent years evidence for NS kicks and NS asymmetry has become much stronger. The observations that support (or even require) NS kicks fall into three categories:
[*(1) Large NS Velocities ($\gg$ the progenitors’ velocities $\sim 30$ kms$^{-1}$):*]{}
$\bullet$ The study of pulsar proper motion give a mean birth velocity $200-500$ kms$^{-1}$ (Lorimer et al. 1997; Hansen & Phinney 1997; Cordes & Chernoff 1998; Arzoumanian et al 2002), with possibly a significant population having $V\go 1000$ kms$^{-1}$. While velocity of $\sim 100$ kms$^{-1}$ may in principle come from binary breakup in a supernova (without kick), higher velocities would require exceedingly tight presupernova binary. Statistical analysis seems to favor a bimodal pulsar velocity distribution, with peaks around $100~\kms$ and $500~\kms$ (see Arzoumanian et al. 2002).
$\bullet$ Observations of bow shock from the Guitar nebula pulsar (B2224+65) implies $V \simgreat 1000$ km s$^{-1}$ (Cordes et al. 1993; Chatterjee & Cordes 2002).
$\bullet$ The studies of NS – SNR associations have, in some cases, implied large NS velocities, up to $\sim 10^3$ km s$^{-1}$ (e.g., NS in Cas A SNR has $V>330$ km s$^{-1}$; Thorstensen et al. 2001).
[*(2) Characteristics of NS Binaries:*]{} While large space velocities can in principle be accounted for by binary break-up (see Iben & Tutukov 1996), many observed characteristics of NS binaries demonstrate that binary break-up can not be solely responsible for pulsar velocities, and that kicks are required. Examples include:
$\bullet$ The spin-orbit misalignment in PSR J0045-7319/B-star binary, as manifested by the orbital plane precession (Kaspi et al. 1996; Lai et al. 1995) and fast orbital decay (which indicates retrograde rotation of the B star with respect to the orbit; Lai 1996; Kumar & Quataert 1997) require that the NS received a kick at birth (see Lai 1996b). Similar precession of orbital plane has been observed in PSR J1740-3052 system (Stairs et al. 2003).
$\bullet$ The detection of geodetic precession in binary pulsar PSR 1913+16 implies that the pulsar’s spin is misaligned with the orbital angular momentum; this can result from the aligned pulsar-He star progenitor only if the explosion of the He star gave a kick to the NS that misalign the orbit (Kramer 1998; Wex et al. 1999). $\bullet$ The system radial velocity ($430\,\kms$) of X-ray binary Circinus X-1 requires $V_{\rm kick}\simgreat 500$kms$^{-1}$ (Tauris et al. 1999). Also, PSR J1141-6545 has $V_{\rm sys}\simeq 125~\kms$.
$\bullet$ High eccentricities of Be/X-ray binaries cannot be explained without kicks (van den Heuvel & van Paradijs 1997; but see Pfahl et al. 2002).
$\bullet$ Evolutionary studies of NS binary population (in particular the double NS systems) imply the existence of pulsar kicks (e.g., Fryer et al. 1998).
[*(3) Observations of SNe and SNRs:*]{} There are many direct observations, detailed in other contributions to this proceedings, of nearby supernovae (e.g., spectropolarimetry, Wang et al. 2003; X-ray and gamma-ray observations and emission line profiles of SN1987A) and supernova remnants which support the notion that supernova explosions are not spherically symmetric.
The Problem of Core-Collapse Supernovae and NS Kicks
----------------------------------------------------
The current paradigm for core-collapse supernovae leading to NS formation is that these supernovae are neutrino-driven (see, e.g., Burrows & Thompson 2002; Janka et al. 2002 for a recent review): As the central core of a massive star collapses to nuclear density, it rebounds and sends off a shock wave, leaving behind a proto-NS. The shock stalls at several 100’s km because of neutrino loss and nuclear dissociation in the shock. A fraction of the neutrinos emitted from the proto-NS get absorbed by nucleons behind the shock, thus reviving the shock, leading to an explosion on the timescale several 100’s ms — This is the so-called “delayed mechanism”. It has been argued that neutrino-driven convection in the proto-NS and that in the shocked mantle are central to the explosion mechanism (e.g., Mezzacappa et al. 1998), although current 2D simulations with the state-of-the-art neutrino interaction and transport have not produced a successful explosion model (Buras et al. 2003; see Fryer & Warren 2003 for simulations in 3D that use more approximate neutrino physics/transport). What is even more uncertain is the role of rotation and magnetic field on the explosion (see Rampp, Müller & Ruffert 1998; Fryer & Heger 2000; Ott et al. 2004 for simulations of collapse/explosion with rotation, and Thompson & Norman 2001, Wheeler et al. 2002 and Akiyama et al. 2003 and references therein for discussions of magnetic effects). It is clear that our understanding of the physical mechanisms of core-collapse supernovae remains rather incomplete. The prevalence of neutron star kicks poses a significant mystery, and indicates that large-scale, global deviation from spherical symmetry is an important ingredient in our understanding of core-collapse supernovae. In the following sections, we review different classes of physical mechanisms for generating NS kicks, and then discuss possible observational constraints and astrophysical implications.
Kick Mechanisms
---------------
### Hydrodynamically Driven Kicks
[*(1) Can Convections lead to NS Kicks?*]{} The collapsed stellar core and its surrounding mantle are susceptible to a variety of hydrodynamical (convective) instabilities (e.g., Herant et al. 1994; Burrows et al. 1995; Janka & Müller 1996; Mezzacappa et al. 1998). It is natural to expect that the asymmetries in the density, temperature and velocity distributions associated with the instabilities can lead to asymmetric matter ejection and/or asymmetric neutrino emission. Most numerical simulations in the 1990s indicate that the local, post-collapse instabilities are not adequate to account for kick velocities $\simgreat 100$ km s$^{-1}$. Recently, Scheck et al. (2003) reported computer experiments in which they adjust the neutrino luminosity from the neutrinosphere (the inner boundary of the simulation domain) to obtain successful explosions. They found that for long-duration (more than a second) explosions, neutrino-driven convection behind the expanding shock can lead to global ($l=1,2$) asymmetries, accelerating the remnant NS to a range of velocities up to several hundreds of $\kms$ (cf. Thompson 2000). This result is encouraging. But note that, like most other SN simulations, the Scheck et al. simulations were done in 2D, the proton-NS was fixed on the grids (with the kick calculated by adding up the momentum flux across the inner boundary), and the explosions were obtained in an ad hoc manner. It is also not clear whether kick velocities of 500-1000 $\kms$ can be easily obtained.
-3.4cm
[*(2)Asymmetries in Pre-Supernova Cores:*]{} It has been recognized that one way to produce large kicks is to have global asymmetric perturbations prior to core collapse (Goldreich et al. 1996; Burrows & Hayes 1996). One possible origin for the pre-SN asymmetry is the overstable oscillations in the pre-SN core (Goldreich et al. 1996). The idea is the following. A few hours prior to core collapse, the central region of the progenitor star consists of a Fe core surrounded by Si-O burning shells and other layers of envelope. This configuration is overstable to nonspherical oscillation modes. It is simplest to see this by considering a $l=1$ mode: If we perturb the core to the right, the right-hand-side of the shell will be compressed, resulting in an increase in temperature; since the shell nuclear burning rate depends sensitively on temperature (power-law index $\sim 47$ for Si burning and $\sim 33$ for O burning), the nuclear burning is greatly enhanced; this generates a large local pressure, pushing the core back to the left. The result is an oscillating g-mode with increasing amplitude. There are also damping mechanisms for these modes, the most important one being leakage of mode energy: The local (WKB) dispersion relation for nonradial waves is k\_r\^2=(\^2c\_s\^2)\^[-1]{}(\^2-L\_l\^2)(\^2-N\^2), where $k_r$ is the radial wavenumber, $L_l=\sqrt{l(l+1)}c_s/r$ ($c_s$ is the sound speed) and $N$ are the acoustic cut-off (Lamb) frequency and the Brunt-Väisälä frequency, respectively. Since acoustic waves whose frequencies lie above the acoustic cutoff can propagate through convective regions, each core g-mode will couple to an outgoing acoustic wave, which drains energy from the core g-modes (see Fig. \[fig1\]). In another word, the g-mode is not exactly trapped in the core. Our calculations (based on the $15M_\odot$ and $25M_\odot$ presupernova models of Weaver & Woosley) indicate that a large number of g-modes are overstable, although for low-order modes (small $l$ and $n$) the results depend sensitively on the detailed structure and burning rates of the presupernova models (see Lai 2001). The typical mode periods are $\simgreat 1$ s, the growth time $\sim 10-50$ s, and the lifetime of the Si shell burning is $\sim$ hours. Thus there could be a lot of e-foldings for the nonspherical g-modes to grow. Our preliminary calculations based on the recent models of A. Heger and S. Woosley (Heger et al. 2001) give similar results (work in progress). Our tentative conclusion is that overstable g-modes may potentially grow to large amplitudes prior to core implosion, although several issues remain to be understood better. For example, the O-Si burning shell is highly convective, with convective speed reaching $1/4$ of the sound speed, and hydrodynamical simulation may be needed to properly modeled such convection zones (see Asida & Arnett 2000).
So now we have a plausible way of generating initial asymmetric perturbations before core collapse. During the collapse, the asymmetries are amplified by a factor of 5-10 (Lai & Goldreich 2000; see also Lai 2000). How do we get the kick? The numerical simulations by Burrows & Hayes (1996) illustrate the effect. Suppose the right-hand-side of the collapsing core is denser than the left-hand side. As the shock wave comes out after bounce, it will see different densities in different directions, and it will move preferentially on the direction where the density is lower. So we have an asymmetric shock propagation and mass ejection, a “mass rocket” (cf. Fryer 2003).
### Neutrino – Magnetic Field Driven Kicks
The second class of kick mechanisms rely on asymmetric neutrino emission induced by strong magnetic fields. Since $99\%$ of the NS binding energy (a few times $10^{53}$ erg) is released in neutrinos, tapping the neutrino energy would appear to be an efficient means to kick the newly-formed NS. The fractional asymmetry $\alpha$ in the radiated neutrino energy required to generate a kick velocity $V_{\rm kick}$ is $\alpha=MV_{\rm kick}c/E_{\rm tot}$ ($=0.028$ for $V_{\rm kick}=1000$ km s$^{-1}$, NS mass $M=1.4\,M_\odot$ and total neutrino energy radiated $E_{\rm tot}
=3\times 10^{53}$ erg). There are several possible effects:
[*(1) Parity Violation:*]{} Because weak interaction is parity violating, the neutrino opacities and emissivities in a magnetized nuclear medium depend asymmetrically on the directions of neutrino momenta with respect to the magnetic field, and this can give rise to asymmetric neutrino emission from the proto-NS. Calculations indicate that to generate interesting kicks with this effect requires the proto-NS to have a large-scale, ordered magnetic field of at least a few $\times 10^{15}$ G (see Arras & Lai 1999a,b and references therein).
[*(2) Asymmetric Field Topology:*]{} Another effect relies on the asymmetric magnetic field distribution in proto-NSs: Since the cross section for $\nu_e$ ($\bar\nu_e$) absorption on neutrons (protons) depends on the local magnetic field strength, the local neutrino fluxes emerged from different regions of the stellar surface are different. Calculations indicate that to generate a kick velocity of $\sim 300$ km s$^{-1}$ using this effect alone would require that the difference in the field strengths at the two opposite poles of the star be at least $10^{16}$ G (see Lai & Qian 1998). Note that only the magnitude of the field matters here.
[*(3) Dynamical Effect of Magnetic Fields:*]{} A superstrong magnetic field may also play a dynamical role in the proto-NS. For example, it has been suggested that a locally strong magnetic field can induce “dark spots” (where the neutrino flux is lower than average) on the stellar surface by suppressing neutrino-driven convection (Duncan & Thompson 1992). While it is difficult to quantify the kick velocity resulting from an asymmetric distribution of dark spots, order-of-magnitude estimate indicates that a local magnetic field of at least $10^{15}$ G is needed for this effect to be of importance.
[*(4) Exotic Neutrino Physics:*]{} There have also been several ideas of pulsar kicks which rely on nonstandard neutrino physics. For example, it was suggested (Kusenko & Segre 1996, 1998) that asymmetric $\nu_\tau$ emission could result from the MSW flavor transformation between $\nu_\tau$ and $\nu_e$ inside a magnetized proto-NS because a magnetic field changes the resonance condition for MSW effect. This mechanism requires neutrino mass of order $100$ eV. A similar idea (Akhmedov et al. 1997; Grasso et al. 1998) relies on both the neutrino mass and the neutrino magnetic moment to facilitate the flavor transformation (resonant neutrino spin-flavor precession). Fuller et al. (2003) discussed the effect of sterile neutrinos. Analysis of neutrino transport (Janka & Raffelt 1998) indicates that even with favorable neutrino parameters, strong magnetic fields $B\gg 10^{15}$ G are required to obtain a $100$ km s$^{-1}$ kick.
### Electromagnetically Drievn Kicks
Harrison & Tademaru (1975) showed that electromagnetic (EM) radiation from an off-centered rotating magnetic dipole imparts a kick to the pulsar along its spin axis. The kick is attained on the initial spindown timescale of the pulsar (i.e., this really is a gradual acceleration), and comes at the expense of the spin kinetic energy. A reexamination of this effect (Lai et al. 2001) showed that the force on the pulsar due to asymmetric EM radiation is larger than the original Harrison & Tademaru expression by a factor of four. Thus, the maximum possible velocity is $V_{\rm kick}^{(\rm max)}\simeq
1400\,\left({1\,{\rm kHz}/P_i}\right)^2~\kms$. Nevertheless, to generate interesting kicks using this mechanism requires the initial spin period $P_i$ of the NS to be less than 1-2 ms. Gravitational radiation may also affect the net velocity boost.
### Other Possibilities
\(1) If rotation and magnetic fields play a dominate role in the explosion, bipolar jets may be produced. A slight asymmetry between the two jets will naturally lead to large kick (e.g., Khokhlov et al. 1999; Akiyama et al. 2003). While difficult to calculate, this is a serious possibility given the increasing observational evidence for bipolar explosions in many SNe (see other contributions to this proceedings).
\(2) Colpi and Wasserman (2003) considered the formation of double proton-NS binary in a rapidly rotating core collapse; the lighter NS explodes after reaching its minimum mass limit (via mass transfer), giving the remaining NS a large kick ($\sim 10^3\,\kms$). A related suggestion relies on the coalescence of proto-NS binary as providing the kick (Davies et al. 2002). The biggest uncertainty for such scenarios is that it is not clear core fragmentation can take place in the collapse (and numerical simulations seem to say no; see Fryer & Warren 2003).
Astrophysical Constraints on Kick Mechanisms
--------------------------------------------
The review in previous section clearly shows that NS kick is not only a matter of curiosity, it is intimately connected to the other fundamental parameters of young NSs (initial spin and magnetic field). For example, the neutrino-magnetic field driven mechanisms are of relevance only for $B\simgreat 10^{15}$ G. While recent observations have lent strong support that some neutron stars (“magnetars”) are born with such a superstrong magnetic field, it is not clear (perhaps unlikely) that ordinary radio pulsars (for which large velocities have been measured) had initial magnetic fields of such magnitude.
One of the reasons that it has been difficult to pin down the kick mechanisms is the lack of correlation between NS velocity and the other properties of NSs. The situation has changed with the recent X-ray observations of the compact X-ray nebulae of the Crab and Vela pulsars, which have a two sided asymmetric jet at a position angle coinciding with the position angle of the pulsar’s proper motion (Pavlov et al. 2000; Helfand et al. 2001). The symmetric morphology of the nebula with respect to the jet direction strongly suggests that the jet is along the pulsar’s spin axis. Analysis of the polarization angle of Vela’s radio emission corroborates this interpretation (Lai et al. 2001). Thus, while statistical analysis of pulsar population neither support nor rule out any spin-kick correlation, at least for the Vela and Crab pulsars (and perhaps for several other pulsars; see Ng & Romani 2003), the proper motion and the spin axis appear to be aligned.
The apparent alignment between the spin axis and proper motion raises an interesting question: Under what conditions is the spin-kick alignment expected for different kick mechanisms? Let us look at the three classes of mechanisms discussed before (Lai et al. 2001): (1) For the electromagnetically driven kicks, the spin-kick slignment is naturally produced. (Again, note that $P_i\sim 1-2$ ms is required to generate sufficiently large $V_{\rm kick}$). (2) For the neutrino–magnetic field driven kicks: The kick is imparted to the NS near the neutrinosphere (at 10’s of km) on the neutrino diffusion time, $\tau_{\rm kick}\sim 10$ seconds. As long as the initial spin period $P_i$ is much less than a few seconds, spin-kick alignment is naturally expected. (3) For the hydrodynamically driven kicks: because the kick is imparted at a large radius ($\go 100$ km), to get effective rotational averaging, we require that the rotation period at $\sim 100$ km to be shorter than the kick timescale ($\sim 100$ ms). This translates to $P_{\rm NS}\lo
1$ ms, which means that rotation must be dynamically important [^1]. On the otherhand, if rotation indeed plays a dynamically important role, the basic collapse and explosion may be qualitatively different (e.g., core bounce may occur at subnuclear density, the explosion is weaker and takes the form of two-sided jets (e.g. Khokhlov et al. 1999; Fryer & Heger 2000). The possibility of a kick in such systems has not been studied, but it is conceivable that an asymmetric dipolar perturbation may be coupled to rotation, thus producing spin-kick alignment. Current observations, however, seem to suggest that most pulsars are born rotating slowly ($P_i>10$ ms), with rotation palying a negligible role in the dynamics (see also Heger et al. 2003 for preSN evolution of rotating stars).
Currently we do not know whether spin-kick alignment is a generic feature of all pulsars; if it is, then it can provide powerful constraint on the kick mechanisms and the SN explosion mechanisms in general.
Finally, it is worth noting that recent observations showed that black hole (BH) formation can be accompanied by SN explosion: The companion of the BH X-ray binary GRO J1655-40 (Nova Sco) and that of SAX J1819.3-2525 (V4641 SGR) have high abundance of $\alpha$-elements (Israelian et al. 1999; Orosz et al. 2001), which can only be produced in a SN explosion (see Podsiadlowski et al. 2002). Apparently, the BH forms in an indirect process where a shock wave successfully makes an explosion and a NS forms temporarily followed by fall-back, or loss of angular momentum and thermal energy in the proto-NS which then collapses to a BH. This indirect process may explain the the relatively large space velocity of GRO J1655-40.
[*Acknowledgements*]{}: Support for this work is provided in part by NASA NAG 5-12034 and NSF AST 0307252. I thank my collaborators P. Arras, D. Chernoff, J. Cordes, P. Goldreich, A. Heger, Y.-Z. Qian and A. Shirakawa for their important contributions. I also thank the conference organizers for a stimulating meeting and travel support.
[99]{}
Akhmedov, E. K., Lanza, A., & Sciama, D. W. 1997, Phys. Rev. D, 56, 6117
Akiyama, S., et al. 2003, ApJ, 584, 954
Arras, P., & Lai, D. 1999a, ApJ, 519, 745 Arras, P., & Lai, D. 1999b, Phys. Rev. D60, 043001 Arzoumanian, Z., Chernoff, D.F., & Cordes, J.M. 2002, ApJ, 568, 289
Asida, S.M., & Arnett, D. 2000, ApJ, 545, 435 Buras, R. et al. 2003, Phys. Rev. Lett. 90, 241101
Burrows, A., Hayes, J., & Fryxell, B.A. 1995, ApJ, 450, 830
Burrows, A., & Hayes, J. 1996, Phys. Rev. Lett., 76, 352
Burrows, A., & Thompson, T.A. 2002, astro-ph/0210212 Chatterjee, S., & Cordes, J. M. 2002, ApJ, 575, 407 Colpi, M., & Wasserman, I. 2002, ApJ, 581, 1271. Cordes, J.M., Romani, R.W., & Lundgren, S.C. 1993, Nature, 362, 133
Davies, M.B., et al. 2002, ApJ, 579, L63 Duncan, R.C., & Thompson, C. 1992, ApJ, 392, L9.
Fryer, C.L. 2003, astro-ph/0312265
Fryer, C., Burrows, A., & Benz, W. 1998, ApJ, 498, 333
Fryer, C. L., & Heger, A. 2000, 541, 1033 Fryer, C.L., & Warren, M.S. 2003, ApJ, in press (astro-ph/0309539)
Fuller, G.M., et al. 2003, astro-ph/0307267 Goldreich, P., Lai, D., & Sahrling, M. 1996, in “Unsolved Problems in Astrophysics", ed. J.N. Bahcall and J.P. Ostriker (Princeton Univ. Press)
Grasso, D., Nunokawa, H., & Valle, J.W.F. 1998, Phys. Rev. Lett., 81, 2412 Hansen, B.M.S., & Phinney, E.S. 1997, MNRAS, 291, 569
Heger, A., Woosley, S.E., Martinez-Pinedo, G., & Langanke, K. 2001, ApJ, 560, 307
Heger, A., Woosley, S.E., Langer, N., & Spruit, H. 2003, in IAU 215 “Stellar Rotation” (astro-ph/0301374)
Herant, M., et al. 1994, ApJ, 435, 339
Harrison, E.R., & Tademaru, E. 1975, ApJ, 201, 447
Iben, I., & Tutukov, A. V. 1996, ApJ, 456, 738
Israelian, G, et al. 1999, Nature, 401, 6749 GRO J1655 - 40
Janka, H.-T., & Müller, E. 1996, A&A, 306, 167
Janka, H.-T., & Raffelt, G.G. 1998, Phys. Rev. D59, 023005 Janka, H.-Th., et al. 2002, in “Core Collapse of Massive Stars” (astro-ph/0212316)
Kaspi, V.M., et. al. 1996, Nature, 381, 583
Khokhlov, A.M., et al. 1999, ApJ, 524, L107 Kramer, M. 1998, ApJ, 509, 856
Kumar, P., & Quataert, E.J. 1997, 479, L51
Kusenko, A., & Segré, G. 1996, Phys. Rev. Lett., 77, 4872 Lai, D. 1996, ApJ, 466, L35
Lai, D. 2000, ApJ, 540, 946 Lai, D., Bildsten, L., & Kaspi, V.M. 1995, ApJ, 452, 819
Lai, D., Chernoff, D.F., & Cordes, J.M. 2001, ApJ, 549, 1111
Lai, D., & Goldreich, P. 2000, ApJ, 535, 402
Lai, D., & Qian, Y.-Z. 1998, ApJ, 505, 844 Lorimer, D.R., Bailes, M., & Harrison, P.A. 1997, MNRAS, 289, 592
Mezzacappa, A., et al. 1998, ApJ, 495, 911.
Ng, C.-Y., & Romani, R.W. 2003, ApJ, in press (astro-ph/0310155)
Orosz, J. et al. 2001, ApJ, 555, 489 Ott, C.D., et al. 2004, ApJ, 600 (astro-ph/0307472)
Pavlov, G.G., et al. 2001, ApJ, 552, L129 Pfahl, E., et al. 2002, ApJ, 574, 364
Podsiadlowski, Ph., et al. 2002, ApJ, submitted (astro-ph/0109244) Rampp, M., Müller, E., & Ruffert, M. 1998, A&A, 332, 969
Scheck, L. et al. 2003, PRL, in press (astro-ph/0307352) Spruit, H., & Phinney, E.S. 1998, Nature, 393, 139
Tauris, T., et al. 1999, MNRAS, 310, 1165 Thompson, C. 2000, ApJ, 534, 915
Thompson, C., & Murray, N. 2001, ApJ, 560, 339
Thorstensen, J.R., et al. 2001, AJ, 122, 297
van den Heuvel, E.P.J., & van Paradijs, J. 1997, ApJ, 483, 399.
Wang, L., Baade, D., Höflich, P., & Wheeler, J.C. 2003, ApJ, 592, 457 Weaver, T.A., & Woosley, S.E. 1993, Phys. Rep., 227, 65
Wex, N., Kalogera, V., & Kramer, M. 2000, ApJ, 528, 401 Wheeler, J.C., Meier, D.L., & Wilson, J.R. 2002, ApJ, 568, 807
[^1]: It has been suggested that even with zero initial angular momentum, aligned spin-kick may be possible if the kick is the result of many off-centered small thrusts which are appropriately oriented (Spruit & Phinney 1998). However, in all the mechanisms studied so far, the kick is the result of a single or a few thrusts, and one would expect the final spin to be roughly perpendicular the kick.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With no CSI at the users, transmission over the two-user Gaussian Multiple Access Channel with fading and finite constellation at the input, is not efficient because error rates will be high when the channel conditions are poor. However, perfect CSI at the users is an unrealistic assumption in the wireless scenario, as it would involve massive feedback overheads. In this paper we propose a scheme which uses only quantized knowledge of CSI at the transmitters with the overhead being nominal. The users rotate their constellation without varying their transmit power to adapt to the existing channel conditions, in order to meet certain pre-determined minimum Euclidean distance requirement in the equivalent constellation at the destination. The optimal modulation scheme has been described for the case when both the users use symmetric $M$-PSK constellations at the input, where $ M=2^\lambda $, $ \lambda $ being a positive integer. The strategy has been illustrated by considering examples where both users use QPSK or 8-PSK signal sets at the input. It is shown that the proposed scheme has better throughput and error performance compared to the conventional non-adaptive scheme, at the cost of a feedback overhead of just $\left\lceil \log _2 \left(\frac{M^2}{8}-\frac{M}{4}+2\right)\right\rceil + 1 $ bits, for the $M$-PSK case.'
author:
-
title: 'An Adaptive Modulation Scheme for Two-user Fading MAC with Quantized Fade State Feedback'
---
INTRODUCTION
============
A multiple access channel (MAC) consists of multiple users transmitting independent information to a common destination. There is no cooperation among the users. The capacity region for a discrete memoryless MAC is well known [@Cover] [@Gallager]. For a two-user MAC with additive white Gaussian noise (AWGN) the capacity achieving input is the continuous Gaussian alphabet. The two-user Gaussian MAC with finite input constellations like $M$-QAM, $M$-PSK was studied in [@HarshanCR] [@HarshanCPA]. It was shown that relative rotation between input constellations [@HarshanCR], or a constellation power allocation scheme [@HarshanCPA] may be employed to maximize the constellation constrained (CC) capacity regions. Trellis based coding schemes were also suggested to achieve any rate pair within the CC capacity region.
In this paper, a two-user MAC with quasi-static fading is considered, as shown in Fig. \[fig:mac\]. The two users transmit information to a common destination. The random variables $h_1$ and $h_2$ are the channel gains for User-1 and User-2 respectively and $h_1$,$h_2$ $\sim \mathcal{CN}(0,1)$, where $ \mathcal{CN}(0,s) $ denotes the circular symmetric complex Gaussian random variable with variance $ s $. AWGN $ z $ gets added to the received signal at the destination, $z$ $\sim \mathcal{CN}(0,\sigma^2)$. User-$i$ transmits a symbol $ x_i $ from a complex finite constellation $ \mathcal{S}_i $ (like $M$-QAM or $M$-PSK) of unit average energy, *i.e*, $\mathbb{E} [\mid x_i \mid ^2]=1$. Let $ P $ be the average power constraint for each user. The received signal at the destination is thus represented by $$\begin{aligned}
\nonumber
y=\sqrt{P}h_1x_1+\sqrt{P}h_2x_2+z.\end{aligned}$$ We assume that perfect CSI *i.e.* the tuple $ (h_1,h_2) $ is available only at the destination.
![Two-user fading MAC with Gaussian noise[]{data-label="fig:mac"}](mac.eps){width="3in"}
At the destination the system can be viewed as a single user AWGN channel with the symbols drawn from a sum constellation $$\begin{aligned}
\nonumber
\mathcal{S}_{\text{sum}} &= \sqrt{P} h_1\mathcal{S}_1+ \sqrt{P} h_2\mathcal{S}_2\\
\nonumber
&= \sqrt{P} h_1(\mathcal{S}_1+\dfrac{h_2}{h_1}\mathcal{S}_2)\\
\label{eqneff}
&= \sqrt{P} h_1\underbrace{(\mathcal{S}_1+\gamma e^{j\theta}\mathcal{S}_2)}_{\mbox{$\mathcal{S}_{\text{eff}}$}},\end{aligned}$$ where $\gamma= | \frac{h_2}{h_1}|$, $\theta = \angle \frac{h_2}{h_1}$ and $\mathcal{S}_{\text{eff}}$ denotes the effective constellation.
Without loss of generality it can be assumed that $ \gamma \geq 1 $, as destination has knowledge of both $h_1$ and $h_2$ separately. If $\vert \frac{h_2}{h_1}\vert < 1$, then at the destination the ratio can be simply reversed to compute $\frac{h_1}{h_2}$. Which one among the two ratios is calculated is made known to the users via a single bit of feedback. For the rest of the paper, we assume that the ratio $\frac{h_2}{h_1}$ is calculated at the destination. However, the results obtained still hold when the ratio calculated is $\frac{h_1}{h_2}$, by interchanging the roles of User-1 and User-2. For the rest of the paper a $M$-PSK constellation refers to a symmetric PSK signal set, with $ M=2^\lambda $, $ \lambda $ being a positive integer. The points in the $ M $-PSK signal set are of the form $ e^{j\frac{(k-1)2\pi}{M}} $, where $ 1\leq k \leq M $. We assume that $ \mathcal{S}_1 = \mathcal{S}_2= \mathcal{S}$, where $\mathcal{S}$ is an $M$-PSK constellation. We refer to the pair $ (\gamma , \theta)$ to represent $\gamma e^{j \theta}$ and call it the fade state throughout the paper. We refer to the complex plane that represents $ \gamma e^{j \theta } $ with $ \gamma \geq 1 $ as the $ (\Gamma, \Theta) $ plane.
Perfect channel state information (CSI) is available at the destination only, which quantizes the $ (\Gamma , \Theta) $ plane into finite number of regions. The quantization obtained is similar to the one used for physical layer network coding in [@Akino], which was subsequently derived analytically in [@Vijay]. This quantized knowledge of the fade state is made available to the users to adapt their modulation scheme via rotation of constellations to compensate for the possibly bad channel conditions. MAC with limited channel state information at transmitter (CSIT) has been studied from an information theoretic point of view in [@Reza],[@Cemal]. In [@Wiese], it was shown that for a two-user discrete memoryless MAC with additional common message, finer CSIT results in increasing the capacity region. To the best of our knowledge, explicit modulation schemes with finite constellations and quantized fade state feedback has not been reported before.
The contributions and organization of this paper are as follows:
- A quantization of the $ (\Gamma , \Theta) $ plane is derived, for the case when both users use $M$-PSK constellations at the input. We illustrate the quantization procedure by taking examples of the QPSK and 8-PSK case. (Section \[sec:chan\_quan\])
- A modulation scheme is proposed for the users, which adapts according to the quantized feedback about the fade state that they receive from the destination, in order to satisfy a certain minimum distance guarantee $\delta$ in $\mathcal{S}_{\text{eff}}$ given in . The fade states which leads to violation of this minimum distance guarantee have been identified. Adaptation involves rotation of the constellation of one user relative to the other, without any change in transmit power, in order to effectively avoid these bad channel conditions. (Section \[sec:adapt\])
- The procedure to obtain the optimal angles for rotation is stated for the $M$-PSK case. The optimal rotation angles are calculated in closed form for the QPSK and 8-PSK case. (Section \[sec:optimal\])
- An upper bound on $\delta$, *i.e.*, the maximum value of the minimum distance in the effective constellation that can be guaranteed, is derived. (Section \[sec:upper\])
- Simulation results are presented to show the extent to which the proposed strategy outperforms the conventional transmission scheme without adaptation. (Section \[sec:results\])
CHANNEL QUANTIZATION FOR $ M $-PSK SIGNAL SETS
==============================================
In this section we obtain a quantization of the $ (\Gamma, \Theta) $ plane into finite number of regions at the destination.
Distance Distribution in the effective constellation {#sec:quandist}
----------------------------------------------------
Without loss of generality we assume that the average power constraint of each user is $ P=1 $. It is known that the error performance for an AWGN channel is determined by the Euclidean distance distribution of the input constellation. In our case, the distance distribution of $ \mathcal{S}_{\text{sum}} $ decides the error performance at the destination. For any value of $ (\gamma,\theta) $, $d^{'2}_{(s_1,s_2)_{\text{sum}}\leftrightarrow (s'_1, s'_2)_{\text{sum}}}$ denotes the distance between the two points $(s_1,s_2)_{\text{sum}}$ and $(s'_1,s'_2)_{\text{sum}}$, where $ (s_1,s_2)_{\text{sum}} $, $(s'_1,s'_2)_{\text{sum}}$ $\in \mathcal{S}_{\text{sum}}$ refer to the points $ \sqrt{P}h_1(s_1+\gamma e^{j \theta} s_2) $ and $\sqrt{P}h_1(s'_1+\gamma e^{j \theta} s'_2)$ respectively with $ s_1,s_2,s'_1,s'_2 \in \mathcal{S} $. It is given by $$\begin{aligned}
\nonumber
d ^{'2}_{(s_1,s_2)_{\text{sum}}\leftrightarrow (s'_1, s'_2)_{\text{sum}}} &= P \vert h_1\vert ^2\vert(s_1-s'_1 )+ \gamma e^{j\theta } (s_2-s'_2) \vert ^2 \\
\label{eqn:sumdistance}
&= P \vert h_1 \vert ^2 d^2_{(s_1,s_2)\leftrightarrow (s'_1, s'_2)},\end{aligned}$$ where , $d^2_{(s_1,s_2)\leftrightarrow (s'_1, s'_2)}$ denotes the distance between the points $(s_1,s_2)$ and $(s'_1,s'_2)$, where $ (s_1,s_2), (s'_1,s'_2) $ refers to the points $ s_1+\gamma e^{j \theta} s_2 $ and $s'_1+\gamma e^{j \theta} s'_2$ in $ \mathcal{S}_{\text{eff}} $. Since $ P \vert h_1 \vert ^2 $ simply scales the distances in $ \mathcal{S}_{\text{eff}} $ we can focus only on $$\begin{aligned}
\label{eqn:dist}
d^2_{(s_1,s_2)\leftrightarrow (s'_1, s'_2)}= \vert(s_1-s'_1 )+ \gamma e^{j\theta } (s_2-s'_2) \vert ^2\end{aligned}$$ as the quantity of interest.
It is clear from that for certain values of $ (\gamma, \theta) $ the distance between points $(s_1,s_2)$ and $(s'_1,s'_2)$ in $\mathcal{S}_{\text{eff}}$ reduces to zero, *i.e.* if $$\begin{aligned}
\label{eqn:sing}
\gamma e^{j \theta} = -\frac{(s_1-s'_1)}{(s_2-s'_2)}\end{aligned}$$ then $ d^2_{(s_1,s_2) \leftrightarrow (s_1\prime, s_2\prime) }=0 $. These values of $ (\gamma, \theta) $ are called the *singular fade states* [@Akino], [@Vijay]. Singular fade states can also be defined as follows:
\[def:sing\] A fade state $ (\gamma, \theta) $ is said to be a *singular fade state* if $ \vert \mathcal{S}_{\text{eff}} \vert < M^2 $.
Clearly, $ \gamma e^{j \theta}=0 $ is a singular fade state, for any arbitrary signal set $ \mathcal{S} $. For any input constellation $\mathcal{S}$ the other non-zero singular fade states are obtained using . For a given input constellation $ \mathcal{S} $, let $ \mathcal{H} $ denote the set of all singular fade states.
When $\mathcal{S}_1=\mathcal{S}_2=\mathcal{S}$, where $\mathcal{S}$ is a QPSK constellation, then the non-zero singular fade states are at $$\begin{aligned}
\nonumber
\gamma &= \sqrt{2} , \hspace{5pt} \theta = 45^\circ , 135^\circ ,225^\circ , 315^\circ \\
\nonumber
\gamma &= 1, \hspace{13pt} \theta = 0^\circ , 90^\circ , 180^\circ , 270^\circ\\
\nonumber
\gamma &= \frac{1}{\sqrt{2}}, \,\theta = 45^\circ , 135^\circ ,225^\circ , 315^\circ\end{aligned}$$
Since for an AWGN channel the error performance at the destination is dominated by the minimum distance of the input constellation, it is sufficient to study the minimum distance of $ \mathcal{S}_{\text{eff}} $. Also from Definition \[def:sing\] minimum distance in $ \mathcal{S}_{eff} $ reduces to zero at the singular fade states. The following lemma provides an upper bound on the minimum distance of the effective constellation $ \mathcal{S}_{\text{eff}} $.
\[lem:uppermin\] When both the users use any arbitrary signal set $\mathcal{S}$ (which includes $M$-PSK, $M$-QAM) at the input, then for any fade state $ (\gamma, \theta) $, the minimum distance $ d_{min}(\gamma , \theta) $ between any two points in $ \mathcal{S}_{\text{eff}} $ is upper bounded by the minimum distance in the input constellation $ d_{min} (\mathcal{S}) $.
From the definition of $d_{min}^2(\gamma ,\theta)$, we have, [$$\begin{aligned}
\nonumber
d_{min}^2(\gamma ,\theta) &=\hspace{-0.5cm}\min_{(s_1,s_2) \neq (s'_1,s'_2) \in \mathcal{S}^2} \vert (s_1-s'_1)+\gamma e^{j \theta} (s_2-s'_2)\vert ^2\\
\nonumber
&\leq \hspace{-0.5cm}\min_{(s_1,s_2) \neq (s'_1,s'_2) \in \mathcal{S}^2} \left\lbrace \vert s_1-s'_1\vert ^2 +\gamma ^2 \vert s_2-s'_2\vert ^2 \right\rbrace \end{aligned}$$ $$\begin{aligned}
\label{Lemma1_1}
\text{Now}\hspace{5pt} d_{min}^2(\gamma ,\theta) &\leq \min_{s_1 \neq s'_1 \in \mathcal{S}}\vert s_1-s'_1\vert ^2 =d_{min}^2(\mathcal{S}).\\
\label{Lemma1_2}
\text{Also} \hspace{5pt} d_{min}^2(\gamma ,\theta) &\leq\min_{s_2 \neq s'_2 \in \mathcal{S}}\gamma ^2\vert s_2-s'_2\vert ^2=\gamma ^2 d_{min}^2 (\mathcal{S}).
%&\leq \min_{s_2 \neq s'_2 \in \mathcal{S}}\vert (s_2-s'_2)\vert ^2 \hspace{5pt} (\because \gamma \geq 1) \tag{B}\end{aligned}$$ $$\begin{aligned}
\nonumber
\text{From \eqref{Lemma1_1} and \eqref{Lemma1_2}, and using the fact that $\gamma \geq 1$, we have} \\
\nonumber
d_{min}^2(\gamma ,\theta) \leq \min \lbrace d_{min}^2(\mathcal{S}),\gamma ^2 d_{min}^2(\mathcal{S})\rbrace =d_{min}^2(\mathcal{S}). \end{aligned}$$ ]{}
In the following lemma, it is proved that in order to study the distance profile in $ \mathcal{S}_{\text{eff}} $ it is sufficient to consider $\theta \in \left[ 0,\pi /M\right] $ when both users use $M$-PSK signal sets. Distance profiles for other values of $\theta$ can be obtained from $\theta \in [0, \pi/M] $. We use the term wedge $[\theta_1, \theta _2]$ to denote the region $\gamma \geq 1$ and $\theta \in [\theta_1, \theta_2]$ on the $ (\Gamma , \Theta) $ plane. The lines $ \theta= \theta_1 $ and $ \theta= \theta _2 $ for $ \gamma \geq 1 $ and the arc $ \gamma=1 $ for $ \theta \in [\theta _1, \theta _2] $ form the boundary of the wedge $ [\theta _1 , \theta _2] $.
\[lemmaangle\] To study the distance profile in $ \mathcal{S}_{\text{eff}} $ when both the users use $M$-PSK constellations, it is sufficient to consider the case $ 0 \leq \theta \leq \pi /M $. All other cases can be obtained from this.
The proof is in two steps. First we show that the distance profile is a repetitive structure with period $ 2\pi /M $. Next, it is shown that within the wedge $ [0,2\pi /M] $ the distance profile is symmetric about the bisector of this wedge *i.e.*, the $ \theta = \pi /M $ line. We have from , $$\begin{aligned}
\nonumber
\mathcal{S}_{\text{eff}}= \mathcal{S}+ \gamma e^{j \theta } \mathcal{S} .\end{aligned}$$ For any arbitrary value of $ \theta = \frac{k 2 \pi }{M} +\theta ' $ where $k \in \mathbb{Z},\, 0 \leq \theta ' < \frac{2 \pi}{M}$, $$\begin{aligned}
\nonumber
\mathcal{S}_{\text{eff}}&= \mathcal{S}+ \gamma e^{j \theta ' } (\mathcal{S}e^{j \frac{2 k \pi}{M}})\\
\nonumber
&= \mathcal{S}+ \gamma e^{j \theta ' } \mathcal{S}.\end{aligned}$$ The last equality follows from the fact that rotating a $M$-PSK constellation by an integral multiple of $2\pi /M$ does not alter the distance profile of the constellation. Thus, whatever distance profiles for $ \mathcal{S}_{\text{eff}} $ are obtained for the wedge $ [0,2\pi /M] $, it is exactly repeated for the remaining $M-1$ wedges to cover the entire range of $ \theta $.
To show that the distance profiles are symmetric about $\theta= \pi /M$, we need to show that $\mathcal{S}+\gamma e^ {j (\frac{\pi}{M} + \alpha)} \mathcal{S}$ and $\mathcal{S}+\gamma e^{j (\frac{\pi}{M} -\alpha )}\mathcal{S}$, where $0\leq \alpha \leq \pi /M$, have the same distance profiles. We have $$\begin{aligned}
\nonumber
\mathcal{S}+\gamma e^{j (\frac{\pi}{M} -\alpha )}\mathcal{S}&=\mathcal{S}+\gamma e^{j (\frac{\pi}{M} -\alpha )}(\mathcal{S}e^ {\frac{-2\pi}{M}})\\
\label{Lemma2_1}
&=\mathcal{S}+\gamma e^{-j (\frac{\pi}{M} +\alpha )}\mathcal{S}.\end{aligned}$$ The first equality is because $\mathcal{S}=\mathcal{S}e^{j k 2\pi /M},$ *i.e.* rotating $\mathcal{S}$ by $2k\pi /M$ gives the same constellation. Thus for $k=-1$, $\mathcal{S} = \mathcal{S}e^ {\frac{-2\pi}{M}}$. Also due to the symmetric nature of $M$-PSK constellation, the distance distribution of the sum constellation depends only on the relative angle of rotation between the input constellations. Thus $\mathcal{S}+\gamma e^ \beta \mathcal{S}$ and $\mathcal{S}+\gamma e^ {-\beta} \mathcal{S}$ have same distance profiles, for any $\beta \in [0,\pi]$. This together with proves the second part of the lemma.
From Lemma \[lemmaangle\], it is clear that when both users use $M$-PSK signal sets, if $ (\gamma',\theta') $ is a singular fade state, then there exists singular fade states at $ (\gamma',\theta' + p \frac{2\pi}{M}) $, where $1\leq p \leq M-1$ because distance distribution in $\mathcal{S}_{\text{eff}}$ is periodic with period $\frac{2\pi}{M}$. The distance distribution of $\mathcal{S}_{\text{eff}}$ is the basis for channel quantization. Also from Lemma \[lemmaangle\], it suffices to obtain such a quantization only for the wedge $ [0, \pi /M]$. This can then be reflected along the $\theta = \pi /M$ line, to give the quantization for the wedge $[0, 2\pi /M]$, which when repeated for the remaining $M-1$ wedges will cover the entire $(\Gamma , \Theta)$ plane.
Channel Quantization for the $M$-PSK case {#sec:chan_quan}
-----------------------------------------
In this subsection we propose a technique to obtain the quantization of the $(\Gamma, \Theta)$ plane, when both users use $M$-PSK signal sets. From Lemma 1, when both users use $M$-PSK signal sets at the input, the minimum distance in $\mathcal{S}_{\text{eff}}$ for any value of $(\gamma,\theta)$, $d_{min}(\gamma,\theta) \leq d_{min}(\mathcal{S}) =\sqrt{2\left[1-\cos (\frac{2 \pi}{M})\right]}$. Now the following lemma gives the number of singular fade states in the wedge $[0,\pi /M]$.
{width="3.45in"}
\[periodic\] When both users use $M$-PSK signal sets at the input, the number of singular fade states in the wedge $[0, \pi /M]$, is given by $\frac{M^2}{8}-\frac{M}{4}+1$. Further, these singular fade states lie along the two lines $\theta=0$ and $\theta=\pi /M$.
From [@Vijay], the total number of singular fade states other than zero is $\frac{M^3}{4}-\frac{M^2}{2}+M$. Out of these, $M$ lie on the circle $\gamma=1$. It is also known from [@Vijay], that if $\gamma e^{j \theta}$ is a singular fade state, then $\frac{1}{\gamma}e^{-j \theta}$ is also a singular fade state. Thus, half of the total number of remaining singular fade states lie inside the circle $\gamma=1$ and the other half lies outside it. This along with the fact that singular fade states are periodic, implies the number of singular fade states for the wedge $ [0,\pi /M]$, is given by $$\begin{aligned}
\nonumber
\frac{1}{M}\left( \frac{\frac{M^3}{4}-\frac{M^2}{2}}{2}+M\right) =\frac{M^2}{8}-\frac{M}{4}+1.\end{aligned}$$ Also from [@Vijay], it is clear that these fade states lie along $\theta =0$ and $\theta = \pi /M$ lines.
We denote this set of all singular fade states lying in the wedge $ [0, \pi/M] $ by $ \mathcal{H}_W $. Let $N_W = \vert \mathcal{H}_W \vert $.
Observe from , that the distance between two points in $ \mathcal{S}_{\text{eff}} $ is a function of $ \gamma $ and $ \theta $. Let $ \vert \triangle s_j \vert = \vert s_j-s'_j \vert $ and $ \phi_j = \angle (s_j-s'_j) $ for $ j=1,\,2 $. Now from , the distance between the elements of the pair $ \lbrace (s_1,s_2),(s'_1,s'_2) \rbrace \in \mathcal{S}^2_{\text{eff}} $ is
[$$\begin{aligned}
\label{eq:distclassex}
d^2_{(s_1,s_2)\leftrightarrow (s'_1,s'_2)}= \vert \triangle s_1 \vert ^2 + \gamma ^2 \vert \triangle s_2 \vert ^2 + 2 \gamma \vert \triangle s_1 \triangle s_2 \vert \cos (\theta + \phi_2 -\phi_1 ). \end{aligned}$$ ]{} Consider another other pair $ \lbrace (\hat{s}_1,\hat{s}_2), (\hat{s}'_1,\hat{s}'_2)\rbrace \in \mathcal{S}^2_{\text{eff}} $ with $ \hat{s}_1,\, \hat{s}'_1,\,\hat{s}_2,\,\hat{s}'_2 \in \mathcal{S}$, and let $ \vert \triangle \hat{s}_j \vert = \vert \hat{s}_j - \hat{s}'_j \vert $ and $ \hat{\phi}_j = \angle (\hat{s}_j-\hat{s}'_j) $ for $ j=1,\,2 $. If $ \vert \triangle \hat{s}_1 \vert = \vert \triangle s_1 \vert $, $ \vert \triangle \hat{s}_2 \vert = \vert \triangle s_2 \vert $ and $ \hat{\phi}_2-\hat{\phi}_1 = \phi_2 - \phi_1 $ or $ \hat{\phi}_2-\hat{\phi}_1 = \pi -(\phi_2 - \phi_1) $, then from , $ d^2_{(\hat{s}_1,\hat{s}_2)\leftrightarrow (\hat{s}'_1,\hat{s}'_2)} = d^2_{(s_1,s_2)\leftrightarrow (s'_1,s'_2)} $ for all values of $ (\gamma, \theta) $, even though the value of this distance changes with $ (\gamma, \theta) $.
\[def:distclass\] A distance class denoted by $ \mathcal{C} $, is a subset of $ \mathcal{S}_{\text{eff}}^2 $, which contains the pairs of the form $ \lbrace (s_1,s_2),(s'_1,s'_2)\rbrace $, $ (s_1,s_2)\neq (s'_1,s'_2) $ where $ (s_1,s_2)$ and $(s'_1,s'_2) $ denote the complex points in $ \mathcal{S}_{eff} $, such that the distance between the two elements of a pair is same for all pairs in $ \mathcal{C} $ and this property holds for all values of $ \gamma $ and $ \theta $, though the value of the distance depends on $ (\gamma, \theta) $.
For a given input constellation $ \mathcal{S} $, let $ \bar{\mathcal{C}} $ denote the set of the all distance classes for it.
\[def:classfunc\] Associated with every distance class $ \mathcal{C} $ is a function $ d_{\mathcal{C}}(\gamma ,\theta): (\Gamma,\Theta )\leftrightarrow \mathbb{R} $, called the class distance function, which gives the value of the distance between the two elements of a pair in $ \mathcal{C} $ for any $ (\gamma, \theta) $.
\[def:fadefunc\] For a given fade state $ (\gamma, \theta) $, the function $ d_{\gamma , \theta} (\mathcal{C}): \bar{\mathcal{C}} \rightarrow \mathbb{R} $ gives the value of the distance between the two elements of a pair in $ \mathcal{C} $, for any $ \mathcal{C} \in \bar{\mathcal{C}}$. This is called the fade state distance function.
We use integer $ m $, $ 1\leq m \leq M $ to represent the point $ e^{j\frac{(m-1)2 \pi}{M}} $ in $ \mathcal{S} $ *i.e.* the $ M $-PSK signal set. The integer $ q=m+M(n-1) $, $ 1\leq q \leq M^2 $ denotes the complex point in $ \mathcal{S}_{\text{eff}} $ obtained by combining the points $ m $ and $ n $ of $ \mathcal{S} $ *i.e.* it refers to the point $ e^{j\frac{(m-1)2 \pi}{M}} + \gamma e^{j \theta} e^{j\frac{(n-1)2 \pi}{M}} $ in $ \mathcal{S}_{\text{eff}} $. For each distance class $ \mathcal{C} $, among all the pairs $ (i,j) \in \mathcal{C} $ choose the one with the minimum value of $ i+j $ to be the representative in $ \mathcal{C} $. If more than one pair has the same value of $ i+j $ choose the one with the lowest value of $ i $ as the class representative. When the users $ M $-PSK signal set at input, there are $ \frac{M^2(M^2-1)}{2} $ pairwise distances in $ \mathcal{S}_{\text{eff}} $. These pairwise distances are thus partitioned into distance classes.
Fig. \[fig:qpskeff\], shows $ \mathcal{S}_{\text{eff}} $ when both users use QPSK constellations for $(\gamma, \theta)=(2,14^\circ)$. There are 120 pairwise distances and 20 distance classes. These are listed in Table \[tab:qpskex\], along with the corresponding class distance functions, and the class representatives.
$ k $ $ \mathcal{C}_k $ $ d_{\mathcal{C}_k}(\gamma, \theta) $ Class representative
-------- ----------------------------------------------------------------- -------------------------------------------------------------------- ----------------------
$ 1 $ $(1,2),(1,4),(2,3),(3,4),(5,6),(6,7),(7,8),(5,8),$ $ \sqrt{2} $ $ (1,2) $
$(9,10),(10,11),(11,12),(9,12),(13,14),(14,15),(15,16),(13,16)$
$ 2 $ $(1,3),(2,4),(5,7),(6,8),(9,11),(10,12),(13,15),(14,16)$ $ 2 $ $(1,3)$
$ 3 $ $(1,6),(1,16),(2,15),(4,7),(6,11),(5,12),(11,16),(10,13)$ $ 2 \gamma ^2 +2+ 4 \gamma \cos \theta $ $ (1,6) $
$ 4 $ $(1,8),(2,7),(2,13),(3,16),(6,9),(7,12),(11,14),(12,13)$ $ 2\gamma ^2 + 2+ 4 \gamma \sin \theta$ $ (1,8) $
$ 5 $ $ (1,14),(4,15), (3,6),(4,5),(8,11),(6,10),(9,16),(10,15) $ $2\gamma ^2 + 2- 4 \gamma \sin \theta$ $ (3,6) $
$ 6 $ $ (2,5),(3,8),(3,14),(4,13),(7,10),(8,9),(12,15),(9,14) $ $ 2 \gamma ^2 +2- 4 \gamma \cos \theta$ $ (2,5) $
$ 7 $ $ (1,5), (1,13),(4,8),(2,6),(2,14),(3,7),(3,15),(4,16),$ $ 2 \gamma ^2 $ $ (1,5) $
$ (6,10),(7,11),(8,12),(5,9),(12,16),(11,15),(10,14),(9,13) $
$ 8 $ $ (1,9),(2,10),(3,11),(4,12),(6,14),(5,13),(7,15),(8,16) $ $ 4 \gamma ^2 $ $ (1,9) $
$ 9 $ $ (1,7),(2,16),(6,12),(11,13) $ $ 2 \gamma ^2 +4 + 4 \gamma \cos \theta + 4 \gamma \sin \theta $ $ (1,7) $
$ 10 $ $ (1,15),(4,6),(5,11),(10,16) $ $ 2 \gamma ^2 +4 + 4 \gamma \cos \theta - 4 \gamma \sin \theta $ $ (4,6) $
$ 11 $ $ (3,13),(2,8),(7,9),(12,14)) $ $ 2 \gamma ^2 +4 - 4 \gamma \cos \theta + 4 \gamma \sin \theta $ $ (2,8) $
$ 12 $ $ (3,5),(8,10),(9,15),(4,14) $ $ 2 \gamma ^2 +4 - 4 \gamma \cos \theta - 4 \gamma \sin \theta $ $ (3,5) $
$ 13 $ $ (1,12),(2,11),(6,13),(7,16) $ $ 4 \gamma ^2 +2 + 4 \gamma \cos \theta + 4 \gamma \sin \theta $ $ (1,12) $
$ 14 $ $ (1,10),(6,15),(4,11),(5,16) $ $ 4 \gamma ^2 +2 + 4 \gamma \cos \theta - 4 \gamma \sin \theta $ $ (1,10) $
$ 15 $ $ (2,9),(3,12),(7,14),(8,13) $ $ 4 \gamma ^2 +2 - 4 \gamma \cos \theta + 4 \gamma \sin \theta $ $ (2,9) $
$ 16 $ $ (3,10),(8,15),(4,9),(5,14) $ $ 4 \gamma ^2 +2 - 4 \gamma \cos \theta - 4 \gamma \sin \theta $ $ (3,10) $
$ 17 $ $ (1,11),(6,16) $ $ 4 \gamma ^2 +4 + 8 \gamma \cos \theta $ $ (1,11) $
$ 18 $ $ (2,12),(7,13) $ $ 4 \gamma ^2 +4 + 8 \gamma \sin \theta $ $ (2,12) $
$ 19 $ $ (3,9),(8,14) $ $ 4 \gamma ^2 +4 - 8 \gamma \cos \theta $ $ (3,9) $
$ 20 $ $ (4,10),(5,15) $ $ 4 \gamma ^2 +4 - 8 \gamma \sin \theta $ $ (4,10) $
We define the set of all class distance function, $ d_{\bar{\mathcal{C}}}(\gamma, \theta) $ and the set of all fade state distance functions $ d_{\Gamma, \Theta} (\mathcal{C}) $ as follows, $$\begin{aligned}
\nonumber
d_{\bar{\mathcal{C}}}(\gamma, \theta) &= \lbrace d_{\mathcal{C}}(\gamma, \theta) \vert \mathcal{C} \in \bar{\mathcal{C}} \rbrace \\
\nonumber
d_{\Gamma, \Theta}(\mathcal{C})&= \lbrace d_{\gamma , \theta}(\mathcal{C})\vert (\gamma, \theta) \in (\Gamma , \Theta)\, \text{plane} \rbrace .\end{aligned}$$
From Definition \[def:sing\], at a singular fade state the value of at least one of the class distance functions in $ d_{\bar{\mathcal{C}}}(\gamma, \theta) $ will reduce to zero.
\[lemma:ordering\] Among the set of all class distance functions that reduce to zero at the singular fade state $ (\gamma',\theta') $, there is a particular one which is the minimum among that set, for all values of $ (\gamma , \theta) \neq (\gamma',\theta' ) $.
Let $ L $ be the number of class distance functions that reduce to zero at the singular fade state $ (\gamma',\theta') $. Denote these by $ d_{\mathcal{C}_i}(\gamma, \theta) $, $ 1\leq i \leq L $ and let $ \lbrace (s_{1,i}, s_{2,i}),(s'_{1,i}, s'_{2,i})\rbrace $ be the representative element for the distance class $ \mathcal{C}_i $. From and , we have $$\begin{aligned}
\nonumber
d_{\mathcal{C}_i}^2(\gamma,\theta)&= d^2_{(s_{1,i},s_{2,i})\leftrightarrow (s'_{1,i},s'_{2,i})} \\
\nonumber
&= \vert (s_{1,i}-s'_{1,i}) + \gamma e^{j \theta} (s_{2,i}-s'_{2,i})\vert ^2\\
\nonumber
&= \vert s_{2,i}-s'_{2,i}\vert ^2 \vert \gamma e^{j \theta} + \frac{s_{1,i}-s'_{1,i}}{s_{2,i}-s'_{2,i}}\vert ^2 \\
\label{eqn:ordr1}
&= \vert s_{2,i}-s'_{2,i}\vert ^2 \vert \gamma e^{j \theta } - \gamma' e ^{j \theta'}\vert ^2.\end{aligned}$$ From , these $ L $ class distance functions differ only in the constant coefficient $ \vert s_{2,i}-s'_{2,i}\vert ^2 $. From Definition \[def:distclass\], all these coefficients are different. Let $$\begin{aligned}
\nonumber
l'= \text{arg} \, \min_{1\leq i \leq L} \vert s_{2,i}-s'_{2,i}\vert ^2.\end{aligned}$$ Now, from , $ d_{\mathcal{C}_{l'}}(\gamma, \theta) $ is minimum among all $ d_{\mathcal{C}_i}(\gamma, \theta) $ for all values of $ (\gamma , \theta) \neq (\gamma' , \theta')$.
\[rd\] The region corresponding to distance class $ \mathcal{C} $, $ \mathcal{R}(\mathcal{C}) $ denotes the region in the complex plane $ \lbrace (\Gamma, \Theta)/ \mathcal{H} \rbrace $ for which the class distance function $ d_{\mathcal{C}}(\gamma, \theta) $ gives the minimum distance in $ \mathcal{S}_{\text{eff}}$, *i.e.*, $$\begin{aligned}
\nonumber
\mathcal{R}(\mathcal{C}) = \lbrace &(\gamma ,\theta ) \in \lbrace (\Gamma , \Theta) / \mathcal{H} \rbrace \vert \\
\nonumber
&d_{\mathcal{C}}(\gamma ,\theta) \leq d_{\mathcal{C'}}(\gamma , \theta)\: \text{for all}\: \mathcal{C'} \neq \mathcal{C} \in \bar{\mathcal{C}} \rbrace .\end{aligned}$$
When both the users use $ M $-PSK constellations at the input, $ \mathcal{R}_W (\mathcal{C}) $ denotes the portion of the region $ \mathcal{R}(\mathcal{C}) $ lying in the wedge $ [0, \pi/M] $, *i.e.*, $$\begin{aligned}
\nonumber
\mathcal{R}_W(\mathcal{C}) = \mathcal{R}(\mathcal{C}) \cap \, \text{wedge}\, [0, \pi/M].\end{aligned}$$
Note that, when both the users use $ M $-PSK constellations at the input, for some $ \mathcal{C} \in \bar{\mathcal{C}} $ the corresponding region $ \mathcal{R}(\mathcal{C}) $ can be a null set, because the associated class distance function $ d_{\mathcal{C}}(\gamma, \theta) $ does not give the minimum distance in $ \mathcal{S}_{\text{eff}} $ for any value of $ (\gamma, \theta) $ in $ \lbrace (\Gamma, \Theta) \setminus \mathcal{H} \rbrace $. There is always a distance class $\mathcal{C} \in \bar{\mathcal{C}} $ for which the associated class distance function is $ d_{\mathcal{C}(\gamma, \theta)} = d_{min}(\mathcal{S}) $. We denote this particular distance class as $ \mathcal{C}_{d_{min}(\mathcal{S})} $. From Lemma \[lem:uppermin\], the value of this class distance function is the upper bound for the minimum distance in $ \mathcal{S}_{\text{eff}} $. For example when both users use QPSK signal sets at the input, then from Table \[tab:qpskex\], $ \mathcal{C}_{d_{min}(\mathcal{S})} = \mathcal{C}_1 $ and the associated class distance function is $ d_{\mathcal{C}_1}(\gamma, \theta)= d_{min}(\mathcal{S})= \sqrt{2} $.
The procedure to obtain the quantization of the $(\Gamma,\Theta)$ plane, when both users use $M$-PSK constellations at the input, is as follows:
1. Obtain the $N_W$ singular fade states in $ \mathcal{H}_W $ *i.e.*, lying in the wedge $[0, \pi /M]$. Each of these singular fade state is denoted by $ (\gamma _i,\theta _i) $ where $ 1\leq i\leq N_W $.
2. For the singular fade state $ (\gamma_1, \theta _1) $ in $ \mathcal{H}_W $, identify the set of class distance functions in $ d_{\bar{\mathcal{C}}}(\gamma , \theta) $ that reduces to zero at that singular fade state $ (\gamma_1, \theta_1) $. Choose the one among them, which is minimum in that set for all values of $ (\gamma, \theta) \neq (\gamma_1, \theta_1) $. (From Lemma \[lemma:ordering\], there is always only one such class distance function.) Let this class distance function be $ d_{\mathcal{C}_{k_1}}(\gamma ,\theta) $ corresponding to distance class $ \mathcal{C}_{k_1} $. Repeat this for all $ (\gamma_i, \theta_i) \in \mathcal{H}_W $, to obtain a set of class distance functions $ \lbrace d_{\mathcal{C}_{k_i}}(\gamma, \theta) $, $ 1 \leq i \leq N_W \rbrace $. Each $ d_{\mathcal{C}_{k_i}}(\gamma, \theta) $ reduces to zero at the singular fade state $ (\gamma_i, \theta_i) $. This is the set of all possible class distance functions other than $ d_{min}(\mathcal{S}) $, that can possibly produce the minimum distance in $ \mathcal{S}_{\text{eff}} $.
3. To find the region $ \mathcal{R}_W(\mathcal{C}_{k_i}) $, we need to obtain the values of $ (\gamma, \theta) \in \, \text{wedge}\, [0,\pi/M] $ for which $d^2_{\mathcal{C}_{k_i}}(\gamma, \theta) \leq d^2_{\mathcal{C}_{k_j}}(\gamma, \theta) $ where $1\leq j\neq i \leq N_W$, and $d^2_{\mathcal{C}_{k_i}}(\gamma, \theta) \leq d^2_{min}(\mathcal{S})$. The curves $d^2_{\mathcal{C}_{k_i}}(\gamma, \theta)=d^2_{\mathcal{C}_{k_j}}(\gamma, \theta)$, $1\leq i \neq j\leq N_W$, form the pairwise boundary between the regions corresponding to the two distance classes $ \mathcal{C}_{k_i} $ and $ \mathcal{C}_{k_j} $. The curves $ d^2_{\mathcal{C}_{k_i}}(\gamma ,\theta) = d^2_{min}(\mathcal{S})$ form the pairwise boundary between the regions corresponding to distance classes $ \mathcal{C}_{k_i} $ and $ \mathcal{C}_{d_{min}(\mathcal{S})} $. The region $\mathcal{R}_W(\mathcal{C}_{k_i})$ is that region in the wedge $ [0,\pi/M] $ excluding the complex point $ (\gamma_i, \theta_i) $, which is the innermost region bounded by these pairwise boundaries, enclosing the singular fade state $ (\gamma_i,\theta_i) $. For example, Fig. \[fig:8pskregion\] depicts the region corresponding to the singular fade state at $(1,0)$ when both users use 8-PSK signal sets. In the figure, the curve $d_1^2=d_j^2$ refers to the curve $d_{\mathcal{C}_{k_1}}^2(\gamma,\theta)=d_{\mathcal{C}_{k_k}}^2(\gamma,\theta)$, and $d_1^2= 2-\sqrt{2}$ refers to the curve $d_{\mathcal{C}_{k_1}}^2(\gamma ,\theta) = 2-\sqrt{2}$. It is the innermost region (shaded in the figure) in the wedge $ [0, \pi/8] $ bounded by the pair-wise boundaries $d_{\mathcal{C}_{k_1}}^2(\gamma, \theta)=d_{\mathcal{C}_{k_2}}^2(\gamma, \theta),\, d_{\mathcal{C}_{k_1}}^2(\gamma, \theta)=d_{\mathcal{C}_{k_4}}^2(\gamma, \theta),\, d_{\mathcal{C}_{k_1}}^2(\gamma, \theta)=d_{\mathcal{C}_{k_5}}^2(\gamma, \theta)$ and $ d_{\mathcal{C}_{k_1}}^2(\gamma, \theta)=d_{\mathcal{C}_{k_6}}^2(\gamma, \theta) $ surrounding the singular fade state $ (1,0)$. Once the regions $ \mathcal{R}(\mathcal{C}_{k_i}) $, $ 1\leq i \leq N_W $ are obtained, the region exterior to all these regions, lying within the wedge $ [0, \pi /M] $, is the region where $ d_{min}(\mathcal{S})$ is the minimum distance in $ \mathcal{S}_{\text{eff}} $, *i.e.*, the region $ \mathcal{R}_W(\mathcal{C}_{d_{min}(\mathcal{S})}) $.
4. The quantization obtained in Step 3, for the wedge $[0,\pi/M]$, can now be extended by the procedure suggested in Section \[sec:quandist\] to cover the entire $ (\Gamma,\Theta) $ plane.
We will illustrate the procedure with two examples.
\[exa:qpskquan\] *Channel Quantization for QPSK signal sets.*\
![Quantization of the wedge $[0,\pi/4]$ for QPSK signal sets. []{data-label="fig:qpsksector"}](qpsksector.eps){width="2.25in"}
Here we consider the scenario where both users use QPSK constellations at input, *i.e.* $M=4$. From Lemma \[periodic\], there are two singular fade states in the wedge $ [0,\pi/4]$, *i.e.* $N_W=2$ and these are at $(1,0)$ and $(\sqrt{2},\frac{\pi}{4})$. The class distance functions in $d_{\bar{\mathcal{C}}}(\gamma, \theta)$ which reduce to zero at these singular fade states are identified. For the singular fade state $(\sqrt{2},\frac{\pi}{4})$ the distance $d_{\mathcal{C}_{k_1}}$ falls to zero, and for $(1,0)$ the distances $d_{\mathcal{C}_{k_2}}$ and $d_{\mathcal{C}_{k_3}}$ both fall to zero, as shown in Fig \[fig:qpskeff\]. These are as follows: $$\begin{aligned}
\nonumber
d_{\mathcal{C}_{k_1}}^2(\gamma, \theta) &= 2\gamma ^2 +4 - 4\gamma \cos \theta - 4 \gamma \sin \theta \\
\nonumber
d_{\mathcal{C}_{k_2}}^2(\gamma, \theta) &= 2\gamma ^2 + 2 -4 \gamma \cos \theta\\
\nonumber
d_{\mathcal{C}_{k_3}}^2(\gamma, \theta) &= 4\gamma ^2 + 4 -8 \gamma \cos \theta = 2d_{\mathcal{C}_{k_2}}^2 \geq d_{\mathcal{C}_{k_2}}^2.\end{aligned}$$ As $d_{\mathcal{C}_{k_3}}(\gamma, \theta) > d_{\mathcal{C}_{k_2}}(\gamma, \theta)$ for all values of $(\gamma,\theta) \neq (1,0)$, we consider the class distance function corresponding to the singular fade state $(1,0)$ as $d_{\mathcal{C}_{k_2}}(\gamma, \theta)$. Now we proceed to obtain the regions $ \mathcal{R}_W(\mathcal{C}_{k_1})$, $ \mathcal{R}_W(\mathcal{C}_{k_2})$ and $\mathcal{R}_W(\mathcal{C}_{d_{min}(\mathcal{S})})$. The pair-wise boundaries are obtained as follows,
[ $$\begin{aligned}
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma ,\theta) &= d^2_{\mathcal{C}_{k_2}}(\gamma, \theta) \Rightarrow \gamma \sin \theta = 1/2\\
\nonumber
d^2_{\mathcal{C}_{k_1}} (\gamma, \theta) &= d^2_{min}(\mathcal{S})=2 \Rightarrow (\gamma \cos \theta -1)^2+(\gamma \sin \theta -1)^2 =1\\
\nonumber
d^2_{\mathcal{C}_{k_2}}(\gamma, \theta) &= d^2_{min}(\mathcal{S})=2 \Rightarrow (\gamma \cos \theta -1)^2+\gamma ^2 \sin ^2 \theta =1.\end{aligned}$$ ]{} These regions are shown in Fig. \[fig:qpsksector\] for the wedge $[0,\pi/4] $. This can now be extended to cover the entire range of $ \theta $. For this, the quantization obtained for $ \theta \in [0, \pi/4] $ is reflected along the line $ \theta = \pi/4 $ to obtain the quantization for the wedge $ [0, \pi/2] $. This is now rotated by $ \frac{k\pi}{2} $, $1\leq k\leq 3 $, to obtain the quantization for $ \theta \in \left[ \frac{k\pi}{2}, \frac{(k+1)\pi}{2} \right] $, thus covering the entire $ (\Gamma,\Theta) $ plane. This has been shown in Fig. \[fig:qpskquan\].
![Quantization of the entire $(\Gamma, \Theta)$ plane for QPSK signal sets []{data-label="fig:qpskquan"}](qpskquan.eps){width="2.2in"}
*Channel quantization for 8-PSK signal sets*
Here we consider the scenario when both users use 8-PSK signal sets at the input. The number of singular fade states lying in the wedge $[0, \pi/8]$ is, $N_W=7$. The singular fade states $(\gamma _i,\theta _i)$, $1\leq i \leq 7$ are as follows: $$\begin{aligned}
\nonumber
&(1, 0),\: (\sqrt{2},0), \:(1+\sqrt{2} ,0), \\
\nonumber
&\left(\sqrt{4-2\sqrt{2}} ,\frac{\pi}{8}\right),\: \left(\sqrt{1+\frac{1}{\sqrt{2}}} ,\frac{\pi}{8}\right),\\
\nonumber
&\left(\sqrt{2+\sqrt{2} } ,\frac{\pi}{8}\right),\: \left(\sqrt{4+2\sqrt{2}} ,\frac{\pi}{8}\right).\end{aligned}$$ The class distance functions in $d_{\bar{\mathcal{C}}}(\gamma, \theta)$ which reduce to zero for each of the above singular fade states are identified. When more than one class distance function reduces to zero at a singular fade state $ (\gamma_i, \theta_i) $, the one which is minimum among them for any other $ (\gamma, \theta) \neq (\gamma_i, \theta_i) $ is chosen. The class distance functions $ d_{\mathcal{C}_{k_i}}(\gamma, \theta) $, $ 1 \leq i \leq 7 $ are identified as shown in Fig. \[fig:8pskeff\]. In the figure $ d_i $ denotes the class distance function $ d_{\mathcal{C}_{k_i}}(\gamma , \theta) $. These are as follows
[ $$\begin{aligned}
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma ,\theta) &= (2-\sqrt{2})(\gamma ^2 - 2\gamma \cos \theta +1)\\
\nonumber
d^2_{\mathcal{C}_{k_2}}(\gamma ,\theta) &= 2\gamma ^2 + 4 - 4\sqrt{2}\gamma \cos \theta \\
\nonumber
d^2_{\mathcal{C}_{k_3}}(\gamma ,\theta) &= (2-\sqrt{2})\gamma ^2 + 2 +\sqrt{2}- 2\sqrt{2}\gamma \cos \theta \\
\nonumber
d^2_{\mathcal{C}_{k_4}}(\gamma ,\theta) &= (2+\sqrt{2})\gamma ^2 +4 -2(2+\sqrt{2})\gamma \cos \theta -2\sqrt{2}\gamma\sin \theta \\
\nonumber
d^2_{\mathcal{C}_{k_5}}(\gamma ,\theta) &= 2\gamma ^2 +2 + \sqrt{2}-2(1+\sqrt{2})\gamma\cos\theta - 2\gamma\sin\theta \\
\nonumber
d^2_{\mathcal{C}_{k_6}}(\gamma ,\theta) &= (2-\sqrt{2})\gamma ^2 +2 -2(\sqrt{2}-1)\gamma\sin\theta - 2\gamma\cos\theta \\
\nonumber
d^2_{\mathcal{C}_{k_7}}(\gamma ,\theta) &= (2-\sqrt{2})\gamma ^2 +4 -2(2-\sqrt{2})\gamma\sin\theta -2\sqrt{2}\gamma\cos\theta.\end{aligned}$$ ]{} Now to obtain the region $\mathcal{R}_W(\mathcal{C}_{k_1})$ we need to obtain the curves $d^2_{\mathcal{C}_{k_1}}(\gamma,\theta)=d^2_{\mathcal{C}_{k_j}}(\gamma ,\theta) $, $2\leq j\leq7$ and $d^2_{\mathcal{C}_{k_1}}(\gamma ,\theta) = d^2_{min}(\mathcal{S})= 2-\sqrt{2}$. These are as follows:
[ $$\begin{aligned}
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma,\theta) &= d^2_{\mathcal{C}_{k_2}}(\gamma,\theta) \Rightarrow (\gamma\cos\theta -(3-\sqrt{2}))^2+\gamma ^2 \sin ^2 \theta = 10 -7\sqrt{2}\\
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma,\theta) &= d^2_{\mathcal{C}_{k_3}}(\gamma , \theta) \Rightarrow \gamma\cos\theta = \frac{1}{ 2-\sqrt{2}}\\
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma, \theta) &= d^2_{\mathcal{C}_{k_4}}(\gamma, \theta) \Rightarrow (\gamma\cos\theta -1)^2+(\gamma \sin\theta - \frac{1}{2})^2= \frac{3 -2\sqrt{2}}{4}\\
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma,\theta) &= d^2_{\mathcal{C}_{k_5}}(\gamma, \theta) \Rightarrow (\gamma\cos\theta -2+\frac{1}{\sqrt{2}})^2+(\gamma\sin \theta - \frac{1}{\sqrt{2}})^2 \\
\nonumber
& \hspace{2in} = 3 -2\sqrt{2}\\
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma ,\theta) &= d^2_{\mathcal{C}_{k_6}}(\gamma,\theta) \Rightarrow \gamma\cos\theta +\gamma \sin \theta = 1 +\frac{1}{\sqrt{2}}\\
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma,\theta) &= d^2_{\mathcal{C}_{k_7}}(\gamma, \theta) \Rightarrow \sqrt{2}\gamma\cos\theta +\gamma \sin \theta = \frac{3}{2}+ \sqrt{2}\\
\nonumber
d^2_{\mathcal{C}_{k_1}}(\gamma, \theta) &= 2-\sqrt{2} \Rightarrow (\gamma\cos\theta -1)^2+\gamma ^2 \sin ^2 \theta = 1.\end{aligned}$$ ]{}
![Effective constellation $\mathcal{S}_{\text{eff}}$ for $(\gamma,\theta) = (2.9,10^\circ)$ when both users use 8-PSK signal set.[]{data-label="fig:8pskeff"}](8psksum.eps){width="2.5in"}
All of the above curves are shown in Fig. \[fig:8pskregion\]. In the figure, the curve $d_1^2=d_j^2$ refers to the curve $d_{\mathcal{C}_{k_1}}^2(\gamma,\theta)=d_{\mathcal{C}_{k_k}}^2(\gamma,\theta)$, and $d_1^2= 2-\sqrt{2}$ refers to the curve $d_{\mathcal{C}_{k_1}}^2(\gamma ,\theta) = 2-\sqrt{2}$. The region $\mathcal{R}_W(\mathcal{C}_{k_1})$ is the innermost region in the wedge $ [0,\pi/8] $ bounded by these curves, surrounding the point $ (1,0) $. It is the shaded region in the Fig. \[fig:8pskregion\]. All the regions $\mathcal{R}_W(\mathcal{C}_{k_i})$, $1\leq i \leq 7$ can be obtained by the same procedure. The region exterior to all these regions, inside the wedge $[0, \pi/8] $ is the region $\mathcal{R}_W(\mathcal{C}_{d_{min}(\mathcal{S})})$, where $d_{min}(\mathcal{S}) = \sqrt{2-\sqrt{2}}$ is the minimum distance in $\mathcal{S}_{\text{eff}}$. Thus, the quantization of the fade states for the wedge $[0, \pi /8]$ is obtained. It is shown is Fig. \[fig:8psksector\] in the next page. This can now be extended to cover the entire $ (\Gamma, \Theta) $ plane by the similar technique used for the QPSK case in the Example \[exa:qpskquan\].
THE ADAPTIVE MODULATION SCHEME
==============================
In this section the fade states which results in reducing the minimum distance in $ \mathcal{S}_{\text{eff}} $ below a minimum distance guarantee of $ \delta $ are identified. Then a modulation scheme is proposed for the users to avoid these fade states by suitable relative rotation between the signal sets used by the two users.
Adaptive Modulation Scheme {#sec:adapt}
--------------------------
It is clear from Section \[sec:chan\_quan\] that the minimum distance in $ \mathcal{S}_{\text{eff}} $ falls to zero at the singular fade states. For fade states $ (\gamma ,\theta) $ lying close to a singular fade state, the minimum distance in $ \mathcal{S}_{\text{eff}} $ is very low, resulting in degradation of error performance at the destination. Hence, such values of $ (\gamma ,\theta)$ have to be avoided to provide better performance.
Our goal is to provide a minimum distance guarantee of $ \delta $ in $ \mathcal{S}_{\text{eff}} $ *i.e.*, not allow minimum distance in $ \mathcal{S}_{\text{eff}} $ to fall below $ \delta $. In the previous subsection, the regions $ \mathcal{R}(\mathcal{C}_{k_i}) $ on the $ (\Gamma , \Theta) $ plane was identified, in which the class distance function $ d_{\mathcal{C}_{k_i}}(\gamma, \theta) $, $ 1\leq i \leq N_W $ gives the minimum distance in $ \mathcal{S}_{\text{eff}} $. In order to satisfy the minimum distance guarantee of $ \delta $ in $ \mathcal{S}_{\text{eff}} $, it is thus required to avoid the fade states $ (\gamma,\theta) $ for which $ d_{\mathcal{C}_{k_i}}(\gamma ,\theta) < \delta $, $ 1\leq i \leq N_W $.
At the singular fade state $ (\gamma_i, \theta _i) $, the class distance function $ d_{\mathcal{C}_{k_i}}(\gamma, \theta) $ reduces to zero. If $ \lbrace (s_{1,i},s_{2,i}), (s'_{1,i},s'_{2,i})\rbrace $ is the representative element for the distance class $\mathcal{C}_{k_i}$, then at the singular fade state $ (\gamma _i, \theta _i) $ the two points $ (s_{1,i},s_{2,i}) $ and $ (s'_{1,i},s'_{2,i}) $ collapse to a single point in $ \mathcal{S}_{\text{eff}} $. From the definition of singular fade state , we have $$\begin{aligned}
\nonumber
\gamma _i e^{j \theta _i} = -\frac{s_{1,i}-s'_{1,i}}{s_{2,i}-s'_{2,i}}.\end{aligned}$$ From , we have $$\begin{aligned}
\label{eqn:di-eff}
d_{\mathcal{C}_{k_i}}(\gamma , \theta) = \vert s_{2,i}-s'_{2,i}\vert \vert \gamma e^{j \theta } - \gamma _i e ^{j \theta _i}\vert.\end{aligned}$$ Only those fade states $ (\gamma ,\theta) $ which results in $ d_{\mathcal{C}_{k_i}}(\gamma, \theta) < \delta $, $1\leq i \leq N_W$ have to be avoided, *i.e.*, from , we need to avoid the fade states $ (\gamma ,\theta) $, for $ 1\leq i \leq N_W $, where $$\begin{aligned}
\label{eqn:violation}
\vert \gamma e^{j\theta} - \gamma _i e^{j \theta _i}\vert < \frac{\delta}{\vert s_{2,i}-s'_{2,i}\vert}.\end{aligned}$$ The above equation represents a circular region in the complex plane $ (\Gamma, \Theta) $ centred at the singular fade state $ (\gamma _i, \theta _i) $ and radius $ \delta /\vert s_{2,i}-s'_{2,i}\vert $. We call these circular regions, the *violation circles* because when the fade state lies inside them the minimum distance requirement of $ \mathcal{S}_{\text{eff}} $ is violated. For fade states outside the violation circles the minimum distance in $ \mathcal{S}_{\text{eff}} $ is always greater than $ \delta $. The radius of the violation circle centred at the singular fade state $ (\gamma_i, \theta _i) $ by is denoted by $ \rho (\gamma_i, \theta_i) $. Formally violation circles are defined as follows:
*Violation Circles* are circular regions on the $(\Gamma ,\Theta)$ plane with centres at the singular fade states $ (\gamma _i,\theta _i)= -\frac{s_{1,i}-s'_{1,i}}{s_{2,i}-s'_{2,i}}$ with radius $ \rho(\gamma_i,\theta_i)=\delta /\vert s_{2,i}-s'_{2,i}\vert $ for $1\leq i \leq N_W$.
It is observed that the the violation circles centred at $ (\gamma_i,\theta_i+p\frac{2\pi}{M}) $, $ 1\leq i \leq N_W \, \text{and} \, 1\leq p \leq M-1 $ have the same radius as the one centred at $ (\gamma_i,\theta_i) $, where $ (\gamma_i , \theta_i) \in \mathcal{H}_W $. This is because from Lemma \[lemmaangle\], the corresponding effective constellations are the same.
{width="5in"}
{width="6in"}
When both users use QPSK constellations at the input, the violation circles are as follows. At singular fade state $ (\sqrt{2},\pi /4) $ the class distance function $ d_{\mathcal{C}_{k_1}}(\gamma,\theta) $ reduces to zero, *i.e.* the points $ 3 $ and $ 5 $ in $ \mathcal{S}_{\text{eff}} $ collapse to a single point, as shown in Fig. \[fig:qpskeff\]. The $ 3 $ and $ 5 $ are obtained after combining $ 3,\,1 \in \mathcal{S} $ and $ 1,\,2 \in \mathcal{S} $ respectively. Thus, $ \vert s_{2,1}-s'_{2,1} \vert = \sqrt{2}$. The violation circle corresponding to class distance function $ d_{\mathcal{C}_{k_1}}(\gamma,\theta) $ is the circular region centred at $ (\sqrt{2}, \pi/4) $ and radius $\delta/\sqrt{2}$. Similarly, for the singular fade state at $ (1,0) $ the violation circle has a radius of $\delta/\sqrt{2}$. These are shown by dotted circles around the singular fade states in Fig. \[fig:qpskquan\].
When both users are using 8-PSK constellations at the input, the centres and radii of the violation circles corresponding to the distances $ d_{\mathcal{C}_{k_i}},\, 1\leq i \leq 7 $ are tabulated in Table \[table1\].
$ i $ Centre Radius
------- ----------------------------------------------------------- ------------------------------------
$ 1 $ $(1,0)$ $\frac{\delta}{\sqrt{2-\sqrt{2}}}$
$ 2 $ $(\sqrt{2},0)$ $\frac{\delta}{\sqrt{2}}$
$ 3 $ $(\sqrt{2}+1 ,0)$ $\frac{\delta}{\sqrt{2-\sqrt{2}}}$
$ 4 $ $\left(\sqrt{4-2\sqrt{2}},\frac{\pi}{8}\right)$ $\frac{\delta}{\sqrt{2+\sqrt{2}}}$
$ 5 $ $\left(\sqrt{1+\frac{1}{\sqrt{2}}} ,\frac{\pi}{8}\right)$ $\frac{\delta}{\sqrt{2}}$
$ 6 $ $\left(\sqrt{2+\sqrt{2}} ,\frac{\pi}{8}\right)$ $\frac{\delta}{\sqrt{2-\sqrt{2}}}$
$ 7 $ $\left(\sqrt{4+2\sqrt{2}},\frac{\pi}{8}\right)$ $\frac{\delta}{\sqrt{2-\sqrt{2}}}$
: Centre and Radius of violation circles for 8-PSK case[]{data-label="table1"}
When the fade state $ (\gamma ,\theta) $ lies inside any of the violation circles the users need to adapt their transmission, in order to avoid these fade states effectively. One way to achieve this without increasing the transmit power, is to rotate the constellation of User-2. Rotation can be interpreted as simply altering the phase of the fade state.
\[relative\] When the fade state is $(\gamma , \theta )$, rotation of the constellation of User-2 by an angle $\alpha$ with respect to the constellation of User-1 in an anticlockwise direction, results in effectively altering the phase of the fade state from $ \theta $ to $ \theta + \alpha $.
Let $\mathcal{S}$ be the constellation being used by both the users at the input. Now the User-2 rotates its constellation by angle $ \alpha $ in the anticlockwise direction, such that it can now be represented as $ e^{j \alpha} \mathcal{S} $. The effective constellation $ \mathcal{S}_{\text{eff}} $ can be written as $$\begin{aligned}
\nonumber
\mathcal{S}+ \gamma e^ {j \theta}\lbrace e^{j \alpha} \mathcal{S}\rbrace =\mathcal{S}+ \gamma e^ {j (\theta +\alpha)} \mathcal{S}.\end{aligned}$$ Hence, the fade state $(\gamma , \theta)$ is transformed to $(\gamma, \theta + \alpha)$ after rotation.
The proposed strategy is thus, to rotate the signal set of User-2 whenever the fade state $(\gamma ,\theta)$ lies within any of the violation circles such that the transformed fade state lies outside the violation circles, in order to satisfy the minimum distance guarantee in $ \mathcal{S}_{\text{eff}} $. For fade states outside the violation circles no rotation is required. The destination which has complete CSI sends feedback of $\lceil \log _2(N_W+1)\rceil$ bits to the users to indicate in which one of the violation circles the fade state lies, or if it lies outside all of them.
Optimal Angle of Rotation for the $M$-PSK case {#sec:optimal}
----------------------------------------------
An optimal rotation angle, for a violation circle with centre at singular fade state $(\gamma_i,\theta_i)$ , $ 1\leq i \leq N_W $ is that angle of rotation which maximizes the minimum distance in $\mathcal{S}_{\text{eff}}$ for the same transmit power, when fade state $ (\gamma ,\theta ) = (\gamma_i, \theta _i)$.
It should be noted for non-singular fade states inside the violation circle, the minimum distance in $ \mathcal{S}_{\text{eff}} $ after rotation will be less than what it could have been if the fade state $ (\gamma ,\theta) $ was exactly known at the users. When both users use $M$-PSK constellations at the input, it is sufficient to concentrate on the violation circles centred at $(\gamma_i, \theta_i) \in \mathcal{H}_W $ , because the optimal rotation angles for these, is also optimal for other such circles centred at $(\gamma_i, \theta _i +p \frac{2 \pi}{M})$, where $ 1 \leq p \leq M-1$. This follows from the fact that the corresponding effective constellations are equivalent.
From Lemma \[relative\], rotation of the constellation of User-2 relative to User-1, results in effectively altering the phase of the fade state. Rotation thus results in moving the violation circle with centre at singular fade state $(\gamma _i, \theta _i) $, along a circular arc such that its centre always lie on the curve $\gamma=\gamma_i$ and within the wedge $ [0, \pi/M] $. In order to obtain the optimal angle of rotation for the violation circle centred at $(\gamma_i, \theta _i)$, it is required to calculate for fixed $\gamma=\gamma_i$ the value of phase $\theta = \theta_{i,opt} $, $ \theta \in [0,\pi/M] $ which maximizes the minimum distance in $\mathcal{S}_{\text{eff}}$. Thus after rotation the violation circle centred $ (\gamma_i ,\theta _i) $ is shifted such that its new centre is the point $ (\gamma_i, \theta_{i,opt}) $ on the $ (\Gamma, \Theta) $ plane. We refer to this rotated violation circle as the effective shifted circle. We now prove two lemmas to obtain the value $ \theta_{i,opt} $ when both users use $M$-PSK signal sets.
\[increasedi\] Let $ d_{\mathcal{C}_{k_i}}(\gamma, \theta) $ be the class distance function which reduces to zero at the singular fade state $ (\gamma_i, \theta_i) \in \mathcal{H}_W $. For a fixed $ \gamma_0 $ and $ (\gamma_0, \theta) $ lying within the wedge $ [0, \pi/M] $, the value of $ d_{\mathcal{C}_{k_i}}(\gamma_0,\theta) $ increases as the difference $ \vert \theta_i - \theta \vert $ increases.
From , we have [ $$\begin{aligned}
\nonumber
d_{\mathcal{C}_{k_i}}^2(\gamma_0, \theta) &= \vert s_{2,i}-s'_{2,i}\vert ^2 \vert \gamma _0 e^{j\theta} - \gamma _i e^{j \theta _i}\vert ^2\\
\nonumber
&= \vert s_{2,i} - s'_{2,i}\vert ^2 \gamma _0 ^2 \vert 1 -\gamma' e^{j(\theta_i - \theta )}\vert ^2 \,\text{where}\, \gamma'=\gamma_i /\gamma _0 >0\\
\nonumber
&= \vert s_{2,i} - s'_{2,i}\vert ^2 \gamma _0 ^2 \left\lbrace 1+ {\gamma '}^2 - 2\gamma ' \cos \vert \theta _i - \theta \vert \right\rbrace. \end{aligned}$$ ]{} As $ \cos \phi $ is a decreasing function of $\phi$, $ 0\leq \phi \leq \pi/M $, $ d_{\mathcal{C}_{k_i}}^2(\gamma_0, \theta) $ increases as $ \vert \theta_i - \theta \vert $ increases. This proves the lemma.
\[lem:optimalboundary\] Let $(\gamma_i, \theta_i)$ be a singular fade state. Let $\mathcal{R}_W(\mathcal{C}_{k_a})$ and $\mathcal{R}_W(\mathcal{C}_{k_b})$ be the regions surrounding the singular fade states $(\gamma_a, \theta_a)$ and $(\gamma_b, \theta_b)$ respectively. Consider the arc traced by the point $(\gamma_i, \theta)$ that lies within the wedge $[0,\pi/M]$ as $\theta$ varies in the direction to move away from the singular fade state $(\gamma_i, \theta_i)$. Let the region $\mathcal{R}(\mathcal{C}_{k_a})$ is encountered before $\mathcal{R}(\mathcal{C}_{k_b})$ as $\theta$ varies. We have the following:
- The minimum distance in $\mathcal{S}_{\text{eff}}$ is maximized at one of the points of intersection of this arc and the boundaries between the regions $\mathcal{R}(\mathcal{C}_{k_a})$ and $\mathcal{R}(\mathcal{C}_{k_b})$.
- Among all the points in (i), those that lie on the boundary between $\mathcal{R}(\mathcal{C}_{k_a})$ and $\mathcal{R}(\mathcal{C}_{k_b})$ with $\theta_i = \theta_a = \theta_b$ can never correspond to the maximum value.
The proof for part (i) is as follows. Let $\mathbf{A}$ and $\mathbf{B}$ be the first and the second points of intersection of the arc with the boundary of the region $\mathcal{R}_W(\mathcal{C}_{k_a})$. From Section II-B, the value of class distance function $d_{\mathcal{C}_{k_a}}(\gamma, \theta)$ gives the minimum distance in $\mathcal{S}_{\text{eff}}$ when $ (\gamma, \theta) \in \mathcal{R}_W({\mathcal{C}_{k_a}})$. Now if $\theta_a = \theta_i$, then from Lemma \[increasedi\], $d_{\mathcal{C}_{k_a}}(\gamma_i, \theta)$ increases as $(\gamma_i, \theta)$ moves from $\mathbf{A}$ to $\mathbf{B}$. Likewise, if $\theta_a \neq \theta_i$, then again from Lemma \[increasedi\], $d_{\mathcal{C}_{k_a}}(\gamma_i, \theta)$ decreases as one moves from $\mathbf{A}$ to $\mathbf{B}$. Thus the minimum distance in $\mathcal{S}_{\text{eff}}$ can never be maximum for $(\gamma_i, \theta)$ lying inside the regions $\mathcal{R}_W(\mathcal{C}_{k_a})$. It can only be maximized at the points of intersection of the arc with the boundary of the region.
To prove the second part of the lemma, we assume $\theta_i= \pi/M$. Consider the scenario shown in Fig. 8. To prove part (ii), we need to show that the minimum distance in $\mathcal{S}_{\text{eff}}$ can never be maximum at the point $\mathbf{B}$. In the region $\mathcal{R}_W(\mathcal{C}_{k_a})$ the value of class distance function $d_{\mathcal{C}_{k_a}}(\gamma, \theta)$ gives the minimum distance in $\mathcal{S}_{\text{eff}}$. As $(\gamma_i,\theta)$ moves from $\mathbf{A}$ to $\mathbf{B}$, from Lemma \[increasedi\], the minimum distance in $\mathcal{S}_{\text{eff}}$, *i.e.* the value of $d_{\mathcal{C}_{k_a}}(\gamma_i, \theta)$, increases since $\vert \theta_a- \theta \vert$ increases. At $\mathbf{B}$, we have $d_{\mathcal{C}_{k_a}}(\gamma_i, \theta)= d_{\mathcal{C}_{k_b}}(\gamma_i, \theta)$. Beyond $\mathbf{B}$, in the region $\mathcal{R}_W(\mathcal{C}_{k_b})$, the class distance function $d_{\mathcal{C}_{k_b}}(\gamma, \theta)$ gives the minimum distance in $\mathcal{S}_{\text{eff}}$. As $(\gamma_i, \theta)$ moves from $\mathbf{B}$ to $\mathbf{C}$, $\vert \theta_b - \theta \vert $ increases, thus from Lemma \[increasedi\], the minimum distance in $\mathcal{S}_{\text{eff}}$, *i.e.* the value of $d_{\mathcal{C}_{k_b}}(\gamma_i, \theta)$ continues to increase. So $\mathbf{B}$ can never correspond to the point where the minimum distance in $\mathcal{S}_{\text{eff}}$ is maximized. The proof for the case when $\theta_i=0$ is exactly similar to the above proof. This completes the proof.
\[t\] {width="3.2in"} \[pointersect\]
The procedure to obtain the optimal phase $ \theta_{i,opt} $ of the fade state, for the violation circle centred at singular fade state $ (\gamma_i, \theta _i) \in \mathcal{H}_W $ is stated as follows:
- Find the points of intersections of the arc $ \gamma = \gamma_i $, $ \theta \in [0, \pi/M] $, with the boundaries that satisfy the conditions mentioned in Lemma \[lem:optimalboundary\].
- If there is only one such point of intersection, say with the boundary between the regions $\mathcal{R}_W(\mathcal{C}_{k_a})$ and $\mathcal{R}_W(\mathcal{C}_{k_b})$, $ \theta_{i,opt} $ is obtained by solving the equation $ d_{\mathcal{C}_{k_a}}^2(\gamma,\theta)=d_{\mathcal{C}_{k_b}}^2(\gamma,\theta) \vert _{\gamma=\gamma_i} $. On the other hand, if there are $L$ such points of intersections, say with boundaries between regions $\mathcal{R}_W(\mathcal{C}_{k_{a,l}})$ and $\mathcal{R}_W(\mathcal{C}_{k_{b,l}})$, $1\leq l \leq L$, calculate the phase of each of these points of intersection, $ \theta_{l,\text{intersect}} $, $ 1\leq l \leq L $ by solving the equation $ d_{\mathcal{C}_{k_{a,l}}}^2(\gamma,\theta)=d_{\mathcal{C}_{k_{b,l}}}^2(\gamma, \theta) \vert _{\gamma=\gamma_i}$. Then compute the minimum distances in $ \mathcal{S}_{\text{eff}} $ for the fade state corresponding to the point of intersection $ (\gamma_i , \theta_{l,\text{intersect}}) $ as $ d_{min}(\gamma_i, \theta_{l,\text{intersect}})= d_{\mathcal{C}_{k_{a,l}}}^2 (\gamma, \theta)\vert_{\gamma=\gamma_i, \theta= \theta_{l,\text{intersect}}} $. Choose $$\begin{aligned}
\nonumber
l'&= \text{arg}\,\max_{1\leq l \leq L} d_{min}(\gamma_i , \theta_{l,\text{intersect}}).\end{aligned}$$ Then we have, $$\begin{aligned}
\nonumber
\theta_{i,opt}&= \theta_{l',\text{intersect}}.\end{aligned}$$
The optimal rotation angles for the violation circle with centres at $ (\gamma_i,\theta_i) $, can now be easily calculated from $ \theta_{i,opt} $. The optimal rotation angle, $ \alpha_{i,opt} $, for the violation circle centred at $ (\gamma_i, \theta_i) $ is that rotation angle that transforms the fade state from $ (\gamma_i, \theta_i) $ to $ (\gamma_i,\theta_{i,opt}) $. (See Lemma \[relative\].)
For violation circles with centre at $ (\gamma_i ,\theta_i) $ , $ 1\leq i \leq N_W $, the optimal rotation angles for the User-2 are as follows:
- If $ \theta_i = \pi/M $, from Lemma \[relative\], the optimal rotation is $ \alpha_{i,opt} = \pi/M - \theta_{i,opt} $ in a clockwise direction.
- If $ \theta_i = 0 $, from Lemma \[relative\], the optimal rotation is $ \alpha_{\text{i,opt}} = \theta_{\text{i,opt}} $ in an anticlockwise direction.
*Optimal angles for QPSK signal sets* When both users use QPSK constellations at input, the channel quantization is shown in Fig. \[fig:qpsksector\]. The optimal rotation angle are calculated as shown below. These are shown in Fig. \[fig:optimal\].
![Optimal rotation angles for the QPSK case[]{data-label="fig:optimal"}](angles.eps){width="2.5in"}
- For violation circle centred at $ (\sqrt{2},\frac{\pi}{4}) $, the optimal phase $ \theta_{1,opt} $ is the phase of point of intersection of the arc $ \gamma = \sqrt{2} $ and pairwise boundary $ d_{\mathcal{C}_{k_1}}^2(\gamma,\theta)=d_{\mathcal{C}_{k_2}}^2(\gamma, \theta) $ *i.e.* $ \gamma \sin \theta = 0.5 $. Therefore, $ \theta_{1,opt} = \sin ^{-1} (\frac{1}{2\sqrt{2}}) \approx 20.7 ^\circ$. Thus the optimal rotation angle for the constellation of User-2 relative to User-1 is $ \alpha_{1,opt} = 45^\circ - 20.7^\circ = 24.3 ^\circ $ in a clockwise direction.
- For violation circle centred at $ (1,0) $, the optimal phase corresponds to the point of intersection of the arc $ \gamma =1 $ and boundary $ d_{\mathcal{C}_{k_1}}^2(\gamma, \theta) = d_{\mathcal{C}_{k_2}}^2(\gamma,\theta) $, *i.e.* $ \gamma \sin \theta = 0.5 $. Thus $ \theta _{2,opt} = \sin ^{-1}(0.5) = 30^\circ $. Thus the optimal rotation angle for User-2 is $ \alpha_{2,opt} = 30^\circ $ in an anticlockwise direction relative to User-1.
*Optimal Angles for 8-PSK signal sets* When both users are using 8-PSK constellations at the input, the channel quantization is obtained as shown in Fig. \[fig:8psksector\]. The optimal angles of rotation for User-2 relative to User-1, $ \alpha _{i,opt} $ can be calculated using the same technique for all the violation circles centred at the singular fade states $ (\gamma_i,\theta_i) $, $ 1\leq i \leq 7 $. For each $ (\gamma_i, \theta_i) $, $ 1\leq i \leq 7 $, the boundaries corresponding to the optimal phase, $ \theta_{i,opt} $ and $ \alpha _{i,opt} $ are tabulated in Table \[table2\]. The letters (C) or (A) in the column corresponding to $ \alpha_{i,opt} $ indicates the direction of rotation as clockwise and anticlockwise respectively.
$ i $ $ (\gamma_i, \theta_i) $ Boundary $ \theta _{i,opt} $ $ \alpha _{i,opt} $ (C/A)
------- ----------------------------------------------------------- -------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------ ---------------------------
$ 1 $ $(1,0)$ $ d^2_{\mathcal{C}_{k_1}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_4}}(\gamma, \theta) $ $ \tan ^{-1} \left( \frac{1}{2}\right)- \cos ^{-1} \left( \frac{\gamma^2_1 + \frac{1}{2}+\frac{1}{\sqrt{2}}}{\sqrt{5}\gamma_1}\right) \approx 17.3^\circ $ $ 17.3 $ (A)
$ 2 $ $(\sqrt{2},0)$ $ d^2_{\mathcal{C}_{k_1}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_5}}(\gamma, \theta) $ $ \tan ^{-1} \left( \frac{1}{2\sqrt{2}-1}\right)- \cos ^{-1} \left( \frac{\gamma^2_2 + 2}{2 \gamma _2 \sqrt{5- 2 \sqrt{2}}}\right) \approx 12.4^ \circ $ $ 12.4^\circ $ (A)
$ 3 $ $(\sqrt{2}+1 ,0)$ $ d^2_{\mathcal{C}_{k_3}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_7}}(\gamma, \theta) $ $ \sin ^{-1}\left(\frac{1}{2 \gamma_3} \right) \approx 12^\circ$ $ 12^\circ $ (A)
$ 4 $ $\left(\sqrt{4-2\sqrt{2}},\frac{\pi}{8}\right)$ $ d^2_{\mathcal{C}_{k_1}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_4}}(\gamma, \theta) $ $ \tan ^{-1} \left( \frac{1}{2}\right)- \cos ^{-1} \left( \frac{\gamma^2_4 + \frac{1}{2}+\frac{1}{\sqrt{2}}}{\sqrt{5}\gamma_4}\right) \approx 15.9^\circ$ $ 6.6^\circ $ (C)
$ 5 $ $\left(\sqrt{1+\frac{1}{\sqrt{2}}} ,\frac{\pi}{8}\right)$ $ d^2_{\mathcal{C}_{k_1}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_5}}(\gamma, \theta) $ $ \tan ^{-1} \left( \frac{1}{2\sqrt{2}-1}\right)- \cos ^{-1} \left( \frac{\gamma^2_5 + 2}{2 \gamma_5 \sqrt{5-2\sqrt{2}}}\right) \approx 13^\circ$ $ 9.5^\circ $ (C)
$ 6 $ $\left(\sqrt{2+\sqrt{2}} ,\frac{\pi}{8}\right)$ $ d^2_{\mathcal{C}_{k_3}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_6}}(\gamma, \theta) $ $ \cos ^{-1} \left(\frac{1+\sqrt{2}}{2\gamma_6}\right)-\frac{\pi}{4} \approx 4.2^\circ$ $ 18.3^\circ $ (C)
$ 7 $ $\left(\sqrt{4+2\sqrt{2}},\frac{\pi}{8}\right)$ $ d^2_{\mathcal{C}_{k_3}}(\gamma, \theta)= d^2_{\mathcal{C}_{k_7}}(\gamma, \theta) $ $ \sin^{-1} \left(\frac{1}{2\gamma_7} \right) \approx 11^\circ$ $ 11.5^\circ $ (C)
Upper Bound on $\delta$ {#sec:upper}
-----------------------
In this subsection we obtain an upper bound on $ \delta $ for which the proposed rotation scheme can be employed. It is necessary that the violation circle corresponding to any of the singular fade states must not overlap with any of the effective shifted circles, otherwise the minimum distance guarantee will be violated. This is illustrated in Fig. \[violation\_delta\]. It is clear from the figure, to avoid the overlap it is necessary that the distance between the centre of each of the effective shifted circle $ (\gamma_i,\theta_{i,opt}) $, $ 1 \leq i \leq N_W $ and the singular fade states $ (\gamma_j, \theta_j) $, $ 1\leq j \leq N_W $ should be at least equal to sum of the radius of the shifted circle and the radius of the violation circles centred at $ (\gamma_j,\theta_j) $. It is required, for each $i$, $1\leq i\leq N_W $ , $$\begin{aligned}
\label{eqn:upperbound}
\rho(\gamma_i,\theta_i)+\rho(\gamma_j,\theta_j) &\leq d_{(\gamma _i, \theta _{i,opt}) \leftrightarrow (\gamma_j, \theta_j)} \\
\nonumber
&\text{for all }j\,,1\leq j\leq N_W, \end{aligned}$$ where $d_{(\gamma _i, \theta _{i,opt}) \leftrightarrow (\gamma_j, \theta_j)}$ is the Euclidean distance between the points $(\gamma_j, \theta_j)$ and $(\gamma_i, \theta_{i,opt})$ in the $(\Gamma, \Theta)$ plane. Since $ \rho (\gamma_k, \theta_k) $, $ 1\leq k \leq N_W $, is a function of $ \delta $, provides an upper bound on $ \delta $.
When both users use QPSK constellations at the input, both the violation circles centred at $ (1,0^\circ) $ and $ (\sqrt{2},45^\circ) $ has radius $ \delta /\sqrt{2} $ *i.e.* $\rho(1,0^\circ) = \rho (\sqrt{2},45^\circ) = \frac{\delta}{\sqrt{2}}$. Now, $$\begin{aligned}
\nonumber
&d_{(\sqrt{2},45^\circ)\leftrightarrow (\sqrt{2},20.7^\circ) }=d_{(1,0^\circ)\leftrightarrow (\sqrt{2},20.7^\circ) } \approx 0.5936\\
\nonumber
&d_{(\sqrt{2},45^\circ) \leftrightarrow (1,30^\circ) }= d_{(1,0^\circ) \leftrightarrow (1,30^\circ) } \approx 0.5176.\end{aligned}$$ To avoid overlap, from , we have $$\begin{aligned}
\nonumber
&2(\frac{\delta}{\sqrt{2}}) \leq \min \lbrace 0.5936, 0.5176 \rbrace \\
\nonumber
& \therefore \delta \leq 0.365 \approx \delta_{max}.\end{aligned}$$
For $ \delta > \delta _{max} $ there always exist some fade states for which the minimum distance in $ \mathcal{S}_{\text{eff}} $ cannot be increased beyond $ \delta $ using the proposed scheme. For example, in Fig. \[violation\_delta\], the fade state corresponding to the point $ P $ is transferred to $ P' $ after rotation. But $ P' $ still lies within the violation circle corresponding to singular fade state $ (\sqrt{2},\frac{\pi}{4}) $, thus minimum distance guarantee is violated.
![Diagram illustrating the necessity for an upper bound on $\delta$ []{data-label="violation_delta"}](violation_delta.eps){width="3in"}
SIMULATION RESULTS AND DISCUSSIONS {#sec:results}
==================================
All through the previous section we have assumed that the ratio $\frac{h_2}{h_1}$ is calculated at the destination. But even if is not so, *i.e.* actually the ratio $\frac{h_1}{h_2}$ is used for channel quantization at the destination, then exactly the same scheme would work except that instead of rotating the constellation of User-2 we have to rotate the signal set of User-1. The optimal angles of rotation will still be same as calculated before. The feedback that the destination sends back to the users indicates if the fade state $ (\gamma,\theta) $ lies in any of the violation circles or not, and if it does, then identifies in which one of the $ N_W $ violation circles does it lie in. It also needs to indicate which one among the two ratios is calculated for fade state quantization at the destination. Thus the total feedback overhead is $\lceil \log_2 (N_W +1)\rceil +1$ bits. For example, the feedback overhead for the QPSK and 8-PSK case are 3 and 4 bits respectively. This feedback overhead is very nominal.
![Probability of error vs SNR plot, when both users use QPSK signal sets[]{data-label="PeSNR"}](newPesnr.eps){width="3.1in"}
The system is simulated for the case when both users use QPSK signal sets at the input. The probability of error is plotted against SNR in Fig. \[PeSNR\] both without and with adaptive modulation for different values of $ \delta $. The gains obtained increases on increasing $ \delta $ as expected. For a $ P_e= 10^{-3} $ there is a 4dB gain obtained with adaptive modulation with $ \delta=0.35 $, as shown in the figure.
CONCLUSIONS
===========
In this paper, we have proposed a modulation scheme for the two-user MAC with fading which adapts according to the fade states. For this purpose we have obtained a quantization of all possible fade states based on a minimum distance criteria when both users use $M$-PSK constellations at input. We have identified the regions, called violation circles, such that when the fade state lies in them the minimum distance requirement in $\mathcal{S}_{\text{eff}}$ is violated. The quantized fade state knowledge is fed back to the users using just $\lceil \log_2 (N_W +1)\rceil +1 $ bits. Based on this quantized feedback, one of the users rotates it’s constellation to effectively overcome the ‘bad channel conditions’. We have shown the extent to which the error performance of this proposed scheme is better than the conventional scheme without adaptation. The case when both users use QAM constellations at input has not been considered in this paper, is a natural topic for future work. Other cases with more than two users communicating with a single destination is also an interesting direction for future work.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work was supported partly by the DRDO-IISc program on Advanced Research in Mathematical Engineering through a research grant as well as the INAE Chair Professorship grant to B. S. Rajan.
Thomas M. Cover and Joy A. Thomas *Elements of Information Theory*,Second Edition, New York: Wiley 1999
R. Gallager, “A perspective on multiaccess channels”, *IEEE Trans. on Inf. Theory*, Vol. 31, No. 2, pp. 124-142, Mar. 1985
J. Harshan and B. Sundar Rajan, “On Two-User Gaussian Multiple Access Channels with Finite Input Constellations ”, *IEEE Trans.on Inf. Theory*, Vol.57, No. 3,pp. 1299-1327,Mar. 2011
J. Harshan and B. Sundar Rajan, “A Constellation Power Allocation Scheme for Two-User Gaussian MAC ”, *Proc of IEEE PIMRC’11*
T. Koike-Akino, P. Popovski and V. Tarokh, “Optimized constellation for two-way wireless relaying with physical network coding”, *IEEE Journal on selected Areas in Comm.*, Vol.27, pp. 773- - 787, June 2009.
Vijayvaradharaj T Muralidharan, Vishnu Namboodiri and B. Sundar Rajan, “Channel Quantization for Physical-Layer Network Coded Two-Way Relaying”, *available online at arXiv:1109.6101v1* \[cs.IT\], 28 Sept.2011.
Haghi, A. Khosravi-Farsani, R. Aref, M. R. Marvasti, “The capacity region of fading Multiple Access Channels with cooperative encoders and partial CSIT”,*Proc. of IEEE ISIT,2010*, pp.485-489, June 2010
Y Cemal, Y. Steinberg “The Multiple Access Channel with Partial State Information at the Encoders” ,*IEEE Trans. on Inf Theory* Vol 51, No.11, pp. 3392-4003,2005
Moritz Wiese, Holger Boche and Igor Bjelakovic “The compound MAC with common message and partial channel state information”,*IEEE ISITA 2010*, Vol 1, pp. 416-421
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Working in the context of symmetric spectra, we describe and study a homotopy completion tower for algebras and left modules over operads in the category of modules over a commutative ring spectrum (e.g., structured ring spectra). We prove a strong convergence theorem that for $0$-connected algebras and modules over a $(-1)$-connected operad, the homotopy completion tower interpolates (in a strong sense) between topological Quillen homology and the identity functor.
By systematically exploiting strong convergence, we prove several theorems concerning the topological Quillen homology of algebras and modules over operads. These include a theorem relating finiteness properties of topological Quillen homology groups and homotopy groups that can be thought of as a spectral algebra analog of Serre’s finiteness theorem for spaces and H.R. Miller’s boundedness result for simplicial commutative rings (but in reverse form). We also prove absolute and relative Hurewicz theorems and a corresponding Whitehead theorem for topological Quillen homology. Furthermore, we prove a rigidification theorem, which we use to describe completion with respect to topological Quillen homology (or ${{ \mathsf{TQ} }}$-completion). The ${{ \mathsf{TQ} }}$-completion construction can be thought of as a spectral algebra analog of Sullivan’s localization and completion of spaces, Bousfield-Kan’s completion of spaces with respect to homology, and Carlsson’s and Arone-Kankaanrinta’s completion and localization of spaces with respect to stable homotopy. We prove analogous results for algebras and left modules over operads in unbounded chain complexes.
address:
- 'Department of Mathematics, University of Western Ontario, London, Ontario, N6A 5B7, Canada'
- 'MATHGEOM, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland'
- 'MATHGEOM, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland'
author:
- 'John E. Harper'
- Kathryn Hess
bibliography:
- 'HomotopyCompletion.bib'
title: 'Homotopy completion and topological [Q]{}uillen homology of structured ring spectra'
---
Introduction
============
Associated to each non-unital commutative ring $X$ is the completion tower arising in commutative ring theory $$\begin{aligned}
\label{eq:NUCA_completion_tower}
X/X^2 \leftarrow X/X^3 \leftarrow \cdots \leftarrow
X/X^{n} \leftarrow X/X^{n+1} \leftarrow \cdots\end{aligned}$$ of non-unital commutative rings. The limit of the tower is the completion $X^\wedge$ of $X$, which is sometimes also called the $X$-adic completion of $X$. Here, $X/X^n$ denotes the quotient of $X$ in the underlying category by the image of the multiplication map $X^{\otimes n}{{ \longrightarrow }}X$. In algebraic topology, algebraic $K$-theory, and derived algebraic geometry, it is common to encounter objects that are naturally equipped with algebraic structures more general than, for example, commutative rings, but that share certain formal similarities with these classical algebraic structures. A particularly useful and interesting class of such generalized algebraic structures are those that can be described as algebras and modules over operads; see Fresse [@Fresse_lie_theory], Goerss-Hopkins [@Goerss_Hopkins_moduli_spaces], Kriz-May [@Kriz_May], Mandell [@Mandell], and McClure-Smith [@McClure_Smith_conjecture].
These categories of (generalized) algebraic structures can often be equipped with an associated homotopy theory, or Quillen model category structure, which allows one to construct and calculate derived functors on the associated homotopy category. In [@Quillen II.5], Quillen defines “homology” in the general context of a model category—now called Quillen homology—to be the left derived functor of abelianization, if it exists. Quillen homology often behaves very much like the ordinary homology of topological spaces, which it recovers as a special case. Quillen [@Quillen_rings] and André [@Andre] originally developed and studied a particular case of Quillen’s notion of homology for the special context of commutative rings, now called André-Quillen homology. A useful introduction to Quillen homology is given in Goerss-Schemmerhorn [@Goerss_Schemmerhorn]; see also Goerss [@Goerss_f2_algebras] and H.R. Miller [@Miller] for a useful development (from a homotopy viewpoint) in the case of augmented commutative algebras.
In this paper we are primarily interested in the topological analog of Quillen homology, called topological Quillen homology, for (generalized) algebraic structures on spectra. The topological analog for commutative ring spectra, called topological André-Quillen homology, was originally studied by Basterra [@Basterra]; see also Baker-Gilmour-Reinhard [@Baker_Gilmour_Reinhard], Baker-Richter [@Baker_Richter], Basterra-Mandell [@Basterra_Mandell; @Basterra_Mandell_thh], Goerss-Hopkins [@Goerss_Hopkins], Lazarev [@Lazarev], Mandell [@Mandell_TAQ], Richter [@Richter], Rognes [@Rognes_topological_Galois; @Rognes_logarithmic] and Schwede [@Schwede_cotangent; @Schwede_algebraic].
\[assumption:commutative\_ring\_spectrum\] From now on in this paper, we assume that ${{ \mathcal{R} }}$ is any commutative ring spectrum; i.e., we assume that ${{ \mathcal{R} }}$ is any commutative monoid object in the category $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ of symmetric spectra [@Hovey_Shipley_Smith; @Schwede_book_project]. Here, the tensor product ${{ \otimes }}_S$ denotes the usual smash product [@Hovey_Shipley_Smith 2.2.3] of symmetric spectra (Remark \[rem:smash\_product\_and\_tensor\_product\]).
Among *structured ring spectra* we include many different types of algebraic structures on spectra (resp. ${{ \mathcal{R} }}$-modules) including (i) associative ring spectra, which we simply call ring spectra, (ii) commutative ring spectra, (iii) all of the $E_n$ ring spectra for $1\leq n\leq \infty$ that interpolate between these two extremes of non-commutativity and commutativity, together with (iv) any generalized algebra spectra (resp. generalized ${{ \mathcal{R} }}$-algebras) that can be described as algebras over operads in spectra (resp. ${{ \mathcal{R} }}$-modules). It is important to note that the generalized class of algebraic structures in (iv) includes as special cases all of the others (i)–(iii). The area of stable homotopy theory that focuses on problems arising from constructions involving different types of structured ring spectra, their modules, and their homotopy invariants, is sometimes called *brave new algebra* or *spectral algebra*.
In this paper we describe and study a (homotopy invariant) spectral algebra analog of the completion tower arising in commutative ring theory. The tower construction is conceptual and provides a sequence of refinements of the Hurewicz map for topological Quillen homology. More precisely, if ${{ \mathcal{O} }}$ is an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial (i.e., ${{ \mathcal{O} }}$-algebras are *non-unital*), we associate to ${{ \mathcal{O} }}$ itself a tower $$\begin{aligned}
\tau_1{{ \mathcal{O} }}\leftarrow
\tau_2{{ \mathcal{O} }}\leftarrow \cdots \leftarrow
\tau_{k-1}{{ \mathcal{O} }}\leftarrow
\tau_{k}{{ \mathcal{O} }}\leftarrow \cdots\end{aligned}$$ of $({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodules, which for any ${{ \mathcal{O} }}$-algebra $X$ induces the *completion* tower $$\begin{aligned}
\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X) \leftarrow
\tau_2{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X) \leftarrow \cdots \leftarrow
\tau_{k-1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X) \leftarrow
\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X) \leftarrow \cdots\end{aligned}$$ of ${{ \mathcal{O} }}$-algebras whose limit is the *completion* $X^\wedge$ of $X$. There is a homotopy theory of algebras over operads (Theorem \[thm:positive\_flat\_stable\_AlgO\]) and this construction is homotopy invariant if applied to cofibrant ${{ \mathcal{O} }}$-algebras. We sometimes refer to the completion tower of a cofibrant replacement $X^c$ of $X$ as the *homotopy completion tower* of $X$ whose homotopy limit is denoted $X^{{ {{ \mathsf{h} }}\wedge }}$. By construction, $\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X^c)$ is the topological Quillen homology ${{ \mathsf{TQ} }}(X)$ of $X$. Hence the homotopy completion tower of $X$ interpolates between ${{ \mathsf{TQ} }}(X)$, which is the bottom term of the tower, and the homotopy completion $X^{{ {{ \mathsf{h} }}\wedge }}$ of $X$.
By systematically exploiting the strong convergence properties of this tower (Theorem \[MainTheorem\] and its proof), we prove a selection of theorems concerning the topological Quillen homology of structured ring spectra. We also prove analogous results for left modules over operads (Definition \[defn:algebras\_and\_modules\]). The first main theorem in this paper is the following finiteness theorem for topological Quillen homology. It can be thought of as a structured ring spectra analog of Serre’s finiteness theorem for spaces (e.g., for the homotopy groups of spheres) and H.R. Miller’s [@Miller 4.2] boundedness result for simplicial commutative rings (but in reverse form); for a related but different type of finiteness result in the algebraic context of augmented commutative algebras over a field of non-zero characteristic, see Turner [@Turner]. The ${{ \mathsf{TQ} }}$ finiteness theorem provides conditions under which topological Quillen homology detects certain finiteness properties.
In this paper, we say that a symmetric sequence $X$ of symmetric spectra is $n$-connected if each symmetric spectrum $X[\mathbf{t}]$ is $n$-connected. We say that an algebra (resp. left module) over an operad is $n$-connected if the underlying symmetric spectrum (resp. symmetric sequence of symmetric spectra) is $n$-connected, and similarly for operads.
\[thm:finiteness\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Let $X$ be a $0$-connected ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and assume that ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected and $\pi_k{{ \mathcal{O} }}[\mathbf{r}],\pi_k{{ \mathcal{R} }}$ are finitely generated abelian groups for every $k,r$.
- If the topological Quillen homology groups $\pi_k{{ \mathsf{TQ} }}(X)$ (resp. $\pi_k{{ \mathsf{TQ} }}(X)[\mathbf{r}]$) are finite for every $k,r$, then the homotopy groups $\pi_k X$ (resp. $\pi_k X[\mathbf{r}]$) are finite for every $k,r$.
- If the topological Quillen homology groups $\pi_k{{ \mathsf{TQ} }}(X)$ (resp. $\pi_k{{ \mathsf{TQ} }}(X)[\mathbf{r}]$) are finitely generated abelian groups for every $k,r$, then the homotopy groups $\pi_k X$ (resp. $\pi_k X[\mathbf{r}]$) are finitely generated abelian groups for every $k,r$.
Since the sphere spectrum $S$ is $(-1)$-connected and $\pi_kS$ is a finitely generated abelian group for every $k$, we obtain the following immediate corollary.
\[cor:finiteness\_commutative\_ring\_spectra\] Let $X$ be a $0$-connected non-unital commutative ring spectrum. If the topological Quillen homology groups $\pi_k{{ \mathsf{TQ} }}(X)$ are finite (resp. finitely generated abelian groups) for every $k$, then the homotopy groups $\pi_k X$ are finite (resp. finitely generated abelian groups) for every $k$.
Since all of the theorems in this section apply to the special case of non-unital commutative ring spectra, it follows that each theorem below specializes to a corollary about non-unital commutative ring spectra, similar to the corollary above. To avoid repetition, we usually leave the formulation to the reader.
We also prove the following Hurewicz theorem for topological Quillen homology. It can be thought of as a structured ring spectra analog of Schwede’s [@Schwede_algebraic 5.3] simplicial algebraic theories result, Goerss’ [@Goerss_f2_algebras 8.3] algebraic result for augmented commutative $\mathbb{F}_2$-algebras, Livernet’s [@Livernet 2.13] rational algebraic result for algebras over operads in non-negative chain complexes over a field of characteristic zero, and Chataur-Rodriguez-Scherer’s [@Chataur_Rodriguez_Scherer 2.1] algebraic result for algebras over cofibrant operads in non-negative chain complexes over a commutative ring. The ${{ \mathsf{TQ} }}$ Hurewicz theorem provides conditions under which topological Quillen homology detects $n$-connected structured ring spectra. It also provides conditions under which the first non-trivial homotopy group agrees via the Hurewicz map with the first non-trivial topological Quillen homology group.
\[thm:hurewicz\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Let $X$ be a $0$-connected ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), $n\geq 0$, and assume that ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected.
- Topological Quillen homology ${{ \mathsf{TQ} }}(X)$ is $n$-connected if and only if $X$ is $n$-connected.
- If topological Quillen homology ${{ \mathsf{TQ} }}(X)$ is $n$-connected, then the natural Hurewicz map $
\pi_k X{{ \longrightarrow }}\pi_k{{ \mathsf{TQ} }}(X)
$ is an isomorphism for $k\leq 2n+1$ and a surjection for $k=2n+2$.
Note that one implication of Theorem \[thm:hurewicz\](a) follows from Theorem \[thm:hurewicz\](b). We also prove the following relative Hurewicz theorem for topological Quillen homology, which we regard as the second main theorem in this paper. It can be thought of as a structured ring spectra analog of the relative Hurewicz theorem for spaces. It provides conditions under which topological Quillen homology detects $n$-connected maps.
\[thm:relative\_hurewicz\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) and $n\geq 0$. Assume that ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected.
- If $X,Y$ are $0$-connected, then $f$ is $n$-connected if and only if $f$ induces an $n$-connected map ${{ \mathsf{TQ} }}(X){{ \longrightarrow }}{{ \mathsf{TQ} }}(Y)$ on topological Quillen homology.
- If $X,Y$ are $(-1)$-connected and $f$ is $(n-1)$-connected, then $f$ induces an $(n-1)$-connected map ${{ \mathsf{TQ} }}(X){{ \longrightarrow }}{{ \mathsf{TQ} }}(Y)$ on topological Quillen homology.
- If $f$ induces an $n$-connected map ${{ \mathsf{TQ} }}(X){{ \longrightarrow }}{{ \mathsf{TQ} }}(Y)$ on topological Quillen homology between $(-1)$-connected objects, then $f$ induces an $(n-1)$-connected map $X^{{ {{ \mathsf{h} }}\wedge }}{{ \longrightarrow }}Y^{{ {{ \mathsf{h} }}\wedge }}$ on homotopy completion.
- If topological Quillen homology ${{ \mathsf{TQ} }}(X)$ is $(n-1)$-connected, then homotopy completion $X^{{ {{ \mathsf{h} }}\wedge }}$ is $(n-1)$-connected.
Here, ${{ \mathsf{TQ} }}(X){{ \longrightarrow }}{{ \mathsf{TQ} }}(Y)$, $X^{{ {{ \mathsf{h} }}\wedge }}{{ \longrightarrow }}Y^{{ {{ \mathsf{h} }}\wedge }}$ denote the natural induced zigzags in the category of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) with all backward facing maps weak equivalences.
\[rem:preservation\_of\_connectivity\] It is important to note Theorem \[thm:relative\_hurewicz\](b) implies that the conditions in Theorem \[thm:relative\_hurewicz\](c) are satisfied if $X,Y$ are $(-1)$-connected and $f$ is $n$-connected.
As a corollary we obtain the following Whitehead theorem for topological Quillen homology. It can be thought of as a structured ring spectra analog of Schwede’s [@Schwede_algebraic 5.4] simplicial algebraic theories result, Goerss’ [@Goerss_f2_algebras 8.1] algebraic result for augmented commutative $\mathbb{F}_2$-algebras, and Livernet’s [@Livernet_thesis] rational algebraic result for algebras over Koszul operads in non-negative chain complexes over a field of characteristic zero. As a special case, it recovers Kuhn’s [@Kuhn] result for non-unital commutative ring spectra, and more generally, Lawson’s [@Lawson] original structured ring spectra result (which is based on [@Harper_Bar]). The ${{ \mathsf{TQ} }}$ Whitehead theorem provides conditions under which topological Quillen homology detects weak equivalences.
\[cor:whitehead\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules). Assume that ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected. If $X,Y$ are $0$-connected, then $f$ is a weak equivalence if and only if $f$ induces a weak equivalence $
{{ \mathsf{TQ} }}(X){{ \ \simeq \ }}{{ \mathsf{TQ} }}(Y)
$ on topological Quillen homology.
Associated to the homotopy completion tower is the *homotopy completion spectral sequence*, which goes from topological Quillen homology to homotopy completion (Theorem \[MainTheorem\]). It can be thought of as a structured ring spectra analog of Quillen’s fundamental spectral sequence [@Quillen_rings 6.9] for commutative rings and the corresponding spectral sequence studied by Goerss [@Goerss_f2_algebras 6.2] for augmented commutative $\mathbb{F}_2$-algebras. As a special case, it recovers the spectral sequence in Minasian [@Minasian] for non-unital commutative ring spectra. Under the conditions of Theorem \[MainTheorem\](b), the homotopy completion spectral sequence is a second quadrant homologically graded spectral sequence and arises from the exact couple of long exact sequences associated to the homotopy completion tower and its homotopy fibers; this is the homotopy spectral sequence of a tower of fibrations [@Bousfield_Kan], reindexed as a homologically graded spectral sequence. For ease of notational purposes, in Theorem \[MainTheorem\] and Remark \[rem:strong\_convergence\], we regard such towers $\{A_s\}$ of fibrations as indexed by the integers such that $A_s=*$ for each $s<0$.
The third main theorem in this paper is the following strong convergence theorem for homotopy completion of structured ring spectra. It can be thought of as a structured ring spectra analog of Johnson-McCarthy’s [@Johnson_McCarthy] rational algebraic tower results for non-unital commutative differential graded algebras over a field of characteristic zero. As a special case, it recovers Kuhn’s [@Kuhn] and Minasian’s [@Minasian] tower results for non-unital commutative ring spectra. For a very restricted class of cofibrant operads in simplicial sets, which they call primitive operads, McCarthy-Minasian [@McCarthy_Minasian_preprint] describe a tower that agrees with the completion tower in the special case of non-unital commutative ring spectra, but that is different for most operads.
\[MainTheorem\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules).
- If $X$ is $0$-connected and ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected, then the natural coaugmentation $X{{ \ \simeq \ }}X^{{ {{ \mathsf{h} }}\wedge }}$ is a weak equivalence.
- If topological Quillen homology ${{ \mathsf{TQ} }}(X)$ is $0$-connected and ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected, then the homotopy completion spectral sequence $$\begin{aligned}
E^1_{-s,t} &= \pi_{t-s}\Bigl(i_{s+1}{{ \mathcal{O} }}\circ^{{{ \mathsf{h} }}}_{\tau_1{{ \mathcal{O} }}}\bigl({{ \mathsf{TQ} }}(X)\bigr)\Bigr)
\Longrightarrow
\pi_{t-s}\bigl(X^{{{ {{ \mathsf{h} }}\wedge }}}\bigr)\\
\text{resp.}\quad
E^1_{-s,t}[\mathbf{r}] &= \pi_{t-s}\Bigl(\bigl(i_{s+1}{{ \mathcal{O} }}\circ^{{{ \mathsf{h} }}}_{\tau_1{{ \mathcal{O} }}}{{ \mathsf{TQ} }}(X)\bigr)[\mathbf{r}]\Bigr)
\Longrightarrow
\pi_{t-s}\bigl(X^{{{ {{ \mathsf{h} }}\wedge }}}[\mathbf{r}]\bigr),\quad\text{$r\geq 0$},\end{aligned}$$ converges strongly (Remark \[rem:strong\_convergence\]).
- If $f$ induces a weak equivalence ${{ \mathsf{TQ} }}(X){{ \ \simeq \ }}{{ \mathsf{TQ} }}(Y)$ on topological Quillen homology, then $f$ induces a weak equivalence $X^{{ {{ \mathsf{h} }}\wedge }}{{ \ \simeq \ }}Y^{{ {{ \mathsf{h} }}\wedge }}$ on homotopy completion.
\[rem:strong\_convergence\] By *strong convergence* of $\{E^r\}$ to $\pi_*(X^{{ {{ \mathsf{h} }}\wedge }})$ we mean that (i) for each $(-s,t)$, there exists an $r$ such that $E^r_{-s,t}=E^\infty_{-s,t}$ and (ii) for each $i$, $E^\infty_{-s,s+i}=0$ except for finitely many $s$. Strong convergence implies that for each $i$, $\{E^\infty_{-s,s+i}\}$ is the set of filtration quotients from a finite filtration of $\pi_i(X^{{ {{ \mathsf{h} }}\wedge }})$; see, for instance, Bousfield-Kan [@Bousfield_Kan IV.5.6, IX.5.3, IX.5.4] and Dwyer [@Dwyer_strong_convergence].
Regard the homotopy completion tower as a tower of functors on the category of ${{ \mathcal{O} }}$-algebras, and consider the case when ${{ \mathcal{O} }}[\mathbf{1}]=I[\mathbf{1}]$ (Definition \[defn:operad\]). Then it follows easily that (i) the bottom term (or first stage) ${{ \mathsf{TQ} }}$ of the tower is $1$-excisive in the sense of [@Goodwillie_calc3; @Kuhn_survey], (ii) by Theorem \[thm:calculating\_fiber\_of\_induced\_map\](c), the $n$-th layer of the tower has the form ${{ \mathcal{O} }}[\mathbf{n}]{{ \,\wedge\, }}^{{ \mathsf{L} }}_{\Sigma_n}{{ \mathsf{TQ} }}^{\wedge^{{ \mathsf{L} }}n}$, and (iii) by the connectivity estimates in the proof of Theorem \[thm:hurewicz\], the identity functor and the $n$-th stage of the tower *agree to order $n$* in the sense of [@Goodwillie_calc3 1.2]; more precisely, they satisfy $O_n(0,1)$ as defined in [@Goodwillie_calc3 1.2]. Here, $\wedge^{{ \mathsf{L} }}_{\Sigma_n}$, $\wedge^{{ \mathsf{L} }}$ are the total left derived functors of $\wedge_{\Sigma_n}$, $\wedge$, respectively. Properties (i)–(iii) illustrate that the homotopy completion tower is the analog, in the context of ${{ \mathcal{O} }}$-algebras, of Goodwillie’s Taylor tower of the identity functor. More precisely, according to [@Goodwillie_calc3 1.6, proof of 1.8] and the results in [@Goodwillie_calc2] on cubical diagrams, it follows immediately from (i)–(iii) that there are maps of towers (under the constant tower $\{{{ \mathrm{id} }}(-)^c\}$) of levelwise weak equivalences of the form $\{P_n{{ \mathrm{id} }}(-)^c\}\rightarrow\{P_n\tau_n{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(-)^c\}\leftarrow\{\tau_n{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(-)^c\}$ where $(-)^c$ denotes functorial cofibrant replacement (see Definition \[defn:homotopy\_completion\]), and hence the homotopy completion tower is weakly equivalent to the Taylor tower of the identity functor on ${{ \mathcal{O} }}$-algebras, provided that the analogs of the appropriate constructions and results in [@Goodwillie_calc2; @Goodwillie_calc3] remain true in the category of ${{ \mathcal{O} }}$-algebras; this is the subject of current work, and will not be further elaborated here (but see [@Kuhn_survey]).
Since in the calculation of the layers in (ii) the operad ${{ \mathcal{O} }}$ plays a role analogous to that of the Goodwillie derivatives of the identity functor (see [@Goodwillie_calc3; @Kuhn_survey]), this sheds some positive light on a conjecture of Arone-Ching [@Arone_Ching] that an appropriate model of the Goodwillie derivatives of the identity functor on ${{ \mathcal{O} }}$-algebras is weakly equivalent as an operad to ${{ \mathcal{O} }}$ itself.
The following relatively weak cofibrancy condition is exploited in the proofs of the main theorems above. The statements of these theorems do not require this cofibrancy condition since a comparison theorem (Theorem \[thm:comparing\_homotopy\_completion\_towers\], Proposition \[prop:replacement\_of\_operads\]) shows that the operad ${{ \mathcal{O} }}$ can always be replaced by a weakly equivalent operad ${{ \mathcal{O} }}'$ that satisfies this cofibrancy condition and such that the corresponding homotopy completion towers are naturally weakly equivalent.
\[CofibrancyCondition\] If ${{ \mathcal{O} }}$ is an operad in ${{ \mathcal{R} }}$-modules, consider the unit map ${{ {\eta}\colon\thinspace{I}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ of the operad ${{ \mathcal{O} }}$ (Definition \[defn:operad\]) and assume that $I[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}[\mathbf{r}]$ is a flat stable cofibration (Subsection \[sec:model\_structures\_on\_capR\_modules\]) between flat stable cofibrant objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$.
This is the same as assuming that $I[\mathbf{1}]{{ \longrightarrow }}{{ \mathcal{O} }}[\mathbf{1}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$. It can be thought of as the structured ring spectra analog of the following cofibrancy condition: if $X$ is a pointed space, assume that $X$ is well-pointed; i.e., assume that the unique map $*\rightarrow X$ of pointed spaces is a cofibration.
Most operads appearing in homotopy theoretic settings in mathematics already satisfy Cofibrancy Condition \[CofibrancyCondition\] and therefore require no replacement in the proofs of the theorems. For instance, Cofibrancy Condition \[CofibrancyCondition\] is satisfied by every operad in simplicial sets that is regarded as an operad in ${{ \mathcal{R} }}$-modules via adding a disjoint basepoint and tensoring with ${{ \mathcal{R} }}$ (Subsection \[sec:simplicial\_bar\_and\_homotopy\_completion\]).
In this paper, the homotopy groups $\pi_*Y$ of a symmetric spectrum $Y$ denote the *derived* homotopy groups (or true homotopy groups) [@Schwede_book_project; @Schwede_homotopy_groups]; i.e., $\pi_*Y$ always denotes the homotopy groups of a stable fibrant replacement of $Y$, and hence of a flat stable fibrant replacement of $Y$. See Schwede [@Schwede_homotopy_groups] for several useful properties enjoyed by the true homotopy groups of a symmetric spectrum.
Organization of the paper
-------------------------
In Section \[sec:preliminaries\] we recall some preliminaries on algebras and modules over operads. The purpose of Section \[sec:homotopy\_completion\] is to describe homotopy completion (Definition \[defn:homotopy\_completion\]) and ${{ \mathsf{TQ} }}$-completion, or less concisely, completion with respect to topological Quillen homology (Definition \[defn:quillen\_homology\_completion\]) and to establish a comparison theorem for homotopy completion towers (Theorem \[thm:comparing\_homotopy\_completion\_towers\]). In Section \[sec:homotopical\_analysis\_of\_the\_completion\_tower\] we prove our main theorems, which involves a homotopical analysis of the completion tower. We establish several necessary technical results on the homotopical properties of the forgetful functors in Section \[sec:homotopical\_analysis\_forgetful\_functors\], and on simplicial structures and the homotopical properties of the simplicial bar constructions in Section \[sec:homotopical\_analysis\_bar\_constructions\]. The results in these two sections lie at the heart of the proofs of the main theorems. The purpose of Section \[sec:model\_structures\] is to improve the main results in [@Harper_Spectra; @Harper_Bar] on model structures, homotopy colimits and simplicial bar constructions from the context of operads in symmetric spectra to the more general context of operads in ${{ \mathcal{R} }}$-modules. This amounts to establishing certain technical propositions for ${{ \mathcal{R} }}$-modules sufficient for the proofs of the main results in [@Harper_Spectra; @Harper_Bar] to remain valid in the more general context of ${{ \mathcal{R} }}$-modules; these results play a key role in this paper. In Section \[sec:chain\_complexes\_over\_a\_commutative\_ring\] we observe that the analogs of the main theorems stated above remain true in the context of unbounded chain complexes over a commutative ring.
Acknowledgments {#acknowledgments .unnumbered}
---------------
The authors would like to thank Greg Arone, Michael Ching, Bill Dwyer, Emmanuel Farjoun, Rick Jardine, Nick Kuhn, Haynes Miller, and Stefan Schwede for useful suggestions and remarks and Kristine Bauer, Mark Behrens, Bjorn Dundas, Benoit Fresse, Paul Goerss, Tom Goodwillie, Jens Hornbostel, Brenda Johnson, Tyler Lawson, Muriel Livernet, Ib Madsen, Mike Mandell, Randy McCarthy, Jack Morava, and Charles Rezk for helpful comments. The first author is grateful to Jens Hornbostel and Stefan Schwede for a stimulating and enjoyable visit to the Mathematisches Institut der Universität Bonn in summer 2010, and to Mark Behrens and Haynes Miller for a stimulating and enjoyable visit to the Department of Mathematics at the Massachusetts Institute of Technology in summer 2011, and for their invitations which made this possible. The authors would like to thank the anonymous referee for his or her detailed suggestions and comments, which have resulted in a significant improvement.
Preliminaries {#sec:preliminaries}
=============
The purpose of this section is to recall various preliminaries on algebras and modules over operads. In this paper the following two contexts will be of primary interest. Denote by $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ the closed symmetric monoidal category of ${{ \mathcal{R} }}$-modules (Basic Assumption \[assumption:commutative\_ring\_spectrum\], Remark \[rem:dropping\_the\_adjective\_left\]), and by $({{ \mathsf{Ch} }}_{{ {\mathcal{K}} }},{{ \otimes }},{{ {\mathcal{K}} }})$ the closed symmetric monoidal category of unbounded chain complexes over ${{ {\mathcal{K}} }}$ [@Hovey; @MacLane_homology]; here, ${{ {\mathcal{K}} }}$ is any commutative ring. Both categories have all small limits and colimits, and the null object is denoted by $*$. It will be useful in this paper, both for establishing certain results and for ease of notational purposes, to sometimes work in the following more general context; see [@MacLane_categories VII] followed by [@MacLane_categories VII.7].
From now on in this section we assume that $({{ \mathsf{C} }},{{ \,\wedge\, }},S)$ is a closed symmetric monoidal category with all small limits and colimits. In particular, ${{ \mathsf{C} }}$ has an initial object $\emptyset$ and a terminal object $*$.
By *closed* we mean there exists a functor ${{ \mathsf{C} }}^{{ \mathrm{op} }}\times{{ \mathsf{C} }}{{ \longrightarrow }}{{ \mathsf{C} }}: (Y,Z)\longmapsto \operatorname{Map}(Y,Z)$, which we call the *mapping object*, which fits into $
\hom(X{{ \,\wedge\, }}Y,Z){{ \ \cong \ }}\hom(X,\operatorname{Map}(Y,Z))
$ isomorphisms natural in $X,Y,Z$, where $\hom$ denotes the set of morphisms in ${{ \mathsf{C} }}$. Define the sets $\mathbf{n}:=\{1,\dots,n\}$ for each $n\geq 0$, where $\mathbf{0}:=\emptyset$ denotes the empty set. If $T$ is a finite set, we denote by $|T|$ the number of elements in $T$.
\[defn:symmetric\_sequences\] Let $n\geq 0$.
- $\Sigma$ is the category of finite sets and their bijections.
- A *symmetric sequence* in ${{ \mathsf{C} }}$ is a functor ${{ {A}\colon\thinspace{\Sigma^{{{ \mathrm{op} }}}}{{ \longrightarrow }}{{{ \mathsf{C} }}} }}$. Denote by ${{ \mathsf{SymSeq} }}$ the category of symmetric sequences in ${{ \mathsf{C} }}$ and their natural transformations.
- A symmetric sequence $A$ is *concentrated at $n$* if $A[\mathbf{r}]=\emptyset$ for all $r\neq n$.
For a more detailed development of the material that follows, see [@Harper_Spectra; @Harper_Modules].
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $A_1,\dotsc,A_t\in{{ \mathsf{SymSeq} }}$. Their *tensor product* $A_1{{ \check{{{ \otimes }}} }}\dotsb{{ \check{{{ \otimes }}} }}A_t\in{{ \mathsf{SymSeq} }}$ is the left Kan extension of objectwise smash along coproduct of sets $$\begin{aligned}
\xymatrix{
(\Sigma^{{{ \mathrm{op} }}})^{\times t}
\ar[rr]^-{A_1\times\dotsb\times A_t}\ar[d]^{\coprod} & &
{{ \mathsf{C} }}^{\times t}\ar[r]^-{{{ \,\wedge\, }}} & {{ \mathsf{C} }}\\
\Sigma^{{{ \mathrm{op} }}}\ar[rrr]^{A_1{{ \check{{{ \otimes }}} }}\dotsb{{ \check{{{ \otimes }}} }}A_t}_{\text{left Kan extension}} & & & {{ \mathsf{C} }}}\end{aligned}$$
If $X$ is a finite set and $A$ is an object in ${{ \mathsf{C} }}$, we use the usual dot notation $A\cdot X$ ([@MacLane_categories], [@Harper_Modules 2.3]) to denote the copower $A\cdot X$ defined by $
A\cdot X := \coprod_X A
$, the coproduct in ${{ \mathsf{C} }}$ of $|X|$ copies of $A$. Recall the following useful calculations for tensor products.
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $A_1,\dotsc,A_t\in{{ \mathsf{SymSeq} }}$ and $R\in\Sigma$, with $r:=|R|$. There are natural isomorphisms $$\begin{aligned}
\notag
(A_1{{ \check{{{ \otimes }}} }}\dotsb{{ \check{{{ \otimes }}} }}A_t)[R]&{{ \ \cong \ }}\
\coprod_{\substack{{{ {\pi}\colon\thinspace{R}{{ \longrightarrow }}{\mathbf{t}} }}\\ \text{in ${{ \mathsf{Set} }}$}}}
A_1[\pi^{-1}(1)]{{ \,\wedge\, }}\dotsb{{ \,\wedge\, }}A_t[\pi^{-1}(t)],\\
\label{eq:tensor_check_calc}
&{{ \ \cong \ }}\coprod_{r_1+\dotsb +r_t=r}A_1[\mathbf{r_1}]{{ \,\wedge\, }}\dotsb{{ \,\wedge\, }}A_t[\mathbf{r_t}]\underset{{\Sigma_{r_1}\times\dotsb\times
\Sigma_{r_t}}}{\cdot}\Sigma_{r}\end{aligned}$$
Here, ${{ \mathsf{Set} }}$ is the category of sets and their maps, and displays the tensor product $(A_1{{ \check{{{ \otimes }}} }}\dotsb{{ \check{{{ \otimes }}} }}A_t)[R]$ as a coproduct of $\Sigma_{r_1}\times\dotsb\times\Sigma_{r_t}$-orbits. It will be conceptually useful to extend the definition of tensor powers $A^{{{ \check{{{ \otimes }}} }}t}$ to situations in which the integers $t$ are replaced by a finite set $T$.
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $A\in{{ \mathsf{SymSeq} }}$ and $R,T\in\Sigma$. The *tensor powers* $A^{{{ \check{{{ \otimes }}} }}T}\in{{ \mathsf{SymSeq} }}$ are defined objectwise by $$\begin{aligned}
(A^{{{ \check{{{ \otimes }}} }}\emptyset})[R]:=
\coprod_{\substack{{{ {\pi}\colon\thinspace{R}{{ \longrightarrow }}{\emptyset} }}\\ \text{in ${{ \mathsf{Set} }}$}}}
S,\quad\quad
&(A^{{{ \check{{{ \otimes }}} }}T})[R]:=
\coprod_{\substack{{{ {\pi}\colon\thinspace{R}{{ \longrightarrow }}{T} }}\\ \text{in ${{ \mathsf{Set} }}$}}}
\bigwedge_{t\in T} A[\pi^{-1}(t)]\quad(T\neq\emptyset).\end{aligned}$$ Note that there are no functions ${{ {\pi}\colon\thinspace{R}{{ \longrightarrow }}{\emptyset} }}$ in ${{ \mathsf{Set} }}$ unless $R=\emptyset$. We will use the abbreviation $A^{{{ \check{{{ \otimes }}} }}0}:=A^{{{ \check{{{ \otimes }}} }}\emptyset}$.
\[defn:circle\_product\] Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $A,B,C\in{{ \mathsf{SymSeq} }}$, and $r,t\geq 0$. The *circle product* (or composition product) $A\circ B\in{{ \mathsf{SymSeq} }}$ is defined objectwise by the coend $$\begin{aligned}
\label{eq:circle_product_calc}
(A\circ B)[\mathbf{r}] := A{{ \,\wedge\, }}_\Sigma (B^{{{ \check{{{ \otimes }}} }}-})[\mathbf{r}]
&{{ \ \cong \ }}\coprod_{t\geq 0}A[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t}
(B^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}].\end{aligned}$$ The *mapping sequence* $\operatorname{Map}^\circ(B,C)\in{{ \mathsf{SymSeq} }}$ and the *mapping object* $\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,C)\in{{ \mathsf{SymSeq} }}$ are defined objectwise by the ends $$\begin{aligned}
\operatorname{Map}^\circ(B,C)[\mathbf{t}] &:= \operatorname{Map}((B^{{{ \check{{{ \otimes }}} }}\mathbf{t}})[-],C)^\Sigma {{ \ \cong \ }}\prod_{r\geq 0}\operatorname{Map}((B^{{{ \check{{{ \otimes }}} }}\mathbf{t}})[\mathbf{r}],
C[\mathbf{r}])^{\Sigma_r},\\
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,C)[\mathbf{t}] &:= \operatorname{Map}(B,C[\mathbf{t}\amalg -])^\Sigma {{ \ \cong \ }}\prod_{r\geq 0}\operatorname{Map}(B[\mathbf{r}],C[\mathbf{t}
\boldsymbol{+}\mathbf{r}])^{\Sigma_r}.\end{aligned}$$
These mapping sequences and mapping objects fit into isomorphisms $$\begin{aligned}
\label{eq:circle_mapping_sequence_adjunction}
\hom(A\circ B,C)&{{ \ \cong \ }}\hom(A,\operatorname{Map}^\circ(B,C)),\\
\label{eq:tensorcheck_mapping_sequence_adjunction}
\hom(A{{ \check{{{ \otimes }}} }}B,C)&{{ \ \cong \ }}\hom(A,\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,C)),\end{aligned}$$ natural in symmetric sequences $A,B,C$. Here, the $\hom$ notation denotes the indicated set of morphisms in ${{ \mathsf{SymSeq} }}$.
\[prop:closed\_monoidal\_on\_symmetric\_sequences\] Consider symmetric sequences in ${{ \mathsf{C} }}$.
- $({{ \mathsf{SymSeq} }},{{ \check{{{ \otimes }}} }},1)$ has the structure of a closed symmetric monoidal category with all small limits and colimits. The unit for ${{ \check{{{ \otimes }}} }}$ denoted “$1$” is the symmetric sequence concentrated at $0$ with value $S$.
- $({{ \mathsf{SymSeq} }},\circ,I)$ has the structure of a closed monoidal category with all small limits and colimits. The unit for $\circ$ denoted “$I$” is the symmetric sequence concentrated at $1$ with value $S$. Circle product is not symmetric.
\[defn:hat\_construction\_embed\_at\_zero\] Let $Z\in{{ \mathsf{C} }}$. Define $\hat{Z}\in{{ \mathsf{SymSeq} }}$ to be the symmetric sequence concentrated at $0$ with value $Z$.
The functor ${{ {\hat{-}}\colon\thinspace{{{ \mathsf{C} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ fits into the adjunction $
\xymatrix@1{
\hat{-}\colon{{ \mathsf{C} }}\ar@<0.5ex>[r] &
{{ \mathsf{SymSeq} }}:{{ \mathrm{Ev} }}_0\ar@<0.5ex>[l]
}
$ with left adjoint on top and ${{ \mathrm{Ev} }}_0$ the *evaluation* functor defined objectwise by ${{ \mathrm{Ev} }}_0(B):=B[\mathbf{0}]$. Note that $\hat{-}$ embeds ${{ \mathsf{C} }}$ in ${{ \mathsf{SymSeq} }}$ as the full subcategory of symmetric sequences concentrated at $0$.
\[defn:corresponding\_functor\] Consider symmetric sequences in ${{ \mathsf{C} }}$. Let ${{ \mathcal{O} }}$ be a symmetric sequence and $Z\in{{ \mathsf{C} }}$. The corresponding functor ${{ {{{ \mathcal{O} }}}\colon\thinspace{{{ \mathsf{C} }}}{{ \longrightarrow }}{{{ \mathsf{C} }}} }}$ is defined objectwise by $
{{ \mathcal{O} }}(Z):={{ \mathcal{O} }}\circ(Z):=\amalg_{t\geq 0}{{ \mathcal{O} }}[\mathbf{t}]
{{ \,\wedge\, }}_{\Sigma_t}Z^{\wedge t}.
$
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let ${{ \mathcal{O} }},A\in{{ \mathsf{SymSeq} }}$ and $Z\in{{ \mathsf{C} }}$. There are natural isomorphisms $$\begin{aligned}
\label{eq:circ_product_and_evaluate_at_zero}
\widehat{{{ \mathcal{O} }}(Z)}=
\widehat{{{ \mathcal{O} }}\circ(Z)}{{ \ \cong \ }}{{ \mathcal{O} }}\circ\hat{Z},\quad\quad
{{ \mathrm{Ev} }}_0({{ \mathcal{O} }}\circ A){{ \ \cong \ }}{{ \mathcal{O} }}\circ\bigl({{ \mathrm{Ev} }}_0(A)\bigr).\end{aligned}$$
This follows from and .
\[defn:operad\] Consider symmetric sequences in ${{ \mathsf{C} }}$. An *operad* in ${{ \mathsf{C} }}$ is a monoid object in $({{ \mathsf{SymSeq} }},\circ,I)$ and a *morphism of operads* is a morphism of monoid objects in $({{ \mathsf{SymSeq} }},\circ,I)$.
If ${{ \mathcal{O} }}$ is an operad, then the associated functor ${{ \mathcal{O} }}: {{ \mathsf{C} }}\to {{ \mathsf{C} }}$ is a monad.
\[defn:algebras\_and\_modules\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$.
- A *left ${{ \mathcal{O} }}$-module* is an object in $({{ \mathsf{SymSeq} }},\circ,I)$ with a left action of ${{ \mathcal{O} }}$ and a *morphism of left ${{ \mathcal{O} }}$-modules* is a map that respects the left ${{ \mathcal{O} }}$-module structure. Denote by ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ the category of left ${{ \mathcal{O} }}$-modules and their morphisms.
- A *right ${{ \mathcal{O} }}$-module* is an object in $({{ \mathsf{SymSeq} }},\circ,I)$ with a right action of ${{ \mathcal{O} }}$ and a *morphism of right ${{ \mathcal{O} }}$-modules* is a map that respects the right ${{ \mathcal{O} }}$-module structure. Denote by ${{ {{ \mathsf{Rt} }}_{{ \mathcal{O} }}}}$ the category of right ${{ \mathcal{O} }}$-modules and their morphisms.
- An *$({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodule* is an object in $({{ \mathsf{SymSeq} }},\circ,I)$ with compatible left ${{ \mathcal{O} }}$-module and right ${{ \mathcal{O} }}$-module structures and a *morphism of $({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodules* is a map that respects the $({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodule structure. Denote by ${{ \mathsf{Bi} }}_{({{ \mathcal{O} }},{{ \mathcal{O} }})}$ the category of $({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodules and their morphisms.
- An *${{ \mathcal{O} }}$-algebra* is an algebra for the monad ${{ {{{ \mathcal{O} }}}\colon\thinspace{{{ \mathsf{C} }}}{{ \longrightarrow }}{{{ \mathsf{C} }}} }}$ and a *morphism of ${{ \mathcal{O} }}$-algebras* is a map in ${{ \mathsf{C} }}$ that respects the ${{ \mathcal{O} }}$-algebra structure. Denote by ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ the category of ${{ \mathcal{O} }}$-algebras and their morphisms.
It follows easily from that an ${{ \mathcal{O} }}$-algebra is the same as an object $Z$ in ${{ \mathsf{C} }}$ with a left ${{ \mathcal{O} }}$-module structure on $\hat{Z}$, and if $Z$ and $Z'$ are ${{ \mathcal{O} }}$-algebras, then a morphism of ${{ \mathcal{O} }}$-algebras is the same as a map ${{ {f}\colon\thinspace{Z}{{ \longrightarrow }}{Z'} }}$ in ${{ \mathsf{C} }}$ such that ${{ {\hat{f}}\colon\thinspace{\hat{Z}}{{ \longrightarrow }}{\hat{Z'}} }}$ is a morphism of left ${{ \mathcal{O} }}$-modules. In other words, an algebra over an operad ${{ \mathcal{O} }}$ is the same as a left ${{ \mathcal{O} }}$-module that is concentrated at $0$, and ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ embeds in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ as the full subcategory of left ${{ \mathcal{O} }}$-modules concentrated at $0$, via the functor ${{ {\hat{-}}\colon\thinspace{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} }}$, $Z\longmapsto \hat{Z}$. Define the *evaluation* functor ${{ {{{ \mathrm{Ev} }}_0}\colon\thinspace{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}} }}$ objectwise by ${{ \mathrm{Ev} }}_0(B):=B[\mathbf{0}]$.
\[prop:basic\_properties\_LTO\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$. There are adjunctions $$\begin{aligned}
\label{eq:free_forgetful_adjunction}
\xymatrix{
{{ \mathsf{C} }}\ar@<0.5ex>[r]^-{{{ \mathcal{O} }}\circ(-)} & {{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]^-{U}
}\quad\quad
\xymatrix{
{{ \mathsf{SymSeq} }}\ar@<0.5ex>[r]^-{{{ \mathcal{O} }}\circ-} & {{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]^-{U}
}\quad\quad
\xymatrix{
{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{\hat{-}} & {{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]^-{{{ \mathrm{Ev} }}_0}
}\end{aligned}$$ with left adjoints on top and $U$ the forgetful functor. All small colimits exist in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ and ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and both reflexive coequalizers and filtered colimits are preserved (and created) by the forgetful functors. All small limits exist in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ and ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and are preserved (and created) by the forgetful functors.
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let ${{ \mathsf{D} }}$ be a small category, and let $X,Y\in{{ \mathsf{SymSeq} }}^{{ \mathsf{D} }}$. Denote by $\operatorname{Map}^\circ(X,Y)$ the indicated composition of functors ${{ \mathsf{D} }}^{{ \mathrm{op} }}\times{{ \mathsf{D} }}{{ \longrightarrow }}{{ \mathsf{SymSeq} }}$. The *mapping sequence* of ${{ \mathsf{D} }}$-shaped diagrams is defined by the end $\operatorname{Map}^\circ(X,Y)^{{ \mathsf{D} }}\in{{ \mathsf{SymSeq} }}$.
By the universal property of ends, it follows easily that for all ${{ \mathcal{O} }}\in{{ \mathsf{SymSeq} }}$, there are isomorphisms $$\begin{aligned}
\label{eq:circle_adjunction_for_diagrams}
\hom_{{ \mathsf{D} }}\bigl({{ \mathcal{O} }}\circ X,Y)&{{ \ \cong \ }}\hom({{ \mathcal{O} }},\operatorname{Map}^\circ(X,Y)^{{ \mathsf{D} }}\bigr)\end{aligned}$$ natural in ${{ \mathcal{O} }},X,Y$ and that $\operatorname{Map}^\circ(X,Y)^{{ \mathsf{D} }}$ may be calculated by an equalizer in ${{ \mathsf{SymSeq} }}$ of the form $$\begin{aligned}
\operatorname{Map}^\circ(X,Y)^{{ \mathsf{D} }}{{ \ \cong \ }}\lim
\biggl(
\xymatrix{
\prod_{\alpha\in{{ \mathsf{D} }}}\limits
\operatorname{Map}^\circ(X_\alpha,Y_\alpha)\ar@<2.0ex>[r]\ar@<1.0ex>[r] &
\prod_{({{ {\xi}\colon{\alpha}\rightarrow{\alpha'} }})\in{{ \mathsf{D} }}}\limits
\operatorname{Map}^\circ(X_\alpha,Y_{\alpha'})
}
\biggr).\end{aligned}$$ Here, ${{ \mathcal{O} }}\circ X$ denotes the indicated composition of functors ${{ \mathsf{D} }}{{ \longrightarrow }}{{ \mathsf{SymSeq} }}$, the $\hom_{{ \mathsf{D} }}$ notation on the left-hand side of denotes the indicated set of morphisms in ${{ \mathsf{SymSeq} }}^{{ \mathsf{D} }}$, and the $\hom$ notation on the right-hand side of denotes the indicated set of morphisms in ${{ \mathsf{SymSeq} }}$.
Let ${{ \mathsf{D} }}$ be a small category and $X\in{{ \mathsf{C} }}^{{ \mathsf{D} }}$ (resp. $X\in{{ \mathsf{SymSeq} }}^{{ \mathsf{D} }}$) a ${{ \mathsf{D} }}$-shaped diagram. The *endomorphism operad* $\operatorname{End}(X)$ of $X$ is defined by $$\begin{aligned}
\operatorname{End}(X):=\operatorname{Map}^\circ(\hat{X},\hat{X})^{{ \mathsf{D} }}\quad\quad
\Bigl(\text{resp.}\quad
\operatorname{End}(X):=\operatorname{Map}^\circ(X,X)^{{ \mathsf{D} }}\Bigr)\end{aligned}$$ with its natural operad structure; i.e., such that for each $\alpha\in{{ \mathsf{D} }}$, the natural map $\operatorname{End}(X){{ \longrightarrow }}\operatorname{Map}^\circ(\hat{X}_\alpha,\hat{X}_\alpha)$ (resp. $\operatorname{End}(X){{ \longrightarrow }}\operatorname{Map}^\circ(X_\alpha,X_\alpha)$) is a morphism of operads.
Let $X$ be a ${{ \mathsf{D} }}$-shaped diagram in ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$). It follows easily from and that giving a map of operads ${{ {m}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{\operatorname{End}(X)} }}$ is the same as giving $X_\alpha$ an ${{ \mathcal{O} }}$-algebra structure (resp. left ${{ \mathcal{O} }}$-module structure) for each $\alpha\in{{ \mathsf{D} }}$, such that $X$ is a diagram of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules). Note that if ${{ \mathsf{D} }}$ is the terminal category (with exactly one object and no non-identity morphisms), then $\operatorname{End}(X){{ \ \cong \ }}\operatorname{Map}^\circ(\hat{X},\hat{X})$ (resp. $\operatorname{End}(X){{ \ \cong \ }}\operatorname{Map}^\circ(X,X)$), which recovers the usual endomorphism operad of an object $X$ in ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$) [@Harper_Modules; @Kriz_May].
Homotopy completion and ${{ \mathsf{TQ} }}$-completion {#sec:homotopy_completion}
======================================================
The purpose of this section is to describe two notions of completion for structured ring spectra: (i) homotopy completion (Definition \[defn:homotopy\_completion\]) and (ii) ${{ \mathsf{TQ} }}$-completion, or less concisely, completion with respect to topological Quillen homology (Definition \[defn:quillen\_homology\_completion\]). We will also establish a rigidification theorem for derived ${{ \mathsf{TQ} }}$-resolutions (Theorem \[thm:rigidification\]), which is required to define ${{ \mathsf{TQ} }}$-completion, and we will prove Theorem \[thm:comparing\_homotopy\_completion\_towers\] which compares homotopy completion towers along a map of operads.
Let ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a map of operads in ${{ \mathcal{R} }}$-modules. Recall that the change of operads adjunction $$\begin{aligned}
\label{eq:quillen_adjunction_change_of_operads_nice}
\xymatrix{
{{ \mathsf{Alg} }}_{{{ \mathcal{O} }}}\ar@<0.5ex>[r]^-{f_*} & {{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}\ar@<0.5ex>[l]^-{f^*}
}
\quad\quad
\Bigl(\text{resp.}\quad
\xymatrix{
{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}\ar@<0.5ex>[r]^-{f_*} & {{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}\ar@<0.5ex>[l]^-{f^*}
}
\Bigr)\end{aligned}$$ is a Quillen adjunction with left adjoint on top and $f^*$ the forgetful functor (more accurately, but less concisely, also called the “restriction along $f$ of the operad action”) [@Harper_Spectra; @Harper_Modules]; note that this is a particular instance of the usual change of monoids adjunction.
In this paper we always regard ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ and ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ with the positive flat stable model structure (Theorem \[thm:positive\_flat\_stable\_AlgO\]), unless otherwise specified.
Let ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a map of operads in ${{ \mathcal{R} }}$-modules. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and define the ${{ \mathcal{O} }}$-algebra ${{ \mathcal{O} }}'\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X)$ (resp. left ${{ \mathcal{O} }}$-module ${{ \mathcal{O} }}'\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X$) by $$\begin{aligned}
{{ \mathcal{O} }}'\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X):={{ \mathsf{R} }}f^*({{ \mathsf{L} }}f_* (X)) =
{{ \mathsf{R} }}f^* \bigl({{ \mathcal{O} }}'\circ^{{ \mathsf{L} }}_{{ \mathcal{O} }}(X)\bigr)\\
\Bigl(\text{resp.}\quad
{{ \mathcal{O} }}'\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X:={{ \mathsf{R} }}f^*({{ \mathsf{L} }}f_* (X)) =
{{ \mathsf{R} }}f^* ({{ \mathcal{O} }}'\circ^{{ \mathsf{L} }}_{{ \mathcal{O} }}X)
\Bigr).\end{aligned}$$ Here, ${{ \mathsf{R} }}f^*,{{ \mathsf{L} }}f_*$ are the total right (resp. left) derived functors of $f^*,f_*$, respectively.
\[rem:compute-htpy-orbits\] Note that ${{ \mathsf{Alg} }}_I={{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and ${{ \mathsf{Lt} }}_I={{ \mathsf{SymSeq} }}$ (since $I$ is the initial operad) and that for any map of operads ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$, there are weak equivalences $$\begin{aligned}
{{ \mathcal{O} }}'\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X){{ \ \simeq \ }}{{ \mathsf{L} }}f_* (X) =
{{ \mathcal{O} }}'\circ^{{ \mathsf{L} }}_{{ \mathcal{O} }}(X)\quad\quad
\Bigl(\text{resp.}\quad
{{ \mathcal{O} }}'\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X{{ \ \simeq \ }}{{ \mathsf{L} }}f_* (X) =
{{ \mathcal{O} }}'\circ^{{ \mathsf{L} }}_{{ \mathcal{O} }}X
\Bigr)\end{aligned}$$ in the underlying category ${{ \mathsf{Alg} }}_I$ (resp. ${{ \mathsf{SymSeq} }}$), natural in $X$; this follows from the property that the forgetful functor to the underlying category preserves weak equivalences.
The *truncation* functor ${{ {\tau_k}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ is defined objectwise by $$\begin{aligned}
(\tau_k X)[\mathbf{r}]:=
\left\{
\begin{array}{rl}
X[\mathbf{r}],&\text{for $r\leq k$,}\\
*,&\text{otherwise},
\end{array}
\right.\end{aligned}$$ for each $k\geq 1$. In other words, $\tau_k X$ is the symmetric sequence obtained by truncating $X$ above level $k$. Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. It is easy to verify that the canonical map of operads ${{ \mathcal{O} }}{{ \longrightarrow }}\tau_1{{ \mathcal{O} }}$ factors through each truncation $\tau_k{{ \mathcal{O} }}$, and hence gives rise to a commutative diagram of operads $$\begin{aligned}
\label{eq:towers_of_operads}
\xymatrix{
\{\tau_k{{ \mathcal{O} }}\}: &
\tau_1{{ \mathcal{O} }}& \tau_2{{ \mathcal{O} }}\ar[l] & \tau_3{{ \mathcal{O} }}\ar[l] &
\dotsb \ar[l] & & \\
\{{{ \mathcal{O} }}\}:\ar@<0.5ex>[u] &
\ \ \ {{ \mathcal{O} }}\ar@<-1.5ex>[u]\ar@/_0.5pc/[ur]\ar@/_0.5pc/[urr]
\ar@{}[urrrrr]_(.4){\dotsb}
}\end{aligned}$$ and $({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodules. In other words, associated to each such operad ${{ \mathcal{O} }}$ is a coaugmented tower $\{{{ \mathcal{O} }}\}{{ \longrightarrow }}\{\tau_k{{ \mathcal{O} }}\}$ of operads and $({{ \mathcal{O} }},{{ \mathcal{O} }})$-bimodules, where $\{{{ \mathcal{O} }}\}$ denotes the constant tower with value ${{ \mathcal{O} }}$. This tower underlies the following definition of completion for ${{ \mathcal{O} }}$-algebras and left ${{ \mathcal{O} }}$-modules, which plays a key role in this paper.
Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$.
- The canonical maps $\tau_1{{ \mathcal{O} }}{{ \longrightarrow }}{{ \mathcal{O} }}{{ \longrightarrow }}\tau_1{{ \mathcal{O} }}$ of operads factor the identity map.
- Note that ${{ \mathcal{O} }}[\mathbf{0}]=*$ and ${{ \mathcal{O} }}[\mathbf{1}]=I[\mathbf{1}]$ if and only if $\tau_1{{ \mathcal{O} }}=I$, i.e., if and only if the operad ${{ \mathcal{O} }}$ agrees with the initial operad $I$ at levels $0$ and $1$.
Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module). The *completion tower* of $X$ is the coaugmented tower of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) $$\begin{aligned}
\label{eq:towers_of_algebras}
\{X\}{{ \longrightarrow }}\{\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X)\}
\quad\quad
\Bigl(\text{resp.}\quad
\{X\}{{ \longrightarrow }}\{\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X\}
\Bigr)\end{aligned}$$ obtained by applying $-\circ_{{ \mathcal{O} }}(X)$ (resp. $-\circ_{{ \mathcal{O} }}X$) to the coaugmented tower . The *completion* $X^\wedge$ of $X$ is the ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) defined by $$\begin{aligned}
\label{eq:completion_of_algebras}
X^\wedge:=\lim\nolimits^{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X)\bigr)
\quad\quad
\Bigl(\text{resp.}\quad
X^\wedge:=\lim\nolimits^{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X\bigr)
\Bigr),
\end{aligned}$$ i.e., the limit of the completion tower of $X$. Here, $\{X\}$ denotes the constant tower with value $X$. Thus, completion defines a coaugmented functor on ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
We often suppress the forgetful functors ${{ \mathsf{Alg} }}_{\tau_k{{ \mathcal{O} }}}{{ \longrightarrow }}{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ and ${{ \mathsf{Lt} }}_{\tau_k{{ \mathcal{O} }}}{{ \longrightarrow }}{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ from the notation, as in .
Homotopy completion and topological [Q]{}uillen homology
--------------------------------------------------------
The purpose of this subsection is to introduce homotopy completion (Definition \[defn:homotopy\_completion\]) and topological Quillen homology (Definition \[defn:quillen\_homology\]).
In this paper we will primarily be interested in a homotopy invariant version of the completion functor, which involves the following homotopy invariant version of the limit functor on towers.
\[defn:model\_structure\_on\_towers\] Let ${{ \mathsf{M} }}$ be a model category with all small limits and let ${{ \mathsf{D} }}$ be the category $\{0\leftarrow 1\leftarrow 2\leftarrow\cdots\}$ with objects the non-negative integers and a single morphism $i\leftarrow j$ for each $i\leq j$. Consider the category ${{ \mathsf{M} }}^{{ \mathsf{D} }}$ of ${{ \mathsf{D} }}$-shaped diagrams (or towers) in ${{ \mathsf{M} }}$ with the injective model structure [@Goerss_Jardine VI.1.1]. The *homotopy limit* functor ${{ {\operatorname{holim}}\colon\thinspace{{{ \mathsf{Ho} }}({{{ \mathsf{M} }}^{{ \mathsf{D} }}})}{{ \longrightarrow }}{{{ \mathsf{Ho} }}({{ \mathsf{M} }})} }}$ is the total right derived functor of the limit functor ${{ {\lim}\colon\thinspace{{{ \mathsf{M} }}^{{ \mathsf{D} }}}{{ \longrightarrow }}{{{ \mathsf{M} }}} }}$.
We are now in a good position to define homotopy completion.
\[defn:homotopy\_completion\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module). The *homotopy completion* $X^{{ {{ \mathsf{h} }}\wedge }}$ of $X$ is the ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) defined by $$\begin{aligned}
X^{{ {{ \mathsf{h} }}\wedge }}:=\operatorname{holim}^{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X^c)\bigr)
\quad\quad
\Bigl(\text{resp.}\quad
X^{{ {{ \mathsf{h} }}\wedge }}:=\operatorname{holim}^{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X^c\bigr)
\Bigr),\end{aligned}$$ the homotopy limit of the completion tower of the functorial cofibrant replacement $X^c$ of $X$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
It is easy to check that if $X$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module), then the weak equivalence $X^c{{ \longrightarrow }}X$ induces zigzags of weak equivalences $$\begin{aligned}
X^{{ {{ \mathsf{h} }}\wedge }}&{{ \ \simeq \ }}\operatorname{holim}^{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X)\bigr){{ \ \simeq \ }}\operatorname{holim}^{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X)\bigr)\\
\Bigl(\text{resp.}\quad
X^{{ {{ \mathsf{h} }}\wedge }}&{{ \ \simeq \ }}\operatorname{holim}^{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X\bigr){{ \ \simeq \ }}\operatorname{holim}^{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X\bigr)
\Bigr)\end{aligned}$$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), natural in $X$. Hence the homotopy completion $X^{{ {{ \mathsf{h} }}\wedge }}$ of a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module) $X$ may be calculated by taking the homotopy limit of the completion tower of $X$.
In this paper we consider topological Quillen homology of an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) as an object in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) via the forgetful functor as follows.
\[defn:quillen\_homology\] If ${{ \mathcal{O} }}$ is an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$, and $X$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), then the *topological Quillen homology* ${{ \mathsf{TQ} }}(X)$ of $X$ is the ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) $\tau_1{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X)$ (resp. $\tau_1{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X$).
In particular, when applied to a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module) $X$, the completion tower interpolates between topological Quillen homology ${{ \mathsf{TQ} }}(X)$ and homotopy completion $X^{{ {{ \mathsf{h} }}\wedge }}$.
${{ \mathsf{TQ} }}$-completion {#subsec:TQ_completion}
------------------------------
The purpose of this subsection is to introduce a second naturally occurring notion of completion for structured ring spectra, called ${{ \mathsf{TQ} }}$-completion, or less concisely, completion with respect to topological Quillen homology (Definition \[defn:quillen\_homology\_completion\]). Defining ${{ \mathsf{TQ} }}$-completion requires the construction of a rigidification of the derived ${{ \mathsf{TQ} }}$-resolution from a diagram in the homotopy category to a diagram in the model category. This rigidification problem is solved in Theorem \[thm:rigidification\].
The ${{ \mathsf{TQ} }}$-completion construction is conceptual and can be thought of as a spectral algebra analog of Sullivan’s [@Sullivan_MIT_notes; @Sullivan_genetics] localization and completion of spaces, Bousfield-Kan’s [@Bousfield_Kan I.4] completion of spaces with respect to homology, and Carlsson’s [@Carlsson_equivariant II.4] and Arone-Kankaanrinta’s [@Arone_Kankaanrinta 0.1] completion and localization of spaces with respect to stable homotopy.
Here is the idea behind the construction. We want to define ${{ \mathsf{TQ} }}$-completion $X_{{ \mathsf{TQ} }}^\wedge$ of a structured ring spectrum $X$ to be the structured ring spectrum defined by (showing only the coface maps) the homotopy limit of $$\begin{aligned}
X^{\wedge}_{{{ \mathsf{TQ} }}}:=
\operatorname{holim}\limits_{\Delta}
\Bigl(
\xymatrix{
{{ \mathsf{TQ} }}(X)\ar@<-0.5ex>[r]\ar@<0.5ex>[r] &
({{ \mathsf{TQ} }})^2 (X)
\ar@<-1.0ex>[r]\ar[r]\ar@<1.0ex>[r] &
({{ \mathsf{TQ} }})^3 (X)\cdots
}
\Bigr)\end{aligned}$$ the cosimplicial resolution (or Godement resolution) with respect to the monad (or triple) ${{ \mathsf{TQ} }}$. However, there are technical details that one needs to resolve in order to make sense of this definition for ${{ \mathsf{TQ} }}$-completion. This is because ${{ \mathsf{TQ} }}$ naturally arises as a functor on the level of the homotopy categories, and to work with and make sense of the homotopy limit $\operatorname{holim}_\Delta$ we need a point-set level construction of the derived ${{ \mathsf{TQ} }}$-cosimplicial resolution , or more precisely, a construction on the level of model categories. Successfully resolving this issue is the purpose of the rest of this subsection, and amounts to solving a rigidification problem (Theorem \[thm:rigidification\]) for the derived cosimplicial resolution with respect to ${{ \mathsf{TQ} }}$.
Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Then the canonical map of operads ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{\tau_1{{ \mathcal{O} }}} }}$ induces a Quillen adjunction as in and hence induces a corresponding adjunction $$\begin{aligned}
\label{eq:derived_adjunction_quillen_homology}
\xymatrix{
{{ \mathsf{Ho} }}({{ \mathsf{Alg} }}_{{{ \mathcal{O} }}})\ar@<0.5ex>[r]^-{{{ \mathsf{L} }}f_*} & {{ \mathsf{Ho} }}({{ \mathsf{Alg} }}_{\tau_1{{ \mathcal{O} }}})\ar@<0.5ex>[l]^-{{{ \mathsf{R} }}f^*}
}
\quad\quad
\Bigl(\text{resp.}\quad
\xymatrix{
{{ \mathsf{Ho} }}({{ \mathsf{Lt} }}_{{{ \mathcal{O} }}})\ar@<0.5ex>[r]^-{{{ \mathsf{L} }}f_*} & {{ \mathsf{Ho} }}({{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}})\ar@<0.5ex>[l]^-{{{ \mathsf{R} }}f^*}
}
\Bigr)\end{aligned}$$ on the homotopy categories. Hence topological Quillen homology ${{ \mathsf{TQ} }}$ is the monad (or triple) on the homotopy category ${{ \mathsf{Ho} }}({{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}})$ (resp. ${{ \mathsf{Ho} }}({{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}})$) associated to the derived adjunction . Denote by ${{ \mathsf{K} }}$ the corresponding comonad (or cotriple) $$\begin{aligned}
{{ \mathrm{id} }}{{ \longrightarrow }}{{ \mathsf{TQ} }}&\quad\text{(unit)},\quad\quad\quad
&{{ \mathrm{id} }}{{ \longleftarrow }}{{ \mathsf{K} }}& \quad\text{(counit)}, \\
{{ \mathsf{TQ} }}{{ \mathsf{TQ} }}{{ \longrightarrow }}{{ \mathsf{TQ} }}&\quad\text{(multiplication)},\quad\quad\quad
&{{ \mathsf{K} }}{{ \mathsf{K} }}{{ \longleftarrow }}{{ \mathsf{K} }}& \quad\text{(comultiplication)},\end{aligned}$$ on ${{ \mathsf{Ho} }}({{ \mathsf{Alg} }}_{\tau_1{{ \mathcal{O} }}})$ (resp. ${{ \mathsf{Ho} }}({{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}})$). Then ${{ \mathsf{TQ} }}={{ \mathsf{R} }}f^*{{ \mathsf{L} }}f_*$ and ${{ \mathsf{K} }}={{ \mathsf{L} }}f_*{{ \mathsf{R} }}f^*$, and it follows that for any ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) X, the adjunction determines a cosimplicial resolution of $X$ with respect to topological Quillen homology ${{ \mathsf{TQ} }}$ of the form $$\begin{aligned}
\label{eq:derived_QH_resolution}
\xymatrix{
X\ar[r] &
{{ \mathsf{TQ} }}(X)\ar@<-0.5ex>[r]\ar@<0.5ex>[r] &
{{ \mathsf{TQ} }}^2(X)
\ar@<-1.0ex>[r]\ar[r]\ar@<1.0ex>[r]\ar@<-2.0ex>[l] &
{{ \mathsf{TQ} }}^3(X)\cdots\ar@<-2.5ex>[l]\ar@<-3.5ex>[l]
}\end{aligned}$$ This derived ${{ \mathsf{TQ} }}$-resolution can be thought of as encoding what it means for ${{ \mathsf{TQ} }}(X)$ to have the structure of a ${{ \mathsf{K} }}$-coalgebra. More precisely, the extra structure on ${{ \mathsf{TQ} }}(X)$ is the ${{ \mathsf{K} }}$-coalgebra structure on the underlying object ${{ \mathsf{L} }}f_*(X)$ of ${{ \mathsf{TQ} }}(X)$. One difficulty in working with the diagram is that it lives in the homotopy category ${{ \mathsf{Ho} }}({{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}})$ (resp. ${{ \mathsf{Ho} }}({{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}})$). The purpose of the rigidification theorem below is to construct a model of that lives in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
Consider any factorization of the canonical map ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{\tau_1{{ \mathcal{O} }}} }}$ in the category of operads as ${{ \mathcal{O} }}\xrightarrow{g}J_1\xrightarrow{h}\tau_1{{ \mathcal{O} }}$, a cofibration followed by a weak equivalence (Definition \[defn:model-cat-operads\]) with respect to the positive flat stable model structure on ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (Definition \[defn:lets\_define\_flat\_model\_structure\]); it is easy to verify that such factorizations exist using a small object argument (Proposition \[prop:functorial\_factorizations\_of\_maps\_of\_operads\]). The corresponding change of operads adjunctions have the form $$\begin{aligned}
\label{eq:factored_adjunctions}
\xymatrix{
{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{g_*} & {{ \mathsf{Alg} }}_{J_1}\ar@<0.5ex>[l]^-{g^*}\ar@<0.5ex>[r]^-{h_*} &
{{ \mathsf{Alg} }}_{\tau_1{{ \mathcal{O} }}}\ar@<0.5ex>[l]^-{h^*}
}\quad
\Bigl(\text{resp.}\quad
\xymatrix{
{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{g_*} & {{ \mathsf{Lt} }}_{J_1}\ar@<0.5ex>[l]^-{g^*}\ar@<0.5ex>[r]^-{h_*} &
{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}\ar@<0.5ex>[l]^-{h^*}
}
\Bigr)\end{aligned}$$ with left adjoints on top and $g^*,h^*$ the forgetful functors (more accurately, but less concisely, also called the “restriction along $g,h$, respectively, of the operad action”). These are Quillen adjunctions and since $h$ is a weak equivalence it follows that the $(h_*,h^*)$ adjunction is a Quillen equivalence (Theorem \[thm:comparing\_homotopy\_categories\]). We defer the proof of the following rigidification theorem to Section \[sec:homotopical\_analysis\_forgetful\_functors\] (just after Theorem \[thm:cofibration\_property\_needed\_for\_homology\_completion\]).
\[thm:rigidification\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$. If $X$ is a cofibrant ${{ \mathcal{O} }}$–algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module) and $n\geq 1$, then there are weak equivalences $
(g^*g_*)^n(X){{ \ \simeq \ }}{{ \mathsf{TQ} }}^n(X)
$ natural in such $X$.
The following description of ${{ \mathsf{TQ} }}$-completion is closely related to [@Carlsson] and [@Hess].
\[defn:quillen\_homology\_completion\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module). The *${{ \mathsf{TQ} }}$-completion* (or completion with respect to topological Quillen homology) $X_{{ \mathsf{TQ} }}^\wedge$ of $X$ is the ${{ \mathcal{O} }}$–algebra (resp. left ${{ \mathcal{O} }}$-module) defined by (showing only the coface maps) the homotopy limit of the cosimplicial resolution $$\begin{aligned}
\label{eq:homology_completion}
X^{\wedge}_{{{ \mathsf{TQ} }}}:=
\operatorname{holim}\limits_{\Delta}
\Bigl(
\xymatrix{
(g^*g_*)(X^c)\ar@<-0.5ex>[r]\ar@<0.5ex>[r] &
(g^*g_*)^2(X^c)
\ar@<-1.0ex>[r]\ar[r]\ar@<1.0ex>[r] &
(g^*g_*)^3(X^c)\cdots
}
\Bigr)\end{aligned}$$ (or Godement resolution) of the functorial cofibrant replacement $X^c$ of $X$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) with respect to the monad $g^*g_*$. Here, $\operatorname{holim}_\Delta$ is calculated in the category of ${{ \mathcal{O} }}$–algebras (resp. left ${{ \mathcal{O} }}$-modules).
The $(g^*g_*)$-resolution can be thought of as encoding what it means for ${{ \mathsf{TQ} }}(X)$ to have the structure of a ${{ \mathsf{K} }}$-coalgebra. More precisely, the extra structure on $g^*g_*(X^c){{ \ \simeq \ }}{{ \mathsf{TQ} }}(X)$ is the $(g_*g^*)$-coalgebra structure on the underlying object $g_*(X^c)$ of $g^*g_*(X^c)$. In particular, the comonad $(g_*g^*)$ provides a point-set model for the derived comonad ${{ \mathsf{K} }}$ that coacts on ${{ \mathsf{TQ} }}(X)$ (up to a Quillen equivalence). This point-set model of ${{ \mathsf{K} }}$ is conjecturally related to the Koszul dual cooperad associated to ${{ \mathcal{O} }}$ (see, for instance, [@Ching_duality; @Fresse; @Ginzburg_Kapranov]).
It follows that the cosimplicial resolution in provides a rigidification of the derived cosimplicial resolution . One of our motivations for introducing the homotopy completion tower was its role as a potentially useful tool in analyzing ${{ \mathsf{TQ} }}$-completion defined above, but an investigation of these properties and the ${{ \mathsf{TQ} }}$-completion functor will be the subject of other papers and will not be elaborated here.
Comparing homotopy completion towers
------------------------------------
The purpose of this subsection is to prove Theorem \[thm:comparing\_homotopy\_completion\_towers\], which compares homotopy completion towers along a map of operads.
Let ${{ {g}\colon\thinspace{{{ \mathcal{O} }}'}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ be a map of operads in ${{ \mathcal{R} }}$-modules, and for each ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) $X$, consider the corresponding ${{ \mathcal{O} }}'$-algebra (resp. left ${{ \mathcal{O} }}'$-module) $X$ given by forgetting the left ${{ \mathcal{O} }}$-action along the map $g$; here we have dropped the forgetful functor $g^*$ from the notation. Consider the map $\emptyset{{ \longrightarrow }}X$ in ${{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}$ (resp. ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$) and use functorial factorization in ${{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}$ (resp. ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$) to obtain $$\begin{aligned}
\label{eq:functorial_factorization_prime_notation}
\emptyset{{ \longrightarrow }}X'{{ \longrightarrow }}X,\end{aligned}$$ a cofibration followed by an acyclic fibration.
In the next theorem we establish that replacing an operad ${{ \mathcal{O} }}$ by a weakly equivalent operad ${{ \mathcal{O} }}'$ changes the homotopy completion tower of $X$ only up to natural weak equivalence. In particular, the homotopy completion of $X$ as an ${{ \mathcal{O} }}'$-algebra is weakly equivalent to its homotopy completion as an ${{ \mathcal{O} }}$-algebra.
\[thm:comparing\_homotopy\_completion\_towers\] Let ${{ {g}\colon\thinspace{{{ \mathcal{O} }}'}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ be a map of operads in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}'[\mathbf{0}]=*$ and ${{ \mathcal{O} }}[\mathbf{0}]=*$. If $X$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), then there are maps of towers $$\begin{aligned}
\label{eq:comparing_completion_towers}
\xymatrix{
\{X'\}\ar[d]\ar@{=}[r] & \{X'\}\ar[d]^{(\sharp)}
\ar[r] & \{X\}\ar[d]\\
\{\tau_k{{ \mathcal{O} }}'\circ_{{{ \mathcal{O} }}'}(X')\}\ar[r]^-{(*)} &
\{\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}(X')\}\ar[r]^-{(**)} &
\{\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}}(X)\}
}\end{aligned}$$ $$\begin{aligned}
\label{eq:comparing_completion_towers_left_modules}
\text{resp.}\quad
\xymatrix{
\{X'\}\ar[d]\ar@{=}[r] & \{X'\}\ar[d]^{(\sharp)}
\ar[r] & \{X\}\ar[d]\\
\{\tau_k{{ \mathcal{O} }}'\circ_{{{ \mathcal{O} }}'}X'\}\ar[r]^-{(*)} &
\{\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}X'\}\ar[r]^-{(**)} &
\{\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}}X\}
}\end{aligned}$$ of ${{ \mathcal{O} }}'$-algebras (resp. left ${{ \mathcal{O} }}'$-modules), natural in $X$. If, furthermore, $g$ is a weak equivalence in the underlying category ${{ \mathsf{SymSeq} }}$, and $X$ is fibrant and cofibrant in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then the maps $(*)$ and $(**)$ are levelwise weak equivalences; here, we are using the notation to denote functorial cofibrant replacement of $X$ as an ${{ \mathcal{O} }}'$-algebra (resp. left ${{ \mathcal{O} }}'$-module).
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. The map of operads ${{ \mathcal{O} }}'{{ \longrightarrow }}{{ \mathcal{O} }}$ induces a commutative diagram of towers $$\begin{aligned}
\label{eq:diagram_of_towers_associated_to_map_of_operads}
\xymatrix{
\{{{ \mathcal{O} }}'\}\ar[d]\ar[r] & \{{{ \mathcal{O} }}\}\ar[d]\\
\{\tau_k{{ \mathcal{O} }}'\}\ar[r] & \{\tau_k{{ \mathcal{O} }}\}
}\end{aligned}$$ of operads and $({{ \mathcal{O} }}',{{ \mathcal{O} }}')$-bimodules; here, $\{{{ \mathcal{O} }}'\}$ and $\{{{ \mathcal{O} }}\}$ denote the constant towers with values ${{ \mathcal{O} }}'$ and ${{ \mathcal{O} }}$, respectively.
Consider the map of towers $(*)$. Each map $\tau_k{{ \mathcal{O} }}'\circ_{{{ \mathcal{O} }}'} X'{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}X'$ in $(*)$ is obtained by applying $-\circ_{{{ \mathcal{O} }}'}X'$ to the map $\tau_k{{ \mathcal{O} }}'{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}$. By , this map is isomorphic to the composite $
\tau_k{{ \mathcal{O} }}'\circ_{{{ \mathcal{O} }}'}X'\xrightarrow{\eta}
\tau_k{{ \mathcal{O} }}\circ_{\tau_k{{ \mathcal{O} }}'}\tau_k{{ \mathcal{O} }}'\circ_{{{ \mathcal{O} }}'}\circ X'
{{ \ \cong \ }}\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}X'
$ where ${{ {\eta}\colon\thinspace{{{ \mathrm{id} }}}{{ \longrightarrow }}{\tau_k{{ \mathcal{O} }}\circ_{\tau_k{{ \mathcal{O} }}'}-} }}$ is the unit map associated to the change of operads adjunction $
\xymatrix@1{
{{ \mathsf{Lt} }}_{\tau_k{{ \mathcal{O} }}'}\ar@<0.5ex>[r] &
{{ \mathsf{Lt} }}_{\tau_k{{ \mathcal{O} }}}\ar@<0.5ex>[l]
}
$. If, furthermore, $g$ is a weak equivalence in ${{ \mathsf{SymSeq} }}$, then the map $\tau_k{{ \mathcal{O} }}'{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}$ is a weak equivalence, and since $X'$ is cofibrant in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$ it follows from \[thm:comparing\_homotopy\_categories\] and \[prop:unit\_map\_is\_weak\_equivalence\] that $(*)$ is a levelwise weak equivalence.
Consider the map of towers $(**)$ and the change of operads adjunction $
\xymatrix@1{
{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}\ar@<0.5ex>[r] &
{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}\ar@<0.5ex>[l]
}
$. The weak equivalence $X'{{ \longrightarrow }}X$ of left ${{ \mathcal{O} }}'$-modules in has corresponding adjoint map $
{{ {\xi}\colon\thinspace{{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}X'}{{ \longrightarrow }}{X} }}
$. Each map $\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}X'{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}}X$ in $(**)$ is obtained by applying $\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}}-$ to the map $\xi$. If, furthermore, $g$ is a weak equivalence in ${{ \mathsf{SymSeq} }}$, and $X$ is fibrant and cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, then by \[thm:comparing\_homotopy\_categories\] the map $\xi$ is a weak equivalence between cofibrant objects in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and hence $(**)$ is a levelwise weak equivalence. To finish the proof, it suffices to describe the map of towers $(\sharp)$ in . Each map $X'{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ_{{{ \mathcal{O} }}'}X'$ is obtained by applying $-\circ_{{{ \mathcal{O} }}'}X'$ to the map ${{ \mathcal{O} }}'{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}$.
We defer the proof of the following proposition to Section \[sec:homotopical\_analysis\_forgetful\_functors\].
\[prop:replacement\_of\_operads\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Then there exists a map of operads ${{ {g}\colon\thinspace{{{ \mathcal{O} }}'}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ such that ${{ \mathcal{O} }}'[\mathbf{0}]=*$, and
- $g$ is a weak equivalence in the underlying category ${{ \mathsf{SymSeq} }}$,
- ${{ \mathcal{O} }}'$ satisfies Cofibrancy Condition \[CofibrancyCondition\].
Later in this paper, we need the following observation that certain homotopy limits commute with the forgetful functor.
\[prop:forgetful\_functor\_commutes\_with\_holim\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Consider any tower $B_0\leftarrow B_1\leftarrow B_2\leftarrow\cdots$ of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules). There are natural zigzags $$\begin{aligned}
U\operatorname{holim}^{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}_k B_k{{ \ \simeq \ }}\operatorname{holim}_k UB_k
\quad\quad
\Bigl(\text{resp.}\quad
U\operatorname{holim}^{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}_k B_k{{ \ \simeq \ }}\operatorname{holim}_k UB_k
\Bigr)\end{aligned}$$ of weak equivalences. Here, $U$ is the forgetful functor .
This follows from the dual of [@Harper_Bar proof of 3.15], together with the observation that the forgetful functor $U$ preserves weak equivalences and that fibrant towers are levelwise fibrant.
Homotopical analysis of the completion tower {#sec:homotopical_analysis_of_the_completion_tower}
============================================
The purpose of this section is to prove the main theorems stated in the introduction (Theorems \[thm:finiteness\], \[thm:hurewicz\], \[thm:relative\_hurewicz\], and \[MainTheorem\]). The unifying approach behind each of these theorems is to systematically exploit induction “up the homotopy completion tower” together with explicit calculations of the layers in terms of simplicial bar constructions (Theorem \[thm:calculating\_fiber\_of\_induced\_map\] and Proposition \[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\]). An important property of these layer calculations, which we fully exploit in the proofs of the main theorems, is that the simplicial bar constructions are particularly amenable to systematic connectivity and finiteness estimates (Propositions \[prop:connectivity\_of\_simplicial\_maps\_spectra\], \[prop:connectivity\], and \[prop:homotopy\_spectral\_sequence\]–\[prop:useful\_finiteness\_properties\_dervied\_smash\]).
The first step to proving the main theorems is to establish conditions under which the homotopy completion tower of $X$ converges strongly to $X$. This is accomplished in Theorem \[MainTheorem\], which necessarily is the first of the main theorems to be proved. Establishing strong convergence amounts to verifying that the connectivity of the natural maps from X into each stage of the tower increase as you go up the tower, and verifying this essentially reduces to understanding the implications of the connectivity estimates in Propositions \[prop:connectivity\_of\_simplicial\_maps\_spectra\] and \[prop:connectivity\] when studied in the context of the calculations in Propositions \[prop:fattened\_version\_of\_tower\] and \[prop:pushout\_diagram\_for\_bar\_construction\_tower\_coaugmented\] (see Proposition \[prop:connectivity\_estimates\]).
The upshot of strong convergence is that to calculate $\pi_i X$ for a fixed $i$, one only needs to calculate $\pi_i$ of a (sufficiently high but) finite stage of the tower. Having to only go “finitely high up the tower” to calculate $\pi_iX$, together with the explicit layer calculations in Theorem \[thm:calculating\_fiber\_of\_induced\_map\] and Proposition \[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\], are the key technical properties underlying our approach to the main theorems. For instance, our approach to the ${{ \mathsf{TQ} }}$ finiteness theorem (Theorem \[thm:finiteness\]) is to (i) start with an assumption about the finiteness properties of $\pi_i$ of ${{ \mathsf{TQ} }}$-homology (which is the bottom stage of the tower), (ii) to use explicit calculations of the layers of the tower to prove that these same finiteness properties are inherited by $\pi_i$ of the layers, and (iii) to conclude that these finiteness properties are inherited by $\pi_i$ of each stage of the tower. Strong convergence of the homotopy completion tower then finishes the proof of the ${{ \mathsf{TQ} }}$ finiteness theorem. It is essentially in this manner that we systematically exploit induction “up the homotopy completion tower” to prove each of the main theorems stated in the introduction.
Simplicial bar constructions and the homotopy completion tower {#sec:simplicial_bar_and_homotopy_completion}
--------------------------------------------------------------
Recall that ${{ \mathcal{R} }}$ is any commutative ring spectrum (Basic Assumption \[assumption:commutative\_ring\_spectrum\]) and that $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ denotes the closed symmetric monoidal category of ${{ \mathcal{R} }}$-modules (Definition \[defn:left\_R\_modules\]). Denote by ${{ \mathsf{S} }}$ (resp. ${{ \mathsf{S} }}_*$) the category of simplicial sets (resp. pointed simplicial sets). There are adjunctions $
\xymatrix@1{
{{ \mathsf{S} }}\ar@<0.5ex>[r]^-{(-)_+} &
{{ \mathsf{S} }}_*\ar@<0.5ex>[l]^-{U}\ar@<0.5ex>[r]^-{{{ \mathcal{R} }}{{ \otimes }}G_0} &
{{ \mathsf{Mod}_{{ \mathcal{R} }}}},\ar@<0.5ex>[l]
}
$ with left adjoints on top and $U$ the forgetful functor (see Proposition \[prop:closed\_symmetric\_monoidal\_structure\_on\_sym\_sequences\_pointed\_ssets\] for the tensor product ${{ \otimes }}$ notation together with ). The functor ${{ \mathcal{R} }}{{ \otimes }}G_0$ is left adjoint to “evaluation at $0$”; the notation agrees with Subsection \[sec:model\_structures\_on\_capR\_modules\] and [@Hovey_Shipley_Smith after 2.2.5]. Note that if $X\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and $K\in{{ \mathsf{S} }}_*$, then there are natural isomorphisms $X{{ \,\wedge\, }}K{{ \ \cong \ }}X{{ \,\wedge\, }}({{ \mathcal{R} }}{{ \otimes }}G_0 K)$ in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$; in other words, taking the objectwise smash product of $X$ with $K$ (as pointed simplicial sets) is the same as taking the smash product of $X$ with ${{ \mathcal{R} }}{{ \otimes }}G_0 K$ (as ${{ \mathcal{R} }}$-modules).
Recall the usual realization functor on simplicial ${{ \mathcal{R} }}$-modules and simplicial symmetric sequences; see also [@Goerss_Jardine IV.1, VII.1].
\[defn:realization\] Consider symmetric sequences in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. The *realization* functors $|-|$ for simplicial ${{ \mathcal{R} }}$-modules and simplicial symmetric sequences are defined objectwise by the coends $$\begin{aligned}
{{ {|-|}\colon\thinspace{{{ \mathsf{sMod}_{{ \mathcal{R} }}}}}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }},
&\quad\quad
X\longmapsto |X|:=X{{ \,\wedge\, }}_{\Delta}\Delta[-]_+\ ,\\
{{ {|-|}\colon\thinspace{{{ \mathsf{sSymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }},
&\quad\quad
X\longmapsto |X|:=X{{ \,\wedge\, }}_{\Delta}\Delta[-]_+\ .\end{aligned}$$
\[prop:realzns\_fit\_into\_adjunctions\] The realization functors fit into adjunctions $$\begin{aligned}
\label{eq:realization_mapping_object_adjunction_underlying}
\xymatrix{
{{ \mathsf{sMod}_{{ \mathcal{R} }}}}\ar@<0.5ex>[r]^-{|-|} & {{ \mathsf{Mod}_{{ \mathcal{R} }}}},\ar@<0.5ex>[l]
}\quad\quad
\xymatrix{
{{ \mathsf{sSymSeq} }}\ar@<0.5ex>[r]^-{|-|} & {{ \mathsf{SymSeq} }},\ar@<0.5ex>[l]
}\end{aligned}$$ with left adjoints on top.
Consider the case of ${{ \mathcal{R} }}$-modules (resp. symmetric sequences). Using the universal property of coends, it is easy to verify that the functor given objectwise by $
\operatorname{Map}({{ \mathcal{R} }}{{ \otimes }}G_0\Delta[-]_+,Y)
$ is a right adjoint of $|-|$.
The following is closely related to [@Goerss_Jardine IV.1.7] and [@EKMM X.2.4]; see also [@Dugger_Isaksen A] and [@Hirschhorn Chapter 18].
\[prop:realization\_monomorphisms\_weak\_equivalences\] Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a morphism of simplicial ${{ \mathcal{R} }}$-modules. If $f$ is a monomorphism (resp. objectwise weak equivalence), then ${{ {|f|}\colon\thinspace{|X|}{{ \longrightarrow }}{|Y|} }}$ is a monomorphism (resp. weak equivalence).
This is verified exactly as in [@Harper_Bar proof of 4.8, 4.9], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$.
The following is closely related to [@EKMM X.1.3].
\[prop:realization\_respects\_smash\_tensor\_and\_circle\] Consider symmetric sequences in ${{ \mathcal{R} }}$-modules.
- If $X,Y$ are simplicial ${{ \mathcal{R} }}$-modules, then there is a natural isomorphism\
$|X{{ \,\wedge\, }}Y|{{ \ \cong \ }}|X|{{ \,\wedge\, }}|Y|$.
- If $X,Y$ are simplicial symmetric sequences, then there are natural isomorphisms $|X{{ \check{{{ \otimes }}} }}Y|{{ \ \cong \ }}|X|{{ \check{{{ \otimes }}} }}|Y|$ and $|X\circ Y|{{ \ \cong \ }}|X|\circ|Y|$.
- If ${{ \mathcal{O} }}$ is a symmetric sequence, and $B$ is a simplicial symmetric sequence, then there is a natural isomorphism $|{{ \mathcal{O} }}[\mathbf{k}]{{ \,\wedge\, }}_{\Sigma_k}B^{{{ \check{{{ \otimes }}} }}k}|{{ \ \cong \ }}{{ \mathcal{O} }}[\mathbf{k}]{{ \,\wedge\, }}_{\Sigma_k}|B|^{{{ \check{{{ \otimes }}} }}k}$ for every $k\geq 2$.
Here, smash products, tensor products and circle products of simplicial objects are defined objectwise.
\[rem:realization\_of\_simplicial\_pointed\_spaces\] If $X\in{{ \mathsf{sS} }}_*$, denote by $|X|:=X{{ \,\wedge\, }}_\Delta\Delta[-]_+$ the realization of $X$. There is a natural isomorphism $X\times_\Delta\Delta[-]{{ \ \cong \ }}|X|$.
Consider part (a). Let $X,Y$ be simplicial objects in ${{ \mathsf{S} }}_*$. By Remark \[rem:realization\_of\_simplicial\_pointed\_spaces\], together with [@Goerss_Jardine IV.1.4], there is a natural isomorphism $|X\times Y|{{ \ \cong \ }}|X|\times|Y|$. Since realization ${{ {|-|}\colon\thinspace{{{ \mathsf{sS} }}_*}{{ \longrightarrow }}{{{ \mathsf{S} }}_*} }}$ is a left adjoint it commutes with colimits, and thus there is a natural isomorphism $|X{{ \,\wedge\, }}Y|{{ \ \cong \ }}|X|{{ \,\wedge\, }}|Y|$. Let $X,Y$ be simplicial ${{ \mathcal{R} }}$-modules and recall that $X{{ \,\wedge\, }}Y{{ \ \cong \ }}X{{ \otimes }}_{{ \mathcal{R} }}Y$. It follows that there are natural isomorphisms $
|X{{ \,\wedge\, }}Y|
{{ \ \cong \ }}\operatorname{colim}\Bigl(
\xymatrix@1{
|X|{{ \otimes }}|Y| & |X|{{ \otimes }}|{{ \mathcal{R} }}|{{ \otimes }}|Y|
\ar@<-0.5ex>[l]\ar@<0.5ex>[l]
}
\Bigr)
{{ \ \cong \ }}|X|{{ \,\wedge\, }}|Y|.
$ Parts (b) and (c) follow from part (a), together with the property that realization $|-|$ is a left adjoint and hence commutes with colimits.
\[rem:realization\_induces\_functor\_on\_algebras\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. It follows easily from Proposition \[prop:realization\_respects\_smash\_tensor\_and\_circle\] that if $X$ is a simplicial ${{ \mathcal{O} }}$-algebra (resp. simplicial left ${{ \mathcal{O} }}$-module), then the realization of its underlying simplicial object $|X|$ has an induced ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) structure; it follows that realization of the underlying simplicial objects induces functors $
{{ {|-|}\colon\thinspace{{{ {{ \mathsf{sAlg} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}} }}
$ and $
{{ {|-|}\colon\thinspace{{{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} }}.
$
In this paper we use the notation $\operatorname{Bar}$, as in Proposition \[prop:natural\_map\_is\_weak\_equivalence\] below, to denote the simplicial bar construction (with respect to circle product) defined in [@Harper_Bar 5.30].
\[prop:natural\_map\_is\_weak\_equivalence\] Let ${{ \mathcal{O} }}{{ \longrightarrow }}{{ \mathcal{O} }}'$ be a morphism of operads in ${{ \mathcal{R} }}$-modules. Let $X$ be a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module). If the simplicial bar construction $\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)$ is objectwise cofibrant in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then the natural map $$\begin{aligned}
&|\operatorname{Bar}({{ \mathcal{O} }}',{{ \mathcal{O} }},X)|\xrightarrow{{{ \ \simeq \ }}}{{ \mathcal{O} }}'\circ_{{ \mathcal{O} }}(X)
\quad\quad
\Bigl(\text{resp.}\quad
&|\operatorname{Bar}({{ \mathcal{O} }}',{{ \mathcal{O} }},X)|\xrightarrow{{{ \ \simeq \ }}}{{ \mathcal{O} }}'\circ_{{ \mathcal{O} }}X
\Bigr)\end{aligned}$$ is a weak equivalence.
This follows easily from Theorem \[thm:bar\_calculates\_derived\_circle\] and its proof.
The following theorem illustrates some of the good properties of the (positive) flat stable model structures (Section \[sec:model\_structures\]). We defer the proof to Section \[sec:homotopical\_analysis\_forgetful\_functors\].
\[thm:homotopical\_analysis\_of\_forgetful\_functors\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration between cofibrant objects in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then $j$ is a positive flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$).
- If $A$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module) and ${{ \mathcal{O} }}[\mathbf{0}]=*$, then $A$ is positive flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$).
If $X$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), then under appropriate cofibrancy conditions the coaugmented tower $
\{|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|\}{{ \longrightarrow }}\{|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\}
$ obtained by applying $|\operatorname{Bar}(-,{{ \mathcal{O} }}, X)|$ to the coaugmented tower , provides a weakly equivalent “fattened version” of the completion tower of $X$.
\[CofibrancyConditionWeaker\] If ${{ \mathcal{O} }}$ is an operad in ${{ \mathcal{R} }}$-modules, assume that ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$.
\[prop:fattened\_version\_of\_tower\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyConditionWeaker\]. If $X$ is a cofibrant left ${{ \mathcal{O} }}$-module, then in the following commutative diagram of towers in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ $$\begin{aligned}
\xymatrix{
\{|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|\}\ar[d]^{{{ \ \simeq \ }}}\ar[r] &
\{|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\}\ar[d]^{{{ \ \simeq \ }}}\\
\{X\}\ar[r] & \{\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X\},
}\end{aligned}$$ the vertical maps are levelwise weak equivalences.
It follows from Remark \[rem:realization\_induces\_functor\_on\_algebras\] that this diagram is a diagram of towers of left ${{ \mathcal{O} }}$-modules.
Since $X$ is a cofibrant left ${{ \mathcal{O} }}$-module, by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] the simplicial bar construction $\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)$ is objectwise cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and Proposition \[prop:natural\_map\_is\_weak\_equivalence\] finishes the proof.
Homotopy fiber sequences and the homotopy completion tower
----------------------------------------------------------
The purpose of this subsection is to prove Theorem \[MainTheorem\](c). We begin by introducing the following useful notation. For each $k\geq 0$, the functor ${{ {i_k}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ is defined objectwise by $$\begin{aligned}
(i_k X)[\mathbf{r}]:=
\left\{
\begin{array}{rl}
X[\mathbf{k}],&\text{for $r=k$,}\\
*,&\text{otherwise}.
\end{array}
\right.\end{aligned}$$ In other words, $i_k X$ is the symmetric sequence concentrated at $k$ with value $X[\mathbf{k}]$.
\[prop:pushout\_diagram\_for\_bar\_construction\_tower\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and $k\geq 2$. Then the left-hand pushout diagram $$\begin{aligned}
\label{eq:pushout_diagram_bar_constructions}
\xymatrix{
i_{k}{{ \mathcal{O} }}\ar[r]^-{{{ \ \subset\ }}}\ar[d] & \tau_{k}{{ \mathcal{O} }}\ar[d] \\
{*}\ar[r] & \tau_{k-1}{{ \mathcal{O} }}}
\quad\quad
\xymatrix{
|\operatorname{Bar}(i_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]\ar[r]^-{(*)} &
|\operatorname{Bar}(\tau_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]\\
{*}\ar[r] &
|\operatorname{Bar}(\tau_{k-1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|
}\end{aligned}$$ in ${{ {{ \mathsf{Rt} }}_{{ \mathcal{O} }}}}$ induces the right-hand pushout diagram in ${{ \mathsf{Alg} }}_I$ (resp. ${{ \mathsf{SymSeq} }}$). The map $(*)$ is a monomorphism, the left-hand diagram is a pullback diagram in ${{ \mathsf{Bi} }}_{({{ \mathcal{O} }},{{ \mathcal{O} }})}$, and the right-hand diagram is a pullback diagram in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. The right-hand diagram is obtained by applying $|\operatorname{Bar}(-,{{ \mathcal{O} }}, X)|$ to the left-hand diagram. Since the forgetful functor ${{ {{ \mathsf{Rt} }}_{{ \mathcal{O} }}}}{{ \longrightarrow }}{{ \mathsf{SymSeq} }}$ preserves colimits, the left-hand diagram is also a pushout diagram in ${{ \mathsf{SymSeq} }}$. It follows from the adjunction that applying $\operatorname{Bar}(-,{{ \mathcal{O} }}, X)$ to the left-hand diagram gives a pushout diagram of simplicial symmetric sequences. Noting that the realization functor $|-|$ is a left adjoint and preserves monomorphisms (\[prop:realzns\_fit\_into\_adjunctions\], \[prop:realization\_monomorphisms\_weak\_equivalences\]), together with the fact that pullbacks in ${{ \mathsf{Bi} }}_{({{ \mathcal{O} }},{{ \mathcal{O} }})}$ and ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ are calculated in the underlying category, finishes the proof.
\[prop:retract\_property\_and\_derived\_circle\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$, and let $k\geq 2$.
- The canonical maps $i_k{{ \mathcal{O} }}{{ \longrightarrow }}{{ \mathcal{O} }}{{ \longrightarrow }}i_k{{ \mathcal{O} }}$ in ${{ \mathsf{Rt} }}_{\tau_1{{ \mathcal{O} }}}$ factor the identity map.
- The functors $
{{ {i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}(-)}\colon\thinspace{{{ \mathsf{Alg} }}_{\tau_1{{ \mathcal{O} }}}}{{ \longrightarrow }}{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}} }}
$ and $
{{ {i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}-}\colon\thinspace{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}{{ \longrightarrow }}{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} }}
$ preserve weak equivalences between cofibrant objects, and hence the total left derived functors $i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}(-)$ and $i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}-$ exist [@Dwyer_Spalinski 9.3, 9.5].
Part (a) is clear. To prove part (b), it suffices to consider the case of left $\tau_1{{ \mathcal{O} }}$-modules. Let $B{{ \longrightarrow }}B'$ be a weak equivalence between cofibrant objects in ${{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}$. By part (a) there is a retract of maps of the form $$\begin{aligned}
\xymatrix{
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B\ar[d]^{(*)}\ar[r] &
{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B\ar[d]^{(**)}\ar[r] &
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B\ar[d]^{(*)}\\
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B'\ar[r] &
{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B'\ar[r] &
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B'
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$. Since ${{ {{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}-}\colon\thinspace{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}{{ \longrightarrow }}{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} }}$ is a left Quillen functor (induced by the canonical map $\tau_1{{ \mathcal{O} }}{{ \longrightarrow }}{{ \mathcal{O} }}$ of operads), we know that $(**)$ is a weak equivalence and hence $(*)$ is a weak equivalence.
The following theorem illustrates a few more of the good properties of the (positive) flat stable model structures (Section \[sec:model\_structures\]). We defer the proof to Section \[sec:homotopical\_analysis\_bar\_constructions\].
\[thm:reedy\_cofibrant\_for\_bar\_constructions\] Let ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a morphism of operads in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyCondition\]. Let $Y$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and consider the simplicial bar construction $\operatorname{Bar}({{ \mathcal{O} }}',{{ \mathcal{O} }},Y)$.
- If $Y$ is positive flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), then $\operatorname{Bar}({{ \mathcal{O} }}',{{ \mathcal{O} }},Y)$ is Reedy cofibrant in ${{ \mathsf{sAlg} }}_{{{ \mathcal{O} }}'}$ (resp. ${{ \mathsf{sLt} }}_{{{ \mathcal{O} }}'}$).
- If $Y$ is positive flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), then $|\operatorname{Bar}({{ \mathcal{O} }}',{{ \mathcal{O} }},Y)|$ is cofibrant in ${{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}$ (resp. ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$).
\[prop:cofibrant\_tau\_1\_O\_algebras\_and\_bar\_constructions\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyCondition\]. If $X$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module), then $|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is cofibrant in ${{ \mathsf{Alg} }}_{\tau_1{{ \mathcal{O} }}}$ (resp. ${{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}$).
This follows from Theorems \[thm:reedy\_cofibrant\_for\_bar\_constructions\] and \[thm:homotopical\_analysis\_of\_forgetful\_functors\].
Next we explicitly calculate the $k$-th layer of the homotopy completion tower.
\[thm:calculating\_fiber\_of\_induced\_map\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyCondition\]. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), and let $k\geq 2$.
- There is a homotopy fiber sequence of the form $$\begin{aligned}
&i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\bigl({{ \mathsf{TQ} }}(X)\bigr){{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X){{ \longrightarrow }}\tau_{k-1}{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X)\\
\Bigl(\text{resp.}\quad
&i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}{{ \mathsf{TQ} }}(X){{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X{{ \longrightarrow }}\tau_{k-1}{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X
\Bigr)\end{aligned}$$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), natural in $X$.
- If $X$ is cofibrant in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then there are natural weak equivalences $$\begin{aligned}
&|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}(|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|){{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\bigl({{ \mathsf{TQ} }}(X)\bigr)
\\
\Bigl(\text{resp.}\quad
&|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}{{ \mathsf{TQ} }}(X)
\Bigr).\end{aligned}$$
- If $X$ is cofibrant in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) and ${{ \mathcal{O} }}[\mathbf{1}]=I[\mathbf{1}]$, then there are natural weak equivalences $$\begin{aligned}
&{{ \mathcal{O} }}[\mathbf{k}]{{ \,\wedge\, }}_{\Sigma_{k}}
|\operatorname{Bar}(I,{{ \mathcal{O} }},X)|^{\wedge k}{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\bigl({{ \mathsf{TQ} }}(X)\bigr)
\\
\Bigl(\text{resp.}\quad
&{{ \mathcal{O} }}[\mathbf{k}]{{ \,\wedge\, }}_{\Sigma_{k}}
|\operatorname{Bar}(I,{{ \mathcal{O} }},X)|^{{{ \check{{{ \otimes }}} }}k}{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}{{ \mathsf{TQ} }}(X)
\Bigr).\end{aligned}$$
For useful material related to homotopy fiber sequences, see [@Goerss_Jardine II.8, II.8.20].
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). It is enough to treat the special case where $X$ is a cofibrant left ${{ \mathcal{O} }}$-module. By Proposition \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] there is a homotopy fiber sequence of the form $$\begin{aligned}
\label{eq:homotopy_fiber_sequence_bar_constructions}
|\operatorname{Bar}(i_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)| {{ \longrightarrow }}|\operatorname{Bar}(\tau_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)| {{ \longrightarrow }}|\operatorname{Bar}(\tau_{k-1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\end{aligned}$$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, natural in $X$. By Proposition \[prop:fattened\_version\_of\_tower\] we know that has the form $$\begin{aligned}
|\operatorname{Bar}(i_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X{{ \longrightarrow }}\tau_{k-1}{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X.\end{aligned}$$ Since the right ${{ \mathcal{O} }}$-action map $i_k{{ \mathcal{O} }}\circ{{ \mathcal{O} }}{{ \longrightarrow }}i_k{{ \mathcal{O} }}$ factors as $i_k{{ \mathcal{O} }}\circ{{ \mathcal{O} }}{{ \longrightarrow }}i_k{{ \mathcal{O} }}\circ\tau_1{{ \mathcal{O} }}{{ \longrightarrow }}i_k{{ \mathcal{O} }}$, there are natural isomorphisms $$\begin{aligned}
\label{eq:factoring_the_trivial_module_in_bar_construction}
\operatorname{Bar}(i_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X){{ \ \cong \ }}i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)\end{aligned}$$ of simplicial left ${{ \mathcal{O} }}$-modules. Applying the realization functor to , it follows from Proposition \[prop:realization\_respects\_smash\_tensor\_and\_circle\], Proposition \[prop:cofibrant\_tau\_1\_O\_algebras\_and\_bar\_constructions\], Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\], and Proposition \[prop:fattened\_version\_of\_tower\], that there are natural weak equivalences $$\begin{aligned}
\label{eq:derived_functor_interpretation_of_above}
|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}{{ \mathsf{TQ} }}(X)\end{aligned}$$ which finishes the proof of part (a). Part (b) follows from the proof of part (a) above.
Consider part (c). Proceed as in the proof of part (a) above, and assume furthermore that ${{ \mathcal{O} }}[\mathbf{1}]=I[\mathbf{1}]$. It follows from that $$\begin{aligned}
i_k{{ \mathcal{O} }}\circ |\operatorname{Bar}(I,{{ \mathcal{O} }},X)|{{ \ \simeq \ }}{{ \mathcal{O} }}[\mathbf{k}]{{ \,\wedge\, }}_{\Sigma_{k}}
|\operatorname{Bar}(I,{{ \mathcal{O} }},X)|^{{{ \check{{{ \otimes }}} }}k}\end{aligned}$$ from which we can conclude, by applying the second equivalence in , since $\tau_1{{ \mathcal{O} }}=I$ (Definition \[defn:operad\]).
\[prop:induction\_argument\_cofiber\_sequence\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyCondition\]. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map between cofibrant objects in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$). If the induced map $|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\xrightarrow{{{ \ \simeq \ }}}|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},Y)|$ is a weak equivalence, then the induced map $|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\xrightarrow{{{ \ \simeq \ }}}|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},Y)|$ is a weak equivalence for each $k\geq 2$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider the $$\begin{aligned}
\label{eq:diagram_of_fibration_sequences_bar}
\xymatrix{
|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]\ar[r] &
|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]\ar[r] &
|\operatorname{Bar}(\tau_{k-1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d] \\
|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},Y)|\ar[r] &
|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},Y)|\ar[r] &
|\operatorname{Bar}(\tau_{k-1}{{ \mathcal{O} }},{{ \mathcal{O} }},Y)|
}\end{aligned}$$ commutative diagram in ${{ \mathsf{SymSeq} }}$. It follows from Theorem \[thm:calculating\_fiber\_of\_induced\_map\] that the left-hand vertical map is a weak equivalence for each $k\geq 2$. If $k=2$, then the right-hand vertical map is a weak equivalence by assumption, hence by Proposition \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] and induction on $k$, the middle vertical map is a weak equivalence for each $k\geq 2$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. By Theorem \[thm:comparing\_homotopy\_completion\_towers\] and Propositions \[prop:replacement\_of\_operads\] and \[prop:forgetful\_functor\_commutes\_with\_holim\], we can suppose that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyCondition\]. We can restrict to the following special case. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map of left ${{ \mathcal{O} }}$-modules between cofibrant objects in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ such that the induced map $\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X{{ \longrightarrow }}\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}Y$ is a weak equivalence. We need to verify that the induced map $
{{ {f_*}\colon\thinspace{\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X}{{ \longrightarrow }}{\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}Y} }}
$ is a weak equivalence for each $k\geq 2$. We know by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] that $X,Y$ are positive flat stable cofibrant in ${{ \mathsf{SymSeq} }}$. If $k=1$, the map $f_*$ is a weak equivalence by assumption, and hence the induced map $
|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \longrightarrow }}|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},Y)|
$ is a weak equivalence by Proposition \[prop:fattened\_version\_of\_tower\]. It follows from Propositions \[prop:induction\_argument\_cofiber\_sequence\] and \[prop:fattened\_version\_of\_tower\] that $f_*$ is a weak equivalence for each $k\geq 2$, which finishes the proof.
Strong convergence of the homotopy completion tower
---------------------------------------------------
The purpose of this subsection is to prove Theorem \[MainTheorem\](a). For each $k\geq 0$, the functor ${{ {(-)^{>k}}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ is defined objectwise by $$\begin{aligned}
(X^{>k})[\mathbf{r}]:=
\left\{
\begin{array}{rl}
X[\mathbf{r}],&\text{for $r>k$,}\\
*,&\text{otherwise}.
\end{array}
\right.\end{aligned}$$
\[prop:pushout\_diagram\_for\_bar\_construction\_tower\_coaugmented\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and $k\geq 1$. Then the left-hand pushout diagram $$\begin{aligned}
\label{eq:pushout_diagram_bar_constructions_tower_coaugmented}
\xymatrix{
{{ \mathcal{O} }}^{>k}\ar[d]\ar[r]^-{{{ \ \subset\ }}} & {{ \mathcal{O} }}\ar[d]\\
{*}\ar[r] & \tau_k{{ \mathcal{O} }}}\quad\quad
\xymatrix{
|\operatorname{Bar}({{ \mathcal{O} }}^{>k},{{ \mathcal{O} }},X)|\ar[d]\ar[r]^-{(*)} &
|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]\\
{*}\ar[r] & |\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|
}\end{aligned}$$ in ${{ {{ \mathsf{Rt} }}_{{ \mathcal{O} }}}}$ induces the right-hand pushout diagram in ${{ \mathsf{Alg} }}_I$ (resp. ${{ \mathsf{SymSeq} }}$). The map $(*)$ is a monomorphism, the left-hand diagram is a pullback diagram in ${{ \mathsf{Bi} }}_{({{ \mathcal{O} }},{{ \mathcal{O} }})}$, and the right-hand diagram is a pullback diagram in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. The right-hand diagram is obtained by applying $|\operatorname{Bar}(-,{{ \mathcal{O} }}, X)|$ to the left-hand diagram, and exactly the same argument used in the proof of Proposition \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] allows to conclude.
The following two propositions are well known in stable homotopy theory. For the convenience of the reader, we have included short homotopical proofs in the context of symmetric spectra; see also [@Jardine_generalized_etale 4.3]. We defer the proof of the second proposition to Section \[sec:homotopical\_analysis\_forgetful\_functors\].
\[prop:connectivity\_of\_simplicial\_maps\_spectra\] Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a morphism of simplicial symmetric spectra (resp. simplicial ${{ \mathcal{R} }}$-modules). Let $k\in{{ \mathbb{Z} }}$.
- If $Y$ is objectwise $k$-connected, then $|Y|$ is $k$-connected.
- If $f$ is objectwise $k$-connected, then ${{ {|f|}\colon\thinspace{|X|}{{ \longrightarrow }}{|Y|} }}$ is $k$-connected.
Consider part (b) for the case of symmetric spectra. We need to verify that the realization ${{ {|f|}\colon\thinspace{|X|}{{ \longrightarrow }}{|Y|} }}$ is $k$-connected. By exactly the same argument as in the proof of [@Harper_Bar 9.21], it follows from a filtration of degenerate subobjects (see also [@Jardine_generalized_etale 4.3]) that the induced map ${{ {Df_n}\colon\thinspace{DX_n}{{ \longrightarrow }}{DY_n} }}$ on degenerate subobjects is $k$-connected for each $n\geq 1$. Using exactly the same argument as in the proof of [@Harper_Bar 4.8], it then follows from the skeletal filtration of realization that $|f|$ is $k$-connected. Part (a) follows from part (b) by considering the map $*{{ \longrightarrow }}Y$. The case of ${{ \mathcal{R} }}$-modules reduces to the case of symmetric spectra by applying the forgetful functor.
\[rem:smash\_product\_and\_tensor\_product\] It is important to note (Basic Assumption \[assumption:commutative\_ring\_spectrum\]), particularly in Proposition \[prop:connectivity\] below, that the tensor product ${{ \otimes }}_S$ denotes the usual smash product of symmetric spectra [@Hovey_Shipley_Smith 2.2.3]. For notational convenience, in this paper we use the smash product notation $\wedge$ to denote the smash product of ${{ \mathcal{R} }}$-modules (Definition \[defn:left\_R\_modules\]), since the entire paper is written in this context. In particular, in the special case when ${{ \mathcal{R} }}=S$, the two agree ${{ \,\wedge\, }}={{ \otimes }}_S$.
\[prop:connectivity\] Consider symmetric sequences in ${{ \mathcal{R} }}$-modules. Let $m,n\in{{ \mathbb{Z} }}$ and $t\geq 1$. Assume that ${{ \mathcal{R} }}$ is $(-1)$-connected.
- If $X,Y$ are symmetric spectra such that $X$ is $m$-connected and $Y$ is $n$-connected, then $X{{ \otimes }}_S^{{ \mathsf{L} }}Y$ is $(m+n+1)$-connected.
- If $X,Y$ are ${{ \mathcal{R} }}$-modules such that $X$ is $m$-connected and $Y$ is $n$-connected, then $X\wedge^{{ \mathsf{L} }}Y$ is $(m+n+1)$-connected.
- If $X,Y$ are ${{ \mathcal{R} }}$-modules with a right (resp. left) $\Sigma_t$-action such that $X$ is $m$-connected and $Y$ is $n$-connected, then $X\wedge^{{ \mathsf{L} }}_{\Sigma_t} Y$ is $(m+n+1)$-connected.
- If $X,Y$ are symmetric sequences such that $X$ is $m$-connected and $Y$ is $n$-connected, then $X{{ \check{{{ \otimes }}} }}^{{ \mathsf{L} }}Y$ is $(m+n+1)$-connected.
- If $X,Y$ are symmetric sequences with a right (resp. left) $\Sigma_t$-action such that $X$ is $m$-connected and $Y$ is $n$-connected, then $X{{ \check{{{ \otimes }}} }}^{{ \mathsf{L} }}_{\Sigma_t} Y$ is $(m+n+1)$-connected.
Here, ${{ \otimes }}_S^{{ \mathsf{L} }}$, $\wedge^{{ \mathsf{L} }}$, $\wedge^{{ \mathsf{L} }}_{\Sigma_t}$, ${{ \check{{{ \otimes }}} }}^{{ \mathsf{L} }}$, and ${{ \check{{{ \otimes }}} }}^{{ \mathsf{L} }}_{\Sigma_t}$ are the total left derived functors of ${{ \otimes }}_S$, $\wedge$, $\wedge_{\Sigma_t}$, ${{ \check{{{ \otimes }}} }}$, and ${{ \check{{{ \otimes }}} }}_{\Sigma_t}$ respectively.
\[prop:connectivity\_estimates\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyConditionWeaker\]. Let $X$ be a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module) and $k\geq 1$. If ${{ \mathcal{O} }},{{ \mathcal{R} }}$ are $(-1)$-connected and $X$ is $0$-connected, then $|\operatorname{Bar}(\tau_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is $0$-connected and both $|\operatorname{Bar}({{ \mathcal{O} }}^{>k},{{ \mathcal{O} }},X)|$ and $|\operatorname{Bar}(i_{k+1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ are $k$-connected.
This follows from Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] and Propositions \[prop:connectivity\_of\_simplicial\_maps\_spectra\] and \[prop:connectivity\].
The following Milnor type short exact sequences are well known in stable homotopy theory (for a recent reference, see [@Dwyer_Greenlees_Iyengar]); they can be established as a consequence of [@Bousfield_Kan IX].
\[prop:short\_exact\_sequence\] Consider any tower $B_0\leftarrow B_1\leftarrow B_2\leftarrow\cdots$ of symmetric spectra (resp. ${{ \mathcal{R} }}$-modules). There are natural short exact sequences $$\begin{aligned}
&0\rightarrow\lim\nolimits^1_k\pi_{i+1}B_k\rightarrow
\pi_i\operatorname{holim}_k B_k\rightarrow
\lim\nolimits_k\pi_i B_k\rightarrow 0.\end{aligned}$$
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. By Theorem \[thm:comparing\_homotopy\_completion\_towers\] and Propositions \[prop:replacement\_of\_operads\] and \[prop:forgetful\_functor\_commutes\_with\_holim\], we can restrict to operads ${{ \mathcal{O} }}$ satisfying Cofibrancy Condition \[CofibrancyCondition\]. It is enough to treat the following special case. Let $X$ be a $0$-connected, cofibrant left ${{ \mathcal{O} }}$-module. We need to verify that the natural coaugmentation $
X{{ \ \simeq \ }}\operatorname{holim}_k X {{ \longrightarrow }}\operatorname{holim}_k(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)
$ is a weak equivalence. By Proposition \[prop:fattened\_version\_of\_tower\] it suffices to verify that $
\operatorname{holim}_k|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \longrightarrow }}\operatorname{holim}_k|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|
$ is a weak equivalence. Consider the commutative diagram $$\begin{aligned}
\xymatrix{
\pi_i\operatorname{holim}_k|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]_{{{ \ \cong \ }}}\ar[r]^-{(*)} &
\pi_i\operatorname{holim}_k|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[d]^{(*'')}\\
\lim_k\pi_i|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[r]^-{(*')} &
\lim_k\pi_i|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|
}\end{aligned}$$ for each $i$. Since $\lim^1_k\pi_{i+1}|\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)|=0$, the left-hand vertical map is an isomorphism by Proposition \[prop:short\_exact\_sequence\]. We need to show that the map $(*)$ is an isomorphism, hence it suffices to verify that $(*')$ and $(*'')$ are isomorphisms. First note that Propositions \[prop:pushout\_diagram\_for\_bar\_construction\_tower\_coaugmented\] and \[prop:connectivity\_estimates\] imply that $(*')$ is an isomorphism. Similarly, by Propositions \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] and \[prop:connectivity\_estimates\], it follows that for each $k\geq 1$ the induced map $
\pi_i|\operatorname{Bar}(\tau_{k+1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \longrightarrow }}\pi_i|\operatorname{Bar}(\tau_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|
$ is an isomorphism for $i\leq k$ and a surjection for $i={k+1}$; in particular, for each fixed $i$ the tower of abelian groups $\{\pi_i|\operatorname{Bar}(\tau_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\}$ is eventually constant. Hence $
\lim\nolimits^1_k\pi_{i+1}|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|=0
$ and by Proposition \[prop:short\_exact\_sequence\] the map $(*'')$ is an isomorphism which finishes the proof. By the argument above, note that for each $k\geq 1$ the natural maps $
\pi_i X{{ \longrightarrow }}\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)
$ and $
\pi_i(\tau_{k+1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X){{ \longrightarrow }}\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)
$ are isomorphisms for $i\leq k$ and surjections for $i={k+1}$; we sometimes refer to this as the *strong convergence* of the homotopy completion tower.
On $n$-connected maps and the homotopy completion tower
-------------------------------------------------------
The purpose of this subsection is to prove Theorems \[thm:hurewicz\], \[thm:relative\_hurewicz\], and \[MainTheorem\](b).
\[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Cofibrancy Condition \[CofibrancyCondition\]. Let $X$ be a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module) and $k\geq 2$. There are natural weak equivalences $$\begin{aligned}
\label{eq:refined_bar_calculation_hello}
|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}|\operatorname{Bar}(i_k{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|)|.\end{aligned}$$
Below we give a simple conceptual proof of this proposition using derived functors. An anonymous referee has suggested an alternate proof working directly with (bi)simplicial bar constructions, for which the interested reader may jump directly to Remark \[rem:suggested\_proof\_multisimplicial\]. The following proposition is an easy exercise in commuting certain left derived functors and homotopy colimits; we defer the proof to Section \[sec:homotopical\_analysis\_forgetful\_functors\].
\[prop:commuting\_hocolim\_of\_simplicial\_objects\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Let $k\geq 2$. If $B$ is a simplicial $\tau_1{{ \mathcal{O} }}$-algebra (resp. simplicial left $\tau_1{{ \mathcal{O} }}$-module), then there is a zigzag of weak equivalences $$\begin{aligned}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\bigl(\operatorname{hocolim}\limits^{{{ \mathsf{Alg} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B\bigr)
&{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Alg} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}(B) \\
\Bigl(
\text{resp.}\quad
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B
&{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}B
\Bigr)\end{aligned}$$ natural in $B$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. For notational ease, define $B:=|\operatorname{Bar}(\tau_1{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$. By Theorems \[thm:calculating\_fiber\_of\_induced\_map\] and \[thm:fattened\_replacement\], Proposition \[prop:commuting\_hocolim\_of\_simplicial\_objects\], Proposition \[prop:cofibrant\_tau\_1\_O\_algebras\_and\_bar\_constructions\] and Theorem \[main\_hocolim\_theorem\], there are natural weak equivalences $$\begin{aligned}
|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}B
{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}
\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
\operatorname{Bar}(\tau_1{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},B)\\
{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}
\operatorname{Bar}(\tau_1{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},B)
{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}
\operatorname{Bar}(\tau_1{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},B)\\
{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
\operatorname{Bar}(i_k{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},B)
{{ \ \simeq \ }}|\operatorname{Bar}(i_k{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},B)|.\end{aligned}$$
\[rem:suggested\_proof\_multisimplicial\] Here is an alternate proof of Proposition \[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\] that was suggested by an anonymous referee. It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. For notational ease, define $B:=|\operatorname{Bar}(i_k{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }},\tau_1{{ \mathcal{O} }})|$. The right-hand side of is isomorphic to $|\operatorname{Bar}(B,{{ \mathcal{O} }},X)|$ (they are both realizations of a bisimplicial symmetric sequence). Noting that the natural map $B{{ \longrightarrow }}i_k{{ \mathcal{O} }}$ of right $\tau_1{{ \mathcal{O} }}$-modules (and hence of right ${{ \mathcal{O} }}$-modules) is a weak equivalence ([@Harper_Bar 8.4, 8.3]), together with Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] and Proposition \[prop:realization\_monomorphisms\_weak\_equivalences\], it follows that $|\operatorname{Bar}(B,{{ \mathcal{O} }},X)|{{ \longrightarrow }}|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is a weak equivalence, which finishes the proof.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. By Theorem \[thm:comparing\_homotopy\_completion\_towers\] and Propositions \[prop:replacement\_of\_operads\] and \[prop:forgetful\_functor\_commutes\_with\_holim\], we can restrict to operads ${{ \mathcal{O} }}$ satisfying Cofibrancy Condition \[CofibrancyCondition\]. It is enough to treat the special case where $X$ is a cofibrant left ${{ \mathcal{O} }}$-module.
Consider part (a). Assume that $\tau_{1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X$ is $n$-connected. Then $|\operatorname{Bar}(\tau_{1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is $n$-connected by \[prop:fattened\_version\_of\_tower\], hence by Proposition \[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\], together with Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] and Propositions \[prop:connectivity\_of\_simplicial\_maps\_spectra\] and \[prop:connectivity\], it follows that $|\operatorname{Bar}(i_{k+1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is $((k+1)n+k)$-connected for each $k\geq 1$. Hence it follows from \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] and \[prop:fattened\_version\_of\_tower\] that for each $k\geq 1$ the natural maps $
\pi_i(\tau_{k+1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X){{ \longrightarrow }}\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)
$ are isomorphisms for $i\leq (k+1)n+k$ and surjections for $i=(k+1)(n+1)$. In particular, for each $i\leq 2n+1$ the tower $\{\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)\}$ is a tower of isomorphisms, and since $\tau_{1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X$ is $n$-connected, it follows that each stage in the tower $\{\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X\}$ is $n$-connected. Since $X$ is $0$-connected by assumption, it follows from strong convergence of the homotopy completion tower (proof of Theorem \[MainTheorem\](a)) that the map $\pi_i X{{ \longrightarrow }}\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)$ is an isomorphism for every $i\leq k$. Hence taking $k$ sufficiently large ($k\geq n$) verifies that $X$ is $n$-connected.
Conversely, assume that $X$ is $n$-connected. Then by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] and Propositions \[prop:connectivity\_of\_simplicial\_maps\_spectra\] and \[prop:connectivity\], it follows that $|\operatorname{Bar}(\tau_{k}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is $n$-connected and both $|\operatorname{Bar}({{ \mathcal{O} }}^{>k},{{ \mathcal{O} }},X)|$ and $|\operatorname{Bar}(i_{k+1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ are $((k+1)n+k)$-connected for each $k\geq 1$. It follows from \[prop:pushout\_diagram\_for\_bar\_construction\_tower\], \[prop:pushout\_diagram\_for\_bar\_construction\_tower\_coaugmented\], and \[prop:fattened\_version\_of\_tower\] that for each $k\geq 1$ the natural maps $
\pi_i X{{ \longrightarrow }}\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)
$ and $
\pi_i(\tau_{k+1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X){{ \longrightarrow }}\pi_i(\tau_{k}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)
$ are isomorphisms for $i\leq (k+1)n+k$ and surjections for $i=(k+1)(n+1)$. Consequently, $\pi_i X{{ \longrightarrow }}\pi_i(\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)$ is an isomorphism for $i\leq 2n+1$ and a surjection for $i=2n+2$. Since $X$ is $n$-connected, it follows that $\tau_{1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X$ is $n$-connected.
Consider part (b). Assume that $\tau_{1}{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X$ is $n$-connected. Then it follows from the proof of part (a) above that $\pi_i X{{ \longrightarrow }}\pi_i(\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)$ is an isomorphism for $i\leq 2n+1$ and a surjection for $i=2n+2$.
The homotopy completion spectral sequence is the homotopy spectral sequence [@Bousfield_Kan] associated to the tower of fibrations (of fibrant objects) of a fibrant replacement (Definition \[defn:model\_structure\_on\_towers\]) of the homotopy completion tower, reindexed as a (second quadrant) homologically graded spectral sequence. Strong convergence (Remark \[rem:strong\_convergence\]) follows immediately from the first part of the proof of Theorem \[thm:hurewicz\] by taking $n=0$.
We defer the proof of the following to Section \[sec:homotopical\_analysis\_forgetful\_functors\].
\[prop:connectivity\_of\_smash\_powers\_of\_maps\] Consider symmetric sequences in ${{ \mathcal{R} }}$-modules. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ be a map between $(-1)$-connected objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$). Let $m\in{{ \mathbb{Z} }}$, $n\geq -1$, and $t\geq 1$. Assume that ${{ \mathcal{R} }}$ is $(-1)$-connected.
- If $X,Z$ are flat stable cofibrant and $f$ is $n$-connected, then $X^{\wedge t}{{ \longrightarrow }}Z^{\wedge t}$ (resp. $X^{{{ \check{{{ \otimes }}} }}t}{{ \longrightarrow }}Z^{{{ \check{{{ \otimes }}} }}t}$) is $n$-connected.
- If $B\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t^{{ \mathrm{op} }}}$ (resp. $B\in{{ \mathsf{SymSeq} }}^{\Sigma_t^{{ \mathrm{op} }}}$) is $m$-connected, $X,Z$ are positive flat stable cofibrant and $f$ is $n$-connected, then $B{{ \,\wedge\, }}_{\Sigma_t}X^{\wedge t}{{ \longrightarrow }}B{{ \,\wedge\, }}_{\Sigma_t}Z^{\wedge t}$ (resp. $B{{ \check{{{ \otimes }}} }}_{\Sigma_t}X^{{{ \check{{{ \otimes }}} }}t}{{ \longrightarrow }}B{{ \check{{{ \otimes }}} }}_{\Sigma_t}Z^{{{ \check{{{ \otimes }}} }}t}$) is $(m+n+1)$-connected.
\[prop:holim\_and\_n\_connected\_maps\] Let $n\in{{ \mathbb{Z} }}$. If $\{A_k\}{{ \longrightarrow }}\{B_k\}$ is a map of towers in symmetric spectra (resp. ${{ \mathcal{R} }}$-modules) that is levelwise $n$-connected, then the induced map $\operatorname{holim}_k A_k{{ \longrightarrow }}\operatorname{holim}_k B_k$ is $(n-1)$-connected.
This follows from the short exact sequences in Proposition \[prop:short\_exact\_sequence\].
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. By Theorem \[thm:comparing\_homotopy\_completion\_towers\] and Propositions \[prop:replacement\_of\_operads\] and \[prop:forgetful\_functor\_commutes\_with\_holim\], we can restrict to operads ${{ \mathcal{O} }}$ satisfying Cofibrancy Condition \[CofibrancyCondition\].
We first prove part (c), where it is enough to consider the following special case. Let $X{{ \longrightarrow }}Y$ be a map of left ${{ \mathcal{O} }}$-modules between cofibrant objects in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ such that the induced map $\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X{{ \longrightarrow }}\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}Y$ is an $n$-connected map between $(-1)$-connected objects. Consider the corresponding commutative diagram in ${{ \mathsf{SymSeq} }}$. If $k=2$, then the right-hand vertical map is $n$-connected by Proposition \[prop:fattened\_version\_of\_tower\]. It follows from Proposition \[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\], Proposition \[prop:cofibrant\_tau\_1\_O\_algebras\_and\_bar\_constructions\], and Propositions \[prop:connectivity\], \[prop:connectivity\_of\_smash\_powers\_of\_maps\], and \[prop:connectivity\_of\_simplicial\_maps\_spectra\] that the left-hand vertical map is $n$-connected for each $k\geq 2$. Hence by Proposition \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] and induction on $k$, the middle vertical map is $n$-connected for each $k\geq 2$, and Proposition \[prop:holim\_and\_n\_connected\_maps\] finishes the proof of part (b).
Consider part (b). It is enough to consider the following special case. Let $X{{ \longrightarrow }}Y$ be an $(n-1)$-connected map of left ${{ \mathcal{O} }}$-modules between $(-1)$-connected cofibrant objects in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Consider the corresponding commutative diagram in ${{ \mathsf{SymSeq} }}$. It follows from Propositions \[prop:connectivity\], \[prop:connectivity\_of\_smash\_powers\_of\_maps\], and \[prop:connectivity\_of\_simplicial\_maps\_spectra\] that the right-hand vertical map is $(n-1)$-connected for $k=2$, and hence by Proposition \[prop:fattened\_version\_of\_tower\] the induced map $\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X{{ \longrightarrow }}\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}Y$ is $(n-1)$-connected.
Consider part (a). Proceeding as above for part (c), we know that for each $k\geq 1$ the induced map $
\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X {{ \longrightarrow }}\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}Y
$ is $n$-connected, and hence the bottom horizontal map in the $$\begin{aligned}
\xymatrix{
\pi_iX\ar[r]\ar[d] & \pi_iY\ar[d]\\
\pi_i(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)\ar[r] &
\pi_i(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}Y)
}\end{aligned}$$ commutative diagram is an isomorphism for every $i<n$ and a surjection for $i=n$. Since $X,Y$ are $0$-connected by assumption, it follows from strong convergence of the homotopy completion tower (proof of Theorem \[MainTheorem\](a)) that the vertical maps are isomorphisms for $k\geq i$, and hence the top horizontal map is an isomorphism for every $i<n$ and a surjection for $i=n$. Part (b) implies the converse.
Consider part (d). By arguing as in the proof of Theorem \[thm:hurewicz\], it follows that the layers of the homotopy completion tower are $(n-1)$-connected. Hence by Proposition \[prop:short\_exact\_sequence\] the homotopy limit of this tower is $(n-1)$-connected, which finishes the proof.
Finiteness and the homotopy completion tower
--------------------------------------------
The purpose of this subsection is to prove Theorem \[thm:finiteness\]. The following homotopy spectral sequence for a simplicial symmetric spectrum is well known; for a recent reference, see [@EKMM X.2.9] and [@Jardine_generalized_etale 4.3].
\[prop:homotopy\_spectral\_sequence\] Let $Y$ be a simplicial symmetric spectrum. There is a natural homologically graded spectral sequence in the right-half plane such that $$\begin{aligned}
E^2_{p,q} = H_p(\pi_q(Y))\Longrightarrow\pi_{p+q}(|Y|)\end{aligned}$$ Here, $\pi_q(Y)$ denotes the simplicial abelian group obtained by applying $\pi_q$ levelwise to $Y$.
The following finiteness properties for realization will be useful.
\[prop:realzn\_preserves\_finiteness\_properties\] Let $Y$ be a simplicial symmetric spectrum. Let $m\in{{ \mathbb{Z} }}$. Assume that $Y$ is levelwise $m$-connected.
- If $\pi_k Y_n$ is finite for every $k,n$, then $\pi_k |Y|$ is finite for every $k$.
- If $\pi_k Y_n$ is a finitely generated abelian group for every $k,n$, then $\pi_k|Y|$ is a finitely generated abelian group for every $k$.
This follows from Proposition \[prop:homotopy\_spectral\_sequence\].
Recall the following Eilenberg-Moore type spectral sequences; for a recent reference, see [@EKMM IV.4–IV.6].
\[prop:eilenberg\_moore\] Let $t\geq 1$. Let $X,Y$ be ${{ \mathcal{R} }}$-modules with a right (resp. left) $\Sigma_t$-action. There is a natural homologically graded spectral sequence in the right-half plane such that $$\begin{aligned}
E^2_{p,q} &= \operatorname{Tor}^{\pi_*{{ \mathcal{R} }}[\Sigma_t]}_{p,q}(\pi_*X,\pi_*Y)
\Longrightarrow\pi_{p+q}(X\wedge^{{ \mathsf{L} }}_{\Sigma_t} Y).\end{aligned}$$ Here, ${{ \mathcal{R} }}[\Sigma_t]$ is the group algebra spectrum and $\wedge^{{ \mathsf{L} }}_{\Sigma_t}$ is the total left derived functor of $\wedge_{\Sigma_t}$.
The following proposition, which is well known to the experts, will be needed in the proof of Proposition \[prop:useful\_finiteness\_properties\_dervied\_smash\] below; since it is a key ingredient in the proof of Theorem \[thm:finiteness\], and since we are unaware of an appropriate reference in literature, we give a concise homotopy theoretic proof in Section \[sec:homotopical\_analysis\_forgetful\_functors\].
\[prop:finiteness\_derived\_tensor\_uses\_bar\_construction\] Let ${{ \mathcal{A} }}$ be any monoid object in $({{ \mathsf{Ch} }}_{{ \mathbb{Z} }},{{ \otimes }},{{ \mathbb{Z} }})$. Let $M,N$ be unbounded chain complexes over ${{ \mathbb{Z} }}$ with a right (resp. left) action of ${{ \mathcal{A} }}$. Let $m\in{{ \mathbb{Z} }}$. Assume that ${{ \mathcal{A} }}$ is $(-1)$-connected, $M,N$ are $m$-connected, and $H_kM,H_k{{ \mathcal{A} }}$ are finitely generated abelian groups for every $k$.
- If $H_kN$ is finite for every $k$, then $H_k(M{{ \otimes }}^{{ \mathsf{L} }}_{{ \mathcal{A} }}N)$ is finite for every $k$.
- If $H_kN$ is a finitely generated abelian group for every $k$, then $H_k(M{{ \otimes }}^{{ \mathsf{L} }}_{{ \mathcal{A} }}N)$ is a finitely generated abelian group for every $k$.
Here, ${{ \otimes }}^{{ \mathsf{L} }}_{{ \mathcal{A} }}$ is the total left derived functor of ${{ \otimes }}_{{ \mathcal{A} }}$.
\[prop:useful\_finiteness\_properties\_dervied\_smash\] Let $t\geq 1$. Let $X,Y$ be ${{ \mathcal{R} }}$-modules with a right (resp. left) $\Sigma_t$-action. Let $m\in{{ \mathbb{Z} }}$. Assume that ${{ \mathcal{R} }}$ is $(-1)$-connected, $X,Y$ are $m$-connected, and $\pi_k X,\pi_k{{ \mathcal{R} }}$ are finitely generated abelian groups for every $k$.
- If $\pi_k Y$ is finite for every $k$, then $\pi_k(X\wedge^{{ \mathsf{L} }}_{\Sigma_t} Y)$ is finite for every $k$.
- If $\pi_k Y$ is a finitely generated abelian group for every $k$, then $\pi_k(X\wedge^{{ \mathsf{L} }}_{\Sigma_t} Y)$ is a finitely generated abelian group for every $k$.
Here, $\wedge^{{ \mathsf{L} }}_{\Sigma_t}$ is the total left derived functor of $\wedge_{\Sigma_t}$.
Part (a) follows from Propositions \[prop:eilenberg\_moore\] and \[prop:finiteness\_derived\_tensor\_uses\_bar\_construction\], and the proof of part (b) is similar.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. By Theorem \[thm:comparing\_homotopy\_completion\_towers\] and Propositions \[prop:replacement\_of\_operads\] and \[prop:forgetful\_functor\_commutes\_with\_holim\], we can restrict to operads ${{ \mathcal{O} }}$ satisfying Cofibrancy Condition \[CofibrancyCondition\]. We first prove part (a), for which it suffices to consider the following special case. Let $X$ be a cofibrant left ${{ \mathcal{O} }}$-module such that $\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X$ is $0$-connected and $\pi_i(\tau_1{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X)$ is objectwise finite for every $i$. Consider the cofiber sequences $$\begin{aligned}
\xymatrix{
|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[r] &
|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|\ar[r] &
|\operatorname{Bar}(\tau_{k-1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$. We know by Proposition \[prop:fattened\_version\_of\_tower\] that $\pi_i|\operatorname{Bar}(\tau_{1}{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is objectwise finite for every $i$, hence by Proposition \[prop:refined\_bar\_construction\_calculation\_for\_homotopy\_fiber\], Proposition \[prop:cofibrant\_tau\_1\_O\_algebras\_and\_bar\_constructions\], and Propositions \[prop:realzn\_preserves\_finiteness\_properties\] and \[prop:useful\_finiteness\_properties\_dervied\_smash\], $\pi_i|\operatorname{Bar}(i_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is objectwise finite for every $i$. By Proposition \[prop:pushout\_diagram\_for\_bar\_construction\_tower\] and induction on $k$, it follows that $\pi_i|\operatorname{Bar}(\tau_k{{ \mathcal{O} }},{{ \mathcal{O} }},X)|$ is objectwise finite for every $i$ and $k$. Hence by the first part of the proof of Theorem \[thm:hurewicz\] (by taking $n=0$) it follows easily that $\pi_i(X^{{ {{ \mathsf{h} }}\wedge }})$ is objectwise finite for every $i$. If furthermore $X$ is $0$-connected, then by Theorem \[MainTheorem\](a) the natural coaugmentation $X{{ \ \simeq \ }}X^{{ {{ \mathsf{h} }}\wedge }}$ is a weak equivalence which finishes the proof of part (a). The proof of part (b) is similar.
Homotopical analysis of the forgetful functors {#sec:homotopical_analysis_forgetful_functors}
==============================================
The purpose of this section is to prove Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] together with several closely related technical results on the homotopical properties of the forgetful functors. We will also prove Theorem \[thm:rigidification\] and Propositions \[prop:replacement\_of\_operads\], \[prop:connectivity\], \[prop:connectivity\_of\_smash\_powers\_of\_maps\], and \[prop:finiteness\_derived\_tensor\_uses\_bar\_construction\], each of which uses constructions or results established below in Section \[sec:homotopical\_analysis\_forgetful\_functors\]. It will be useful to work in the following context.
From now on in this section we assume that $({{ \mathsf{C} }},{{ \,\wedge\, }},S)$ is a closed symmetric monoidal category with all small limits and colimits. In particular, ${{ \mathsf{C} }}$ has an initial object $\emptyset$ and a terminal object $*$.
In some of the propositions that follow involving homotopical properties of ${{ \mathcal{O} }}$-algebras and left ${{ \mathcal{O} }}$-modules, we will explicitly assume the following.
\[HomotopicalAssumption\] If ${{ \mathcal{O} }}$ is an operad in ${{ \mathsf{C} }}$, assume that
- ${{ \mathsf{C} }}$ is a cofibrantly generated model category in which the generating cofibrations and acyclic cofibrations have small domains [@Schwede_Shipley 2.2], and that with respect to this model structure $({{ \mathsf{C} }},{{ \,\wedge\, }},S)$ is a monoidal model category [@Schwede_Shipley 3.1]; and
- the following model structure exists on ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$): the model structure on ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) has weak equivalences and fibrations created by the forgetful functor $U$ ; i.e., the weak equivalences are the underlying weak equivalences and the fibrations are the underlying fibrations.
The main reason for working in the generality of a monoidal model category $({{ \mathsf{C} }},{{ \,\wedge\, }})$ is because when we start off with arguments using the properties of a particular monoidal model category, say, $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }})$, we are naturally led to need the corresponding results in the diagram category $({{ \mathsf{SymSeq} }},{{ \check{{{ \otimes }}} }})$, and in the diagram category $({{ \mathsf{SymArray} }},{{ \tilde{{{ \otimes }}} }})$ (e.g., Proposition \[prop:cofibrant\_operad\_symmetric\_array\_underlying\]). So working in the generality of a monoidal model category allows us to give a single proof that works for several different contexts. For instance, we also use the results in this section in the contexts of both symmetric spectra and unbounded chain complexes, even when proving the main theorems only in the context of symmetric spectra (e.g., in the proof of Proposition \[prop:finiteness\_derived\_tensor\_uses\_bar\_construction\]).
\[def:symmetric\_array\] Consider symmetric sequences in ${{ \mathsf{C} }}$. A *symmetric array* in ${{ \mathsf{C} }}$ is a symmetric sequence in ${{ \mathsf{SymSeq} }}$; i.e., a functor ${{ {A}\colon\thinspace{\Sigma^{{ \mathrm{op} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$. Denote by ${{ \mathsf{SymArray} }}:={{ \mathsf{SymSeq} }}^{\Sigma^{{ \mathrm{op} }}}$ the category of symmetric arrays in ${{ \mathsf{C} }}$ and their natural transformations.
Recall from [@Harper_Spectra] the following proposition.
\[prop:coproduct\_modules\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$, $A\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $A\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and $Y\in{{ \mathsf{C} }}$ (resp. $Y\in{{ \mathsf{SymSeq} }}$). Consider any coproduct in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) of the form $A\amalg{{ \mathcal{O} }}\circ(Y)$ (resp. $A\amalg({{ \mathcal{O} }}\circ Y)$). There exists a symmetric sequence ${{ \mathcal{O} }}_A$ (resp. symmetric array ${{ \mathcal{O} }}_A$) and natural isomorphisms $$\begin{aligned}
A\amalg{{ \mathcal{O} }}\circ(Y) {{ \ \cong \ }}\coprod\limits_{q\geq 0}{{ \mathcal{O} }}_A[\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_q}Y^{\wedge q}\quad
\Bigl(\text{resp.}\quad
A\amalg({{ \mathcal{O} }}\circ Y) {{ \ \cong \ }}\coprod\limits_{q\geq 0}{{ \mathcal{O} }}_A[\mathbf{q}]
{{ \check{{{ \otimes }}} }}_{\Sigma_q}Y^{{{ \check{{{ \otimes }}} }}q}
\Bigr)\end{aligned}$$ in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$). If $q\geq 0$, then ${{ \mathcal{O} }}_A[\mathbf{q}]$ is naturally isomorphic to a colimit of the form $$\begin{aligned}
{{ \mathcal{O} }}_A[\mathbf{q}]&{{ \ \cong \ }}\operatorname{colim}\biggl(
\xymatrix{
\coprod\limits_{p\geq 0}{{ \mathcal{O} }}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}A^{\wedge p} &
\coprod\limits_{p\geq 0}{{ \mathcal{O} }}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}({{ \mathcal{O} }}\circ (A))^{\wedge p}\ar@<-0.5ex>[l]^-{d_1}
\ar@<-1.5ex>[l]_-{d_0}
}
\biggl),\\
\text{resp.}\quad
{{ \mathcal{O} }}_A[\mathbf{q}]&{{ \ \cong \ }}\operatorname{colim}\biggl(
\xymatrix{
\coprod\limits_{p\geq 0}{{ \mathcal{O} }}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}A^{{{ \check{{{ \otimes }}} }}p} &
\coprod\limits_{p\geq 0}{{ \mathcal{O} }}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}({{ \mathcal{O} }}\circ A)^{{{ \check{{{ \otimes }}} }}p}\ar@<-0.5ex>[l]^-{d_1}
\ar@<-1.5ex>[l]_-{d_0}
}
\biggl),\end{aligned}$$ in ${{ \mathsf{C} }}^{\Sigma_q^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_q^{{ \mathrm{op} }}}$), with $d_0$ induced by operad multiplication and $d_1$ induced by the left ${{ \mathcal{O} }}$-action map ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ (A)}{{ \longrightarrow }}{A} }}$ (resp. ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ A}{{ \longrightarrow }}{A} }}$).
Other possible notations for ${{ \mathcal{O} }}_A$ include $\operatorname{U}_{{ \mathcal{O} }}(A)$ or $\operatorname{U}(A)$; these are closer to the notation used in [@Elmendorf_Mandell; @Mandell] and are not to be confused with the forgetful functors. It is interesting to note—although we will not use it in this paper—that in the context of ${{ \mathcal{O} }}$-algebras the symmetric sequence ${{ \mathcal{O} }}_A$ has the structure of an operad; it parametrizes ${{ \mathcal{O} }}$-algebras under $A$ and is sometimes called the enveloping operad for $A$.
\[prop:OA\_commutes\_with\_certain\_colimits\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$ and let $q\geq 0$. Then the functor $
{{ {{{ \mathcal{O} }}_{(-)}[\mathbf{q}]}\colon\thinspace{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ \mathsf{C} }}^{\Sigma_q^{{ \mathrm{op} }}}} }}
$ (resp. $
{{ {{{ \mathcal{O} }}_{(-)}[\mathbf{q}]}\colon\thinspace{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}^{\Sigma_q^{{ \mathrm{op} }}}} }}
$) preserves reflexive coequalizers and filtered colimits.
This follows from Proposition \[prop:basic\_properties\_LTO\] and [@Harper_Modules 5.7].
\[prop:relating\_the\_OA\_constructions\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$ and $A$ an ${{ \mathcal{O} }}$-algebra. For each $q\geq 0$, ${{ \mathcal{O} }}_{\hat{A}}[\mathbf{q}]$ is concentrated at $0$ with value ${{ \mathcal{O} }}_A[\mathbf{q}]$; i.e., ${{ \mathcal{O} }}_{\hat{A}}[\mathbf{q}]{{ \ \cong \ }}\widehat{{{ \mathcal{O} }}_A[\mathbf{q}]}$.
This follows from Proposition \[prop:coproduct\_modules\], together with and .
\[def:filtration\_setup\_modules\] Let ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a morphism in ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$) and $t\geq 1$. Define $Q_0^t:=X^{\wedge t}$ (resp. $Q_0^t:=X^{{{ \check{{{ \otimes }}} }}t}$) and $Q_t^t:=Y^{\wedge t}$ (resp. $Q_t^t:=Y^{{{ \check{{{ \otimes }}} }}t}$). For $0<q<t$ define $Q_q^t$ inductively by the left-hand (resp. right-hand) pushout diagrams $$\begin{aligned}
\xymatrix{
\Sigma_t\cdot_{\Sigma_{t-q}\times\Sigma_{q}}X^{\wedge(t-q)}
{{ \,\wedge\, }}Q_{q-1}^q\ar[d]^{i_*}\ar[r]^-{{{ \mathrm{pr} }}_*} & Q_{q-1}^t\ar[d]\\
\Sigma_t\cdot_{\Sigma_{t-q}\times\Sigma_{q}}X^{\wedge(t-q)}
{{ \,\wedge\, }}Y^{\wedge q}\ar[r] & Q_q^t
}\quad
\xymatrix{
\Sigma_t\cdot_{\Sigma_{t-q}\times\Sigma_{q}}X^{{{ \check{{{ \otimes }}} }}(t-q)}
{{ \check{{{ \otimes }}} }}Q_{q-1}^q\ar[d]^{i_*}\ar[r]^-{{{ \mathrm{pr} }}_*} & Q_{q-1}^t\ar[d]\\
\Sigma_t\cdot_{\Sigma_{t-q}\times\Sigma_{q}}X^{{{ \check{{{ \otimes }}} }}(t-q)}
{{ \check{{{ \otimes }}} }}Y^{{{ \check{{{ \otimes }}} }}q}\ar[r] & Q_q^t
}\end{aligned}$$ in ${{ \mathsf{C} }}^{\Sigma_t}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_t}$). We sometimes denote $Q_q^t$ by $Q_q^t(i)$ to emphasize in the notation the map ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$. The maps ${{ \mathrm{pr} }}_*$ and $i_*$ are the obvious maps induced by $i$ and the appropriate projection maps.
The following proposition is proved in [@Harper_Spectra] and is closely related to a similar construction in [@Elmendorf_Mandell]; for other approaches to these types of filtrations compare [@Fresse_modules; @Schwede_Shipley].
\[prop:small\_arg\_pushout\_modules\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$, $A\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $A\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$). Consider any pushout diagram in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) of the form $$\begin{aligned}
\label{eq:small_arg_pushout_modules}
\xymatrix{
{{ \mathcal{O} }}\circ (X)\ar[r]^-{f}\ar[d]^{{{ \mathrm{id} }}\circ (i)} & A\ar[d]^{j}\\
{{ \mathcal{O} }}\circ (Y)\ar[r] & B.
}\quad\quad\text{resp.}\quad
\xymatrix{
{{ \mathcal{O} }}\circ X\ar[r]^-{f}\ar[d]^{{{ \mathrm{id} }}\circ i} & A\ar[d]^{j}\\
{{ \mathcal{O} }}\circ Y\ar[r] & B.
}\end{aligned}$$ The pushout in is naturally isomorphic to a filtered colimit of the form $
B{{ \ \cong \ }}\operatorname{colim}\bigl(
\xymatrix@1{
A_0\ar[r]^{j_1} & A_1\ar[r]^{j_2} & A_2\ar[r]^{j_3} & \dotsb
}
\bigr)
$ in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$), with $A_0:={{ \mathcal{O} }}_A[\mathbf{0}]{{ \ \cong \ }}A$ and $A_t$ defined inductively by pushout diagrams in ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$) of the form $$\begin{aligned}
\label{eq:good_filtration_modules}
\xymatrix{
{{ \mathcal{O} }}_A[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t}Q_{t-1}^t\ar[d]^{{{ \mathrm{id} }}\wedge_{\Sigma_t}i_*}
\ar[r]^-{f_*} & A_{t-1}\ar[d]^{j_t}\\
{{ \mathcal{O} }}_A[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t}Y^{\wedge t}\ar[r]^-{\xi_t} & A_t
}\quad\quad\text{resp.}\quad
\xymatrix{
{{ \mathcal{O} }}_A[\mathbf{t}]{{ \check{{{ \otimes }}} }}_{\Sigma_t}Q_{t-1}^t\ar[d]^{{{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*}
\ar[r]^-{f_*} & A_{t-1}\ar[d]^{j_t}\\
{{ \mathcal{O} }}_A[\mathbf{t}]{{ \check{{{ \otimes }}} }}_{\Sigma_t}Y^{{{ \check{{{ \otimes }}} }}t}\ar[r]^-{\xi_t} & A_t
}\end{aligned}$$
We are now in a good position to prove Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\].
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). Let ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a generating cofibration in ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure, and consider the pushout diagram $$\begin{aligned}
\label{eq:gluing_on_cells_proof_of_forgetful}
\xymatrix{
{{ \mathcal{O} }}\circ X\ar[r]\ar[d] & Z_0\ar[d]^{i_0}\\
{{ \mathcal{O} }}\circ Y\ar[r] & Z_1
}\end{aligned}$$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Assume $Z_0$ is cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$; let’s verify that $i_0$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. Let $A:=Z_0$. By Proposition \[prop:small\_arg\_pushout\_modules\], we know $Z_1$ is naturally isomorphic to a filtered colimit of the form $
Z_1{{ \ \cong \ }}\operatorname{colim}\bigl(
\xymatrix@1{
A_0\ar[r]^{j_1} & A_1\ar[r]^{j_2} & A_2\ar[r]^{j_3} & \dotsb
}
\bigr)
$ in the underlying category ${{ \mathsf{SymSeq} }}$, and hence it suffices to verify each $j_t$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. By the construction of $j_t$ in Proposition \[prop:small\_arg\_pushout\_modules\], it is enough to check that each ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ in is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. The generating cofibrations in ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure have cofibrant domains, and by Proposition \[prop:generating\_cofibration\] we know that $i_*$ is a cofibration between cofibrant objects in ${{ \mathsf{SymSeq} }}^{\Sigma_t}$ with the positive flat stable model structure. We need therefore only show that ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$.
Suppose ${{ {p}\colon\thinspace{C}{{ \longrightarrow }}{D} }}$ is a flat stable acyclic fibration in ${{ \mathsf{SymSeq} }}$. We want to verify ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ has the left lifting property with respect to $p$. Consider any such lifting problem; we want to verify that the corresponding solid commutative diagram $$\begin{aligned}
\label{eq:final_lifting_argument}
\xymatrix{
Q_{t-1}^t\ar[d]_{i_*}\ar[r] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}({{ \mathcal{O} }}_A[\mathbf{t}], C)
\ar[d]^{(*)}\\
Y^{{{ \check{{{ \otimes }}} }}t}\ar[r]\ar@{.>}[ur] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}({{ \mathcal{O} }}_A[\mathbf{t}], D)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{\Sigma_t^{{ \mathrm{op} }}}$ has a lift. We know that $i_*$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_t^{{ \mathrm{op} }}}$, hence it is enough to verify that $(*)$ is a flat stable acyclic fibration in ${{ \mathsf{SymSeq} }}$. By Proposition \[prop:analysis\_of\_OA\_symmetric\_spectra\] below, ${{ \mathcal{O} }}_A[\mathbf{t}]$ is flat stable cofibrant in ${{ \mathsf{SymSeq} }}$, hence we know that $(*)$ has the desired property by [@Harper_Modules 6.1], which finishes the argument that $i_0$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. Consider a sequence $
\xymatrix@1{
Z_0\ar[r]^{i_0} & Z_1\ar[r]^{i_1} & Z_2\ar[r]^{i_2} & \dotsb
}
$ of pushouts of maps as in , and let $Z_\infty:=\operatorname{colim}_k Z_k$. Consider the naturally occurring map ${{ {i_\infty}\colon\thinspace{Z_0}{{ \longrightarrow }}{Z_\infty} }}$, and assume $Z_0$ is cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. By the argument above, we know this is a sequence of positive flat stable cofibrations in ${{ \mathsf{SymSeq} }}$, hence $i_\infty$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. Since every cofibration $A{{ \longrightarrow }}B$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ is a retract of a (possibly transfinite) composition of pushouts of maps as in , starting with $Z_0=A$, where $A$ is assumed to be cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, finishes the proof of part (a). Part (b) follows from part (a) by taking $A={{ \mathcal{O} }}\circ\emptyset$, together with the natural isomorphism ${{ \mathcal{O} }}\circ\emptyset{{ \ \cong \ }}\widehat{{{ \mathcal{O} }}[\mathbf{0}]}$.
Homotopical analysis of the ${{ \mathcal{O} }}_A$ constructions
---------------------------------------------------------------
The purpose of this subsection is to prove the following proposition, which we used in the proof of Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\]. It provides a homotopical analysis of the ${{ \mathcal{O} }}_A$ constructions, and a key ingredient in its proof is a filtration of ${{ \mathcal{O} }}_A$ (Proposition \[prop:filtering\_OA\]). We will also prove Proposition \[prop:analysis\_of\_OA\_monoidal\_model\_category\] and Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\_monoidal\], which are analogs of Proposition \[prop:analysis\_of\_OA\_symmetric\_spectra\] and Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\], respectively. These analogous results are applicable to a general class of monoidal model categories, but at the cost of requiring stronger assumptions.
The following proposition is motivated by [@Mandell 13.6].
\[prop:analysis\_of\_OA\_symmetric\_spectra\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$. If $A$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module), then ${{ \mathcal{O} }}_A[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) for each $r\geq 0$.
The following proposition is closely related to [@Mandell 13.6].
\[prop:analysis\_of\_OA\_monoidal\_model\_category\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$. Suppose that Homotopical Assumption \[HomotopicalAssumption\] is satisfied, and assume that ${{ \mathcal{O} }}[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$. If $A$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module), then ${{ \mathcal{O} }}_A[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$.
\[thm:homotopical\_analysis\_of\_forgetful\_functors\_monoidal\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$. Suppose that Homotopical Assumption \[HomotopicalAssumption\] is satisfied, and assume that ${{ \mathcal{O} }}[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration between cofibrant objects in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then $j$ is a cofibration in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$).
- If $A$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module), then $A$ is cofibrant in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$).
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). This follows exactly as in the proof of Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\], except using Proposition \[prop:analysis\_of\_OA\_monoidal\_model\_category\] instead of Proposition \[prop:analysis\_of\_OA\_symmetric\_spectra\], and replacing the lifting problem with a lifting problem of the form $$\begin{aligned}
\xymatrix{
\emptyset\ar[r]\ar[d] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Y^{{{ \check{{{ \otimes }}} }}t}, C)\ar[d]^{(*)}\\
{{ \mathcal{O} }}_A[\mathbf{t}]\ar[r]\ar@{.>}[ur] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Q_{t-1}^t,C)
\times_{\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Q_{t-1}^t,D)}
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Y^{{{ \check{{{ \otimes }}} }}t},D)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{\Sigma_t^{{ \mathrm{op} }}}$. Part (b) follows from part (a) by taking $A={{ \mathcal{O} }}\circ\emptyset$, together with the natural isomorphism ${{ \mathcal{O} }}\circ\emptyset{{ \ \cong \ }}\widehat{{{ \mathcal{O} }}[\mathbf{0}]}$, since ${{ \mathcal{O} }}[\mathbf{0}]$ is cofibrant in ${{ \mathsf{C} }}$.
When working with certain arguments involving left modules over an operad, we are naturally led to replace $({{ \mathsf{C} }},{{ \,\wedge\, }},S)$ with $({{ \mathsf{SymSeq} }},{{ \check{{{ \otimes }}} }},1)$ as the underlying closed symmetric monoidal category. In particular, we will consider symmetric sequences in $({{ \mathsf{SymSeq} }},{{ \check{{{ \otimes }}} }},1)$, i.e., symmetric arrays (Defintion \[def:symmetric\_array\]), together with the corresponding tensor product and circle product. To avoid notational confusion, we will use ${{ \tilde{{{ \otimes }}} }}$ to denote the tensor product of symmetric arrays and ${{ \,\tilde{\circ}\, }}$ to denote the circle product of symmetric arrays. We summarize their structure and properties in the following propositions.
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $A_1,\dotsc,A_t$ and $A,B$ be symmetric arrays in ${{ \mathsf{C} }}$. Then the tensor product $A_1{{ \tilde{{{ \otimes }}} }}\dotsb{{ \tilde{{{ \otimes }}} }}A_t\in{{ \mathsf{SymArray} }}$ and the circle product $A{{ \,\tilde{\circ}\, }}B\in{{ \mathsf{SymArray} }}$ satisfy objectwise the natural isomorphisms $$\begin{aligned}
\label{eq:tensor_tilde_calc}
(A_1{{ \tilde{{{ \otimes }}} }}\dotsb{{ \tilde{{{ \otimes }}} }}A_t)[\mathbf{r}]
&{{ \ \cong \ }}\coprod_{r_1+\dotsb +r_t=r}A_1[\mathbf{r_1}]{{ \check{{{ \otimes }}} }}\dotsb{{ \check{{{ \otimes }}} }}A_t[\mathbf{r_t}]\underset{{\Sigma_{r_1}\times\dotsb\times
\Sigma_{r_t}}}{\cdot}\Sigma_{r},\\
\label{eq:circle_tilde_calc}
(A{{ \,\tilde{\circ}\, }}B)[\mathbf{r}]
&{{ \ \cong \ }}\coprod_{t\geq 0}A[\mathbf{t}]{{ \check{{{ \otimes }}} }}_{\Sigma_t}
(B^{{{ \tilde{{{ \otimes }}} }}t})[\mathbf{r}].\end{aligned}$$
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $Z\in{{ \mathsf{SymSeq} }}$. Define $\tilde{Z}\in{{ \mathsf{SymArray} }}$ to be the symmetric array such that $\tilde{Z}[\mathbf{t}]\in{{ \mathsf{SymSeq} }}^{\Sigma_t^{{ \mathrm{op} }}}$ is concentrated at $0$ with value $Z[\mathbf{t}]$; i.e., $\tilde{Z}[\mathbf{t}]:=\widehat{Z[\mathbf{t}]}$ and hence $\tilde{Z}[\mathbf{t}][\mathbf{0}]=Z[\mathbf{t}]$.
The adjunction immediately below Definition \[defn:hat\_construction\_embed\_at\_zero\] induces objectwise the adjunction $
\xymatrix@1{
\tilde{-}\colon{{ \mathsf{SymSeq} }}\ar@<0.5ex>[r] &
{{ \mathsf{SymArray} }}:{{ \mathrm{Ev} }}_0\ar@<0.5ex>[l]
}
$ with left adjoint on top and ${{ \mathrm{Ev} }}_0$ the functor defined objectwise by ${{ \mathrm{Ev} }}_0(B)[\mathbf{t}]:={{ \mathrm{Ev} }}_0(B[\mathbf{t}])=B[\mathbf{t}][\mathbf{0}]$; i.e., ${{ \mathrm{Ev} }}_0(B)=B[-][\mathbf{0}]$. Note that $\tilde{-}$ embeds ${{ \mathsf{SymSeq} }}$ in ${{ \mathsf{SymArray} }}$ as a full subcategory.
Consider symmetric sequences in ${{ \mathsf{C} }}$. Let ${{ \mathcal{O} }},A,B\in{{ \mathsf{SymSeq} }}$ and $X,Y\in{{ \mathsf{SymArray} }}$. There are natural isomorphisms $$\begin{aligned}
\label{eq:tilde_respects_monoidal_products}
&\widetilde{A{{ \check{{{ \otimes }}} }}B}{{ \ \cong \ }}\tilde{A}{{ \tilde{{{ \otimes }}} }}\tilde{B},
\quad\quad
\widetilde{A\circ B}{{ \ \cong \ }}\tilde{A}{{ \,\tilde{\circ}\, }}\tilde{B},
\quad\quad
{{ \mathrm{Ev} }}_0(\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}Y){{ \ \cong \ }}{{ \mathcal{O} }}\circ{{ \mathrm{Ev} }}_0(Y),\\
\label{eq:evaluate_at_zero_respects_monoidal_products}
&{{ \mathrm{Ev} }}_0(X{{ \tilde{{{ \otimes }}} }}Y){{ \ \cong \ }}{{ \mathrm{Ev} }}_0(X){{ \check{{{ \otimes }}} }}{{ \mathrm{Ev} }}_0(Y),
\quad\quad
{{ \mathrm{Ev} }}_0(X{{ \,\tilde{\circ}\, }}Y){{ \ \cong \ }}{{ \mathrm{Ev} }}_0(X)\circ{{ \mathrm{Ev} }}_0(Y).\end{aligned}$$
Consider symmetric sequences in ${{ \mathsf{C} }}$.
- $({{ \mathsf{SymArray} }},{{ \tilde{{{ \otimes }}} }},\tilde{1})$ is a closed symmetric monoidal category with all small limits and colimits. The unit for ${{ \tilde{{{ \otimes }}} }}$, denoted “$\tilde{1}$”, is the symmetric array concentrated at $0$ with value the symmetric sequence $1$.
- $({{ \mathsf{SymArray} }},{{ \,\tilde{\circ}\, }},\tilde{I})$ is a closed monoidal category with all small limits and colimits. The unit for ${{ \,\tilde{\circ}\, }}$, denoted “$\tilde{I}$”, is the symmetric array concentrated at $1$ with value the symmetric sequence $1$. Circle product is not symmetric.
Since all of the statements and constructions in earlier sections that were previously described in terms of $({{ \mathsf{C} }},{{ \,\wedge\, }},S)$ are equally true for $({{ \mathsf{SymSeq} }},{{ \check{{{ \otimes }}} }},1)$, we will cite and use the appropriate statements and constructions without further comment.
Consider symmetric sequences in ${{ \mathsf{C} }}$.
- If ${{ \mathcal{O} }}$ is an operad in ${{ \mathsf{C} }}$, then $\tilde{{{ \mathcal{O} }}}$ is an operad in ${{ \mathsf{SymSeq} }}$.
- If $A$ is a left ${{ \mathcal{O} }}$-module, then $\tilde{A}$ is a left $\tilde{{{ \mathcal{O} }}}$-module.
- There are adjunctions $$\begin{aligned}
\label{eq:evaluate_at_zero_adjunction}
\xymatrix{
{{ \mathsf{SymArray} }}\ar@<0.5ex>[r]^-{\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}-} & {{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}},\ar@<0.5ex>[l]^-{U}
}\quad
\xymatrix{
{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{\tilde{-}} &
{{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}\ar@<0.5ex>[l]^-{{{ \mathrm{Ev} }}_0},
}\quad
\xymatrix{
{{ \mathsf{Op} }}({{ \mathsf{C} }})\ar@<0.5ex>[r]^-{\tilde{-}} &
{{ \mathsf{Op} }}({{ \mathsf{SymSeq} }})\ar@<0.5ex>[l]^-{{{ \mathrm{Ev} }}_0}
}\end{aligned}$$ with left adjoints on top, $U$ the forgetful functor and ${{ \mathrm{Ev} }}_0$ the functor defined objectwise by ${{ \mathrm{Ev} }}_0(B)[\mathbf{t}]:={{ \mathrm{Ev} }}_0(B[\mathbf{t}])=B[\mathbf{t}][\mathbf{0}]$, i.e., ${{ \mathrm{Ev} }}_0(B)=B[-][\mathbf{0}]$.
Here, ${{ \mathsf{Op} }}({{ \mathsf{C} }})$ denotes the category of operads in ${{ \mathsf{C} }}$, and similarly for ${{ \mathsf{Op} }}({{ \mathsf{SymSeq} }})$.
The following two propositions are exercises left to the reader. They will be needed in the proof of Proposition \[prop:analysis\_of\_OA\_for\_coproducts\] below.
\[prop:tilde\_commutes\_with\_OA\_constructions\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$ and $A$ a left ${{ \mathcal{O} }}$-module. For each $q,r\geq 0$, $\tilde{{{ \mathcal{O} }}}_{\tilde{A}}[\mathbf{q}][\mathbf{r}]$ is concentrated at $0$ with value ${{ \mathcal{O} }}_A[\mathbf{q}][\mathbf{r}]$ (see ); i.e., $\tilde{{{ \mathcal{O} }}}_{\tilde{A}}[\mathbf{q}]{{ \ \cong \ }}\widetilde{{{ \mathcal{O} }}_A[\mathbf{q}]}$.
\[prop:symmetric\_sequence\_decompositions\] Consider symmetric sequences in ${{ \mathsf{C} }}$. Let $B$ be a symmetric sequence (resp. symmetric array) and $r,t\geq 0$. There are natural isomorphisms $$\begin{aligned}
B[\mathbf{t}]{{ \ \cong \ }}\Bigl(
\coprod\limits_{q\geq 0}\widehat{B[\mathbf{q}]}
{{ \check{{{ \otimes }}} }}_{\Sigma_q}I^{{{ \check{{{ \otimes }}} }}q}
\Bigr)[\mathbf{t}]
\quad\quad
\biggl(\text{resp.}\quad
B[\mathbf{t}][\mathbf{r}]{{ \ \cong \ }}\Bigl(
\coprod\limits_{q\geq 0}\widetilde{B[\mathbf{q}]}
{{ \tilde{{{ \otimes }}} }}_{\Sigma_q}\hat{I}^{{{ \tilde{{{ \otimes }}} }}q}
\Bigr)[\mathbf{r}][\mathbf{t}]
\biggr).\end{aligned}$$ Here, $\hat{I}$ is the symmetric array concentrated at $0$ with value $I$.
The following will be needed in the proof of Proposition \[prop:filtering\_OA\] below.
\[prop:analysis\_of\_OA\_for\_coproducts\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$, $A\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $A\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), $Y\in{{ \mathsf{C} }}$ (resp. $Y\in{{ \mathsf{SymSeq} }}$) and $q\geq 0$. Consider any coproduct in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) of the form $A\amalg{{ \mathcal{O} }}\circ (Y)$ (resp. $A\amalg({{ \mathcal{O} }}\circ Y)$). There are natural isomorphisms $$\begin{aligned}
{{ \mathcal{O} }}_{A\amalg{{ \mathcal{O} }}\circ (Y)}[\mathbf{q}] &{{ \ \cong \ }}\coprod\limits_{p\geq 0}{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}Y^{\wedge p},
\quad
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ (Y)}[\mathbf{q}] {{ \ \cong \ }}\coprod\limits_{p\geq 0}{{ \mathcal{O} }}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}Y^{\wedge p}\\
\Bigl(\text{resp.}\quad
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{q}] &{{ \ \cong \ }}\coprod\limits_{p\geq 0}{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \check{{{ \otimes }}} }}_{\Sigma_p}Y^{{{ \check{{{ \otimes }}} }}p},
\quad
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ Y}[\mathbf{q}] {{ \ \cong \ }}\coprod\limits_{p\geq 0}{{ \mathcal{O} }}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}Y^{{{ \check{{{ \otimes }}} }}p}
\Bigr)\end{aligned}$$ in ${{ \mathsf{C} }}^{\Sigma_q^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_q^{{ \mathrm{op} }}}$). In particular, there are natural isomorphisms $$\begin{aligned}
\label{eq:calculating_OA_for_initial_algebras_and_modules}
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ (\emptyset)}[\mathbf{q}]{{ \ \cong \ }}{{ \mathcal{O} }}[\mathbf{q}]
\quad\quad
\Bigl(\text{resp.}\quad
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ \emptyset}[\mathbf{q}]{{ \ \cong \ }}\widehat{{{ \mathcal{O} }}[\mathbf{q}]}
\Bigr)\end{aligned}$$ in ${{ \mathsf{C} }}^{\Sigma_q^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_q^{{ \mathrm{op} }}}$).
Consider the left-hand natural isomorphisms. Since the case for left ${{ \mathcal{O} }}$-modules is more involved, it is useful to consider first the case of ${{ \mathcal{O} }}$-algebras. Let $A$ be an ${{ \mathcal{O} }}$-algebra and $Y\in{{ \mathsf{C} }}$. Let $Z\in{{ \mathsf{SymSeq} }}$, and consider the corresponding left ${{ \mathcal{O} }}$-module $\hat{A}$ and the corresponding symmetric sequence $\hat{Y}$. It follows easily from Proposition \[prop:coproduct\_modules\] and [@Harper_Spectra proof of 4.7] that there are natural isomorphisms $$\begin{aligned}
\label{eq:first_calculation_proof}
\hat{A}\amalg({{ \mathcal{O} }}\circ \hat{Y})\amalg({{ \mathcal{O} }}\circ Z)&{{ \ \cong \ }}\coprod\limits_{q\geq 0}{{ \mathcal{O} }}_{\hat{A}\amalg({{ \mathcal{O} }}\circ \hat{Y})}[\mathbf{q}]
{{ \check{{{ \otimes }}} }}_{\Sigma_q}Z^{{{ \check{{{ \otimes }}} }}q},\\
\label{eq:second_calculation_proof}
\hat{A}\amalg({{ \mathcal{O} }}\circ \hat{Y})\amalg({{ \mathcal{O} }}\circ Z)
&{{ \ \cong \ }}\coprod\limits_{q\geq 0}
\Bigl(
\coprod\limits_{p\geq 0}{{ \mathcal{O} }}_{\hat{A}}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \check{{{ \otimes }}} }}_{\Sigma_p}\hat{Y}^{{{ \check{{{ \otimes }}} }}p}
\Bigr)
{{ \check{{{ \otimes }}} }}_{\Sigma_q}Z^{{{ \check{{{ \otimes }}} }}q},\end{aligned}$$ in the underlying category ${{ \mathsf{SymSeq} }}$. Comparing with and taking $Z=I$, together with Proposition \[prop:relating\_the\_OA\_constructions\] and Proposition \[prop:symmetric\_sequence\_decompositions\], gives a natural isomorphism of symmetric sequences of the form $$\begin{aligned}
{{ \mathcal{O} }}_{A\amalg{{ \mathcal{O} }}\circ(Y)}[\mathbf{q}]{{ \ \cong \ }}\coprod_{p\geq 0}{{ \mathcal{O} }}_{A}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \,\wedge\, }}_{\Sigma_p}Y^{\wedge p},\quad\quad q\geq 0,\end{aligned}$$ which finishes the proof of the left-hand natural isomorphisms for the case of ${{ \mathcal{O} }}$-algebras.
Consider the case of left ${{ \mathcal{O} }}$-modules. Let $A$ be a left ${{ \mathcal{O} }}$-module and $Y\in{{ \mathsf{SymSeq} }}$. Let $Z\in{{ \mathsf{SymArray} }}$ and consider the corresponding operad $\tilde{{{ \mathcal{O} }}}$ in ${{ \mathsf{SymSeq} }}$, the corresponding left $\tilde{{{ \mathcal{O} }}}$-module $\tilde{A}$ and the corresponding symmetric array $\tilde{Y}$. Arguing as above, by Proposition \[prop:coproduct\_modules\] there is a natural isomorphism $$\begin{aligned}
\label{eq:calculation_proof_modules}
\coprod\limits_{q\geq 0}\tilde{{{ \mathcal{O} }}}_{\tilde{A}\amalg(\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}\tilde{Y})}[\mathbf{q}]
{{ \tilde{{{ \otimes }}} }}_{\Sigma_q}Z^{{{ \tilde{{{ \otimes }}} }}q}
{{ \ \cong \ }}\coprod\limits_{q\geq 0}
\Bigl(
\coprod\limits_{p\geq 0}\tilde{{{ \mathcal{O} }}}_{\tilde{A}}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \tilde{{{ \otimes }}} }}_{\Sigma_p}\tilde{Y}^{{{ \tilde{{{ \otimes }}} }}p}
\Bigr)
{{ \tilde{{{ \otimes }}} }}_{\Sigma_q}Z^{{{ \tilde{{{ \otimes }}} }}q},\end{aligned}$$ in the underlying category ${{ \mathsf{SymArray} }}$. By and taking $Z=\hat{I}$, together with Proposition \[prop:tilde\_commutes\_with\_OA\_constructions\] and Proposition \[prop:symmetric\_sequence\_decompositions\], gives a natural isomorphism of symmetric arrays of the form $$\begin{aligned}
\Bigl(
{{ \mathcal{O} }}_{A\amalg{{ \mathcal{O} }}\circ Y}[\mathbf{q}]
\Bigr)[\mathbf{r}]{{ \ \cong \ }}\Bigl(
\coprod\limits_{p\geq 0}{{ \mathcal{O} }}_{A}[\mathbf{p}\boldsymbol{+}\mathbf{q}]
{{ \check{{{ \otimes }}} }}_{\Sigma_p}Y^{{{ \check{{{ \otimes }}} }}p}
\Bigr)[\mathbf{r}],\quad\quad
q,r\geq 0,\end{aligned}$$ which finishes the proof of the left-hand natural isomorphisms for the case of left ${{ \mathcal{O} }}$-modules. The proof of the right-hand natural isomorphisms is similar.
The following filtrations are motivated by [@Mandell 13.7] and generalize the filtered colimit construction of the form $$\begin{aligned}
B{{ \ \cong \ }}{{ \mathcal{O} }}_B[\mathbf{0}]{{ \ \cong \ }}\operatorname{colim}\bigl(
\xymatrix@1{
{{ \mathcal{O} }}_A[\mathbf{0}]\ar[r]^{j_1} & A_1\ar[r]^{j_2} & A_2\ar[r]^{j_3} & \dotsb
}
\bigr)\end{aligned}$$ in Proposition \[prop:small\_arg\_pushout\_modules\] to a filtered colimit construction of ${{ \mathcal{O} }}_B[\mathbf{r}]$ for each $r\geq 0$; for other approaches to these types of filtrations compare [@Fresse_modules; @Schwede_Shipley].
\[prop:filtering\_OA\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$, $A\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $A\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$). Consider any pushout diagram in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) of the form \[eq:small\_arg\_pushout\_modules\]. For each $r\geq 0$, ${{ \mathcal{O} }}_B[\mathbf{r}]$ is naturally isomorphic to a filtered colimit of the form $$\begin{aligned}
\label{eq:filtered_colimit_modules_refined}
{{ \mathcal{O} }}_B[\mathbf{r}]{{ \ \cong \ }}\operatorname{colim}\Bigl(
\xymatrix{
{{ \mathcal{O} }}_A^0[\mathbf{r}]\ar[r]^{j_1} &
{{ \mathcal{O} }}_A^1[\mathbf{r}]\ar[r]^{j_2} &
{{ \mathcal{O} }}_A^2[\mathbf{r}]\ar[r]^{j_3} & \dotsb
}
\Bigr)\end{aligned}$$ in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$), with ${{ \mathcal{O} }}_A^0[\mathbf{r}]:={{ \mathcal{O} }}_A[\mathbf{r}]$ and ${{ \mathcal{O} }}_A^t[\mathbf{r}]$ defined inductively by pushout diagrams in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) of the form $$\begin{aligned}
\label{eq:good_filtration_modules_refined}
\xymatrix{
{{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]{{ \,\wedge\, }}_{\Sigma_t}Q_{t-1}^t\ar[d]^{{{ \mathrm{id} }}\wedge_{\Sigma_t}i_*}
\ar[r]^-{f_*} & {{ \mathcal{O} }}_A^{t-1}[\mathbf{r}]\ar[d]^{j_t}\\
{{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]{{ \,\wedge\, }}_{\Sigma_t}Y^{\wedge t}\ar[r]^-{\xi_t} & {{ \mathcal{O} }}_A^t[\mathbf{r}]
}\quad\quad
\text{resp.}\quad
\xymatrix{
{{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]{{ \check{{{ \otimes }}} }}_{\Sigma_t}Q_{t-1}^t\ar[d]^{{{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*}
\ar[r]^-{f_*} & {{ \mathcal{O} }}_A^{t-1}[\mathbf{r}]\ar[d]^{j_t}\\
{{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]{{ \check{{{ \otimes }}} }}_{\Sigma_t}Y^{{{ \check{{{ \otimes }}} }}t}\ar[r]^-{\xi_t} & {{ \mathcal{O} }}_A^t[\mathbf{r}]
}\end{aligned}$$
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. The argument is a generalization of the proof given in [@Harper_Spectra 4.20] for the case $r=0$, hence it is enough to describe the constructions and arguments needed for future reference and for a reader of [@Harper_Spectra 4.20] to be able to follow the proof. It is easy to verify that the pushout in may be calculated by a reflexive coequalizer in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ of the form $$\begin{aligned}
\label{eq:reflexive_coequalizer_for_desired_pushout}
B
{{ \ \cong \ }}\operatorname{colim}\Bigl(
\xymatrix{
A\amalg({{ \mathcal{O} }}\circ Y)
& A\amalg({{ \mathcal{O} }}\circ X)\amalg({{ \mathcal{O} }}\circ Y)
\ar@<-0.5ex>[l]_-{{{ \overline{{i}} }}}\ar@<0.5ex>[l]^-{{{ \overline{{f}} }}}
}
\Bigr).\end{aligned}$$ The maps ${{ \overline{{i}} }}$ and ${{ \overline{{f}} }}$ are induced by maps ${{ \mathrm{id} }}\circ i_*$ and ${{ \mathrm{id} }}\circ f_*$, which fit into the commutative diagram $$\begin{aligned}
\label{eq:induced_maps_modules}
\xymatrix{
A\amalg\bigl({{ \mathcal{O} }}\circ(X\amalg Y)\bigr)\ar@<-0.5ex>[d]_{{{ \overline{{i}} }}}\ar@<0.5ex>[d]^{{{ \overline{{f}} }}} &
{{ \mathcal{O} }}\circ(A\amalg X\amalg Y)\ar[l]
\ar@<-0.5ex>[d]_{{{ \mathrm{id} }}\circ i_*}\ar@<0.5ex>[d]^{{{ \mathrm{id} }}\circ f_*} &
{{ \mathcal{O} }}\circ\bigl(({{ \mathcal{O} }}\circ A)\amalg X\amalg Y\bigr)\ar@<-0.5ex>[l]_-{d_0}\ar@<0.5ex>[l]^-{d_1}
\ar@<-0.5ex>[d]_{{{ \mathrm{id} }}\circ i_*}\ar@<0.5ex>[d]^{{{ \mathrm{id} }}\circ f_*}\\
A\amalg({{ \mathcal{O} }}\circ Y) & {{ \mathcal{O} }}\circ(A\amalg Y)\ar[l] &
{{ \mathcal{O} }}\circ\bigl(({{ \mathcal{O} }}\circ A)\amalg Y\bigr)
\ar@<-0.5ex>[l]_-{d_0}\ar@<0.5ex>[l]^-{d_1}
}\end{aligned}$$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, with rows reflexive coequalizer diagrams, and maps $i_*$ and $f_*$ in ${{ \mathsf{SymSeq} }}$ induced by ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ and ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{A} }}$ in ${{ \mathsf{SymSeq} }}$. Here we have used the same notation for both $f$ and its adjoint . Applying ${{ \mathcal{O} }}_{(-)}[\mathbf{r}]$ to and , it follows from Proposition \[prop:OA\_commutes\_with\_certain\_colimits\] that ${{ \mathcal{O} }}_{B}[\mathbf{r}]$ may be calculated by a reflexive coequalizer $$\begin{aligned}
\label{eq:reflexive_coequalizer_for_OA}
&{{ \mathcal{O} }}_B[\mathbf{r}]
{{ \ \cong \ }}\operatorname{colim}\Bigl(
\xymatrix{
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{r}]
& {{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ X)\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{r}]
\ar@<-0.5ex>[l]_-{{{ \overline{{i}} }}}\ar@<0.5ex>[l]^-{{{ \overline{{f}} }}}
}
\Bigr)\\
\label{eq:induced_maps_modules_refined}
&\xymatrix{
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ(X\amalg Y))}[\mathbf{r}]\ar@<-0.5ex>[d]_{{{ \overline{{i}} }}}\ar@<0.5ex>[d]^{{{ \overline{{f}} }}} &
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ(A\amalg X\amalg Y)}[\mathbf{r}]\ar[l]
\ar@<-0.5ex>[d]\ar@<0.5ex>[d] &
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ(({{ \mathcal{O} }}\circ A)\amalg X\amalg Y)}[\mathbf{r}]\ar@<-0.5ex>[l]\ar@<0.5ex>[l]
\ar@<-0.5ex>[d]\ar@<0.5ex>[d]\\
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{r}] &
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ(A\amalg Y)}[\mathbf{r}]\ar[l] &
{{ \mathcal{O} }}_{{{ \mathcal{O} }}\circ(({{ \mathcal{O} }}\circ A)\amalg Y)}[\mathbf{r}]
\ar@<-0.5ex>[l]\ar@<0.5ex>[l]
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ of the form , and that the maps ${{ \overline{{i}} }}$ and ${{ \overline{{f}} }}$ in fit into the commutative diagram in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$, with rows reflexive coequalizer diagrams. By , ${{ \mathcal{O} }}_B[\mathbf{r}]$ may be calculated by the colimit of the left-hand column of in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. By and Proposition \[prop:analysis\_of\_OA\_for\_coproducts\], $f$ induces maps ${{ \overline{{f}} }}_{q,p}$ that make the diagrams $$\begin{aligned}
\xymatrix{
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ(X\amalg Y))}[\mathbf{r}]{{ \ \cong \ }}\coprod\limits_{q\geq 0}\coprod\limits_{p\geq 0}
\Bigl(\ \Bigr)\ar[d]^{{{ \overline{{f}} }}} &
\Bigl(
{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}\boldsymbol{+}\mathbf{r}]{{ \check{{{ \otimes }}} }}_{\Sigma_p\times\Sigma_q}
X^{{{ \check{{{ \otimes }}} }}p}{{ \check{{{ \otimes }}} }}Y^{{{ \check{{{ \otimes }}} }}q}
\Bigr)\ar[l]_-{{{ \mathrm{in} }}_{q,p}}\ar@{.>}[d]^{{{ \overline{{f}} }}_{q,p}}\\
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{r}]{{ \ \cong \ }}\coprod\limits_{t\geq 0}\Bigl(\ \Bigr) &
\Bigl(
{{ \mathcal{O} }}_A[\mathbf{q}\boldsymbol{+}\mathbf{r}]{{ \check{{{ \otimes }}} }}_{\Sigma_q}Y^{{{ \check{{{ \otimes }}} }}q}
\Bigr)\ar[l]_-{{{ \mathrm{in} }}_q}
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ commute. Similarly, $i$ induces maps ${{ \overline{{i}} }}_{q,p}$ that make the diagrams $$\begin{aligned}
\xymatrix{
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ(X\amalg Y))}[\mathbf{r}]{{ \ \cong \ }}\coprod\limits_{q\geq 0}\coprod\limits_{p\geq 0}
\Bigl(\ \Bigr)\ar[d]^{{{ \overline{{i}} }}} &
\Bigl(
{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}\boldsymbol{+}\mathbf{r}]{{ \check{{{ \otimes }}} }}_{\Sigma_p\times\Sigma_q}
X^{{{ \check{{{ \otimes }}} }}p}{{ \check{{{ \otimes }}} }}Y^{{{ \check{{{ \otimes }}} }}q}
\Bigr)\ar[l]_-{{{ \mathrm{in} }}_{q,p}}\ar@{.>}[d]^{{{ \overline{{i}} }}_{q,p}}\\
{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{r}]{{ \ \cong \ }}\coprod\limits_{t\geq 0}\Bigl(\ \Bigr) &
\Bigl(
{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}\boldsymbol{+}\mathbf{r}]
{{ \check{{{ \otimes }}} }}_{\Sigma_{p+q}}Y^{{{ \check{{{ \otimes }}} }}(p+q)}
\Bigr)\ar[l]_-{{{ \mathrm{in} }}_{p+q}}
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ commute.
We can now describe more explicitly what it means to give a cone in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ out of the left-hand column of . Let ${{ {\varphi}\colon\thinspace{{{ \mathcal{O} }}_{A\amalg({{ \mathcal{O} }}\circ Y)}[\mathbf{r}]}{{ \longrightarrow }}{\cdot} }}$ be a morphism in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ and define $\varphi_q:=\varphi{{ \mathrm{in} }}_q$. Then $\varphi{{ \overline{{i}} }}=\varphi{{ \overline{{f}} }}$ if and only if the diagrams $$\begin{aligned}
\label{eq:cone_data_refined}
\xymatrix{
{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}\boldsymbol{+}\mathbf{r}]
{{ \check{{{ \otimes }}} }}_{\Sigma_p\times\Sigma_q}X^{{{ \check{{{ \otimes }}} }}p}
{{ \check{{{ \otimes }}} }}Y^{{{ \check{{{ \otimes }}} }}q}\ar[d]^{{{ \overline{{i}} }}_{q,p}}\ar[r]^-{{{ \overline{{f}} }}_{q,p}} &
{{ \mathcal{O} }}_A[\mathbf{q}\boldsymbol{+}\mathbf{r}]{{ \check{{{ \otimes }}} }}_{\Sigma_q}Y^{{{ \check{{{ \otimes }}} }}q}\ar[d]^{\varphi_q}\\
{{ \mathcal{O} }}_A[\mathbf{p}\boldsymbol{+}\mathbf{q}\boldsymbol{+}\mathbf{r}]
{{ \check{{{ \otimes }}} }}_{\Sigma_{p+q}}Y^{{{ \check{{{ \otimes }}} }}(p+q)}
\ar[r]^-{\varphi_{p+q}} & \cdot
}\end{aligned}$$ commute for every $p,q\geq 0$. Since ${{ \overline{{i}} }}_{q,0}={{ \mathrm{id} }}$ and ${{ \overline{{f}} }}_{q,0}={{ \mathrm{id} }}$, it is sufficient to consider $q\geq 0$ and $p>0$.
The next step is to reconstruct the colimit of the left-hand column of in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ via a suitable filtered colimit in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. The diagrams suggest how to proceed. Define ${{ \mathcal{O} }}_A^0[\mathbf{r}]:={{ \mathcal{O} }}_A[\mathbf{r}]$ and for each $t\geq 1$ define ${{ \mathcal{O} }}_A^t[\mathbf{r}]$ by the pushout diagram in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. The maps $f_*$ and $i_*$ are induced by the appropriate maps ${{ \overline{{f}} }}_{q,p}$ and ${{ \overline{{i}} }}_{q,p}$. Arguing exactly as in [@Harper_Spectra proof of 4.20] for the case $r=0$, it is easy to use the diagrams to verify that is satisfied.
The following proposition is the key result used to prove Proposition \[prop:analysis\_of\_OA\_monoidal\_model\_category\].
\[prop:homotopical\_analysis\_of\_OA\_on\_maps\_monoidal\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$. Suppose that Homotopical Assumption \[HomotopicalAssumption\] is satisfied.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) such that ${{ \mathcal{O} }}_A[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$, then ${{ \mathcal{O} }}_A[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_B[\mathbf{r}]$ is a cofibration in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is an acyclic cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) such that ${{ \mathcal{O} }}_A[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$, then ${{ \mathcal{O} }}_A[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_B[\mathbf{r}]$ is an acyclic cofibration in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. We first prove part (a). Let ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a generating cofibration in ${{ \mathsf{SymSeq} }}$, and consider a pushout diagram of the form in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Assume ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]$ is cofibrant in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$; let’s verify that ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]$ is a cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$. Define $A:=Z_0$, and let $r\geq 0$. By \[prop:filtering\_OA\] we know that ${{ \mathcal{O} }}_{Z_1}[\mathbf{r}]$ is naturally isomorphic to a filtered colimit of the form $
{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]{{ \ \cong \ }}\operatorname{colim}\bigl(
\xymatrix@1{
{{ \mathcal{O} }}_A^0[\mathbf{r}]\ar[r]^{j_1} &
{{ \mathcal{O} }}_A^1[\mathbf{r}]\ar[r]^{j_2} &
{{ \mathcal{O} }}_A^2[\mathbf{r}]\ar[r]^{j_3} & \dotsb
}
\bigr)
$ in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$, hence it is enough to verify each $j_t$ is a cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. By the construction of $j_t$ in Proposition \[prop:filtering\_OA\], we need only show that each ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ in is a cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. Suppose ${{ {p}\colon\thinspace{C}{{ \longrightarrow }}{D} }}$ is an acyclic fibration in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. We need to verify that ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ has the left lifting property with respect to $p$. Consider any such lifting problem; we want to verify that the corresponding solid commutative diagram $$\begin{aligned}
\xymatrix{
\emptyset\ar[r]\ar[d] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Y^{{{ \check{{{ \otimes }}} }}t}, C)\ar[d]^{(*)}\\
{{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]\ar[r]\ar@{.>}[ur] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Q_{t-1}^t,C)
\times_{\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Q_{t-1}^t,D)}
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(Y^{{{ \check{{{ \otimes }}} }}t},D)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{(\Sigma_t\times\Sigma_r)^{{ \mathrm{op} }}}$ has a lift. By assumption, ${{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]$ is cofibrant in ${{ \mathsf{SymSeq} }}^{\Sigma_{t+r}^{{ \mathrm{op} }}}$, hence ${{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]$ is cofibrant in ${{ \mathsf{SymSeq} }}^{(\Sigma_t\times\Sigma_r)^{{ \mathrm{op} }}}$, and it is enough to check that $(*)$ is an acyclic fibration in ${{ \mathsf{SymSeq} }}$. We know that $i_*$ is a cofibration in ${{ \mathsf{SymSeq} }}$ by [@Harper_Modules 7.19], hence we know that $(*)$ has the desired property by [@Harper_Modules 6.1], which finishes the argument that ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]$ is a cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$. Consider a sequence $
\xymatrix@1{
Z_0\ar[r] & Z_1\ar[r] & Z_2\ar[r] & \dotsb
}
$ of pushouts of maps as in . Assume ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]$ is cofibrant in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$. Define $Z_\infty:=\operatorname{colim}_k Z_k$, and consider the natural map $Z_0{{ \longrightarrow }}Z_\infty$. We know from above that $
\xymatrix@1{
{{ \mathcal{O} }}_{Z_0}[\mathbf{r}]\ar[r] &
{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]\ar[r] &
{{ \mathcal{O} }}_{Z_2}[\mathbf{r}]\ar[r] & \dotsb
}
$ is a sequence of cofibrations in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$, hence ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_{Z_\infty}[\mathbf{r}]$ is a cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$. Since every cofibration $A{{ \longrightarrow }}B$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ is a retract of a (possibly transfinite) composition of pushouts of maps as in , starting with $Z_0=A$, and ${{ \mathcal{O} }}_A[\mathbf{r}]$ is cofibrant in ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$, the proof of part (a) is complete. The proof of part (b) is similar.
This follows from Proposition \[prop:homotopical\_analysis\_of\_OA\_on\_maps\_monoidal\](a) by taking $A={{ \mathcal{O} }}\circ(\emptyset)$ (resp. $A={{ \mathcal{O} }}\circ\emptyset$), together with and the assumption that ${{ \mathcal{O} }}[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$.
The following proposition is the key result used to prove Proposition \[prop:analysis\_of\_OA\_symmetric\_spectra\].
\[prop:homotopical\_analysis\_of\_OA\_on\_maps\_symmetric\_spectra\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) such that ${{ \mathcal{O} }}_A[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) for each $r\geq 0$, then ${{ \mathcal{O} }}_A[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_B[\mathbf{r}]$ is a positive flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) for each $r\geq 0$.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is an acyclic cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) such that ${{ \mathcal{O} }}_A[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) for each $r\geq 0$, then ${{ \mathcal{O} }}_A[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_B[\mathbf{r}]$ is a positive flat stable acyclic cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) for each $r\geq 0$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). Let ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a generating cofibration in ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure, and consider a pushout diagram of the form in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Assume ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{SymSeq} }}$ for each $r\geq 0$; let’s verify that ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$ for each $r\geq 0$. Define $A:=Z_0$, and let $r\geq 0$. By Proposition \[prop:filtering\_OA\], ${{ \mathcal{O} }}_{Z_1}[\mathbf{r}]$ is naturally isomorphic to a filtered colimit of the form $
{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]{{ \ \cong \ }}\operatorname{colim}\bigl(
\xymatrix@1{
{{ \mathcal{O} }}_A^0[\mathbf{r}]\ar[r]^{j_1} &
{{ \mathcal{O} }}_A^1[\mathbf{r}]\ar[r]^{j_2} &
{{ \mathcal{O} }}_A^2[\mathbf{r}]\ar[r]^{j_3} & \dotsb
}
\bigr)
$ in ${{ \mathsf{SymSeq} }}$, hence it is enough to verify each $j_t$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. By the construction of $j_t$ in Proposition \[prop:filtering\_OA\], we need only check that each ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ in is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. By Proposition \[prop:generating\_cofibration\], $i_*$ is a cofibration between cofibrant objects in ${{ \mathsf{SymSeq} }}^{\Sigma_t}$ with the positive flat stable model structure. It is thus enough to verify that ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$.
Suppose ${{ {p}\colon\thinspace{C}{{ \longrightarrow }}{D} }}$ is a flat stable acyclic fibration in ${{ \mathsf{SymSeq} }}$. We want to show that ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ has the left lifting property with respect to $p$. By assumption ${{ \mathcal{O} }}_A[\mathbf{t}\boldsymbol{+}\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{SymSeq} }}$, hence by exactly the same argument used in the proof of Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\], ${{ \mathrm{id} }}{{ \check{{{ \otimes }}} }}_{\Sigma_t}i_*$ has the left lifting property with respect to $p$, which finishes the argument that ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$ for each $r\geq 0$. Consider a sequence $
\xymatrix@1{
Z_0\ar[r] & Z_1\ar[r] & Z_2\ar[r] & \dotsb
}
$ of pushouts of maps as in , define $Z_\infty:=\operatorname{colim}_k Z_k$, and consider the naturally occurring map $Z_0{{ \longrightarrow }}Z_\infty$. Assume ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{SymSeq} }}$ for each $r\geq 0$. By the argument above we know that $
\xymatrix@1{
{{ \mathcal{O} }}_{Z_0}[\mathbf{r}]\ar[r] &
{{ \mathcal{O} }}_{Z_1}[\mathbf{r}]\ar[r] &
{{ \mathcal{O} }}_{Z_2}[\mathbf{r}]\ar[r] & \dotsb
}
$ is a sequence of positive flat stable cofibrations in ${{ \mathsf{SymSeq} }}$, hence ${{ \mathcal{O} }}_{Z_0}[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_{Z_\infty}[\mathbf{r}]$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$. Noting that every cofibration $A{{ \longrightarrow }}B$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ is a retract of a (possibly transfinite) composition of pushouts of maps as in , starting with $Z_0=A$, together with the assumption that ${{ \mathcal{O} }}_A[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{SymSeq} }}$ for each $r\geq 0$, finishes the proof of part (a). Consider part (b). By arguing exactly as in part (a), except using generating acyclic cofibrations instead of generating cofibrations, it follows that ${{ \mathcal{O} }}_A[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_B[\mathbf{r}]$ is a monomorphism and a weak equivalence in ${{ \mathsf{SymSeq} }}$; for instance, this follows from exactly the same argument used in the proof of Proposition \[prop:homotopical\_analysis\_of\_certain\_pushouts\]. Noting by part (a) that ${{ \mathcal{O} }}_A[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}_B[\mathbf{r}]$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$ finishes the proof.
This follows from Proposition \[prop:homotopical\_analysis\_of\_OA\_on\_maps\_symmetric\_spectra\](a) by taking $A={{ \mathcal{O} }}\circ(\emptyset)$ (resp. $A={{ \mathcal{O} }}\circ\emptyset$), together with and the assumption that ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$.
Homotopical analysis of ${{ \mathcal{O} }}_A$ for cofibrant operads
-------------------------------------------------------------------
The purpose of this subsection is to prove Theorem \[thm:rigidification\]. We will also prove Theorems \[thm:cofibration\_property\_needed\_for\_homology\_completion\], \[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\], and \[thm:forgetful\_functor\_operad\_symmetric\_spectra\_explore\_flat\] (resp. Propositions \[prop:analysis\_of\_OA\_monoidal\_model\_category\_cofibrant\_operad\] and \[prop:analysis\_of\_OA\_monoidal\_model\_category\_cofibrant\_operad\_spectra\]), which are analogs of Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\] (resp. Proposition \[prop:analysis\_of\_OA\_symmetric\_spectra\]). These analogous results, for operads in ${{ \mathcal{R} }}$-modules and operads in a general class of monoidal model categories, require strong assumptions on the (maps of) operads involved, that allow us to replace arguments involving filtrations of ${{ \mathcal{O} }}_A$ with lifting arguments involving maps of endomorphism operads of diagrams.
In the next results, we need to work with operads satisfying good lifting properties, as specified by the definition below.
\[defn:model-cat-operads\] Suppose that ${{ \mathsf{C} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\](i). A morphism of operads in ${{ \mathsf{C} }}$ is a *fibration* (resp. *weak equivalence*) of operads if the underlying morphism of symmetric sequences is a fibration (resp. weak equivalence) in the corresponding projective model stucture on ${{ \mathsf{SymSeq} }}$. A *cofibration* of operads in ${{ \mathsf{C} }}$ is a morphism of operads that satisfies the left lifting property with respect to all fibrations of operads that are weak equivalences. An operad ${{ \mathcal{O} }}$ in ${{ \mathsf{C} }}$ is *cofibrant* if the unique map from the initial operad to ${{ \mathcal{O} }}$ is a cofibration of operads.
While we have found it convenient to use model category terminology in the definition above, none of the results in this paper require a model structure to exist on the category of operads in ${{ \mathsf{C} }}$, and we will not establish one in this paper. The following proposition was used in Subsection \[subsec:TQ\_completion\].
\[prop:functorial\_factorizations\_of\_maps\_of\_operads\] Let ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a map of operads in ${{ \mathsf{C} }}$. Suppose that ${{ \mathsf{C} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\](i). Then $f$ has a functorial factorization in the category of operads as ${{ \mathcal{O} }}\xrightarrow{g}J\xrightarrow{h}{{ \mathcal{O} }}'$, a cofibration followed by a weak equivalence which is also a fibration (Definition \[defn:model-cat-operads\]).
Consider symmetric sequences in ${{ \mathsf{C} }}$. Since ${{ \mathsf{C} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\](i), it is easy to verify, using the corresponding adjunctions $(G_p,{{ \mathrm{Ev} }}_p)$ in , that the diagram category ${{ \mathsf{SymSeq} }}$ also satisfies Homotopical Assumption \[HomotopicalAssumption\](i). Consider the free-forgetful adjunction $
\xymatrix@1{
F\colon{{ \mathsf{SymSeq} }}\ar@<0.5ex>[r] &
{{ \mathsf{Op} }}\colon U\ar@<0.5ex>[l]
}
$ with left adjoint on top and $U$ the forgetful functor; here, ${{ \mathsf{Op} }}$ denotes the category of operads. It is easy to verify that the functor $F$ can be constructed by a filtered colimit of the form $$\begin{aligned}
F(A){{ \ \cong \ }}\operatorname{colim}\bigl(I\rightarrow I\amalg A\rightarrow
I\amalg A\circ(I\amalg A)\rightarrow
I\amalg A\circ(I\amalg A\circ(I\amalg A))\rightarrow\dotsc\bigr)\end{aligned}$$ in the underlying category ${{ \mathsf{SymSeq} }}$; this useful description appears in [@Rezk]. Since the forgetful functor $U$ commutes with filtered colimits, it follows from [@Schwede_Shipley Remark 2.4] that the smallness conditions required in [@Schwede_Shipley Lemma 2.3] are satisfied, and the (possibly transfinite) small object argument described in the proof of [@Schwede_Shipley Lemma 2.3] finishes the proof.
The following theorem is motivated by [@Rezk 4.1.14].
\[thm:cofibration\_property\_needed\_for\_homology\_completion\] Let ${{ {g}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a cofibration of operads in ${{ \mathsf{C} }}$. Suppose that ${{ \mathcal{O} }},{{ \mathcal{O} }}'$ and ${{ \mathsf{C} }}$ satisfy Homotopical Assumption \[HomotopicalAssumption\].
- If ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ is a cofibration in ${{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}$ (resp. ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$), and $X$ is cofibrant in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$), then $i$ is a cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
- If the forgetful functor ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}{{ \longrightarrow }}{{ \mathsf{C} }}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}{{ \longrightarrow }}{{ \mathsf{SymSeq} }}$) preserves cofibrant objects, and $Y$ is a cofibrant ${{ \mathcal{O} }}'$-algebra (resp. cofibrant left ${{ \mathcal{O} }}'$-module), then $Y$ is cofibrant in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
It suffices to consider the case of left ${{ \mathcal{O} }}'$-modules. Consider part (b). Let $Y$ be a cofibrant left ${{ \mathcal{O} }}'$-module. The map $\emptyset{{ \longrightarrow }}Y$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ factors functorially in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ as $\emptyset\rightarrow X\xrightarrow{p}Y$ a cofibration followed by an acyclic fibration; here, $\emptyset$ denotes an initial object in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. We first want to show there exists a left ${{ \mathcal{O} }}'$-module structure on $X$ such that $p$ is a map in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$. Consider the solid commutative diagram $$\begin{aligned}
\xymatrix{
{{ \mathcal{O} }}\ar[d]_{g}\ar[r] &
\operatorname{End}(X\xrightarrow{p}Y)
\ar[d]^{(*)}\ar[r]^-{(**)}
& \operatorname{Map}^\circ(X,X)\ar[d]^{({{ \mathrm{id} }},p)}\\
{{ \mathcal{O} }}'\ar[r]^-{m}\ar@{.>}[ur]^-{{{ \overline{{m}} }}} &
\operatorname{Map}^\circ(Y,Y)\ar[r]^{(p,{{ \mathrm{id} }})} & \operatorname{Map}^\circ(X,Y)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$ such that the right-hand square is a pullback diagram. It is easy to verify that the maps $(*)$ and $(**)$ are morphisms of operads. By assumption, $X$ is cofibrant in ${{ \mathsf{SymSeq} }}$, hence we know that $({{ \mathrm{id} }},p)$ is an acyclic fibration by [@Harper_Modules 6.2], and therefore $(*)$ is an acyclic fibration in ${{ \mathsf{SymSeq} }}$. Since $g$ is a cofibration of operads, there exists a morphism of operads ${{ \overline{{m}} }}$ that makes the diagram commute. It follows that the composition $
{{ \mathcal{O} }}'\xrightarrow{{{ \overline{{m}} }}}\operatorname{End}(X\xrightarrow{p}Y)
\xrightarrow{(**)}\operatorname{Map}^\circ(X,X)
$ of operad maps determines a left ${{ \mathcal{O} }}'$-module structure on $X$ such that $p$ is a morphism of left ${{ \mathcal{O} }}'$-modules. To finish the proof, we need to show that $Y$ is cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Consider the solid commutative diagram $$\begin{aligned}
\xymatrix{
\emptyset\ar[d]\ar[r]& X\ar[d]^{p}\\
Y\ar@{.>}[ur]^-{\xi}\ar@{=}[r] & Y
}\end{aligned}$$ in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$, where $\emptyset$ denotes an initial object in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$. Since $Y$ is cofibrant in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$, and $p$ is an acyclic fibration, this diagram has a lift $\xi$ in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$. In particular, $Y$ is a retract of $X$ in ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$, and hence in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Noting that $X$ is cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ finishes the proof of part (b). Part (a) can be established exactly as in the proof of Theorem \[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\](a), by replacing the map $I{{ \longrightarrow }}{{ \mathcal{O} }}$ with the map ${{ \mathcal{O} }}{{ \longrightarrow }}{{ \mathcal{O} }}'$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Since $X$ is cofibrant in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ and $g_*$ is a left Quillen functor, $g_*(X)$ is cofibrant in ${{ \mathsf{Lt} }}_{J_1}$ and hence by \[thm:comparing\_homotopy\_categories\] and \[prop:unit\_map\_is\_weak\_equivalence\] it follows that $g^*g_*(X){{ \ \simeq \ }}{{ \mathsf{TQ} }}(X)$. To iterate the argument, it suffices to verify that the right Quillen functor $g^*$ preserves cofibrant objects: this follows from Theorem \[thm:cofibration\_property\_needed\_for\_homology\_completion\] and Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\].
The following theorem is closely related to [@Rezk 4.1.15].
\[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\] Let ${{ \mathcal{O} }}$ be a cofibrant operad in ${{ \mathsf{C} }}$. Suppose that Homotopical Assumption \[HomotopicalAssumption\] is satisfied.
- If ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ is a cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and $X$ is cofibrant in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$), then $i$ is a cofibration in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$).
- If $Y$ is a cofibrant ${{ \mathcal{O} }}$-algebra (resp. cofibrant left ${{ \mathcal{O} }}$-module), then $Y$ is cofibrant in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$).
- If the unit $S$ is cofibrant in ${{ \mathsf{C} }}$, then ${{ \mathcal{O} }}[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$.
The proof of this result is very similar to that of the previous theorem. It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). Let ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ be a cofibration in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. The map $i$ factors functorially in the underlying category ${{ \mathsf{SymSeq} }}$ as $X\xrightarrow{j}Y\xrightarrow{p}Z$, a cofibration followed by an acyclic fibration. We want first to show there exists a left ${{ \mathcal{O} }}$-module structure on $Y$ such that $j$ and $p$ are maps in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Consider the solid commutative diagram $$\begin{aligned}
\xymatrix{
I\ar[d]\ar[r] &
\operatorname{End}(X\xrightarrow{j}Y\xrightarrow{p}Z)\ar[d]^{(*)}\ar[r]^{(**)} &
\operatorname{Map}^\circ(Y,Y)\ar[d]^{(j,p)}\\
{{ \mathcal{O} }}\ar[r]^-{m}\ar@{.>}[ur]^(0.4){{{ \overline{{m}} }}} & \operatorname{End}(X\xrightarrow{i}Z)\ar[r] &
\operatorname{Map}^\circ(X,Y)\times_{\operatorname{Map}^\circ(X,Z)}\operatorname{Map}^\circ(Y,Z)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$ such that the right-hand square is a pullback diagram. It is easy to verify that the maps $(*)$ and $(**)$ are morphisms of operads. By assumption, $X$ is cofibrant in ${{ \mathsf{SymSeq} }}$, hence we know that the pullback corner map $(j,p)$ is an acyclic fibration by [@Harper_Modules 6.2], and therefore $(*)$ is an acyclic fibration in ${{ \mathsf{SymSeq} }}$. Since ${{ \mathcal{O} }}$ is a cofibrant operad, the map $I{{ \longrightarrow }}{{ \mathcal{O} }}$ is a cofibration of operads, and there exists a morphism of operads ${{ \overline{{m}} }}$ that makes the diagram commute. It follows that the composition $
{{ \mathcal{O} }}\xrightarrow{{{ \overline{{m}} }}}\operatorname{End}(X\xrightarrow{j}Y\xrightarrow{p}Z)
\xrightarrow{(**)}\operatorname{Map}^\circ(Y,Y)
$ of operad maps determines a left ${{ \mathcal{O} }}$-module structure on $Y$ such that $j$ and $p$ are morphisms of left ${{ \mathcal{O} }}$-modules. To finish the proof, we need to show that $i$ is a cofibration in ${{ \mathsf{SymSeq} }}$. Consider the solid commutative diagram $$\begin{aligned}
\xymatrix{
X\ar[d]_{i}\ar[r]^{j} & Y\ar[d]^{p}\\
Z\ar@{=}[r]\ar@{.>}[ur]^{\xi} & Z
}\end{aligned}$$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Since $i$ is a cofibration and $p$ is an acyclic fibration in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, the diagram has a lift $\xi$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. In particular, $i$ is a retract of $j$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and hence in the underlying category ${{ \mathsf{SymSeq} }}$. Noting that $j$ is a cofibration in ${{ \mathsf{SymSeq} }}$ finishes the proof of part (a). Part (b) follows immediately from [@Harper_Bar proof of 10.2], which uses a similar argument; it is also a special case of Theorem \[thm:cofibration\_property\_needed\_for\_homology\_completion\](b). Consider part (c). By assumption, the unit $S$ is cofibrant in ${{ \mathsf{C} }}$, hence the map $\emptyset{{ \longrightarrow }}I$ is a cofibration in ${{ \mathsf{SymSeq} }}$ and therefore ${{ \mathcal{O} }}\circ\emptyset{{ \longrightarrow }}{{ \mathcal{O} }}\circ I$ is a cofibration in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Hence ${{ \mathcal{O} }}{{ \ \cong \ }}{{ \mathcal{O} }}\circ I$ is a cofibrant left ${{ \mathcal{O} }}$-module, and part (b) finishes the proof.
\[thm:forgetful\_functor\_operad\_symmetric\_spectra\_explore\_flat\] Let ${{ \mathcal{O} }}$ be a cofibrant operad in ${{ \mathcal{R} }}$-modules with respect to the positive flat stable model structure.
- ${{ \mathcal{O} }}[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$.
- If ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ is a cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and $X$ is flat stable cofibrant in the underlying category ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), then $i$ is a flat stable cofibration in the underlying category ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$).
Since every flat stable fibration in ${{ \mathsf{SymSeq} }}$ is a positive flat stable fibration in ${{ \mathsf{SymSeq} }}$, it follows that ${{ \mathcal{O} }}$ is also a cofibrant operad in ${{ \mathcal{R} }}$-modules with respect to the flat stable model structure. The proof of Theorem \[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\] finishes the argument.
\[prop:pushout\_description\_of\_OA\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{C} }}$ and $A\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $A\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$). Consider the pushout diagram in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}$) of the form $$\begin{aligned}
\label{eq:pushout_description_of_OA}
\xymatrix{
{{ \mathcal{O} }}\circ \emptyset\ar[r]\ar[d] &
\hat{A}\ar[d]^{j}\\
{{ \mathcal{O} }}\circ I\ar[r] & \hat{A}\amalg({{ \mathcal{O} }}\circ I)
}\quad\quad\text{resp.}\quad
\xymatrix{
\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}\emptyset\ar[r]\ar[d] &
\tilde{A}\ar[d]^{j}\\
\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}\hat{I}\ar[r] &
\tilde{A}\amalg(\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}\hat{I})
}\end{aligned}$$ There are natural isomorphisms $
{{ \mathcal{O} }}_A[\mathbf{t}]{{ \ \cong \ }}\bigl(\hat{A}\amalg({{ \mathcal{O} }}\circ I)\bigr)[\mathbf{t}]
$ (resp. $
{{ \mathcal{O} }}_A[\mathbf{t}][\mathbf{r}]{{ \ \cong \ }}\bigl(\tilde{A}\amalg(\tilde{{{ \mathcal{O} }}}{{ \,\tilde{\circ}\, }}\hat{I})
\bigr)[\mathbf{r}][\mathbf{t}]
$) for each $r,t\geq 0$. Here, $\hat{I}$ is the symmetric array concentrated at $0$ with value $I$.
This follows from Propositions \[prop:coproduct\_modules\], \[prop:relating\_the\_OA\_constructions\], \[prop:tilde\_commutes\_with\_OA\_constructions\], and \[prop:symmetric\_sequence\_decompositions\].
\[prop:cofibrant\_operad\_symmetric\_array\_underlying\] Let ${{ \mathcal{O} }}$ be a cofibrant operad in ${{ \mathsf{C} }}$. Suppose that ${{ \mathcal{O} }},\tilde{{{ \mathcal{O} }}}$ and ${{ \mathsf{C} }}$ satisfy Homotopical Assumption \[HomotopicalAssumption\]. If ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ is a cofibration in ${{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}$ such that $X$ is cofibrant in the underlying category ${{ \mathsf{SymArray} }}$, then $i$ is a cofibration in the underlying category ${{ \mathsf{SymArray} }}$.
This proof is similar to that of Theorem \[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\], except for the following variation on the lifting argument. Let ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Z} }}$ be a cofibration in ${{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}$. The map $i$ factors functorially in the underlying category ${{ \mathsf{SymArray} }}$ as $X\xrightarrow{j}Y\xrightarrow{p}Z$, a cofibration followed by an acyclic fibration. We need to show there exists a left $\tilde{{{ \mathcal{O} }}}$-module structure on $Y$ such that $j$ and $p$ are maps in ${{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}$. Consider the solid diagram $$\begin{aligned}
\xymatrix{
& \operatorname{End}(X\xrightarrow{j}Y\xrightarrow{p}Z)\ar[d]^{(*)}\ar[r]^{(**)} &
\operatorname{Map}^{{ \,\tilde{\circ}\, }}(Y,Y)\ar[d]^{(j,p)}\\
\tilde{{{ \mathcal{O} }}}\ar[r]^-{m}\ar@{.>}[ur]^(0.4){{{ \overline{{m}} }}} & \operatorname{End}(X\xrightarrow{i}Z)\ar[r] &
\operatorname{Map}^{{ \,\tilde{\circ}\, }}(X,Y)\times_{\operatorname{Map}^{{ \,\tilde{\circ}\, }}(X,Z)}\operatorname{Map}^{{ \,\tilde{\circ}\, }}(Y,Z)
}\end{aligned}$$ in ${{ \mathsf{SymArray} }}$, such that the square is a pullback diagram. It is easy to verify that the maps $(*)$ and $(**)$ are morphisms of operads. Since $X$ is cofibrant in ${{ \mathsf{SymArray} }}$, the pullback corner map $(j,p)$ is an acyclic fibration in ${{ \mathsf{SymArray} }}$ by [@Harper_Modules 6.2], and therefore $(*)$ is as well. We need to show there exists a map of operads ${{ \overline{{m}} }}$ that makes the diagram commute. By the right-hand adjunction in , it is enough to show there exists a map ${{ \overline{{m}} }}$ of operads in ${{ \mathsf{C} }}$ that makes the corresponding diagram $$\begin{aligned}
\xymatrix{
& {{ \mathrm{Ev} }}_0\bigl(\operatorname{End}(X\xrightarrow{j}Y\xrightarrow{p}Z)\bigr)
\ar[d]^{{{ \mathrm{Ev} }}_0(*)} \\
{{ \mathcal{O} }}\ar[r]^-{m}\ar@{.>}[ur]^(0.4){{{ \overline{{m}} }}} &
{{ \mathrm{Ev} }}_0\bigl(\operatorname{End}(X\xrightarrow{i}Z)\bigr)
}\end{aligned}$$ of operads in ${{ \mathsf{C} }}$ commute. Since ${{ \mathcal{O} }}$ is a cofibrant operad in ${{ \mathsf{C} }}$, the desired lift ${{ \overline{{m}} }}$ exists. It follows that the composition $(**){{ \overline{{m}} }}$ of operad maps determines a left $\tilde{{{ \mathcal{O} }}}$-module structure on $Y$ such that $j$ and $p$ are morphisms of left $\tilde{{{ \mathcal{O} }}}$-modules. To finish the proof, we need to show that $i$ is a cofibration in ${{ \mathsf{SymArray} }}$, which follows exactly as in the proof of Theorem \[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\].
\[prop:analysis\_of\_OA\_monoidal\_model\_category\_cofibrant\_operad\] Let ${{ \mathcal{O} }}$ be a cofibrant operad in ${{ \mathsf{C} }}$. Suppose that ${{ \mathcal{O} }},\tilde{{{ \mathcal{O} }}}$ and ${{ \mathsf{C} }}$ satisfy Homotopical Assumption \[HomotopicalAssumption\]. If the unit $S$ is cofibrant in ${{ \mathsf{C} }}$, and $A$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) that is cofibrant in the underlying category ${{ \mathsf{C} }}$ (resp. ${{ \mathsf{SymSeq} }}$), then ${{ \mathcal{O} }}_A[\mathbf{r}]$ is cofibrant in ${{ \mathsf{C} }}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$.
This follows from Proposition \[prop:pushout\_description\_of\_OA\], Theorem \[thm:forgetful\_functor\_cofibrant\_operad\_lifting\_argument\], and Proposition \[prop:cofibrant\_operad\_symmetric\_array\_underlying\].
\[prop:analysis\_of\_OA\_monoidal\_model\_category\_cofibrant\_operad\_spectra\] Let ${{ \mathcal{O} }}$ be a cofibrant operad in ${{ \mathcal{R} }}$-modules with respect to the positive flat stable model structure. If $A$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) that is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), then ${{ \mathcal{O} }}_A[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_r^{{ \mathrm{op} }}}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_r^{{ \mathrm{op} }}}$) for each $r\geq 0$.
Since every flat stable fibration in ${{ \mathsf{SymSeq} }}$ is a positive flat stable fibration in ${{ \mathsf{SymSeq} }}$, it follows that ${{ \mathcal{O} }}$ is also a cofibrant operad in ${{ \mathcal{R} }}$-modules with respect to the flat stable model structure. The proof of Proposition \[prop:analysis\_of\_OA\_monoidal\_model\_category\_cofibrant\_operad\] finishes the argument.
Proofs
------
The purpose of this short subsection is to prove Propositions \[prop:replacement\_of\_operads\], \[prop:connectivity\], \[prop:commuting\_hocolim\_of\_simplicial\_objects\], \[prop:connectivity\_of\_smash\_powers\_of\_maps\], and \[prop:finiteness\_derived\_tensor\_uses\_bar\_construction\].
This follows from a small object argument together with an analysis of the functor $F$ appearing in $
\xymatrix@1{
F\colon{{ \mathsf{SymSeq} }}\ar@<0.5ex>[r] &
{{ \mathsf{Op} }}\colon U\ar@<0.5ex>[l]
}
$ the free-forgetful adjunction with left adjoint on top and $U$ the forgetful functor; here, ${{ \mathsf{Op} }}$ denotes the category of operads. It is easy to verify that the functor $F$ can be constructed by a filtered colimit of the form $$\begin{aligned}
F(A){{ \ \cong \ }}\operatorname{colim}\bigl(I\rightarrow I\amalg A\rightarrow
I\amalg A\circ(I\amalg A)\rightarrow
I\amalg A\circ(I\amalg A\circ(I\amalg A))\rightarrow\dotsc\bigr)\end{aligned}$$ in the underlying category ${{ \mathsf{SymSeq} }}$; this useful description appears in [@Rezk]. Using this description of $F$, it is easy to verify that the unit map $I{{ \longrightarrow }}{{ \mathcal{O} }}'$ of the operad ${{ \mathcal{O} }}'$ constructed in the small object argument satisfies the desired property in Cofibrancy Condition \[CofibrancyCondition\].
For a recent reference of part (a) in the context of symmetric spectra, see [@Schwede_book_project]. Consider part (b). It is enough to treat the special case where $X,Y$ are furthermore fibrant and cofibrant in the category of ${{ \mathcal{R} }}$-modules with the flat stable model structure. Let ${{ \mathcal{R} }}'{{ \longrightarrow }}{{ \mathcal{R} }}$ be a cofibrant replacement in the category of monoids in $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ with the flat stable model structure [@Harper_Modules; @Schwede_Shipley]. Since the sphere spectrum $S$ is flat stable cofibrant in ${{ \mathsf{Sp}^\Sigma }}$, we know by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\_monoidal\](a) that ${{ \mathcal{R} }}'$ is flat stable cofibrant in the underlying category ${{ \mathsf{Sp}^\Sigma }}$, and it follows from [@Harper_Spectra; @Harper_Bar] by arguing as in the proof of Theorem \[thm:comparing\_homotopy\_completion\_towers\] that there are natural weak equivalences $
X{{ \,\wedge\, }}^{{ \mathsf{L} }}Y=
X({{ \otimes }}_S)^{{ \mathsf{L} }}_{{ \mathcal{R} }}Y{{ \ \simeq \ }}X'({{ \otimes }}_S)^{{ \mathsf{L} }}_{{{ \mathcal{R} }}'} Y'{{ \ \simeq \ }}|\operatorname{Bar}^{{{ \otimes }}_S}(X',{{ \mathcal{R} }}',Y')|=|B|
$. Here, $X'{{ \longrightarrow }}X$ and $Y'{{ \longrightarrow }}Y$ are functorial flat stable cofibrant replacements in the category of right (resp. left) ${{ \mathcal{R} }}'$-modules. Denote by $B$ the indicated simplicial bar construction with respect to ${{ \otimes }}_S$. We need to verify that $|B|$ is $(m+n+1)$-connected. We know by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\_monoidal\](b) that $X',Y'$ are flat stable cofibrant in the underlying category ${{ \mathsf{Sp}^\Sigma }}$, hence it follows from part (a) that $B$ is objectwise $(m+n+1)$-connected and Proposition \[prop:connectivity\_of\_simplicial\_maps\_spectra\] finishes the proof for part (b). Part (c) is verified exactly as in the proof of part (b) above, except using the group algebra spectrum ${{ \mathcal{R} }}[\Sigma_t]$ instead of ${{ \mathcal{R} }}$. Part (d) follows easily from part (b) together with . Part (e) follows easily from parts (d) and (c) together with .
It suffices to consider the case of simplicial left $\tau_1{{ \mathcal{O} }}$-modules. Consider the map $\emptyset{{ \longrightarrow }}B$ in ${{ \mathsf{sLt} }}_{\tau_1{{ \mathcal{O} }}}$, and use functorial factorization in ${{ \mathsf{sLt} }}_{\tau_1{{ \mathcal{O} }}}$ [@Harper_Bar 3.6] to obtain $\emptyset{{ \longrightarrow }}B^c{{ \longrightarrow }}B$, a cofibration followed by an acyclic fibration. By Proposition \[prop:retract\_property\_and\_derived\_circle\] and [@Harper_Bar 5.6], there is a retract of the form $$\begin{aligned}
\xymatrix{
|i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c|\ar[d]^{(*)}\ar[r] &
|{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c|\ar[d]^{(**)}\ar[r] &
|i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c|\ar[d]^{(*)}\\
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}\operatorname{colim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B^c\ar[r] &
{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}\operatorname{colim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B^c\ar[r] &
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}\operatorname{colim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B^c
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$. Since $B^c$ is cofibrant in ${{ \mathsf{sLt} }}_{\tau_1{{ \mathcal{O} }}}$, the proof of [@Harper_Bar 3.15] implies that ${{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c$ is cofibrant in ${{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}$. It follows therefore from [@Harper_Bar 5.24] that $(**)$ is a weak equivalence, hence $(*)$ is also a weak equivalence. We know from [@Harper_Bar 3.12] that $B^c$ is objectwise cofibrant in ${{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}$, hence there are natural weak equivalences $
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}B^c{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}B.
$ It follows that there are natural weak equivalences $$\begin{aligned}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B^c{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}\operatorname{colim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B^c\\
{{ \ \simeq \ }}i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}\operatorname{colim}\limits^{{{ \mathsf{Lt} }}_{\tau_1{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}B^c
{{ \ \simeq \ }}|i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c|
{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ_{\tau_1{{ \mathcal{O} }}}B^c \\
{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}B^c
{{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}}_{\Delta^{{ \mathrm{op} }}}
i_k{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{\tau_1{{ \mathcal{O} }}}B\end{aligned}$$ which finishes the proof; here we have used \[main\_hocolim\_theorem\].
Consider part (a) and the case of ${{ \mathcal{R} }}$-modules. The map $f$ factors functorially in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ with the flat stable model structure as $X\xrightarrow{g}Y\xrightarrow{h}Z$ a cofibration followed by an acyclic fibration, and hence the map $f^{\wedge t}$ factors as $
X^{\wedge t}\xrightarrow{g^{\wedge t}}Y^{\wedge t}
\xrightarrow{h^{\wedge t}}Z^{\wedge t}
$. Since smashing with a flat stable cofibrant ${{ \mathcal{R} }}$-module preserves weak equivalences, $h^{\wedge t}$ is a weak equivalence, and hence it is enough to check that $g^{\wedge t}$ is $n$-connected. We argue by induction on $t$. Using the pushout diagrams in Definition \[def:filtration\_setup\_modules\] (see, for instance, [@Harper_Spectra 4.15]) together with the natural isomorphisms $Y^{\wedge t}/Q^t_{t-1}{{ \ \cong \ }}(Y/X)^{\wedge t}$, it follows that each of the maps $
X^{\wedge t}{{ \longrightarrow }}Q^t_{1}{{ \longrightarrow }}Q^t_2{{ \longrightarrow }}\dotsb{{ \longrightarrow }}Q^t_{t-1}{{ \longrightarrow }}Y^{\wedge t}
$ is at least $n$-connected, which finishes the proof for the case of ${{ \mathcal{R} }}$-modules. The case of symmetric sequences is similar. Consider part (b). This follows by proceeding as in the proof of part (a), except using the positive flat stable model structure, together with part (a) and Propositions \[prop:cofibration\], \[prop:good\_properties\], \[prop:cofibrations\_to\_mono\], and \[prop:connectivity\].
Propositions \[prop:homotopy\_spectral\_sequence\_chain\_complexes\], \[prop:realzn\_preserves\_finiteness\_properties\_chain\_complexes\], and \[prop:eilenberg\_moore\_chain\_complexes\] will be needed for the proof of Proposition \[prop:finiteness\_derived\_tensor\_uses\_bar\_construction\] below. The following homotopy spectral sequence for a simplicial unbounded chain complex is well known; for a recent reference, see [@Weibel 5.6].
\[prop:homotopy\_spectral\_sequence\_chain\_complexes\] Let $Y$ be a simplicial unbounded chain complex over ${{ {\mathcal{K}} }}$. There is a natural homologically graded spectral sequence in the right-half plane such that $$\begin{aligned}
E^2_{p,q} = H_p(H_q(Y))\Longrightarrow H_{p+q}(|Y|)\end{aligned}$$ Here, $H_q(Y)$ denotes the simplicial ${{ {\mathcal{K}} }}$-module obtained by applying $H_q$ levelwise to $Y$, and ${{ {\mathcal{K}} }}$ is any commutative ring.
\[prop:realzn\_preserves\_finiteness\_properties\_chain\_complexes\] Let $Y$ be a simplicial unbounded chain complex over ${{ \mathbb{Z} }}$. Let $m\in{{ \mathbb{Z} }}$. Assume that $Y$ is levelwise $m$-connected.
- If $H_k Y_n$ is finite for every $k,n$, then $H_k |Y|$ is finite for every $k$.
- If $H_k Y_n$ is a finitely generated abelian group for every $k,n$, then $H_k|Y|$ is a finitely generated abelian group for every $k$.
This follows from Proposition \[prop:homotopy\_spectral\_sequence\_chain\_complexes\].
Recall the following Eilenberg-Moore type spectral sequences; for a recent reference, see [@Weibel 5.7].
\[prop:eilenberg\_moore\_chain\_complexes\] Let $t\geq 1$. Let $A,B$ be unbounded chain complexes over ${{ {\mathcal{K}} }}$ with a right (resp. left) $\Sigma_t$-action. There is a natural homologically graded spectral sequence in the right-half plane such that $$\begin{aligned}
E^2_{p,q} &= \operatorname{Tor}^{{{ {\mathcal{K}} }}[\Sigma_t]}_{p,q}(H_*A,H_*B)
\Longrightarrow H_{p+q}(A{{ \otimes }}^{{ \mathsf{L} }}_{\Sigma_t} B).\end{aligned}$$ Here, ${{ {\mathcal{K}} }}$ is any commutative ring, $({{ \mathsf{Ch} }}_{{ {\mathcal{K}} }},{{ \otimes }},{{ {\mathcal{K}} }})$ denotes the closed symmetric monoidal category of unbounded chain complexes over ${{ {\mathcal{K}} }}$, ${{ {\mathcal{K}} }}[\Sigma_t]$ is the group algebra, and ${{ \otimes }}^{{ \mathsf{L} }}_{\Sigma_t}$ is the total left derived functor of ${{ \otimes }}_{\Sigma_t}$.
Consider part (a). It is enough to treat the special case where $M,N$ are furthermore cofibrant in the category of right (resp. left) ${{ \mathcal{A} }}$-modules. Let ${{ \mathcal{A} }}'{{ \longrightarrow }}{{ \mathcal{A} }}$ be a cofibrant replacement in the category of monoids in $({{ \mathsf{Ch} }}_{{ \mathbb{Z} }},{{ \otimes }},{{ \mathbb{Z} }})$ with the model structure of [@Schwede_Shipley]. Since ${{ \mathbb{Z} }}$ is cofibrant in ${{ \mathsf{Ch} }}_{{ \mathbb{Z} }}$, we know by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\_monoidal\](a) that ${{ \mathcal{A} }}'$ is cofibrant in the underlying category ${{ \mathsf{Ch} }}_{{ \mathbb{Z} }}$, and it follows easily by arguing as in the proof of Theorem \[thm:comparing\_homotopy\_completion\_towers\] that there are natural weak equivalences $
M{{ \otimes }}^{{ \mathsf{L} }}_{{ \mathcal{A} }}N{{ \ \simeq \ }}M'{{ \otimes }}^{{ \mathsf{L} }}_{{{ \mathcal{A} }}'}N'
{{ \ \simeq \ }}|\operatorname{Bar}^{{ \otimes }}(M',{{ \mathcal{A} }}',N')|=|B|.
$ Here, $M'{{ \longrightarrow }}M$ and $N'{{ \longrightarrow }}N$ are functorial cofibrant replacements in the category of right (resp. left) ${{ \mathcal{A} }}'$-modules. Denote by $B$ the indicated simplicial bar construction with respect to ${{ \otimes }}$. We need to verify that $H_k(|B|)$ is finite for every $k$. We know by Theorem \[thm:homotopical\_analysis\_of\_forgetful\_functors\_monoidal\](b) that $M',N'$ are cofibrant in the underlying category ${{ \mathsf{Ch} }}_{{ \mathbb{Z} }}$, hence it follows from Proposition \[prop:eilenberg\_moore\_chain\_complexes\] (with $t=1$) that $H_k(B_n)$ is finite for every $k$ and $n$, and Proposition \[prop:realzn\_preserves\_finiteness\_properties\_chain\_complexes\] finishes the proof for part (a). Part (b) is similar.
Homotopical analysis of the simplicial bar constructions {#sec:homotopical_analysis_bar_constructions}
========================================================
The purpose of this section is to prove Theorem \[thm:reedy\_cofibrant\_for\_bar\_constructions\] together with several closely related technical results on simplicial structures and the simplicial bar constructions. The results established here lie at the heart of the proofs of the main theorems in this paper.
Simplicial structure on ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ and ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$
------------------------------------------------------------------------------------------------------------------
The purpose of this subsection is to describe the simplicial structure on ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) and to prove several related results. The key technical results of this subsection are Proposition \[prop:realizations\_are\_isomorphic\] and Theorem \[thm:simplicial\_model\_category\_structure\]. They are used in the proof of Theorem \[thm:reedy\_cofibrant\_for\_bar\_constructions\] to construct skeletal filtrations in ${{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}$ (resp. ${{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}$) of realizations (Definition \[defn:realization\]) of the simplicial bar constructions (Proposition \[prop:natural\_map\_is\_weak\_equivalence\]).
Consider symmetric sequences in ${{ \mathcal{R} }}$-modules, and let ${{ \mathcal{O} }}\in{{ \mathsf{SymSeq} }}$, $X$ in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), and $K\in{{ \mathsf{S} }}$. Define $\nu$ to be the natural map $$\begin{aligned}
{{ \mathcal{O} }}\circ(X){{ \,\wedge\, }}K_+\xrightarrow{\ \nu\ }{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+)
\quad\quad
\Bigl(\text{resp.}\quad
({{ \mathcal{O} }}\circ X){{ \,\wedge\, }}K_+\xrightarrow{\ \nu\ }{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+)
\Bigr)\end{aligned}$$ in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) induced by the natural maps $K{{ \longrightarrow }}K^{\times t}$ in ${{ \mathsf{S} }}$ for $t\geq 0$; these are the diagonal maps for $t\geq 1$ and the constant map for $t=0$. Here, ${{ \mathsf{S} }}$ denotes the category of simplicial sets. The construction of the tensor product below is motivated by [@EKMM VII.2.10]. Simplicial structures in the context of symmetric spectra have also been exploited in [@Hornbostel; @Schwede_book_project]; see also [@Arone_Ching; @McClure_Schwanzl_Vogt].
\[defn:tensor\_with\_simplicial\_sets\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules, $X$ an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), and $K$ a simplicial set. Define the *tensor product* $X{{ \dot{{{ \otimes }}} }}K$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) by the reflexive coequalizer $$\begin{aligned}
\label{eq:tensor_with_simplicial_sets_algebras}
X{{ \dot{{{ \otimes }}} }}K &:=\operatorname{colim}\Bigl(
\xymatrix{
{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+) &
{{ \mathcal{O} }}\circ\bigl({{ \mathcal{O} }}\circ(X){{ \,\wedge\, }}K_+\bigr)
\ar@<0.5ex>[l]^-{d_1}\ar@<-0.5ex>[l]_-{d_0}
}
\Bigr)\\
\label{eq:tensor_with_simplicial_sets_modules}
\Bigl(\text{resp.}\quad
X{{ \dot{{{ \otimes }}} }}K &:=\operatorname{colim}\Bigl(
\xymatrix{
{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+) &
{{ \mathcal{O} }}\circ\bigl(({{ \mathcal{O} }}\circ X){{ \,\wedge\, }}K_+\bigr)
\ar@<0.5ex>[l]^-{d_1}\ar@<-0.5ex>[l]_-{d_0}
}
\Bigr)
\Bigr)\end{aligned}$$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), with $d_0$ induced by operad multiplication ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ and the map $\nu$, while $d_1$ is induced by the left ${{ \mathcal{O} }}$-action map ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ(X)}{{ \longrightarrow }}{X} }}$ (resp. ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ X}{{ \longrightarrow }}{X} }}$).
Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules, consider $X,Y$ in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), $K\in{{ \mathsf{S} }}$, and recall the isomorphisms $$\begin{aligned}
\label{eq:adjunction_smash_with_simplicial_set_spectra}
\hom_{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}}(X{{ \,\wedge\, }}K_+,Y)&{{ \ \cong \ }}\hom_{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}}(X,\operatorname{Map}(K_+,Y))\\
\label{eq:adjunction_smash_with_simplicial_set_symmetric_sequences}
\Bigl(\text{resp.}\quad
\hom_{{{ \mathsf{SymSeq} }}}(X{{ \,\wedge\, }}K_+,Y)&{{ \ \cong \ }}\hom_{{{ \mathsf{SymSeq} }}}(X,\operatorname{Map}(K_+,Y))
\Bigr)\end{aligned}$$ natural in $X,K,Y$. Here, we are using the useful shorthand notation $\operatorname{Map}(K_+,-)$ to denote $\operatorname{Map}({{ \mathcal{R} }}{{ \otimes }}G_0K_+,-)$; see, just above \[defn:realization\]. If $Y$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), then $\operatorname{Map}(K_+,Y)$ has an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) structure induced by ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ(Y)}{{ \longrightarrow }}{Y} }}$ (resp. ${{ {m}\colon\thinspace{{{ \mathcal{O} }}\circ Y}{{ \longrightarrow }}{Y} }}$). The next proposition is a formal argument left to the reader. We will use it below in several proofs.
\[prop:useful\_commutative\_diagram\_involving\_nu\_map\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Let $X\in{{ \mathsf{SymSeq} }}$, $Y\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and $K\in{{ \mathsf{S} }}$. If ${{ {f}\colon\thinspace{X{{ \,\wedge\, }}K_+}{{ \longrightarrow }}{Y} }}$ is a map in ${{ \mathsf{SymSeq} }}$, then the diagram $$\begin{aligned}
\xymatrix{
({{ \mathcal{O} }}\circ X){{ \,\wedge\, }}K_+\ar[d]_{\nu}\ar[r]^-{{{ \mathrm{id} }}\circ f{{ \,\wedge\, }}{{ \mathrm{id} }}} &
\bigl({{ \mathcal{O} }}\circ\operatorname{Map}(K_+,Y)\bigr){{ \,\wedge\, }}K_+\ar[r]^-{\nu} &
{{ \mathcal{O} }}\circ\bigl(\operatorname{Map}(K_+,Y){{ \,\wedge\, }}K_+\bigr)\ar[d]^{{{ \mathrm{id} }}\circ{{ \mathrm{ev} }}} \\
{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+)\ar[rr]^-{{{ \mathrm{id} }}\circ f} && {{ \mathcal{O} }}\circ Y
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$ commutes. Here, ${{ \mathrm{ev} }}$ denotes the evaluation map, and we have used the same notation for both $f$ and its adjoint .
The following proposition will be useful.
\[prop:natural\_isomorphisms\_tensordot\_and\_mapping\_object\_algebras\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Let $X,Y$ be ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) and $K$ a simplicial set. There are isomorphisms $$\begin{aligned}
\hom_{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}(X{{ \dot{{{ \otimes }}} }}K,Y)&{{ \ \cong \ }}\hom_{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}(X,\operatorname{Map}(K_+,Y))\\
\Bigl(\text{resp.}\quad
\hom_{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}(X{{ \dot{{{ \otimes }}} }}K,Y)&{{ \ \cong \ }}\hom_{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}(X,\operatorname{Map}(K_+,Y))
\Bigr)\end{aligned}$$ natural in $X,K,Y$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. We need to verify that specifying a map $X{{ \dot{{{ \otimes }}} }}K{{ \longrightarrow }}Y$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ is the same as specifying a map $X{{ \longrightarrow }}\operatorname{Map}(K_+,Y)$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$, and that the resulting correspondence is natural. Suppose ${{ {f}\colon\thinspace{X{{ \dot{{{ \otimes }}} }}K}{{ \longrightarrow }}{Y} }}$ is a map of left ${{ \mathcal{O} }}$-modules, and consider the corresponding commutative diagram $$\begin{aligned}
\label{eq:induced_map_on_reflexive_coequalizers_tensor_with_simplicial_set}
\xymatrix{
X{{ \dot{{{ \otimes }}} }}K\ar[d]_{f} &
{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+)\ar[d]\ar[l]\ar[dl]_{f} &
{{ \mathcal{O} }}\circ\bigl(({{ \mathcal{O} }}\circ X){{ \,\wedge\, }}K_+\bigr)
\ar[d]\ar@<0.5ex>[l]^-{d_1}\ar@<-0.5ex>[l]_-{d_0}\\
Y & {{ \mathcal{O} }}\circ Y\ar[l]_-{m} &
{{ \mathcal{O} }}\circ{{ \mathcal{O} }}\circ Y\ar@<0.5ex>[l]^-{{{ \mathrm{id} }}\circ m}
\ar@<-0.5ex>[l]_-{m\circ{{ \mathrm{id} }}}
}\end{aligned}$$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ with rows reflexive coequalizer diagrams. Using the same notation for both ${{ {f}\colon\thinspace{{{ \mathcal{O} }}\circ(X{{ \,\wedge\, }}K_+)}{{ \longrightarrow }}{Y} }}$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ and its adjoints ${{ {f}\colon\thinspace{X{{ \,\wedge\, }}K_+}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{SymSeq} }}$ and ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{\operatorname{Map}(K_+,Y)} }}$ in ${{ \mathsf{SymSeq} }}$ , it follows easily from and Proposition \[prop:useful\_commutative\_diagram\_involving\_nu\_map\] that the diagram $$\begin{aligned}
\xymatrix{
({{ \mathcal{O} }}\circ X){{ \,\wedge\, }}K_+\ar[d]_{{{ \mathrm{id} }}\circ f{{ \,\wedge\, }}{{ \mathrm{id} }}}
\ar[r]^-{m{{ \,\wedge\, }}{{ \mathrm{id} }}} &
X{{ \,\wedge\, }}K_+\ar[r]^-{f{{ \,\wedge\, }}{{ \mathrm{id} }}} &
\operatorname{Map}(K_+,Y){{ \,\wedge\, }}K_+\ar[dd]^{{{ \mathrm{ev} }}}\\
\bigl({{ \mathcal{O} }}\circ\operatorname{Map}(K_+,Y)\bigr){{ \,\wedge\, }}K_+\ar[d]_{\nu}\\
{{ \mathcal{O} }}\circ\bigl(\operatorname{Map}(K_+,Y){{ \,\wedge\, }}K_+\bigr)\ar[r]^-{{{ \mathrm{id} }}\circ{{ \mathrm{ev} }}} &
{{ \mathcal{O} }}\circ Y\ar[r]^-{m} & Y
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$ commutes, which implies that ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{\operatorname{Map}(K_+,Y)} }}$ is a map of left ${{ \mathcal{O} }}$-modules. Conversely, suppose ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{\operatorname{Map}(K_+,Y)} }}$ is a map of left ${{ \mathcal{O} }}$-modules, and consider the corresponding map ${{ {f}\colon\thinspace{X{{ \,\wedge\, }}K_+}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{SymSeq} }}$. We need to verify that the adjoint map ${{ {f}\colon\thinspace{{{ \mathcal{O} }}\circ (X{{ \,\wedge\, }}K_+)}{{ \longrightarrow }}{Y} }}$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ induces a map ${{ {f}\colon\thinspace{X{{ \dot{{{ \otimes }}} }}K}{{ \longrightarrow }}{Y} }}$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. Applying ${{ \mathcal{O} }}\circ-$ to the commutative diagram in Proposition \[prop:useful\_commutative\_diagram\_involving\_nu\_map\], it follows that $fd_0=fd_1$, which finishes the proof.
\[defn:realization\_algebras\_and\_modules\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. The *realization* functors ${{ {|-|_{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}\colon\thinspace{{{ {{ \mathsf{sAlg} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}} }}$ and ${{ {|-|_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}\colon\thinspace{{{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}}{{ \longrightarrow }}{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} }}$ for simplicial ${{ \mathcal{O} }}$-algebras and simplicial left ${{ \mathcal{O} }}$-modules are defined objectwise by the coends $$\begin{aligned}
X\longmapsto |X|_{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}:=X{{ \dot{{{ \otimes }}} }}_{\Delta}\Delta[-]_+\ ,
\quad\quad
X\longmapsto |X|_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}:=X{{ \dot{{{ \otimes }}} }}_{\Delta}\Delta[-]_+\ .\end{aligned}$$
Recall that the realization functors $|-|$ in Definition \[defn:realization\] are the left adjoints in the adjunctions with right adjoints the functors $\operatorname{Map}(\Delta[-]_+,-)$. The following proposition is closely related to [@EKMM VII.3.3]; see also [@Arone_Ching A].
\[prop:realizations\_are\_isomorphic\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules and $X$ a simplicial ${{ \mathcal{O} }}$-algebra (resp. simplicial left ${{ \mathcal{O} }}$-module). The realization functors fit into adjunctions $$\begin{aligned}
\label{eq:realization_adjunction_algebras_modules_nice}
&\xymatrix{
{{ {{ \mathsf{sAlg} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{|-|_{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}} &
{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]
}\quad\quad
\xymatrix{
{{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{|-|_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} &
{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]
}\\
\label{eq:underlying_realization_adjunction_algebras_modules_nice}
&\xymatrix{
{{ {{ \mathsf{sAlg} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{|-|} &
{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]
}\quad\quad
\xymatrix{
{{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}\ar@<0.5ex>[r]^-{|-|} &
{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}},\ar@<0.5ex>[l]
}\end{aligned}$$ with left adjoints on top and right adjoints the functors $\operatorname{Map}(\Delta[-]_+,-)$. In particular, there are isomorphisms $|X|{{ \ \cong \ }}|X|_{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $|X|{{ \ \cong \ }}|X|_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), natural in $X$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Let $X$ be a simplicial left ${{ \mathcal{O} }}$-module. Verifying follows easily from \[prop:natural\_isomorphisms\_tensordot\_and\_mapping\_object\_algebras\] and the universal property of coends. Consider . Suppose ${{ {f}\colon\thinspace{|X|}{{ \longrightarrow }}{Y} }}$ is a map of left ${{ \mathcal{O} }}$-modules, and consider the corresponding left-hand commutative diagram $$\begin{aligned}
\xymatrix{
{{ \mathcal{O} }}\circ|X|{{ \ \cong \ }}|{{ \mathcal{O} }}\circ X|\ar@<5.0ex>[d]_-{{{ \mathrm{id} }}\circ f}\ar[r]^-{|m|} &
|X|\ar[d]^{f}\\
\quad\quad\quad\quad\quad{{ \mathcal{O} }}\circ Y\ar[r]^-{m} & Y
}\quad\quad
\xymatrix{
{{ \mathcal{O} }}\circ X\ar[d]_{(*)}\ar[r]^-{m} & X\ar[d]^{f}\\
\operatorname{Map}(\Delta[-]_+,{{ \mathcal{O} }}\circ Y)\ar[r]^-{({{ \mathrm{id} }},m)} & \operatorname{Map}(\Delta[-]_+,Y)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$. Using the same notation for both ${{ {f}\colon\thinspace{|X|}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{SymSeq} }}$ and its adjoint ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{\operatorname{Map}(\Delta[-]_+,Y)} }}$ in ${{ \mathsf{sSymSeq} }}$ , we know by that the left-hand diagram commutes if and only if its corresponding right-hand diagram in ${{ \mathsf{sSymSeq} }}$ commutes. Since the map $(*)$ factors in ${{ \mathsf{sSymSeq} }}$ as $
{{ \mathcal{O} }}\circ X\xrightarrow{{{ \mathrm{id} }}\circ f}{{ \mathcal{O} }}\circ\operatorname{Map}(\Delta[-]_+,Y)
{{ \longrightarrow }}\operatorname{Map}(\Delta[-]_+,{{ \mathcal{O} }}\circ Y)
$, the proof is complete.
\[prop:properties\_for\_establishing\_simplicial\_category\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Let $X,Y$ be ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) and $K,L$ simplicial sets. Then
- the functor ${{ {X{{ \dot{{{ \otimes }}} }}-}\colon\thinspace{{{ \mathsf{S} }}}{{ \longrightarrow }}{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}} }}$ (resp. ${{ {X{{ \dot{{{ \otimes }}} }}-}\colon\thinspace{{{ \mathsf{S} }}}{{ \longrightarrow }}{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}} }}$) commutes with all colimits and there are natural isomorphisms $X{{ \dot{{{ \otimes }}} }}*{{ \ \cong \ }}X$,
- there are isomorphisms $X{{ \dot{{{ \otimes }}} }}(K\times L){{ \ \cong \ }}(X{{ \dot{{{ \otimes }}} }}K){{ \dot{{{ \otimes }}} }}L$, natural in $X,K,L$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Part (a) follows easily from and . Part (b) follows easily from the Yoneda lemma by verifying there are natural isomorphisms $
\hom_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}\bigl((X{{ \dot{{{ \otimes }}} }}K){{ \dot{{{ \otimes }}} }}L,Y\bigr)
{{ \ \cong \ }}\hom_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}\bigl(X{{ \dot{{{ \otimes }}} }}(K\times L),Y\bigr)
$; this involves several applications of Proposition \[prop:natural\_isomorphisms\_tensordot\_and\_mapping\_object\_algebras\], together with the observation that the natural isomorphism $\operatorname{Map}(K_+,\operatorname{Map}(L_+,Y)){{ \ \cong \ }}\operatorname{Map}(K_+{{ \,\wedge\, }}L_+,Y)$ in ${{ \mathsf{SymSeq} }}$ respects the left ${{ \mathcal{O} }}$-module structures.
\[defn:mapping\_space\_functor\_simplicial\_category\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Let $X,Y$ be ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules). The *mapping space* $\operatorname{\mathbf{Hom}}(X,Y)\in{{ \mathsf{S} }}$ is defined objectwise by $$\begin{aligned}
\operatorname{\mathbf{Hom}}(X,Y)_n := \hom_{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}(X{{ \dot{{{ \otimes }}} }}\Delta[n],Y)
\quad
\Bigl(\text{resp.}\ \
\operatorname{\mathbf{Hom}}(X,Y)_n := \hom_{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}(X{{ \dot{{{ \otimes }}} }}\Delta[n],Y)
\Bigr).\end{aligned}$$
\[prop:simplicial\_category\_algebras\_modules\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Then the category of ${{ \mathcal{O} }}$-algebras and the category of left ${{ \mathcal{O} }}$-modules are simplicial categories (in the sense of [@Goerss_Jardine II.2.1]), where the mapping space functor is that of Definition \[defn:mapping\_space\_functor\_simplicial\_category\].
This follows from Propositions \[prop:natural\_isomorphisms\_tensordot\_and\_mapping\_object\_algebras\] and \[prop:properties\_for\_establishing\_simplicial\_category\], together with [@Goerss_Jardine II.2.4].
\[prop:simplicial\_model\_category\_axiom\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Consider ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) with the model structure of Theorem \[thm:positive\_flat\_stable\_AlgO\] or \[thm:positive\_stable\_AlgO\].
- If ${{ {j}\colon\thinspace{K}{{ \longrightarrow }}{L} }}$ is a cofibration in ${{ \mathsf{S} }}$, and ${{ {p}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a fibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then $
\xymatrix@1{
\operatorname{Map}(L_+,X)\ar[r] &
\operatorname{Map}(K_+,X)\times_{\operatorname{Map}(K_+,Y)}
\operatorname{Map}(L_+,Y)
}
$ is a fibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) that is an acyclic fibration if either $j$ or $p$ is a weak equivalence.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and ${{ {p}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a fibration in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then the pullback corner map is a fibration $
\xymatrix@1{
\operatorname{\mathbf{Hom}}(B,X)\ar[r] &
\operatorname{\mathbf{Hom}}(A,X)\times_{\operatorname{\mathbf{Hom}}(A,Y)}
\operatorname{\mathbf{Hom}}(B,Y)
}
$ in ${{ \mathsf{S} }}$ that is an acyclic fibration if either $j$ or $p$ is a weak equivalence.
Consider the case of left ${{ \mathcal{O} }}$-modules with the positive flat stable model structure. Part (a) follows from the proof of Proposition \[prop:mixing\_flat\_stable\_with\_positive\_flat\_stable\_tensorcheck\], and part (b) follows from part (a) together with [@Goerss_Jardine II.3.13]. The case of ${{ \mathcal{O} }}$-algebras with the positive flat stable model structure is similar. Consider the case of ${{ \mathcal{O} }}$-algebras or left ${{ \mathcal{O} }}$-modules with the positive stable model structure. This follows by exactly the same argument as above together with the fact that ${{ \mathcal{R} }}{{ \otimes }}G_0(-)_+$ applied to a cofibration in ${{ \mathsf{S} }}$ gives a cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ with the stable model structure (Section \[sec:model\_structures\] and [@Schwede_book_project]).
The following theorem states that the simplicial structure respects the model category structure; this has also been observed in the context of symmetric spectra in [@Hornbostel; @Schwede_book_project]; see also [@Arone_Ching; @EKMM; @McClure_Schwanzl_Vogt].
\[thm:simplicial\_model\_category\_structure\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Consider ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) with the model structure of Theorem \[thm:positive\_flat\_stable\_AlgO\] or \[thm:positive\_stable\_AlgO\]. Then ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) is a simplicial model category with the mapping space functor of Definition \[defn:mapping\_space\_functor\_simplicial\_category\].
This follows from Propositions \[prop:simplicial\_category\_algebras\_modules\] and \[prop:simplicial\_model\_category\_axiom\], together with [@Goerss_Jardine II.3.13].
Homotopical analysis of the simplicial bar constructions {#homotopical-analysis-of-the-simplicial-bar-constructions}
--------------------------------------------------------
The purpose of this subsection is to prove Theorem \[thm:reedy\_cofibrant\_for\_bar\_constructions\]. This will require that we establish certain homotopical properties of the tensor product (Proposition \[prop:mixing\_flat\_stable\_with\_positive\_flat\_stable\_tensorcheck\]) and circle product (Theorem \[thm:mixing\_flat\_stable\_with\_positive\_flat\_stable\] and Proposition \[prop:flat\_stable\_cofibration\_properties\_symmetric\_sequences\]) constructions arising in the description of the degenerate subobjects (Proposition \[prop:nice\_description\_of\_degenerate\_subobjects\]).
\[prop:flat\_cofibrations\_and\_positive\_flat\_cofibrations\] Consider symmetric sequences in ${{ \mathcal{R} }}$-modules. Let $A,B$ be symmetric sequences.
- ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and $X_0\xrightarrow{{{ \ \cong \ }}}Y_0$ is an isomorphism if and only if $f$ is a positive flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$.
- ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$ and $X[\mathbf{r}]_0\xrightarrow{{{ \ \cong \ }}}Y[\mathbf{r}]_0$ is an isomorphism for each $r\geq 0$, if and only if $f$ is a positive flat stable cofibration in ${{ \mathsf{SymSeq} }}$.
- If $X,Y\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, then there is a natural isomorphism $(X{{ \,\wedge\, }}Y)_0{{ \ \cong \ }}X_0{{ \,\wedge\, }}_{{{ \mathcal{R} }}_0} Y_0$.
- If $X,Y\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and $Y_0=*$, then $(X{{ \,\wedge\, }}Y)_0=*$.
- If $B[\mathbf{r}]_0=*$ for each $r\geq 0$, then $(A{{ \check{{{ \otimes }}} }}B)[\mathbf{r}]_0=*$ for each $r\geq 0$.
- If $A[\mathbf{0}]_0=*=B[\mathbf{r}]_0$ for each $r\geq 0$, then $(A\circ B)[\mathbf{r}]_0=*$ for each $r\geq 0$.
- If $A[\mathbf{r}]_0=*$ for each $r\geq 0$, then $(A\circ B)[\mathbf{r}]_0=*$ for each $r\geq 0$.
Parts (a) and (b) follow from \[prop:cofibration\_characterization\]. The remaining parts are an easy exercise left to the reader.
\[prop:mixing\_flat\_stable\_with\_positive\_flat\_stable\_tensorcheck\] Consider symmetric sequences in ${{ \mathcal{R} }}$-modules, and consider ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure.
- If ${{ {i}\colon\thinspace{K}{{ \longrightarrow }}{L} }}$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$, and ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration in ${{ \mathsf{SymSeq} }}$, then $
\xymatrix@1{
L{{ \check{{{ \otimes }}} }}A\coprod_{K{{ \check{{{ \otimes }}} }}A} K{{ \check{{{ \otimes }}} }}B\ar[r] &
L{{ \check{{{ \otimes }}} }}B
}
$ is a cofibration in ${{ \mathsf{SymSeq} }}$ that is an acyclic cofibration if either $i$ or $j$ is a weak equivalence.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$, and ${{ {p}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a fibration in ${{ \mathsf{SymSeq} }}$, then $
\xymatrix@1{
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,X)\ar[r] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(A,X)\times_{\operatorname{Map}^{{ \check{{{ \otimes }}} }}(A,Y)}
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,Y)
}
$ is a fibration in ${{ \mathsf{SymSeq} }}$ that is an acyclic fibration if either $j$ or $p$ is a weak equivalence.
- If ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration in ${{ \mathsf{SymSeq} }}$, and ${{ {p}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a fibration in ${{ \mathsf{SymSeq} }}$, then $
\xymatrix@1{
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,X)\ar[r] &
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(A,X)\times_{\operatorname{Map}^{{ \check{{{ \otimes }}} }}(A,Y)}
\operatorname{Map}^{{ \check{{{ \otimes }}} }}(B,Y)
}
$ is a flat stable fibration in ${{ \mathsf{SymSeq} }}$ that is a flat stable acyclic fibration if either $j$ or $p$ is a weak equivalence.
Consider part (a). Suppose ${{ {i}\colon\thinspace{K}{{ \longrightarrow }}{L} }}$ is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$ and ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration in ${{ \mathsf{SymSeq} }}$. The pushout corner map is a flat stable cofibration in ${{ \mathsf{SymSeq} }}$ by [@Harper_Modules 6.1], hence by Proposition \[prop:flat\_cofibrations\_and\_positive\_flat\_cofibrations\] it suffices to verify the pushout corner map $
\xymatrix@1{
(L{{ \check{{{ \otimes }}} }}A)[\mathbf{r}]_0\coprod_{(K{{ \check{{{ \otimes }}} }}A)[\mathbf{r}]_0} (K{{ \check{{{ \otimes }}} }}B)[\mathbf{r}]_0\ar[r] &
(L{{ \check{{{ \otimes }}} }}B)[\mathbf{r}]_0
}
$ is an isomorphism for each $r\geq 0$. We can therefore conclude by together with Proposition \[prop:flat\_cofibrations\_and\_positive\_flat\_cofibrations\]. The other cases are similar. Parts (b) and (c) follow from part (a) and the natural isomorphisms .
\[thm:mixing\_flat\_stable\_with\_positive\_flat\_stable\] Consider symmetric sequences in ${{ \mathcal{R} }}$-modules, and consider ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure.
- If ${{ {i}\colon\thinspace{K}{{ \longrightarrow }}{L} }}$ is a map in ${{ \mathsf{SymSeq} }}$ such that $K[\mathbf{r}]{{ \longrightarrow }}L[\mathbf{r}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 1$, and ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration between cofibrant objects in ${{ \mathsf{SymSeq} }}$, then $
\xymatrix@1{
L\circ A\coprod_{K\circ A} K\circ B\ar[r] & L\circ B
}
$ is a cofibration in ${{ \mathsf{SymSeq} }}$ that is an acyclic cofibration if either $i$ or $j$ is a weak equivalence.
- If ${{ {i}\colon\thinspace{K}{{ \longrightarrow }}{L} }}$ is a map in ${{ \mathsf{SymSeq} }}$ such that $K[\mathbf{r}]{{ \longrightarrow }}L[\mathbf{r}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$, $K[\mathbf{0}]_0\xrightarrow{{{ \ \cong \ }}}L[\mathbf{0}]_0$ is an isomorphism, and $B$ is a cofibrant object in ${{ \mathsf{SymSeq} }}$, then the map $
K\circ B{{ \longrightarrow }}L\circ B
$ is a cofibration in ${{ \mathsf{SymSeq} }}$ that is an acyclic cofibration if $i$ is a weak equivalence.
Consider part (a). Suppose $K[\mathbf{t}]{{ \longrightarrow }}L[\mathbf{t}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $t\geq 1$, and ${{ {j}\colon\thinspace{A}{{ \longrightarrow }}{B} }}$ is a cofibration between cofibrant objects in ${{ \mathsf{SymSeq} }}$. We want to verify each $
\xymatrix@1{
L[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} A^{{{ \check{{{ \otimes }}} }}t}
\coprod_{K[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} A^{{{ \check{{{ \otimes }}} }}t}}
K[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} B^{{{ \check{{{ \otimes }}} }}t}
\ar[r] &
L[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} B^{{{ \check{{{ \otimes }}} }}t}
}
$ is a cofibration in ${{ \mathsf{SymSeq} }}$. If $t=0$, this map is an isomorphism. Let $t\geq 1$. Consider any acyclic fibration ${{ {p}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{SymSeq} }}$. We want to show that the pushout corner map has the left lifting property with respect to $p$. Consider any such lifting problem; we want to verify that the corresponding solid commutative diagram $$\begin{aligned}
\xymatrix{
A^{{{ \check{{{ \otimes }}} }}t}\ar[d]\ar[r] & \operatorname{Map}(L[\mathbf{t}], X)\ar[d]^{(*)}\\
B^{{{ \check{{{ \otimes }}} }}t}\ar[r]\ar@{.>}[ur] &
\operatorname{Map}(K[\mathbf{t}], X)\times_{\operatorname{Map}(K[\mathbf{t}], Y)}
\operatorname{Map}(L[\mathbf{t}], Y)
}\end{aligned}$$ in ${{ \mathsf{SymSeq} }}^{\Sigma_t}$ has a lift. We know that the left-hand vertical map is a cofibration in ${{ \mathsf{SymSeq} }}^{\Sigma_t}$ by Proposition \[prop:cofibration\], hence it suffices to verify that the map $(*)[\mathbf{r}]$ is a positive flat stable acyclic fibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$. By considering symmetric sequences concentrated at $0$, Proposition \[prop:mixing\_flat\_stable\_with\_positive\_flat\_stable\_tensorcheck\] finishes the argument for this case. The other cases are similar. Consider part (b). Suppose $K[\mathbf{t}]{{ \longrightarrow }}L[\mathbf{t}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $t\geq 0$, $K[\mathbf{0}]_0\xrightarrow{{{ \ \cong \ }}}L[\mathbf{0}]_0$ is an isomorphism, and $B$ is a cofibrant object in ${{ \mathsf{SymSeq} }}$. We need to check that each induced map $
K[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t}B^{{{ \check{{{ \otimes }}} }}t}{{ \longrightarrow }}L[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t}B^{{{ \check{{{ \otimes }}} }}t}
$ is a cofibration in ${{ \mathsf{SymSeq} }}$. The proof of part (a) implies this for $t\geq 1$, and Proposition \[prop:flat\_cofibrations\_and\_positive\_flat\_cofibrations\] implies this for $t=0$. The other case is similar.
\[prop:flat\_stable\_cofibration\_properties\_symmetric\_sequences\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$, and let ${{ {\eta}\colon\thinspace{I}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ be its unit map. Assume that $I[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}[\mathbf{r}]$ is a flat stable cofibration between flat stable cofibrant objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$.
- If ${{ {i}\colon\thinspace{K}{{ \longrightarrow }}{L} }}$ is a map in ${{ \mathsf{SymSeq} }}$ such that $K[\mathbf{r}]{{ \longrightarrow }}L[\mathbf{r}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 1$, then the pushout corner map $
\xymatrix@1{
\bigl(L\circ I\coprod_{K\circ I} K\circ {{ \mathcal{O} }}\bigr)[\mathbf{r}]
\ar[r] & (L\circ{{ \mathcal{O} }})[\mathbf{r}]
}
$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$.
- If $t\geq 1$, then the induced map $(I^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]{{ \longrightarrow }}({{ \mathcal{O} }}^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t}$ for each $r\geq 0$.
Consider part (b). The induced map is an isomorphism for $0\leq r\leq t-1$ and the case for $r\geq t$ follows from Proposition \[prop:cofibration\_characterization\] by arguing as in the proof of Proposition \[prop:cofibration\]. Consider part (a). We need to verify that each $$\begin{aligned}
\xymatrix{
L[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} (I^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]
\coprod_{K[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} (I^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]}
K[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} ({{ \mathcal{O} }}^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]
\ar[r] &
L[\mathbf{t}]{{ \,\wedge\, }}_{\Sigma_t} ({{ \mathcal{O} }}^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]
}\end{aligned}$$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. If $t=0$, this map is an isomorphism. Let $t\geq 1$, and let ${{ {p}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a flat stable acyclic fibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. We need to show that the pushout corner map has the left lifting property with respect to $p$. Consider any such lifting problem; we want to verify that the corresponding solid commutative diagram $$\begin{aligned}
\xymatrix{
(I^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]\ar[d]\ar[r] &
\operatorname{Map}(L[\mathbf{t}], X)\ar[d]^{(*)}\\
({{ \mathcal{O} }}^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]\ar[r]\ar@{.>}[ur] &
\operatorname{Map}(K[\mathbf{t}], X)\times_{\operatorname{Map}(K[\mathbf{t}], Y)}
\operatorname{Map}(L[\mathbf{t}], Y)
}\end{aligned}$$ in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t}$ has a lift. The left-hand vertical map is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t}$ by part (b), hence it suffices to verify the map $(*)$ is a flat stable acyclic fibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. By assumption, each $K[\mathbf{t}]{{ \longrightarrow }}L[\mathbf{t}]$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, which finishes the proof.
Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules, $t\geq 1$ and $n\geq 0$.
- ${{ \mathsf{Cube} }}_t$ is the category with objects the vertices $(v_1,\dotsc,v_t)\in\{0,1\}^t$ of the unit $t$-cube. There is at most one morphism between any two objects, and there is a morphism $
(v_1,\dotsc,v_t){{ \longrightarrow }}(v_1',\dotsc,v_t')
$ if and only if $v_i\leq v_i'$ for each $1\leq i\leq t$. In particular, ${{ \mathsf{Cube} }}_t$ is the category associated to a partial order on the set $\{0,1\}^t$.
- The *punctured cube* ${{ \mathsf{pCube} }}_t$ is the full subcategory of ${{ \mathsf{Cube} }}_t$ with all objects except the terminal object $(1,\dotsc,1)$ of ${{ \mathsf{Cube} }}_t$.
- Define the functor ${{ {w}\colon\thinspace{{{ \mathsf{pCube} }}_t}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ objectwise by $$\begin{aligned}
w(v_1,\dotsc,v_t):=c_1\circ\dotsb\circ c_t\quad\quad\text{with}
\quad\quad
c_i :=
\left\{
\begin{array}{rl}
I,&\text{for $v_i=0$,}\\
{{ \mathcal{O} }},&\text{for $v_i=1$,}
\end{array}
\right.\end{aligned}$$ and with morphisms induced by the unit map ${{ {\eta}\colon\thinspace{I}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$.
- If $X$ is an object in ${{ \mathsf{sMod}_{{ \mathcal{R} }}}}$ or ${{ \mathsf{sSymSeq} }}$, denote by $DX_n{{ \ \subset\ }}X_n$ the *degenerate subobject* [@Harper_Bar 9.12] of $X_n$.
The following proposition gives a useful construction of degenerate subobjects.
\[prop:nice\_description\_of\_degenerate\_subobjects\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules, $Y$ an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and $N$ a right ${{ \mathcal{O} }}$-module. Let $t\geq 1$ and $n\geq 0$. Define $X:=\operatorname{Bar}(N,{{ \mathcal{O} }},Y)$ and $Q^t:=\operatorname{colim}_{{{ \mathsf{pCube} }}_t}(N\circ w)$, and consider the induced maps ${{ {\eta_*}\colon\thinspace{Q^0:=*}{{ \longrightarrow }}{N} }}$ and ${{ {\eta_*}\colon\thinspace{Q^t}{{ \longrightarrow }}{N\circ{{ \mathcal{O} }}^{\circ t}} }}$.
- The inclusion map $DX_n{{ \longrightarrow }}X_n$ is isomorphic to the map $
Q^n\circ (Y)\xrightarrow{\eta_*\circ({{ \mathrm{id} }})}
N\circ{{ \mathcal{O} }}^{\circ n}\circ (Y)
$ (resp. $
Q^n\circ Y\xrightarrow{\eta_*\circ{{ \mathrm{id} }}}
N\circ{{ \mathcal{O} }}^{\circ n}\circ Y
$).
- The induced map ${{ {\eta_*}\colon\thinspace{Q^{n+1}}{{ \longrightarrow }}{N\circ{{ \mathcal{O} }}^{\circ(n+1)}} }}$ is isomorphic to the pushout corner map $
(N\circ{{ \mathcal{O} }}^{\circ n}\circ I)\amalg_{(Q^n\circ I)}(Q^n\circ{{ \mathcal{O} }}){{ \longrightarrow }}N\circ{{ \mathcal{O} }}^{\circ (n+1)}
$ induced by ${{ {\eta}\colon\thinspace{I}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ and ${{ {\eta_*}\colon\thinspace{Q^n}{{ \longrightarrow }}{N\circ{{ \mathcal{O} }}^{\circ n}} }}$.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). It follows easily from [@Harper_Bar 9.23], together with the fact that ${{ {-\circ Y}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ commutes with colimits , that there are natural isomorphisms $$\begin{aligned}
DX_0&{{ \ = \ }}*,\quad DX_1{{ \ \cong \ }}N\circ I\circ Y,\\
DX_2&{{ \ \cong \ }}(N\circ{{ \mathcal{O} }}\circ I\circ Y)\amalg_{(N\circ I\circ I\circ Y)}
(N\circ I\circ{{ \mathcal{O} }}\circ Y)\\
&{{ \ \cong \ }}\bigl(
(N\circ{{ \mathcal{O} }}\circ I)\amalg_{(N\circ I\circ I)}
(N\circ I\circ{{ \mathcal{O} }})
\bigr)\circ Y,\, \dotsc\, ,\\
DX_t&{{ \ \cong \ }}\operatorname{colim}_{pCube_t}(N\circ w\circ Y){{ \ \cong \ }}\bigl(\operatorname{colim}_{pCube_t}(N\circ w)\bigr)\circ Y\end{aligned}$$ in ${{ \mathsf{SymSeq} }}$. Consider part (b). Since ${{ {-\circ B}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}$ commutes with colimits for each $B\in{{ \mathsf{SymSeq} }}$, it follows easily that the colimit $Q^{n+1}$ may be computed inductively using pushout corner maps.
\[thm:inclusion\_of\_degenerate\_subobjects\_positive\_flat\_stable\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules such that ${{ \mathcal{O} }}[\mathbf{0}]=*$, $Y$ an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and $N$ a right ${{ \mathcal{O} }}$-module, and consider the unit map ${{ {\eta}\colon\thinspace{I}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$. Assume that $I[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}[\mathbf{r}]$ is a flat stable cofibration between flat stable cofibrant objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$ and that $N[\mathbf{r}]$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $r\geq 0$. Let $X:=\operatorname{Bar}(N,{{ \mathcal{O} }},Y)$. If $Y$ is positive flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) and $N[\mathbf{0}]_0=*$, then the inclusion maps $$\begin{aligned}
*{{ \longrightarrow }}DX_n{{ \longrightarrow }}X_n,\quad\quad
*{{ \longrightarrow }}|\operatorname{Bar}(N,{{ \mathcal{O} }},Y)|,\end{aligned}$$ are positive flat stable cofibrations in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) for each $n\geq 0$. In particular, the simplicial bar construction $\operatorname{Bar}(N,{{ \mathcal{O} }},Y)$ is Reedy cofibrant in ${{ \mathsf{sMod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{sSymSeq} }}$) with respect to the positive flat stable model structure.
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider Proposition \[prop:nice\_description\_of\_degenerate\_subobjects\]; let’s verify that the left-hand induced maps $$\begin{aligned}
\label{eq:cofibration_properties_of_degenerate_subobjects}
*{{ \longrightarrow }}Q^n[\mathbf{r}]{{ \longrightarrow }}(N\circ{{ \mathcal{O} }}^{\circ n})[\mathbf{r}],
\quad\quad
Q^n[\mathbf{0}]_0=*=(N\circ{{ \mathcal{O} }}^{\circ n})[\mathbf{0}]_0\end{aligned}$$ are flat stable cofibrations in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ for each $n,r\geq 0$ and that the right-hand relations are satisfied for each $n\geq 0$. It is easy to check this for $n=0$, and by induction on $n$, the general case follows from Propositions \[prop:flat\_stable\_cofibration\_properties\_symmetric\_sequences\] and \[prop:nice\_description\_of\_degenerate\_subobjects\]. By assumption, $Y$ is positive flat stable cofibrant in ${{ \mathsf{SymSeq} }}$, hence by Proposition \[prop:nice\_description\_of\_degenerate\_subobjects\] and Theorem \[thm:mixing\_flat\_stable\_with\_positive\_flat\_stable\], the inclusion maps $
*{{ \longrightarrow }}DX_n{{ \longrightarrow }}X_n
$ are positive flat stable cofibrations in ${{ \mathsf{SymSeq} }}$ for each $n\geq 0$. Since $DX_n$ and $X_n$ are positive flat stable cofibrant in ${{ \mathsf{SymSeq} }}$ for each $n\geq 0$, we know by \[prop:cofibration\_characterization\] that the relations $
DX_n[\mathbf{r}]_0=*=X_n[\mathbf{r}]_0
$ are satisfied for each $n,r\geq 0$. It then follows easily from the skeletal filtration of realization [@Harper_Bar 9.11, 9.16], together with Proposition \[prop:flat\_cofibrations\_and\_positive\_flat\_cofibrations\], that $|\operatorname{Bar}(N,{{ \mathcal{O} }},Y)|$ is positive flat stable cofibrant in ${{ \mathsf{SymSeq} }}$. It is easy to check that the natural map $DX_n{{ \longrightarrow }}X_n$ is isomorphic to the natural map $L_nX{{ \longrightarrow }}X_n$ described in [@Goerss_Jardine VII.1.8]. Hence, in particular, we have verified that $X$ is Reedy cofibrant [@Goerss_Jardine VII.2.1] in ${{ \mathsf{sSymSeq} }}$.
\[prop:cofibrant\_bar\_constructions\_for\_chain\_complexes\_algebras\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules, $Y$ an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and $N$ a right ${{ \mathcal{O} }}$-module. Consider ${{ \mathsf{SymSeq} }}$ with the flat stable model structure. Assume that the unit map $I{{ \longrightarrow }}{{ \mathcal{O} }}$ is a cofibration between cofibrant objects in ${{ \mathsf{SymSeq} }}$ and that $N$ is cofibrant in ${{ \mathsf{SymSeq} }}$. If $Y$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$), then $|\operatorname{Bar}(N,{{ \mathcal{O} }},Y)|$ is flat stable cofibrant in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$).
Argue as in the proof of Theorem \[thm:inclusion\_of\_degenerate\_subobjects\_positive\_flat\_stable\].
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. Consider part (a). This follows as in the proof of Theorem \[thm:inclusion\_of\_degenerate\_subobjects\_positive\_flat\_stable\], except using the skeletal filtration in [@Goerss_Jardine VII.3.8], Proposition \[prop:nice\_description\_of\_degenerate\_subobjects\] and Theorem \[thm:mixing\_flat\_stable\_with\_positive\_flat\_stable\], together with the fact that ${{ {{{ \mathcal{O} }}'\circ-}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}} }}$ is a left Quillen functor and hence preserves both colimiting cones and cofibrations. Part (b) follows immediately from part (a) together with Proposition \[prop:realizations\_are\_isomorphic\], Theorem \[thm:simplicial\_model\_category\_structure\], and [@Goerss_Jardine VII.3.4].
Model structures {#sec:model_structures}
================
The purpose of this section is to prove Theorems \[thm:positive\_flat\_stable\_AlgO\], \[thm:positive\_stable\_AlgO\], and \[thm:comparing\_homotopy\_categories\], together with Theorems \[thm:bar\_calculates\_derived\_circle\], \[main\_hocolim\_theorem\], and \[thm:fattened\_replacement\] which improve the main results in [@Harper_Spectra; @Harper_Bar] from operads in symmetric spectra to the more general context of operads in ${{ \mathcal{R} }}$-modules. Our approach to this generalization, which is motivated by Hornbostel [@Hornbostel], is to establish only the necessary minimum of technical propositions for ${{ \mathcal{R} }}$-modules needed for the proofs of the main results as described in [@Harper_Spectra; @Harper_Bar] to remain valid in the more general context of ${{ \mathcal{R} }}$-modules.
Smash products and ${{ \mathcal{R} }}$-modules
----------------------------------------------
Denote by $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ the closed symmetric monoidal category of symmetric spectra [@Hovey_Shipley_Smith; @Schwede_book_project]. To keep this section as concise as possible, from now on we will freely use the notation from [@Harper_Spectra Section 2] which agrees (whenever possible) with [@Hovey_Shipley_Smith].
The following is proved in [@Hovey_Shipley_Smith 2.1] and states that tensor product in the category ${{ \mathsf{S} }}^\Sigma_*$ inherits many of the good properties of smash product in the category ${{ \mathsf{S} }}_*$.
\[prop:closed\_symmetric\_monoidal\_structure\_on\_sym\_sequences\_pointed\_ssets\] $({{ \mathsf{S} }}^\Sigma_*,{{ \otimes }},S^0)$ has the structure of a closed symmetric monoidal category. All small limits and colimits exist and are calculated objectwise. The unit $S^0\in{{ \mathsf{S} }}^\Sigma_*$ is given by $S^0[\mathbf{n}]=*$ for each $n\geq 1$ and $S^0[\mathbf{0}]=S^0$.
There are two naturally occurring maps $S{{ \otimes }}S{{ \longrightarrow }}S$ and $S^0{{ \longrightarrow }}S$ in ${{ \mathsf{S} }}^\Sigma_*$ that give $S$ the structure of a commutative monoid in $({{ \mathsf{S} }}^\Sigma_*,{{ \otimes }},S^0)$. Furthermore, for any symmetric spectrum $X$, there is a naturally occurring map ${{ {m}\colon\thinspace{S{{ \otimes }}X}{{ \longrightarrow }}{X} }}$ endowing $X$ with a left action of $S$ in $({{ \mathsf{S} }}^\Sigma_*,{{ \otimes }},S^0)$. The following is proved in [@Hovey_Shipley_Smith 2.2] and provides a useful interpretation of symmetric spectra.
\[prop:symmetric\_spectra\_as\_S\_modules\] Define the category $\Sigma':=\amalg_{n\geq 0}\Sigma_n$, a skeleton of $\Sigma$.
- The sphere spectrum $S$ is a commutative monoid in $({{ \mathsf{S} }}^\Sigma_*,{{ \otimes }},S^0)$.
- The category of symmetric spectra is equivalent to the category of left $S$-modules in $({{ \mathsf{S} }}^\Sigma_*,{{ \otimes }},S^0)$.
- The category of symmetric spectra is isomorphic to the category of left $S$-modules in $({{ \mathsf{S} }}^{\Sigma'}_*,{{ \otimes }},S^0)$.
In this paper we will not distinguish between these equivalent descriptions of symmetric spectra.
\[defn:left\_R\_modules\] Let ${{ \mathcal{R} }}$ be a commutative monoid in $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ (Basic Assumption \[assumption:commutative\_ring\_spectrum\]). A *left ${{ \mathcal{R} }}$-module* is an object in $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ with a left action of ${{ \mathcal{R} }}$ and a *morphism of left ${{ \mathcal{R} }}$-modules* is a map in ${{ \mathsf{Sp}^\Sigma }}$ that respects the left ${{ \mathcal{R} }}$-module structure. Denote by ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ the category of left ${{ \mathcal{R} }}$-modules and their morphisms.
The *smash product* $X{{ \,\wedge\, }}Y\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ of left ${{ \mathcal{R} }}$-modules $X$ and $Y$ is defined by $$\begin{aligned}
X{{ \,\wedge\, }}Y
:=\operatorname{colim}\Bigl(
\xymatrix{
X{{ \otimes }}_S Y & X{{ \otimes }}_S {{ \mathcal{R} }}{{ \otimes }}_S Y
\ar@<-0.5ex>[l]_-{m{{ \otimes }}{{ \mathrm{id} }}}\ar@<0.5ex>[l]^-{{{ \mathrm{id} }}{{ \otimes }}m}
}
\Bigr)
{{ \ \cong \ }}\operatorname{colim}\Bigl(
\xymatrix{
X{{ \otimes }}Y & X{{ \otimes }}{{ \mathcal{R} }}{{ \otimes }}Y
\ar@<-0.5ex>[l]_-{m{{ \otimes }}{{ \mathrm{id} }}}\ar@<0.5ex>[l]^-{{{ \mathrm{id} }}{{ \otimes }}m}
}
\Bigr)\end{aligned}$$ the indicated colimit. Here, $m$ denotes the indicated ${{ \mathcal{R} }}$-action map and since ${{ \mathcal{R} }}$ is a commutative monoid in $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$, a left action of ${{ \mathcal{R} }}$ on $X$ determines a right action ${{ {m}\colon\thinspace{X{{ \otimes }}_S {{ \mathcal{R} }}}{{ \longrightarrow }}{X} }}$, which gives $X$ the structure of an $({{ \mathcal{R} }},{{ \mathcal{R} }})$-bimodule. Hence the smash product $X{{ \,\wedge\, }}Y$ of left ${{ \mathcal{R} }}$-modules, which is naturally isomorphic to $X{{ \otimes }}_{{ \mathcal{R} }}Y$, has the structure of a left ${{ \mathcal{R} }}$-module.
\[rem:dropping\_the\_adjective\_left\] Since ${{ \mathcal{R} }}$ is commutative, we usually drop the adjective “left” and simply refer to the objects of ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ as ${{ \mathcal{R} }}$-modules.
The following is an easy consequence of [@Hovey_Shipley_Smith 2.2].
$({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ has the structure of a closed symmetric monoidal category. All small limits and colimits exist and are calculated objectwise.
Model structures on ${{ \mathcal{R} }}$-modules {#sec:model_structures_on_capR_modules}
-----------------------------------------------
The material below intentionally parallels [@Harper_Spectra Section 4], except that we work in the more general context of ${{ \mathcal{R} }}$-modules instead of symmetric spectra. We need to recall just enough notation so that we can describe and work with the (positive) flat stable model structure on ${{ \mathcal{R} }}$-modules, and the corresponding projective model structures on the diagram categories ${{ \mathsf{SymSeq} }}$ and ${{ \mathsf{SymSeq} }}^G$ of ${{ \mathcal{R} }}$-modules, for $G$ a finite group. The functors involved in such a description are easy to understand when defined as the left adjoints of appropriate functors, which is how they naturally arise in this context.
For each $m\geq 0$ and subgroup $H{{ \ \subset\ }}\Sigma_m$, denote by ${{ {l}\colon\thinspace{H}{{ \longrightarrow }}{\Sigma_m} }}$ the inclusion of groups and define the *evaluation* functor ${{ {{{ \mathrm{ev} }}_m}\colon\thinspace{{{ \mathsf{S} }}^\Sigma_*}{{ \longrightarrow }}{{{ \mathsf{S} }}^{\Sigma_m}_*} }}$ objectwise by ${{ \mathrm{ev} }}_m(X):=X_m$. There are adjunctions $
\xymatrix@1{
{{ \mathsf{S} }}_*\ar@<0.5ex>[r] &
{{ \mathsf{S} }}^{H}_*\ar@<0.5ex>[l]^-{\lim_H}
\ar@<0.5ex>[r]^-{\Sigma_m\cdot_H-} &
{{ \mathsf{S} }}^{\Sigma_m}_*\ar@<0.5ex>[l]^-{l^*}\ar@<0.5ex>[r] &
{{ \mathsf{S} }}^{\Sigma}_*\ar@<0.5ex>[l]^-{{{ \mathrm{ev} }}_m}
}
$ with left adjoints on top. Define ${{ {G^H_m}\colon\thinspace{{{ \mathsf{S} }}_*}{{ \longrightarrow }}{{{ \mathsf{S} }}^\Sigma_*} }}$ to be the composition of the three top functors, and define ${{ {\lim_H {{ \mathrm{ev} }}_m}\colon\thinspace{{{ \mathsf{S} }}^\Sigma_*}{{ \longrightarrow }}{{{ \mathsf{S} }}_*} }}$ to be the composition of the three bottom functors; we have dropped the restriction functor $l^*$ from the notation. It is easy to check that if $K\in{{ \mathsf{S} }}_*$, then $G^H_m(K)$ is the object in ${{ \mathsf{S} }}^\Sigma_*$ that is concentrated at $m$ with value $\Sigma_m\cdot_H K$. Consider the forgetful functors ${{ \mathsf{Sp}^\Sigma }}{{ \longrightarrow }}{{ \mathsf{S} }}^\Sigma_*$ and ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}{{ \longrightarrow }}{{ \mathsf{Sp}^\Sigma }}$. It follows from Proposition \[prop:symmetric\_spectra\_as\_S\_modules\] that there are adjunctions $$\begin{aligned}
\label{eq:new_adjunctions_for_spectra_as_S_modules}
\xymatrix{
{{ \mathsf{S} }}^\Sigma_*\ar@<0.5ex>[r]^-{S{{ \otimes }}-} &
{{ \mathsf{Sp}^\Sigma }}\ar@<0.5ex>[l]\ar@<0.5ex>[r]^-{{{ \mathcal{R} }}{{ \otimes }}_S-} &
{{ \mathsf{Mod}_{{ \mathcal{R} }}}}\ar@<0.5ex>[l]
},
\quad\quad
\xymatrix{
{{ \mathsf{S} }}^\Sigma_*\ar@<0.5ex>[r]^-{{{ \mathcal{R} }}{{ \otimes }}-} &
{{ \mathsf{Mod}_{{ \mathcal{R} }}}}\ar@<0.5ex>[l]
},\end{aligned}$$ with left adjoints on top; the latter adjunction is the composition of the former adjunctions. For each $p\geq 0$, define the *evaluation* functor ${{ {{{ \mathrm{Ev} }}_p}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }}$ objectwise by ${{ \mathrm{Ev} }}_p(A):=A[\mathbf{p}]$, and for each finite group $G$, consider the forgetful functor ${{ \mathsf{SymSeq} }}^G{{ \longrightarrow }}{{ \mathsf{SymSeq} }}$. There are adjunctions $
\xymatrix@1{
{{ \mathsf{Mod}_{{ \mathcal{R} }}}}\ar@<0.5ex>[r]^-{G_p} &
{{ \mathsf{SymSeq} }}\ar@<0.5ex>[l]^-{{{ \mathrm{Ev} }}_p}\ar@<0.5ex>[r]^-{G\cdot-} &
{{ \mathsf{SymSeq} }}^G\ar@<0.5ex>[l]
}
$ with left adjoints on top. It is easy to check that if $X\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, then $G_p(X)$ is the symmetric sequence concentrated at $p$ with value $X\cdot\Sigma_p$. Putting it all together, there are adjunctions $$\begin{aligned}
\label{eq:adjunctions_stable_flat}
\xymatrix{
{{ \mathsf{S} }}_*\ar@<0.5ex>[r]^-{G^H_m} &
{{ \mathsf{S} }}^\Sigma_*\ar@<0.5ex>[l]^-{\ \lim_H {{ \mathrm{ev} }}_m}
\ar@<0.5ex>[r]^-{{{ \mathcal{R} }}{{ \otimes }}-}&
{{ \mathsf{Mod}_{{ \mathcal{R} }}}}\ar@<0.5ex>[l]\ar@<0.5ex>[r]^-{G_p}&
{{ \mathsf{SymSeq} }}\ar@<0.5ex>[l]^-{{{ \mathrm{Ev} }}_p}\ar@<0.5ex>[r]^-{G\cdot-}&
{{ \mathsf{SymSeq} }}^G\ar@<0.5ex>[l]
}\end{aligned}$$ with left adjoints on top. We are now in a good position to describe several useful model structures. It is proved in [@Shipley_comm_ring] that the following two model category structures exist on ${{ \mathcal{R} }}$-modules.
\[defn:lets\_define\_flat\_model\_structure\]
- The *flat stable model structure* on ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ has weak equivalences the stable equivalences, cofibrations the retracts of (possibly transfinite) compositions of pushouts of maps $$\begin{aligned}
{{ \mathcal{R} }}{{ \otimes }}G^H_m \partial\Delta[k]_+{{ \longrightarrow }}{{ \mathcal{R} }}{{ \otimes }}G^H_m\Delta[k]_+\quad
(m\geq 0,\ k\geq 0,\ H{{ \ \subset\ }}\Sigma_m \ \text{subgroup}),\end{aligned}$$ and fibrations the maps with the right lifting property with respect to the acyclic cofibrations.
- The *positive flat stable model structure* on ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ has weak equivalences the stable equivalences, cofibrations the retracts of (possibly transfinite) compositions of pushouts of maps $$\begin{aligned}
{{ \mathcal{R} }}{{ \otimes }}G^H_m \partial\Delta[k]_+{{ \longrightarrow }}{{ \mathcal{R} }}{{ \otimes }}G^H_m\Delta[k]_+\quad
(m\geq 1,\ k\geq 0,\ H{{ \ \subset\ }}\Sigma_m \ \text{subgroup}),\end{aligned}$$ and fibrations the maps with the right lifting property with respect to the acyclic cofibrations.
\[rem:flat\_notation\] In the sets of maps above, it is important to note that $H$ varies over all subgroups of $\Sigma_m$. For ease of notation purposes, we have followed Schwede [@Schwede_book_project] in using the term *flat* (e.g., flat stable model structure) for what is called ${{ \mathcal{R} }}$ (e.g., stable ${{ \mathcal{R} }}$-model structure) in [@Hovey_Shipley_Smith; @Schwede; @Shipley_comm_ring].
Several useful properties of the flat stable model structure are summarized in the following two propositions, which are consequences of as indicated below; see also [@Schwede_book_project]. These properties are used in several sections of this paper.
\[prop:weak\_equivalences\_and\_monos\_are\_preserved\] Consider ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ with the flat stable model structure. If $Z\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ is cofibrant, then the functor ${{ {-{{ \,\wedge\, }}Z}\colon\thinspace{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }}$ preserves (i) weak equivalences and (ii) monomorphisms.
\[prop:sending\_cofibrations\_to\_monos\] If $B\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and $X{{ \longrightarrow }}Y$ is a flat stable cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, then $B{{ \,\wedge\, }}X{{ \longrightarrow }}B{{ \,\wedge\, }}Y$ in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ is a monomorphism.
Part (i) is the ${{ \mathcal{R} }}$-module analog of [@Hovey_Shipley_Smith 5.3.10]. It can also be verified as a consequence of [@Hovey_Shipley_Smith 5.3.10] by arguing exactly as in the proof of [@Harper_Spectra 4.29(b)]. Part (ii) follows from the ${{ \mathcal{R} }}$-module analog of [@Hovey_Shipley_Smith 5.3.7]; see, [@Hovey_Shipley_Smith proof of 5.4.4] or [@Schwede_book_project].
This follows from the ${{ \mathcal{R} }}$-module analog of [@Hovey_Shipley_Smith 5.3.7]; see, [@Hovey_Shipley_Smith proof of 5.4.4] or [@Schwede_book_project].
The *stable model structure* on ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ is defined by fixing $H$ in Definition \[defn:lets\_define\_flat\_model\_structure\](a) to be the trivial subgroup. This is one of several model category structures that is proved in [@Hovey_Shipley_Smith] to exist on ${{ \mathcal{R} }}$-modules. The *positive stable model structure* on ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ is defined by fixing $H$ in Definition \[defn:lets\_define\_flat\_model\_structure\](b) to be the trivial subgroup. This model category structure is proved in [@Mandell_May_Schwede_Shipley] to exist on ${{ \mathcal{R} }}$-modules. It follows immediately that every (positive) stable cofibration is a (positive) flat stable cofibration.
These model structures on ${{ \mathcal{R} }}$-modules enjoy several good properties, including that smash products of ${{ \mathcal{R} }}$-modules mesh nicely with each of the model structures defined above. More precisely, each model structure above is cofibrantly generated, by generating cofibrations and acyclic cofibrations with small domains, and with respect to each model structure $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ is a monoidal model category.
If $G$ is a finite group, it is easy to check that the diagram categories ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^G$, ${{ \mathsf{SymSeq} }}$ and ${{ \mathsf{SymSeq} }}^G$ inherit corresponding projective model category structures, where the weak equivalences (resp. fibrations) are the maps that are underlying objectwise weak equivalences (resp. objectwise fibrations). We refer to these model structures by the names above (e.g., the *positive flat stable* model structure on ${{ \mathsf{SymSeq} }}^G$). Each of these model structures is cofibrantly generated by generating cofibrations and acyclic cofibrations with small domains. Furthermore, with respect to each model structure $({{ \mathsf{SymSeq} }},{{ \otimes }},1)$ is a monoidal model category; this is proved in [@Harper_Modules].
Model structures on ${{ \mathcal{O} }}$-algebras and left ${{ \mathcal{O} }}$-modules
-------------------------------------------------------------------------------------
The purpose of this subsection is to prove the following two theorems. These generalizations are motivated by Hornbostel [@Hornbostel] and improve the corresponding results in [@Harper_Spectra 1.1, 1.3] from operads in symmetric spectra to the more general context involving operads in ${{ \mathcal{R} }}$-modules and play a key role in this paper. An important first step in establishing these theorems was provided by the characterization given by Schwede [@Schwede_book_project] of flat stable cofibrations in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ in terms of objects with an ${{ \mathcal{R} }}_0$-action; see Proposition \[prop:cofibration\_characterization\] below for the needed generalization of this.
\[thm:positive\_flat\_stable\_AlgO\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Then the category of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) has a model category structure with weak equivalences the stable equivalences (resp. objectwise stable equivalences) and fibrations the maps that are positive flat stable fibrations (resp. objectwise positive flat stable fibrations) in the underlying category of ${{ \mathcal{R} }}$-modules (Definition \[defn:lets\_define\_flat\_model\_structure\](b)).
\[thm:positive\_stable\_AlgO\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. Then the category of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) has a model category structure with weak equivalences the stable equivalences (resp. objectwise stable equivalences) and fibrations the maps that are positive stable fibrations (resp. objectwise positive stable fibrations) in the underlying category of ${{ \mathcal{R} }}$-modules (Definition \[defn:lets\_define\_flat\_model\_structure\](b) and below Proposition \[prop:sending\_cofibrations\_to\_monos\]).
We defer the proof of the following two propositions to Subsection \[sec:flat\_stable\_cofibrations\].
\[prop:cofibration\] Let $B\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t^{{ \mathrm{op} }}}$ (resp. $B\in{{ \mathsf{SymSeq} }}^{\Sigma_t^{{ \mathrm{op} }}}$) and $t\geq 1$. If ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a cofibration between cofibrant objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) with the positive flat stable model structure, then
- $X^{\wedge t}{{ \longrightarrow }}Y^{\wedge t}$ (resp. $X^{{{ \check{{{ \otimes }}} }}t}{{ \longrightarrow }}Y^{{{ \check{{{ \otimes }}} }}t}$) is a cofibration between cofibrant objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_t}$) with the positive flat stable model structure, which is a weak equivalence if $i$ is a weak equivalence,
- the map $B{{ \,\wedge\, }}_{\Sigma_t}Q_{t-1}^t{{ \longrightarrow }}B{{ \,\wedge\, }}_{\Sigma_t}Y^{\wedge t}$ (resp. $B{{ \check{{{ \otimes }}} }}_{\Sigma_t}Q_{t-1}^t{{ \longrightarrow }}B{{ \check{{{ \otimes }}} }}_{\Sigma_t}Y^{{{ \check{{{ \otimes }}} }}t}$) is a monomorphism.
\[prop:good\_properties\] Let $G$ be a finite group and consider ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^G$, ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{G^{{ \mathrm{op} }}}$, ${{ \mathsf{SymSeq} }}$, ${{ \mathsf{SymSeq} }}^G$, and ${{ \mathsf{SymSeq} }}^{G^{{ \mathrm{op} }}}$, each with the flat stable model structure.
- If $B\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{G^{{ \mathrm{op} }}}$ (resp. $B\in{{ \mathsf{SymSeq} }}^{G^{{ \mathrm{op} }}}$), then the functor $$\begin{aligned}
{{ {B{{ \,\wedge\, }}_G -}\colon\thinspace{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^G}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }}
\quad\quad
\Bigl(\text{resp.}\quad
{{ {B{{ \check{{{ \otimes }}} }}_G -}\colon\thinspace{{{ \mathsf{SymSeq} }}^G}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}
\Bigr)\end{aligned}$$ preserves weak equivalences between cofibrant objects, and hence its total left derived functor exists.
- If $Z\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^G$ (resp. $Z\in{{ \mathsf{SymSeq} }}^G$) is cofibrant, then the functor $$\begin{aligned}
{{ {-{{ \,\wedge\, }}_G Z}\colon\thinspace{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{G^{{ \mathrm{op} }}}}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }}
\quad\quad
\Bigl(\text{resp.}\quad
{{ {-{{ \check{{{ \otimes }}} }}_G Z}\colon\thinspace{{{ \mathsf{SymSeq} }}^{G^{{ \mathrm{op} }}}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}
\Bigr)\end{aligned}$$ preserves weak equivalences.
\[prop:homotopical\_analysis\_of\_certain\_pushouts\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules, $A\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $A\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), and ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ a generating acyclic cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) with the positive flat stable model structure. Consider any pushout diagram in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) of the form . Then $j$ is a monomorphism and a weak equivalence in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$).
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. This is verified exactly as in [@Harper_Spectra proof of 4.4], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and Propositions \[prop:cofibration\], \[prop:good\_properties\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and [@Harper_Spectra 4.28, 4.29], respectively.
Consider ${{ \mathsf{SymSeq} }}$ and ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, both with the positive flat stable model structure. We will prove that the model structure on ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$) is created by the middle (resp. left-hand) free-forgetful adjunction in .
Define a map $f$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ to be a weak equivalence (resp. fibration) if $U(f)$ is a weak equivalence (resp. fibration) in ${{ \mathsf{SymSeq} }}$. Similarly, define a map $f$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ to be a weak equivalence (resp. fibration) if $U(f)$ is a weak equivalence (resp. fibration) in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. Define a map $f$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$) to be a cofibration if it has the left lifting property with respect to all acyclic fibrations in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$).
Consider the case of ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$. We want to verify the model category axioms (MC1)-(MC5) in [@Dwyer_Spalinski]. Arguing exactly as in [@Harper_Spectra proof of 1.1], this reduces to the verification of Proposition \[prop:homotopical\_analysis\_of\_certain\_pushouts\]. By construction, the model category is cofibrantly generated. Argue similarly for the case of ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ by considering left ${{ \mathcal{O} }}$-modules concentrated at $0$.
Consider ${{ \mathsf{SymSeq} }}$ and ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$, both with the positive stable model structure. We will prove that the model structure on ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$) is created by the middle (resp. left-hand) free-forgetful adjunction in .
Define a map $f$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ to be a weak equivalence (resp. fibration) if $U(f)$ is a weak equivalence (resp. fibration) in ${{ \mathsf{SymSeq} }}$. Similarly, define a map $f$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ to be a weak equivalence (resp. fibration) if $U(f)$ is a weak equivalence (resp. fibration) in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. Define a map $f$ in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$) to be a cofibration if it has the left lifting property with respect to all acyclic fibrations in ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$).
The model category axioms are verified exactly as in the proof of Theorem \[thm:positive\_flat\_stable\_AlgO\]; this reduces to the verification of Proposition \[prop:homotopical\_analysis\_of\_certain\_pushouts\].
Relations between homotopy categories
-------------------------------------
The purpose of this subsection is to prove the following theorem. This generalization improves the corresponding result in [@Harper_Spectra 1.4] from operads in symmetric spectra to the more general context involving operads in ${{ \mathcal{R} }}$-modules. It plays a key role in this paper.
\[thm:comparing\_homotopy\_categories\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules and let ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) be the category of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) with the model structure of Theorem \[thm:positive\_flat\_stable\_AlgO\] or \[thm:positive\_stable\_AlgO\]. If ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ is a map of operads, then the adjunctions $
\xymatrix@1{
f_*\colon{{ \mathsf{Alg} }}_{{{ \mathcal{O} }}}\ar@<0.5ex>[r] &
{{ \mathsf{Alg} }}_{{{ \mathcal{O} }}'}:f^*\ar@<0.5ex>[l]
}
$ and $
\xymatrix@1{
f_*\colon{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}}\ar@<0.5ex>[r] &
{{ \mathsf{Lt} }}_{{{ \mathcal{O} }}'}:f^*\ar@<0.5ex>[l]
}
$ are Quillen adjunctions with left adjoints on top and $f^*$ the forgetful functor. If furthermore, $f$ is an objectwise stable equivalence, then the adjunctions are Quillen equivalences, and hence induce equivalences on the homotopy categories.
First we make the following observation.
\[prop:preserves\_weak\_equivalences\] Consider ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ and ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure. If $W\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. $W\in{{ \mathsf{SymSeq} }}$) is cofibrant, then the functor $
{{ {-\circ(W)}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }}
$ (resp. $ {{ {-\circ W}\colon\thinspace{{{ \mathsf{SymSeq} }}}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}
$) preserves weak equivalences.
It suffices to consider the case of symmetric sequences. This is verified exactly as in [@Harper_Spectra proof of 5.3], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and Propositions \[prop:cofibration\], \[prop:good\_properties\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and [@Harper_Spectra 4.28, 4.29], respectively.
\[prop:unit\_map\_is\_weak\_equivalence\] Let ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a map of operads in ${{ \mathcal{R} }}$-modules and consider ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) with the positive flat stable model structure. If $Z\in{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. $Z\in{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) is cofibrant and $f$ is a weak equivalence in the underlying category ${{ \mathsf{SymSeq} }}$ with the positive flat stable model structure, then the natural map $Z{{ \longrightarrow }}f^*f_*Z$ is a weak equivalence in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$).
It suffices to consider the case of left ${{ \mathcal{O} }}$-modules. This is verified exactly as in [@Harper_Spectra proof of 5.2], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and Propositions \[prop:cofibration\], \[prop:good\_properties\], \[prop:preserves\_weak\_equivalences\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and [@Harper_Spectra 4.28, 4.29, 5.3], respectively.
This is verified exactly as in [@Harper_Spectra proof of 1.4], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and Proposition \[prop:unit\_map\_is\_weak\_equivalence\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and [@Harper_Spectra 5.2], respectively.
Homotopy colimits and simplicial bar constructions
--------------------------------------------------
The following theorems play a key role in this paper. They improve the corresponding results in [@Harper_Bar] from operads in symmetric spectra to the more general context involving operads in ${{ \mathcal{R} }}$-modules, and are verified exactly as in the proof of [@Harper_Bar 1.10, 1.6, 1.8], respectively.
\[thm:bar\_calculates\_derived\_circle\] Let ${{ {f}\colon\thinspace{{{ \mathcal{O} }}}{{ \longrightarrow }}{{{ \mathcal{O} }}'} }}$ be a morphism of operads in ${{ \mathcal{R} }}$-modules. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and consider ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) with the model structure of Theorem \[thm:positive\_flat\_stable\_AlgO\] or \[thm:positive\_stable\_AlgO\]. If the simplicial bar construction $\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)$ is objectwise cofibrant in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$), then there is a zigzag of weak equivalences $
{{ \mathsf{L} }}f_*(X){{ \ \simeq \ }}|\operatorname{Bar}({{ \mathcal{O} }}',{{ \mathcal{O} }},X)|
$ in the underlying category, natural in such $X$. Here, ${{ \mathsf{L} }}f_*$ is the total left derived functor of $f_*$.
\[main\_hocolim\_theorem\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. If $X$ is a simplicial ${{ \mathcal{O} }}$-algebra (resp. simplicial left ${{ \mathcal{O} }}$-module), then there are zigzags of weak equivalences $$\begin{aligned}
U\operatorname{hocolim}\limits^{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}_{\Delta^{{ \mathrm{op} }}}X & {{ \ \simeq \ }}|U X| {{ \ \simeq \ }}\operatorname{hocolim}\limits_{\Delta^{{ \mathrm{op} }}}U X \\
\Bigl(
\text{resp.}\quad
U\operatorname{hocolim}\limits^{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}_{\Delta^{{ \mathrm{op} }}}X & {{ \ \simeq \ }}|U X| {{ \ \simeq \ }}\operatorname{hocolim}\limits_{\Delta^{{ \mathrm{op} }}}U X
\Bigr)\end{aligned}$$ natural in $X$. Here, $U$ is the forgetful functor, ${{ {{ \mathsf{sAlg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}$) is equipped with the projective model structure inherited from the model structure of Theorem \[thm:positive\_flat\_stable\_AlgO\] or \[thm:positive\_stable\_AlgO\].
\[thm:fattened\_replacement\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathcal{R} }}$-modules. If $X$ is an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), then there is a zigzag of weak equivalences in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$) $$\begin{aligned}
X {{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}}_{\Delta^{{ \mathrm{op} }}}\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)
\quad\quad
\Bigl(
\text{resp.}\quad
X {{ \ \simeq \ }}\operatorname{hocolim}\limits^{{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}}_{\Delta^{{ \mathrm{op} }}}\operatorname{Bar}({{ \mathcal{O} }},{{ \mathcal{O} }},X)
\Bigr)\end{aligned}$$ natural in $X$. Here, ${{ {{ \mathsf{sAlg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{sLt} }}_{{ \mathcal{O} }}}}$) is equipped with the projective model structure inherited from the model structure of Theorem \[thm:positive\_flat\_stable\_AlgO\] or \[thm:positive\_stable\_AlgO\].
Flat stable cofibrations {#sec:flat_stable_cofibrations}
------------------------
The purpose of this subsection is to prove Propositions \[prop:cofibration\] and \[prop:good\_properties\]. This requires several calculations (\[ex:calculation\] and \[calculation\_example\]) together with a characterization of flat stable cofibrations (Proposition \[prop:cofibration\_characterization\]). This characterization is motivated by the characterization given in Schwede [@Schwede_book_project], in terms of left ${{ \mathcal{R} }}_0$–modules, of flat stable cofibrations in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$.
Since ${{ \mathcal{R} }}$ is a commutative monoid in $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$, it follows that ${{ \mathcal{R} }}_0$ is a commutative monoid in $({{ \mathsf{S} }}_*,{{ \,\wedge\, }},S^0)$. In particular, by [@Harper_Modules 2.4] we can regard ${{ \mathcal{R} }}_0$ as a commutative monoid in $({{ \mathsf{S} }}_*^{\Sigma_n},{{ \,\wedge\, }},S^0)$ with the trivial $\Sigma_n$-action.
Let $n\geq 0$. A *left ${{ \mathcal{R} }}_0$-module* is an object in $({{ \mathsf{S} }}_*^{\Sigma_n},{{ \,\wedge\, }},S^0)$ with a left action of ${{ \mathcal{R} }}_0$ and a *morphism of left ${{ \mathcal{R} }}_0$-modules* is a map in ${{ \mathsf{S} }}_*^{\Sigma_n}$ that respects the left ${{ \mathcal{R} }}_0$-module structure. Denote by ${{ \mathcal{R} }}_0-{{ \mathsf{S} }}_*^{\Sigma_n}$ the category of left ${{ \mathcal{R} }}_0$-modules and their morphisms.
For each $n\geq 0$, there is an adjunction $
\xymatrix@1{
{{ \mathsf{S} }}_*^{\Sigma_n}\ar@<0.5ex>[r]^-{{{ \mathcal{R} }}_0{{ \,\wedge\, }}-} &
{{ \mathcal{R} }}_0-{{ \mathsf{S} }}_*^{\Sigma_n}\ar@<0.5ex>[l]
}
$ with left adjoint on top. It is proved in [@Shipley_comm_ring] that the following model category structure exists on left $\Sigma_n$-objects in pointed simplicial sets.
\[def:mixed\_model\_structure\] Let $n\geq 0$.
- The *mixed $\Sigma_n$-equivariant model structure* on ${{ \mathsf{S} }}_*^{\Sigma_n}$ has weak equivalences the underlying weak equivalences of simplicial sets, cofibrations the retracts of (possibly transfinite) compositions of pushouts of maps $$\begin{aligned}
\Sigma_n/H\cdot \partial\Delta[k]_+{{ \longrightarrow }}\Sigma_n/H\cdot \Delta[k]_+ \quad
(k\geq 0,\ H{{ \ \subset\ }}\Sigma_n\ \text{subgroup}),\end{aligned}$$ and fibrations the maps with the right lifting property with respect to the acyclic cofibrations.
Furthermore, it is proved in [@Shipley_comm_ring] that this model structure is cofibrantly generated by generating cofibrations and acyclic cofibrations with small domains, and that the cofibrations are the monomorphisms. It is easy to prove that the category ${{ \mathcal{R} }}_0-{{ \mathsf{S} }}_*^{\Sigma_n}$ inherits a corresponding model structure created by the free-forgetful adjunction above Definition \[def:mixed\_model\_structure\], and that furthermore the diagram category of $(\Sigma_r^{{ \mathrm{op} }}\times G)$-shaped diagrams in ${{ \mathcal{R} }}_0-{{ \mathsf{S} }}^{\Sigma_n}_*$ appearing in the following proposition inherits a corresponding projective model structure. This proposition, whose proof is left to the reader, will be needed for identifying flat stable cofibrations in ${{ \mathsf{SymSeq} }}^G$.
\[prop:mixed\_diagram\_model\_structure\] Let $G$ be a finite group and consider any $n,r\geq 0$. The diagram category $\bigl({{ \mathcal{R} }}_0-{{ \mathsf{S} }}_*^{\Sigma_n}\bigr)^{\Sigma_r^{{ \mathrm{op} }}\times G}$ inherits a corresponding model structure from the mixed $\Sigma_n$-equivariant model structure on ${{ \mathsf{S} }}_*^{\Sigma_n}$. The weak equivalences (resp. fibrations) are the underlying weak equivalences (resp. fibrations) in ${{ \mathsf{S} }}_*^{\Sigma_n}$.
\[defn:overline\_capR\] Define ${{ \overline{{{{ \mathcal{R} }}}} }}\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ such that ${{ \overline{{{{ \mathcal{R} }}}} }}_n:={{ \mathcal{R} }}_n$ for $n\geq 1$ and ${{ \overline{{{{ \mathcal{R} }}}} }}_0:=*$. The structure maps are the naturally occurring ones such that there exists a map of ${{ \mathcal{R} }}$-modules ${{ {i}\colon\thinspace{{{ \overline{{{{ \mathcal{R} }}}} }}}{{ \longrightarrow }}{{{ \mathcal{R} }}} }}$ satisfying $i_n={{ \mathrm{id} }}$ for each $n\geq 1$.
The following calculation, which follows easily from [@Harper_Spectra 2.9], will be needed for characterizing flat stable cofibrations in ${{ \mathsf{SymSeq} }}^G$.
\[ex:calculation\] Let $G$ be a finite group. Let $m,p\geq 0$, $H{{ \ \subset\ }}\Sigma_m$ a subgroup, and $K$ a pointed simplicial set. Recall from the functors $G_p$ and $G^H_m$. Define $X:=G\cdot G_p({{ \mathcal{R} }}{{ \otimes }}G^H_m K)\in{{ \mathsf{SymSeq} }}^G$. Here, $X$ is obtained by applying the indicated functors in to $K$. Then for $r=p$ we have $$\begin{aligned}
({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}X[\mathbf{r}])_n
&{{ \ \cong \ }}\left\{
\begin{array}{rl}
G\cdot\bigl(\Sigma_n\cdot_{\Sigma_{n-m}\times\Sigma_m}{{ \overline{{{{ \mathcal{R} }}}} }}_{n-m}
{{ \,\wedge\, }}(\Sigma_m/H\cdot K)\bigr)\cdot\Sigma_p & \text{for $n>m$,}\\
*&\text{for $n\leq m$,}
\end{array}
\right.\\
X[\mathbf{r}]_n
&{{ \ \cong \ }}\left\{
\begin{array}{rl}
G\cdot\bigl(\Sigma_n\cdot_{\Sigma_{n-m}\times\Sigma_m}{{ \mathcal{R} }}_{n-m}
{{ \,\wedge\, }}(\Sigma_m/H\cdot K)\bigr)\cdot\Sigma_p&\text{for $n>m$,}\\
G\cdot\bigl({{ \mathcal{R} }}_0{{ \,\wedge\, }}(\Sigma_m/H\cdot K)\bigr)\cdot\Sigma_p&\text{for $n=m$,}\\
*&\text{for $n<m$,}
\end{array}
\right.\end{aligned}$$ and for $r\neq p$ we have $X[\mathbf{r}]=*={{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}X[\mathbf{r}]$.
The following characterization of flat stable cofibrations in ${{ \mathsf{SymSeq} }}^G$ is motivated by the characterization given in Schwede [@Schwede_book_project] of flat stable cofibrations in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$. It improves the corresponding characterization given in [@Harper_Spectra 6.6] from the context of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ to the more general context of $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$.
\[prop:cofibration\_characterization\] Let $G$ be a finite group.
- A map ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{SymSeq} }}^G$ with the flat stable model structure is a cofibration if and only if the induced maps $$\begin{aligned}
X[\mathbf{r}]_0{{ \longrightarrow }}Y[\mathbf{r}]_0, &\quad r\geq 0,\ n=0,\\
({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}Y[\mathbf{r}])_n\amalg_{({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}X[\mathbf{r}])_n}X[\mathbf{r}]_n{{ \longrightarrow }}Y[\mathbf{r}]_n, &\quad r\geq 0,\ n\geq 1,\end{aligned}$$ are cofibrations in $\bigl({{ \mathcal{R} }}_0-{{ \mathsf{S} }}^{\Sigma_n}_*\bigr)^{\Sigma_r^{{ \mathrm{op} }}\times G}$ with the model structure in \[prop:mixed\_diagram\_model\_structure\].
- A map ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ in ${{ \mathsf{SymSeq} }}^G$ with the positive flat stable model structure is a cofibration if and only if the maps $
X[\mathbf{r}]_0{{ \longrightarrow }}Y[\mathbf{r}]_0
$, $r\geq 0$, are isomorphisms, and the induced maps $$\begin{aligned}
({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}Y[\mathbf{r}])_n\amalg_{({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}X[\mathbf{r}])_n}X[\mathbf{r}]_n{{ \longrightarrow }}Y[\mathbf{r}]_n, &\quad r\geq 0,\ n\geq 1,\end{aligned}$$ are cofibrations in $\bigl({{ \mathcal{R} }}_0-{{ \mathsf{S} }}^{\Sigma_n}_*\bigr)^{\Sigma_r^{{ \mathrm{op} }}\times G}$ with the model structure in \[prop:mixed\_diagram\_model\_structure\].
This is verified exactly as in [@Harper_Spectra proof of 6.6], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$, Proposition \[prop:mixed\_diagram\_model\_structure\] and Calculation \[ex:calculation\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$, [@Harper_Spectra 6.3] and [@Harper_Spectra 6.5], respectively.
It suffices to consider the case of symmetric sequences. Consider part (b). This is verified exactly as in [@Harper_Spectra proof of 4.29(b)], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and the map $g_*$ obtained by applying the indicated functors in , instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and the map $g_*$ obtained by applying the indicated functors in [@Harper_Spectra (4.1)], respectively. Consider part (a). This is verified exactly as in [@Harper_Spectra proof of 4.29(a)], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$.
\[prop:cofibrations\_to\_mono\] Let $G$ be a finite group. If $B\in{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{G^{{ \mathrm{op} }}}$ (resp. $B\in{{ \mathsf{SymSeq} }}^{G^{{ \mathrm{op} }}}$), then the functor $
{{ {B{{ \,\wedge\, }}_G -}\colon\thinspace{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}^G}{{ \longrightarrow }}{{{ \mathsf{Mod}_{{ \mathcal{R} }}}}} }}
$ (resp. $ {{ {B{{ \check{{{ \otimes }}} }}_G -}\colon\thinspace{{{ \mathsf{SymSeq} }}^G}{{ \longrightarrow }}{{{ \mathsf{SymSeq} }}} }}
$) sends cofibrations in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^G$ (resp. ${{ \mathsf{SymSeq} }}^G$) with the flat stable model structure to monomorphisms.
It suffices to consider the case of symmetric sequences. This is verified exactly as in [@Harper_Spectra proof of 6.11], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and the map $g_*$ obtained by applying the indicated functors in , instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and the map $g_*$ obtained by applying the indicated functors in [@Harper_Spectra (4.1)], respectively.
The following calculation, which follows easily from [@Harper_Spectra 2.9] and , will be needed in the proof of Proposition \[prop:cofibration\] below.
\[calculation\_example\] Let $k,m,p\geq 0$, $H{{ \ \subset\ }}\Sigma_m$ a subgroup, and $t\geq 1$. Let the map ${{ {g}\colon\thinspace{\partial\Delta[k]_+}{{ \longrightarrow }}{\Delta[k]_+} }}$ be a generating cofibration for ${{ \mathsf{S} }}_*$ and define $X{{ \longrightarrow }}Y$ in ${{ \mathsf{SymSeq} }}$ to be the induced map $
\xymatrix@1{
g_*\colon G_p({{ \mathcal{R} }}{{ \otimes }}G^H_m\partial\Delta[k]_+)\ar[r] &
G_p({{ \mathcal{R} }}{{ \otimes }}G^H_m\Delta[k]_+)
}
$. Here, the map $g_*$ is obtained by applying the indicated functors in to the map $g$. For $r=tp$ we have the calculation $$\begin{aligned}
\bigl((Y^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]\bigr)_n
&{{ \ \cong \ }}\left\{
\begin{array}{rl}
\Sigma_n\cdot_{\Sigma_{n-tm}\times H^{\times t}}{{ \mathcal{R} }}_{n-tm}
{{ \,\wedge\, }}(\Delta[k]^{\times t})_+\cdot\Sigma_{tp}&\text{for $n>tm$,}\\
\Sigma_{tm}\cdot_{H^{\times t}}
{{ \mathcal{R} }}_0{{ \,\wedge\, }}(\Delta[k]^{\times t})_+\cdot\Sigma_{tp}&\text{for $n=tm$,}\\
*&\text{for $n<tm$,}
\end{array}
\right.\\
\bigl({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}(Y^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]\bigr)_n
&{{ \ \cong \ }}\left\{
\begin{array}{rl}
\Sigma_n\cdot_{\Sigma_{n-tm}\times H^{\times t}}{{ \overline{{{{ \mathcal{R} }}}} }}_{n-tm}
{{ \,\wedge\, }}(\Delta[k]^{\times t})_+\cdot\Sigma_{tp}&\text{for $n>tm$,}\\
*&\text{for $n\leq tm$,}
\end{array}
\right.
\end{aligned}$$ $$\begin{aligned}
\bigl(Q_{t-1}^t[\mathbf{r}]\bigr)_n
&{{ \ \cong \ }}\left\{
\begin{array}{rl}
\Sigma_n\cdot_{\Sigma_{n-tm}\times H^{\times t}}{{ \mathcal{R} }}_{n-tm}
{{ \,\wedge\, }}\partial(\Delta[k]^{\times t})_+\cdot\Sigma_{tp}&\text{for $n>tm$,}\\
\Sigma_{tm}\cdot_{H^{\times t}}
{{ \mathcal{R} }}_0{{ \,\wedge\, }}\partial(\Delta[k]^{\times t})_+\cdot\Sigma_{tp}&\text{for $n=tm$,}\\
*&\text{for $n<tm$,}
\end{array}
\right.\\
\bigl({{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}Q_{t-1}^t[\mathbf{r}]\bigr)_n
&{{ \ \cong \ }}\left\{
\begin{array}{rl}
\Sigma_n\cdot_{\Sigma_{n-tm}\times H^{\times t}}{{ \overline{{{{ \mathcal{R} }}}} }}_{n-tm}
{{ \,\wedge\, }}\partial(\Delta[k]^{\times t})_+\cdot\Sigma_{tp}&\text{for $n>tm$,}\\
*&\text{for $n\leq tm$,}
\end{array}
\right.
\end{aligned}$$ and for $r\neq tp$ we have $(Y^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]=*={{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}(Y^{{{ \check{{{ \otimes }}} }}t})[\mathbf{r}]$ and $Q_{t-1}^t[\mathbf{r}]=*={{ \overline{{{{ \mathcal{R} }}}} }}{{ \,\wedge\, }}Q_{t-1}^t[\mathbf{r}]$.
It suffices to consider the case of symmetric sequences. Consider part (a). This is verified exactly as in [@Harper_Spectra proof of 4.28(a)], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$, the map $g_*$ obtained by applying the indicated functors in , Proposition \[prop:cofibration\_characterization\], and Calculation \[calculation\_example\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ the map $g_*$ obtained by applying the indicated functors in [@Harper_Spectra (4.1)], [@Harper_Spectra 6.6 and 6.15], respectively. The acyclic cofibration assertion follows immediately from [@Harper_Modules 7.19]. Consider part (b). This is verified exactly as in [@Harper_Spectra proof of 4.28(b)], except using $({{ \mathsf{Mod}_{{ \mathcal{R} }}}},{{ \,\wedge\, }},{{ \mathcal{R} }})$ and Proposition \[prop:cofibrations\_to\_mono\] instead of $({{ \mathsf{Sp}^\Sigma }},{{ \otimes }}_S,S)$ and [@Harper_Spectra 6.11], respectively.
The following will be needed in other sections of this paper.
\[prop:generating\_cofibration\] Let $t\geq 1$. If ${{ {i}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ is a generating cofibration in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}$ (resp. ${{ \mathsf{SymSeq} }}$) with the positive flat stable model structure, then $Q_{t-1}^t{{ \longrightarrow }}Y^{\wedge t}$ (resp. $Q_{t-1}^t{{ \longrightarrow }}Y^{{{ \check{{{ \otimes }}} }}t}$) is a cofibration between cofibrant objects in ${{ \mathsf{Mod}_{{ \mathcal{R} }}}}^{\Sigma_t}$ (resp. ${{ \mathsf{SymSeq} }}^{\Sigma_t}$) with the positive flat stable model structure.
It suffices to consider the case of symmetric sequences. This follows immediately from the proof of Proposition \[prop:cofibration\].
Operads in chain complexes over a commutative ring {#sec:chain_complexes_over_a_commutative_ring}
==================================================
The purpose of this section is to observe that the main results of this paper remain true in the context of unbounded chain complexes over a commutative ring, provided that the desired model category structures exist on algebras (resp. left modules) over operads ${{ \mathcal{O} }}$ and $\tau_k{{ \mathcal{O} }}$. Since the constructions and proofs of the theorems are essentially identical to the arguments above in the context of ${{ \mathcal{R} }}$-modules, modulo the obvious changes, the arguments are left to the reader.
\[assumption:commutative\_ring\_notation\] From now on in this section, we assume that ${{ {\mathcal{K}} }}$ is any commutative ring.
Denote by $({{ \mathsf{Ch} }}_{{ {\mathcal{K}} }},{{ \otimes }},{{ {\mathcal{K}} }})$ the closed symmetric monoidal category of unbounded chain complexes over ${{ {\mathcal{K}} }}$ [@Hovey; @MacLane_homology].
\[HomotopicalAssumption\_chain\_complexes\] If ${{ \mathcal{O} }}$ is an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$, assume that the following model structure exists on ${{ \mathsf{Alg} }}_{\tilde{{{ \mathcal{O} }}}}$ (resp. ${{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}$) for $\tilde{{{ \mathcal{O} }}}={{ \mathcal{O} }}$ and $\tilde{{{ \mathcal{O} }}}=\tau_k{{ \mathcal{O} }}$ for each $k\geq 1$: the model structure on ${{ \mathsf{Alg} }}_{\tilde{{ \mathcal{O} }}}$ (resp. ${{ \mathsf{Lt} }}_{\tilde{{{ \mathcal{O} }}}}$) has weak equivalences the homology isomorphisms (resp. objectwise homology isomorphisms) and fibrations the maps that are dimensionwise surjections (resp. objectwise dimensionwise surjections).
\[CofibrancyCondition\_chain\_complexes\] If ${{ \mathcal{O} }}$ is an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$, consider the unit map ${{ {\eta}\colon\thinspace{I}{{ \longrightarrow }}{{{ \mathcal{O} }}} }}$ of the operad ${{ \mathcal{O} }}$ and assume that $I[\mathbf{r}]{{ \longrightarrow }}{{ \mathcal{O} }}[\mathbf{r}]$ is a cofibration ([@Harper_Bar 3.1]) between cofibrant objects in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}^{\Sigma_r^{{ \mathrm{op} }}}$ for each $r\geq 0$.
If ${{ {\mathcal{K}} }}$ is any field of characteristic zero, then Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\] and Cofibrancy Condition \[CofibrancyCondition\_chain\_complexes\] are satisfied by every operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$ (see [@Harper_Modules; @Hinich]). In the case of algebras over operads, if ${{ {\mathcal{K}} }}$ is any commutative ring and ${{ \mathcal{O} }}'$ is any non-$\Sigma$ operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$, then it is proved in [@Harper_Modules; @Hinich] that the corresponding operad ${{ \mathcal{O} }}={{ \mathcal{O} }}'\cdot\Sigma$ satisfies Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\].
The following is a commutative rings version of Definitions \[defn:homotopy\_completion\] and \[defn:quillen\_homology\].
\[defn:homotopy\_completion\_chain\_complexes\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$ such that ${{ \mathcal{O} }}[\mathbf{0}]=*$. Assume that ${{ \mathcal{O} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\]. Let $X$ be an ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module). The *homotopy completion* $X^{{ {{ \mathsf{h} }}\wedge }}$ of $X$ is the ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) defined by $
X^{{ {{ \mathsf{h} }}\wedge }}:=\operatorname{holim}^{{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}(X^c)\bigr)
$ (resp. $
X^{{ {{ \mathsf{h} }}\wedge }}:=\operatorname{holim}^{{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}_k \bigl(\tau_k{{ \mathcal{O} }}\circ_{{ \mathcal{O} }}X^c\bigr)
$) the homotopy limit of the completion tower of the functorial cofibrant replacement $X^c$ of $X$ in ${{ {{ \mathsf{Alg} }}_{{ \mathcal{O} }}}}$ (resp. ${{ {{ \mathsf{Lt} }}_{{ \mathcal{O} }}}}$). The *Quillen homology complex* (or Quillen homology object) ${{ \mathsf{Q} }}(X)$ of $X$ is the ${{ \mathcal{O} }}$-algebra $\tau_1{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}(X)$ (resp. left ${{ \mathcal{O} }}$-module $\tau_1{{ \mathcal{O} }}\circ^{{ \mathsf{h} }}_{{ \mathcal{O} }}X$).
The following is a commutative rings version of Theorem \[thm:finiteness\].
\[thm:finiteness\_commutative\_rings\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$ such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Assume that ${{ \mathcal{O} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\] and Cofibrancy Condition \[CofibrancyCondition\_chain\_complexes\]. Let $X$ be a $0$-connected ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module) and assume that ${{ \mathcal{O} }}$ is $(-1)$-connected and $H_k{{ \mathcal{O} }}[\mathbf{r}],U{{ {\mathcal{K}} }}$ are finitely generated abelian groups for every $k,r$.
- If the Quillen homology groups $H_k{{ \mathsf{Q} }}(X)$ (resp. $H_k{{ \mathsf{Q} }}(X)[\mathbf{r}]$) are finite for every $k,r$, then the homology groups $H_k X$ (resp. $H_k X[\mathbf{r}]$) are finite for every $k,r$.
- If the Quillen homology groups $H_k{{ \mathsf{Q} }}(X)$ (resp. $H_k{{ \mathsf{Q} }}(X)[\mathbf{r}]$) are finitely generated abelian groups for every $k,r$, then the homology groups $H_k X$ (resp. $H_k X[\mathbf{r}]$) are finitely generated abelian groups for every $k,r$.
Here, $U$ denotes the forgetful functor from commutative rings to abelian groups.
The following is a commutative rings version of Theorem \[thm:hurewicz\].
\[thm:hurewicz\_commutative\_rings\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$ such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Assume that ${{ \mathcal{O} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\] and Cofibrancy Condition \[CofibrancyCondition\_chain\_complexes\]. Let $X$ be a $0$-connected ${{ \mathcal{O} }}$-algebra (resp. left ${{ \mathcal{O} }}$-module), $n\geq 0$, and assume that ${{ \mathcal{O} }}$ is $(-1)$-connected.
- The Quillen homology complex ${{ \mathsf{Q} }}(X)$ is $n$-connected if and only if $X$ is $n$-connected.
- If the Quillen homology complex ${{ \mathsf{Q} }}(X)$ is $n$-connected, then the natural Hurewicz map $H_k X{{ \longrightarrow }}H_k{{ \mathsf{Q} }}(X)$ is an isomorphism for $k\leq 2n+1$ and a surjection for $k=2n+2$.
The following is a commutative rings version of Theorem \[thm:relative\_hurewicz\].
\[thm:relative\_hurewicz\_commutative\_rings\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$ such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Assume that ${{ \mathcal{O} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\] and Cofibrancy Condition \[CofibrancyCondition\_chain\_complexes\]. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) and $n\geq 0$. Assume that ${{ \mathcal{O} }}$ is $(-1)$-connected.
- If $X,Y$ are $0$-connected, then $f$ is $n$-connected if and only if $f$ induces an $n$-connected map ${{ \mathsf{Q} }}(X){{ \longrightarrow }}{{ \mathsf{Q} }}(Y)$ on Quillen homology complexes.
- If $X,Y$ are $(-1)$-connected and $f$ is $(n-1)$-connected, then $f$ induces an $(n-1)$-connected map ${{ \mathsf{Q} }}(X){{ \longrightarrow }}{{ \mathsf{Q} }}(Y)$ on Quillen homology complexes.
- If $f$ induces an $n$-connected map ${{ \mathsf{Q} }}(X){{ \longrightarrow }}{{ \mathsf{Q} }}(Y)$ on Quillen homology complexes between $(-1)$-connected objects, then $f$ induces an $(n-1)$-connected map $X^{{ {{ \mathsf{h} }}\wedge }}{{ \longrightarrow }}Y^{{ {{ \mathsf{h} }}\wedge }}$ on homotopy completion.
- If the Quillen homology complex ${{ \mathsf{Q} }}(X)$ is $(n-1)$-connected, then homotopy completion $X^{{ {{ \mathsf{h} }}\wedge }}$ is $(n-1)$-connected.
Here, ${{ \mathsf{Q} }}(X){{ \longrightarrow }}{{ \mathsf{Q} }}(Y)$, $X^{{ {{ \mathsf{h} }}\wedge }}{{ \longrightarrow }}Y^{{ {{ \mathsf{h} }}\wedge }}$ denote the natural induced zigzags in the category of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules) with all backward facing maps weak equivalences.
The following is a commutative rings version of Theorem \[MainTheorem\].
\[MainTheorem\_commutative\_rings\] Let ${{ \mathcal{O} }}$ be an operad in ${{ \mathsf{Ch} }}_{{ {\mathcal{K}} }}$ such that ${{ \mathcal{O} }}[\mathbf{0}]$ is trivial. Assume that ${{ \mathcal{O} }}$ satisfies Homotopical Assumption \[HomotopicalAssumption\_chain\_complexes\] and Cofibrancy Condition \[CofibrancyCondition\_chain\_complexes\]. Let ${{ {f}\colon\thinspace{X}{{ \longrightarrow }}{Y} }}$ be a map of ${{ \mathcal{O} }}$-algebras (resp. left ${{ \mathcal{O} }}$-modules).
- If $X$ is $0$-connected and ${{ \mathcal{O} }}$ is $(-1)$-connected, then the natural coaugmentation $X{{ \ \simeq \ }}X^{{ {{ \mathsf{h} }}\wedge }}$ is a weak equivalence.
- If the Quillen homology complex ${{ \mathsf{Q} }}(X)$ is $0$-connected and ${{ \mathcal{O} }}$ is $(-1)$-connected, then the homotopy completion spectral sequence $$\begin{aligned}
E^1_{-s,t} &= H_{t-s}\Bigl(i_{s+1}{{ \mathcal{O} }}\circ^{{{ \mathsf{h} }}}_{\tau_1{{ \mathcal{O} }}}\bigl({{ \mathsf{Q} }}(X)\bigr)\Bigr)
\Longrightarrow
H_{t-s}\bigl(X^{{{ {{ \mathsf{h} }}\wedge }}}\bigr)\\
\text{resp.}\quad
E^1_{-s,t}[\mathbf{r}] &= H_{t-s}\Bigl(\bigl(i_{s+1}{{ \mathcal{O} }}\circ^{{{ \mathsf{h} }}}_{\tau_1{{ \mathcal{O} }}}{{ \mathsf{Q} }}(X)\bigr)[\mathbf{r}]\Bigr)
\Longrightarrow
H_{t-s}\bigl(X^{{{ {{ \mathsf{h} }}\wedge }}}[\mathbf{r}]\bigr),\quad\text{$r\geq 0$},\end{aligned}$$ converges strongly.
- If $f$ induces a weak equivalence ${{ \mathsf{Q} }}(X){{ \ \simeq \ }}{{ \mathsf{Q} }}(Y)$ on Quillen homology complexes, then $f$ induces a weak equivalence $X^{{ {{ \mathsf{h} }}\wedge }}{{ \ \simeq \ }}Y^{{ {{ \mathsf{h} }}\wedge }}$ on homotopy completion.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An explicitly energy–conserving full orbit code CUEBIT, developed originally to describe energetic particle effects in laboratory fusion experiments, has been applied to the problem of proton acceleration in solar flares. The model fields are obtained from solutions of the linearised MHD equations for reconnecting modes at an X–type neutral point, with the additional ingredient of a longitudinal magnetic field component. To accelerate protons to the highest observed energies on flare timescales, it is necessary to invoke anomalous resistivity in the MHD solution. It is shown that the addition of a longitudinal field component greatly increases the efficiency of ion acceleration, essentially because it greatly reduces the magnitude of drift motions away from the vicinity of the X-point, where the accelerating component of the electric field is largest. Using plasma parameters consistent with flare observations, we obtain proton distributions extending up to $\gamma$-ray-emitting energies ($>1\,$MeV). In some cases the energy distributions exhibit a bump–on–tail in the MeV range. In general, the shape of the distribution is sensitive to the model parameters.'
---
\#1[=2 0pt 442pt 8pt 434pt \#1]{} \#1[=2 15mm 125mm 15mm 125mm \#1]{}
**FIELD–GUIDED PROTON ACCELERATION AT**
0.2cm
**RECONNECTING X–POINTS IN FLARES**
0.5cm
B. Hamilton$^1$, K. G. McClements$^2$, L. Fletcher$^1$, A. Thyagaraja$^2$
0.5cm
$^1$Department of Physics and Astronomy, University of Glasgow, Glasgow,
G12 8QQ, UK
0.2cm
$^2$UKAEA Culham Division, Culham Science Centre, Abingdon, Oxfordshire,
OX14 3BD, UK
0.5cm
0.2cm
Accepted for publication in [*Solar Physics*]{} February 20 2003
0.2cm
Introduction
============
The process of magnetic reconnection is believed to be intrinsic to solar flares. It is invoked as the mechanism whereby energy-loaded magnetic fields can reconfigure to a lower energy state, liberating the energy which powers particle acceleration, heating, and mass motions. As the resistivity of solar coronal material is very low, the large reconnection rate required to power a flare demands the presence of reconnecting structures with very small length scales – namely, current sheets. Both in solar and magnetospheric physics, the study of particle acceleration in such structures has received considerable attention over the past few decades. Under non-steady state conditions, a reconnecting magnetic field produces an inductive electric field $\mathbf{E}$: the resistive magnetohydrodynamic (MHD) form of Ohm’s law indicates that in general this $\mathbf{E}$ has a component parallel to the local magnetic field, and can thus accelerate charged particles. An example of such a non-steady state is a perturbed reconnecting X-type neutral point, for which Craig and McClymont (1991) calculated the normal modes in a strictly 2-D field, ${\bf B}(x,y)$. In this paper we demonstrate that the analysis of Craig and McClymont is applicable when a $z$-invariant component of the magnetic field is included, enabling it to be extended to configurations more likely to be representative of conditions in the flaring solar corona.
In the context of solar flare physics, a current sheet is often envisaged as arising in a two-ribbon flare, where reconnection of oppositely-directed, predominantly vertical magnetic fields results in a sheet structure having a length (parallel to the solar surface) on the order of the length of a post-flare arcade ($\sim 10^7$m), a vertical extent (height perpendicular to the solar surface) comparable to this, and a thickness on the order of a few times the ion gyroradius. Particle acceleration - in particular proton acceleration - in just such a (collisionless) reconnecting current sheet was addressed by Martens (1998). Following from the magnetotail work of Speiser (1965) he considered the effects of the inclusion of a component of magnetic field perpendicular to the plane of the current sheet, deriving expressions for the maximum energy attained before the gyromotion around this perpendicular field causes the protons to exit the sheet, and hence the acceleration region. With reconnection electric field strengths comparable with those observed (e.g. Kopp and Poletto, 1986), it was found that the proton energy and flux budgets for flares were compatible with those arising from acceleration in a macroscopic current sheet (for example formed in the wake of a rising filament, with vertical extent on the order of $10^7$m, comparable to its length).
Litvinenko and Somov (1993) also considered the current sheet geometry, demonstrating the effects of the addition of a further magnetic field component parallel to the direction of the reconnection electric field. A component of the charged particles’ motion is then gyration about this parallel field, which allows them to stay longer in the current sheet, and attain higher energies. It is this fact which also becomes important in our studies.
While a reconnecting current sheet with finite (and large) vertical extent is frequently used in considering acceleration in large two-ribbon flares, where such a configuration is expected on the basis of numerical simulations of filament lift-off, we concern ourselves here with a current sheet of zero vertical extent, i.e. a coronal X–line, which corresponds to the initial stages of reconnection behind a filament. In the simplest case, this configuration is identical in cross-section \[the $(x,y)$ plane\] to a 2–D X–type neutral point, but is invariant in the $z$–direction. In reality, true magnetic nulls ($B \to 0$) are likely to be rare in the corona. The elaboration of this scenario which we study here, and find to have important consequences for particle acceleration, is the inclusion of a component of magnetic field also in the $z$–direction.
Proton acceleration at an X–point with finite $B_z$ was investigated by Bulanov (1980), Mori, Sakai, and Zhao (1998), Browning and Vekstein (2001), and Bruhwiler and Zweibel (1992). The present study differs from that of previous workers in that the accelerating electric field $E_z$ is obtained self–consistently from a magnetic flux function corresponding to a reconnecting eigenmode of the X–point. This perturbation solution was invoked in the context of particle acceleration also by Petkaki and MacKinnon (1997), but without the addition of the finite $B_z$ component.
A major challenge for any flare acceleration mechanism is to obtain sufficient high energy particles in a short timescale; new observations from the RHESSI satellite are putting ever tighter constraints on timescales, spectra and total energies. The typical requirements for protons, as summarised by Miller (1998) are as follows: they are accelerated up to energies of $\sim$ 100 MeV on timescales of about a second, and to about a GeV on timescales of a few seconds. Proton acceleration lasts for several tens of seconds, at a rate of $\sim 10^{35}{\rm s}^{-1}$, such that the total energy content in protons above an MeV is $\sim 10^{24}$ J. Given a coronal density of $10^{15}-10^{16}{\rm m^{-3}}$ and volume of $\sim
10^{21}{\rm m^{-3}}$, this implies that each second around 1-10% of all coronal protons must be accelerated to MeV energies (and therefore require rapid replenishing - we do not address this here). We find that the inclusion of a moderate longitudinal field component greatly assists in this process.
This paper is structured in the following way. In Section 2 we describe our new algorithm for particle calculations. Section 3 describes the model employed in the simulations. Some sample simulation results and the effect of varying parameters of the simulation, such as the strength of the $z$ component of the magnetic field, are described in Section 4 and we end with discussions and conclusions in Section 5.
Energy–Conserving Algorithm
===========================
The nonrelativistic Lorentz force equations $$m\frac{d\mathbf{v}}{dt} = Ze\mathbf{v} \times
\mathbf{B}(\mathbf{x})+Ze\mathbf{E}(\mathbf{x}), \;\;\; \frac{d\mathbf{x}}
{dt} = \mathbf{v}, \eqno (1)$$ are approximated in the CUEBIT (CUlham Energy–conserving orBIT) code by the following finite difference equations $$m\frac{\mathbf{v}^{i+1} - \mathbf{v}^i}{\Delta t}
=Ze\left(\frac{\mathbf{v}^{i+1}+\mathbf{v}^i}{2}\right)
\times\mathbf{B}\left(\frac{\mathbf{x}^{i+1} +
\mathbf{x}^i}{2}\right)
+ Ze\mathbf{E}\left(\frac{\mathbf{x}^{i+1} + \mathbf{x}^{i}}{2}\right), \eqno
(2)$$ $$\frac{\mathbf{x}^{i+1} - \mathbf{x}^i}{\Delta t} =
\frac{\mathbf{v}^{i+1} + \mathbf{v}^i}{2}. \eqno (3)$$ Here $m$, $Ze$ denote particle mass and charge. At the start of each timestep $\mathbf{x}^{i+1}$ is set equal to $\mathbf{x}^{i}$ in a first calculation of $\mathbf{v}^{i+1}$. A few iterations are made which converge quickly to a final $\mathbf{v}^{i+1}$.
Mori, Sakai and Zhao (1998) also used a scheme in which ${\bf v}$ was set equal to $({\bf v}^{i+1}+{\bf v}^i)/2$ on the right hand side of the Lortentz force equation. In the special case where $\mathbf{E}=0$, the scalar product of the right hand side of Equation (2) with $\mathbf{v}^{i+1}+\mathbf{v}^i$ is identically zero, so that $(v^{i+1})^2=(v^i)^2$. The scheme thus conserves energy exactly. This makes it possible to obtain accurate results with relatively long timesteps. A modified version of Equation (2) conserves total energy exactly when $\mathbf{E}$ is a finite potential electric field. For non-potential electric fields, such as those arising from magnetic reconnection, we have found the method remains accurate for large timesteps (exceeding the Larmor period in most of the computational domain), allowing large numbers of particles to be simulated for timescales that are relevant for flare acceleration. CUEBIT has also been benchmarked by using it to compute energetic particle orbits in magnetic fusion experiments (Wilson et al., 2002), and will in the future be used to study particle transport under various conditions (e.g. in the presence of turbulent electromagnetic fields) in such experiments.
For the simulations discussed in Section 4 it is not necessary to incorporate relativistic kinematics in Equation (2) as the energies reached are only a small fraction of the particles’ rest mass energy (we consider only proton acceleration). However, a relativistic version of the code is currently being developed for the purpose of describing electron acceleration and transport.
Model of a reconnecting X-type structure
========================================
Craig and McClymont solution
----------------------------
For a simple model of a reconnecting field at an X-type neutral point we use the two–dimensional description of Craig and McClymont (1991), with the additional element of a finite third magnetic field component $B_z$. To determine the conditions under which the Craig and McClymont analysis applies with $B_z \ne 0$, we re-derive below their equations for the evolution of a flux function $\psi$, defined such that the curl of $\mathbf{A}\equiv\psi\hat\mathbf{z}$ is equal to the magnetic field in the $(x, y$) plane. The induction equation $$\frac{\partial \mathbf{B}}{\partial t} = \nabla\times
\left(\mathbf{v}\times\mathbf{B}\right)+\frac{\eta}{\mu_{0}}
\nabla^{2} \mathbf{B}, \eqno (4)$$ can be written in the form $$\frac{\partial \mathbf{A}}{\partial t} = \mathbf{v} \times \left(
\nabla \times\mathbf{A} \right) + \frac{\eta}{\mu_{0}}\nabla^{2} \mathbf{A}.
\eqno (5)$$ Here $\mathbf{v}$ is flow velocity, $\eta$ is resistivity (assumed to be constant) and $\mu_0$ is the permeability of free space. The first term on the right hand side of Equation (5) can be expanded to give $$\mathbf{v}\times\left(\nabla\times\mathbf{A}\right) = \nabla
\left(\mathbf{v}\cdot\mathbf{A}\right)-\left(\mathbf{v} \cdot
\nabla\right)\mathbf{A} - \mathbf{A} \times \left(\nabla
\times \mathbf{v}\right) - \left(\mathbf{A} \cdot
\nabla\right)\mathbf{v}. \eqno (6)$$ We assume that flows only occur in the $(x, y)$ plane and that there are no variations in the $z$-direction. In these circumstances the first, third and fourth terms on the right hand side of Equation (6) are all zero and Equation (5) reduces to $$\frac{\partial\psi}{\partial t} + \left(\mathbf{v}\cdot\nabla\right)\psi=
\frac{\eta}{\mu_{0}}\nabla^{2}\psi, \eqno (7)$$ which is the form of the induction equation used by Craig and McClymont.
Neglecting plasma pressure, viscosity and any external forces such as gravity the momentum equation is $$\frac{\partial\mathbf{v}}{\partial t}+(\mathbf{v}\cdot\nabla)\mathbf{v} =
\frac{1}{\rho}\mathbf{j}\times\mathbf{B}, \eqno (8)$$ where $\mathbf{j}$, $\rho$ denote current and mass density. Now in the present case $$\mathbf{j}=\frac{1}{\mu_0}\nabla\times\mathbf{B} = -\frac{\hat\mathbf{z}}
{\mu_0}\nabla^2\psi. \eqno (9)$$ Substituting this expression into Equation (8) and noting that $\nabla\psi$ is in the $(x,y)$ plane, we obtain $$\frac{\partial\mathbf{v}}{\partial t}+(\mathbf{v}\cdot\nabla)\mathbf{v}=
-\frac{1}{\mu_0\rho}\left(\nabla^2\psi\right)\nabla\psi, \eqno (10)$$ i.e. Equation (2.3) in Craig and MyClymont (1991).
Extension of Craig and McClymont solution to $B_z \ne 0$
--------------------------------------------------------
We now assume that the $\mathbf{B}$ field is of the form $$\mathbf{B} = \nabla\times(\psi\hat\mathbf{z}) + B_z\hat{\mathbf{z}}, \eqno
(11)$$ where $B_z$ is constant and uniform. As before, we assume that there are no variations in the $z$ direction. Substituting Equation (11) into Equation (4) and writing $\mathbf{A}=\psi\hat\mathbf{z}$ as before gives $$\frac{\partial}{\partial t}\left(\nabla\times\mathbf{A}+B_z\hat{\mathbf{z}}
\right) = \nabla\times\left(\mathbf{v}\times\left(\nabla\times\mathbf{A}+B_z
\hat{\mathbf{z}}\right)\right)+\frac{\eta}{\mu_0}\nabla^2
\left(\nabla\times\mathbf{A}+B_z\hat{\mathbf{z}}\right). \eqno (12)$$ Since $B_z$ is constant and uniform this reduces to $$\frac{\partial}{\partial t}\left(\nabla\times\mathbf{A}\right)=\nabla\times
\left(\mathbf{v}\times\left(\nabla\times\mathbf{A}\right)\right)+\nabla\times
\left[\mathbf{v}\times\left(B_z\hat{\mathbf{z}}\right)\right]+
\frac{\eta}{\mu_{0}}\nabla^2\left(\nabla\times\mathbf{A}\right). \eqno (13)$$ This is equivalent to the induction equation of Craig and McClymont except for the term containing $B_z$ on the right hand side. This term can be written in the form $$\nabla\times\left[\mathbf{v}\times\left(B_z\hat{\mathbf{z}}\right)\right]=-
\left( B_{z} \hat{\mathbf{z}} \right)\nabla \cdot \mathbf{v}. \eqno (14)$$ Since $B_z$ is not necessarily zero, the only way for this quantity to be zero is if $\nabla\cdot\mathbf{v} = 0$, i.e. the plasma flow must be incompressible. This assumption is often invoked in studies of magnetic reconnection (e.g. Biskamp, 2000; Priest and Forbes, 2001), and is a reasonable one for solar flare plasmas. Incompressibility is implicit in the model of Craig and McClymont: they use Equation (10) in a dimensionless form that is only valid if the plasma density is constrained to be uniform and time–independent. In the limit of ideal MHD, it is straightforward to show that the flow must be incompressible when, as in the scenario considered here, there are no variations in the $z$–direction and $B_z$ is finite (Strauss, 1976). With the assumption of incompressibility, Equation (7) is valid for finite $B_z$. Moreover, since $B_z$ is curl-free it does not contribute to the current; since the latter is oriented in the $z$ direction \[Equation (9)\], it follows that $B_z$ does not contribute to the Lorentz force either, and thus the momentum equation of Craig and McClymont \[Equation (10)\] also remains valid. The solutions of these equations derived by Craig and McClymont are therefore applicable when the magnetic field has a uniform and static $z$ component.
Perturbation solution
---------------------
The equilibrium X–type neutral point structure invoked by Craig and McClymont (1991) is described by $$\mathbf{B} = \frac{B_{o}}{R_{o}} \left(y \hat{\mathbf{x}} + x
\hat{\mathbf{y}} \right), \eqno (15)$$ where $B_o$ is the field strength at the (circular) boundary of the system where $R=R_{o}$, $R$ being the radial co-ordinate measured from the z-axis outwards. Perturbations to this field geometry will result in normal modes of oscillation. Following Petkaki and MacKinnon (1997), we consider only the fundamental mode, with zero azimuthal and radial mode numbers. Approximate analytical solutions for the perturbed magnetic and electric fields can be obtained by dividing the system into an ideal outer region and a resistive inner region, separated by a critical radius $$R_{c} = R_{o} \left( \frac{2}{S} \right)^{\frac{1}{2}}, \eqno (16)$$ where $S$, the Lundquist number at the boundary $R=R_o$, is equal to $\mu_0R_ov_{A}/\eta$ where $v_A$ is the Alfvén speed at $R = R_{o}$ . The solution for the perturbed magnetic field in cylindrical polar coordinates can be approximated by (Craig and McClymont, 1991) $$\delta B = -dB\frac{R_{o}}{R}\omega\cos
\left(\omega\ln\frac{R}{R_o}\right)\cos\left(\frac{\omega
v_A}{R_o}t\right)\exp\left(-\frac{\alpha v_A}{R_o}t\right)
\hat{\mathbf{\varphi}}, \eqno (17)$$ for $R > R_{c}$. Here $dB$ is an arbitrary scaling parameter, $\omega$ is a dimensionless mode frequency, $\alpha$ is a dimensionless decay constant and $\hat{\varphi}$ is the unit vector in the azimuthal direction. For $R \le R_{c}$ the magnetic field perturbation is zero. The value of $\omega$ is determined by the requirement that Equation (17) yields $\delta B = 0$ at $R = R_c$, and $\alpha
\simeq \omega^2/2$ (Craig and McClymont, 1991). The electric field resulting from this solution is wholly in the $z$–direction and is given by $$E_z = dB v_A \sin \left( \omega \ln \frac{R}{R_{o}}\right)\exp\left(-
\frac{\alpha v_A}{R_{o}}t\right)\left(\alpha\cos\left(\frac{\omega v_{A}}
{R_o}t\right)+\omega\sin\left(\frac{\omega v_{A}}{R_{o}} t \right)\right),
\eqno (18)$$ for $R > R_c$ and $$E_{z} = -dB v_A \exp \left( - \frac{\alpha v_A}{R_{o}} t
\right)\left( \alpha \cos \left( \frac{\omega v_{A}}{R_{o}} t
\right) + \omega \sin \left( \frac{\omega v_{A}}{R_{o}} t \right)
\right), \eqno (19)$$ for $R \le R_{c}$. Equations (17), (18), and (19) remain valid for an incompressible plasma with finite $B_z$ constant in time and space. We show below that the inclusion of a modest $B_z$ dramatically increases the efficiency with which test particles are accelerated in this magnetic geometry.
Acceleration in a reconnecting X-type structure
===============================================
Choice of parameters
--------------------
We consider specifically the acceleration of test particle protons in a prescribed field of the type discussed in the previous section. It is known that protons are accelerated in flares to energies of at least several tens of MeV in a timescale of the order of one second (Miller, 1998; Aschwanden, 2002). A straightforward calculation indicates that this could be achieved with a parallel electric field $E_{\parallel}$ of around 1$\,$Vm$^{-1}$. Such field strengths are, incidentally, consistent with the parallel electric fields implied by observations of the separation rates of flare ribbons, e.g. Kopp and Poletto (1986). However, as noted recently by Craig and Litvinenko (2002), fields of this magnitude are not consistent with classical Spitzer resistivity. In deriving their reconnecting field solutions, Craig and McClymont (1991) used the resistive MHD form of Ohm’s law, i.e. $E_{\parallel}=\eta j_{\parallel}$ where $j_{\parallel}$ is the parallel component of ${\bf j}$. The maximum possible current density in a hydrogen plasma is $j_{\rm
max}=2nec$, where $n$ is particle density and $c$ is the speed of light. So a coronal density of $10^{15} - 10^{16}\ {\rm m}^{-3}$ implies that $j_{\rm max} \sim 10^5 -10^6\,$Am$^{-2}$. However, if the resistivity were determined by classical electron–ion collisions, and the plasma temperature $T$ were equal to $10^7\,$K, a field of 1$\,$Vm$^{-1}$ would produce a steady–state current density of around $2\times 10^7\,$Am$^{-2}$. This high current cannot be sustained by a classical collisional resistivity. We deduce that protons can only be accelerated to $\gamma$–ray–emitting energies in a reconnecting field of the type invoked by Craig and McClymont if $\eta$ is much higher than the Spitzer value, i.e. it is anomalous. Craig and Litvinenko (2002) have pointed out another reason for ruling out Spitzer resistivity in the flaring corona, namely that it implies a resistive scale length that is smaller than the collisional mean free path.
Anomalous resistivity can result from lower hybrid or ion acoustic micro–turbulence. Whether lower hybrid or ion acoustic waves are more likely to produce the required anomalous resistivity depends on several parameters, including the ratio of plasma pressure to magnetic field pressure, the electron to ion temperature ratio, and the direction of the current relative to the local magnetic field (Aparicio et al., 1998; Biskamp, 2000). In either case, turbulence results from the current density exceeding a certain threshold $\kappa nec_s$ where $c_s$ is the sound speed and $\kappa$ is a numerical factor typically ranging from unity up to about the square root of the ion to electron mass ratio (Kulsrud, 1998; Aparicio, [*et al.*]{} 1998). The anomalous resistivity then prevents the current density from exceeding this threshold: the turbulence increases the effective collisionality of the plasma. In the case of ion acoustic turbulence, for example, there is a strong interaction with protons whose velocity component $v$ parallel to the wave propagation direction is of the order of $c_s$ (Ishihara and Hirose, 1981). The strength of the interaction falls off roughly as $1/v^2$, and is negligibly small for particles lying sufficiently far out in the tail of a Maxwellian distribution (cf. Litvinenko and Somov, 1993). Therefore, some small fraction of the initial proton population is effectively collisionless, and can be described using Equation (1): it is this sub-population of protons which we simulate using CUEBIT. The bulk of the plasma is effectively collisional, due to the anomalous resistivity. The use of a test particle approach requires that any net current associated with accelerated particles is small compared to the current corresponding to the reconnecting field.
For simplicity, the particles in most of our simulations were given zero initial velocity. This is permissible, since the thermal spread of velocities corresponding to a typical coronal temperature is very small compared to the near–relativistic speeds required for protons to excite $\gamma$–ray emission, and we will show that that the final proton energy spectrum above $\sim 10\,$keV in a typical simulation does not change significantly when the $\delta$–function initial velocity distribution is replaced with a $10^7$K Maxwellian.
The choice of parameters in the simulations is dictated in part by the limits imposed by anomalous resistivity, combined with the typical parallel field strengths needed. We assume, for definiteness $n=10^{16}$m$^{-3}$, $R_o=10^7$m and $B_o=0.01\,$T. If the temperature is assumed to be 10$^7$K, and the current is assumed to be limited to $nec_s$ (i.e. $\kappa=1$), the anomalous resistivity required for $E_{\parallel}=1\,$Vm$^{-1}$ is then $1.4 \times 10^{-3}\,\Omega$m and the Lundquist number at the boundary $R=R_o$ (a key parameter in the Craig and McClymont model) is $S \simeq 2\times 10^{10}$. Equations (3.6) and (3.8) of Craig and McClymont (1991) then yield $\omega \simeq 0.13$ and $\alpha \simeq 0.0088$. The scaling parameter $dB$ in Equations (17), (18) and (19) was chosen to give $E_z\simeq1\,$Vm$^{-1}$ at $t=0$ and $R < R_c$ (in this region $E_z \simeq E_{\parallel}$) rising to approximately 1.4Vm$^-1$ after 1$\,$s. For comparison, simulations were also carried out with $E_z=0.1\,$Vm$^{-1}$ and $10\,$Vm$^{-1}$. Strictly speaking, a change in $E_z$ implies a change in $\eta$ and hence $S$ if the limiting current is assumed to be a fixed multiple of $nec_s$. However, since $\omega$ and $\alpha$ have only a weak (logarithmic) dependence on $S$, the same values of these parameters were used in all the simulations.
The only remaining parameter to fix is $B_z$. It follows from the conclusions drawn in Section 3.2 that this can be varied freely without any of the other parameters being affected. Simulations were carried out with $B_z=0\,$T, $10^{-5}\,$T, $10^{-4}\,$T, $10^{-3}\,$T and $10^{-2}\,$T: these are all reasonable values for active region coronal magnetic fields.
Results
-------
In each simulation the trajectories were computed of approximately $10^6$ protons with an initially uniform random distribution of positions in the reconnection region $\left(R\le R_{o}\right)$, and, unless otherwise stated, a $\delta$–function velocity distribution. The duration of each simulation, 1$\,$s, was chosen on the basis that protons appear to be accelerated to tens of MeV on timescales of this order (Miller, 1998; Aschwanden, 2002). The timestep $\Delta t$ was of the order of 100 Larmor periods calculated at $R=R_{o}$ (simulations with shorter timesteps produced essentially identical results); there were approximately 1500 timesteps per simulation. The value of $dB$ was chosen such that the magnitude of the magnetic field perturbation at $R=R_{o}$ and $t=0$ was always $\le 10^{-4}$ T.
Figure 1 shows $R$ and particle energy $E$ as functions of time for a proton initially lying close to the X-point ($R=10$m) in field configurations with $B_z=0$ (solid curves) and $B_z=10^{-4}\,$T (dashed curves). The other field parameters are those listed at the end of the previous section. The proton initially lies at azimuthal angle $\varphi=24^{\circ}$ in the ($x,y$) plane, and has velocity components in the cylindrical coordinate system $v_R=v_{\varphi}=v_z=10^5\,$ms$^{-1}$.
It is immediately clear that the addition of even a very modest $B_z$ has a dramatic effect on the trajectory of the particle in phase space. When $B_z=0$ the total magnetic field is very small close to the X-point (since $\delta
B=0$ in this region). The Lorentz force in the ($x,y$) plane is consequently very weak, the particle is effectively unmagnetised, and hence moves rapidly away from the X-point. There is only a short phase of acceleration (in this case to about 1 keV): far from the X–point, the particle becomes magnetized and, since ${\bf B}$ is strictly perpendicular to ${\bf E}$, there is no further acceleration. However, a longitudinal field of $10^{-4}$T is sufficient to confine the particle to within a few tens of metres of the X-point, with the result that acceleration to energies of 100 keV and above can easily occur. Because $B_z$ is assumed to be uniform and the field components in the ($x,y$) plane are very small, the combined effect of grad–$B$, curvature and ${\bf E}\times{\bf B}$ drifts on the particle trajectory is negligible.
(12.0,6.0) (0.0,-11.5)
0.5cm
Figure 2 shows final proton energy spectra for $B_z=10^{-4}$T and initial electric field in the inner resistive region $E_{0z}=1$Vm$^{-1}$. Two different initial proton distributions were used: a $\delta$-function (faint curve) and a $10^7$K Maxwellian (bold curve). In both cases, protons are accelerated to energies of up to several MeV. The only significant differences between the two distributions occur at energies $< 10\,$keV: above this energy, the final distribution is insensitive to the initial conditions.
(8.0,8.0)(0.0,0.0) (2.5,0.0)
(4.0,5.1)[$T=0$]{} (4.7,6.3)[$T=10^7$K]{}
0.5cm
Figures 3 and 4 show final proton energy distributions for several pairs of values of $B_z$ and $E_{0z}$. In every case a $\delta$-function initial distribution was used. In Figure 3 $B_z=10^{-4}$T and $E_{0z}$ ranges from 0.1 Vm$^{-1}$ to 10 Vm$^{-1}$. Not surprisingly, protons are accelerated to higher energies as $E_{0z}$ is increased. In contrast to the results obtained by Mori, Sakai, and Zhao (1998), the shape of the accelerated proton spectrum depends strongly on the model parameters: although in some cases the protons have a power law spectrum at low energy (1–100 keV for $E_{0z}=1$ Vm$^{-1}$), the power law index varies considerably. It should be noted that the frequencies $f$ plotted in Figures 3 and 4 represent the number of particles at the end of each simulation with energies in a fixed range of values of log$_{10}E$, i.e. $$f(E)d({\rm log}_{10}E) \propto F(E)dE, \eqno (20)$$ where $F$ is the true energy distribution. Thus, $F(E) \propto f(E)/E$. At low energy, the distributions in Figure 3 correspond to $F(E) \propto
E^{-\gamma}$ where $\gamma \le 2$: Mori, Sakai, and Zhao (1998) obtained $\gamma \simeq 2.0-2.2$.
Another striking feature of the distributions in Figure 3 is the formation of a bump-on-tail at high energy. This also occurs in the simulations of Petkaki and MacKinnon (1997), but not in those of Mori, Sakai and Zhao (1998). In general, models in which there is preferential acceleration of a small sub-population of protons to MeV energies and above are more efficient than models which predict monotonic decreasing spectra, since it is only at high energy that direct observational evidence for accelerated protons exists. The occurrence of bumps-on-tail in our simulations and those of Petkaki and MacKinnon, and their absence from those of Mori, Sakai and Zhao (1998), appears to be due to the choice of field configuration used in these studies. Mori and co-workers assumed a purely hyperbolic magnetic field in the ($x,y$) plane (i.e. $\delta B = 0$) and a uniform $E_z$. With finite $B_z$, there is then a large parallel electric field component $E_{\parallel}=
E_zB_z/B^2$ throughout the computational domain, and all the test particles are susceptible to strong acceleration. In our case, $E_{\parallel}$ is significantly reduced outside the critical radius $R=R_c$, due to the presence of a perturbation to the hyperbolic magnetic field ($\delta B \ne 0$) and also a fall–off in $E_z$ associated with the spatial profile of the reconnecting mode eigenfunction \[cf. Equation (18)\]. Consequently, particles initially lying inside $R=R_c$ are subject to stronger acceleration than those initially lying outside.
(8.0,8.0)(0.0,0.0) (2.5,0.0)
(4.3,2.3)[$0.1\,$Vm$^{-1}$]{} (6.1,3.2)[$1\,$Vm$^{-1}$]{} (8.6,4.5)[$E_{0z} = 10\,$Vm$^{-1}$]{}
0.5cm
Final proton distributions are shown in Figure 4 for $E_{0z}=1~$Vm$^{-1}$ and four different values of $B_z$. Results are not shown for $B_z=0$ since in this case the number of particles accelerated to energies of more than 1 keV was negligible. It can be seen that protons are accelerated to progressively higher energies as $B_z$ is increased – in the case of $B_z=10^{-2}$T, particle energies of up to 40 MeV are observed. This appears to be due simply to improved particle confinement close to the X-point: the amplitude of the sinusoidal variation in Figure 1(a) varies as $1/B_z$ and the drift speed varies as $1/B_z^2$. For $B_z < 10^{-2}$T bump-on-tail formation at high energy is again apparent.
(8.0,8.0)(0.0,0.0) (2.5,0.0)
(5.0,1.8)[$10^{-5}\,$T]{} (6.2,2.8)[$10^{-4}\,$T]{} (7.5,4.0)[$10^{-3}\,$T]{} (10.0,4.5)[$B_{z} = 10^{-2}\,$T]{}
0.5cm
Discussion and Conclusions
==========================
The addition of a longitudinal field component to a two–dimensional reconnecting magnetic configuration massively increases the efficiency of particle acceleration in such configurations. Physically, this is due to the fact that particles close to the magnetic X-point are strongly magnetized by the longitudinal field and are not subject to strong grad–$B$, curvature or ${\bf E}\times{\bf B}$ drifts. The effect of a small but finite longitudinal field $B_z$ on particle acceleration is so dramatic that $B_z=0$ should be regarded as a singular case, unlikely to be representative of the conditions prevailing in the flaring corona. MeV protons can still be produced when $B_z=0$, but the number of particles accelerated to such energies is only significant if the resistivity $\eta$ is assumed to be extremely large – much larger even than the anomalous values corresponding to ion acoustic or lower hybrid turbulence (Petkaki and MacKinnon, 1997). A key result of the present study is that the presence of a longitudinal guide field makes it possible for protons to reach MeV energies in a plasma with realistic values of $\eta$.
In the case of the simulation with $B_z=10^{-2}$T, $E_{0z}=1~$Vm$^{-1}$, 14% of the protons reach energies exceeding 1 MeV in 1s. However, as discussed in Section 4.1, only a fraction of the initial particle population is modelled using the collisionless equations of motion, and so the simulation results imply a super-MeV proton fraction of much less than 14%. As noted in Section 1, observations imply that up to 1-10% of the protons in the flaring corona are accelerated per second. While the test particle method adopted here yields valuable insights into the physics of proton acceleration in flares, a more self-consistent approach may be required to meet the tight constraints on energetic proton fluxes imposed by recent observations.
The sensitivity of the final distributions to $B_z$ (Figure 4) suggests that it might be possible to set constraints on the value of this parameter using $\gamma$–ray observations. However, it is not clear to what extent the simulation results depend on other features of the model, such as the assumption of an azimuthally-symmetric, centrally-peaked perturbation to the magnetic flux. CUEBIT could be used to study particle acceleration in X-point field configurations with a range of different normal mode perturbations. We also intend to study electron acceleration: the numerical scheme in Equations (2) and (3) allows sufficiently large timesteps (relative to the Larmor period) to make this feasible, although it will probably be essential to use a fully relativistic version of the code in this case. Collisions can also be added to the scheme in a straightforward way, enabling simulations to be carried out on longer timescales.
In analytical studies of particle acceleration at reconnecting current sheets Litvinenko and Somov (1993) and Litvinenko (1996) formulated expressions for critical values of $B_z$, such that particles are efficiently accelerated, and for the typical energies which they can reach. Our results are not directly comparable to those of Litvinenko and Somov, since we have invoked a different field configuration. However, our particle code could be applied to the particular current sheet geometry invoked by these authors and their analytical results compared with simulations.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors are grateful to Gordon Emslie and Alec MacKinnon for helpful discussions. BH was supported by an EPSRC CASE studentship, and the work was also funded partly by the UK Department of Trade and Industry.
References {#references .unnumbered}
==========
\[Aparicio, J., Haines, M.G., Hastie, R.J., and Wainwright, J.P.: 1998, Phys. Plasmas [**5**]{}, 3180.\]
\[Aschwanden, M.J.: 2002, [*Space Science Reviews*]{} [**101**]{}, 1.\]
\[Biskamp, D.: 2000, *Magnetic Reconnection in Plasmas*, Cambridge University Press, Cambridge, p. 36.\]
\[Browning, P.K. and Vekstein, G.E.: 2001: [*J. Geophys. Res.*]{} [**106**]{}, 18677.\]
\[Bruhwiler, D.L. and Zweibel, E.G.: 1992, [*J. Geophys. Res.*]{} [**97**]{}, 10825.\]
\[Bulanov, S.V.: 1980, [*Sov. Astron. Lett.*]{} [**6**]{}, 206.\]
\[Craig, I.J.D. and Litvinenko, Yu.E.: 2002, [*Astrophys. J.*]{} [**570**]{}, 387.\]
\[Craig, I.J.D. and McClymont A.N.: 1991, [*Astrophys. J.*]{} [**371**]{}, L41.\]
\[Ishihara, O. and Hirose, A.: 1981, [*Phys. Rev. Lett.*]{} [**46**]{}, 771.\]
\[Kopp, R. A. and Poletto, G.: 1986, in A. Poland (ed.), *Coronal and Prominence Plasmas*, NASA CP 2442, p. 469.\]
\[Kulsrud, R.M.: 1998, [*Phys. Plasmas*]{} [**5**]{}, 1599.\]
\[Litvinenko, Y.E.: 1996, [*Astrophys. J.*]{} [**462**]{}, 997.\]
\[Litvinenko, Y.E. and Somov, B.V.: 1993, [*Solar Phys.*]{} [**146**]{}, 127.\]
\[Martens, P.C.H.: 1988, [*Astrophys. J.*]{} [**330**]{}, L131.\]
\[Miller, J. A.: 1998, [*Space Science Reviews*]{} [**86**]{}, 79.\]
\[Mori, K., Sakai, J., and Zhao, J.: 1998, ApJ [**494**]{}, 430.\]
\[Petkaki, P. and MacKinnon, A.L.: 1997, [*Solar Phys.*]{} [**172**]{}, 279.\]
\[Priest, E.R and Forbes, T.A.: 2001, *Magnetic Reconnection: MHD Theory and Applications*, Cambridge University Press, Cambridge, p. 180.\]
\[Speiser, T.W.: 1965, [*J. Geophys. Res.*]{}, [**70**]{}, 4219.\]
\[Strauss, H.R.: 1976, [*Phys. Fluids*]{}, [**19**]{}, 134.\]
\[Wilson, H.R., et al.: 2002, *Proceedings of the 19th IAEA Fusion Energy Conference*, International Atomic Energy Agency, Vienna, in press (paper no. FT/1-5).\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The main topic of this paper is two folds. First, we compute the first relative cohomology group of the Lie algebra of smooth vector fields on the projective line, ${\mathrm{Vect}}(\mathbb{RP}^1),$ with coefficients in the space of bilinear differential operators that act on tensor densities, ${\cal
D}_{\lambda, \nu;\mu},$ vanishing on the Lie algebra ${\mathrm{sl}}(2,\mathbb{R}).$ Second, we compute the first cohomology group of the Lie algebra ${\mathrm{sl}}(2,\mathbb{R})$ with coefficients in ${\cal
D}_{\lambda, \nu;\mu}.$
author:
- |
Sofiane Bouarroudj\
[Department of Mathematics, U.A.E. University, Faculty of Science]{}\
[P.O. Box 15551, Al-Ain, United Arab Emirates.]{}\
[e-mail:[email protected]]{}
title: 'Cohomology of the vector fields Lie algebras on $\mathbb{RP}^1$ acting on bilinear differential operators'
---
ł
.1truein \[section\] \[thm\][Lemma]{} \[thm\][Corollary]{} \[thm\][Proposition]{} \[thm\][Example]{} \[thm\][Remark]{} \[thm\][Definition]{}
Introduction
============
Let ${\mathfrak g}$ be a Lie algebra and let ${\cal M}$ and ${\cal
N}$ be two ${\mathfrak g}$-modules. It is well-known that nontrivial extensions of ${\mathfrak g}$-modules: $$0\rightarrow {\cal
M}\rightarrow \cdot \rightarrow {\cal N}\rightarrow 0$$ are classified by the first cohomology group ${\mathrm
H}^1({\mathfrak g}; {\mathrm{Hom}}({\cal N},{\cal M}))$ (see, e.g., [@f]). Any 1-cocycle ${\cal L}$ generates a new action on ${\cal M}\oplus {\cal N}$ as follows: for all $g\in {\mathfrak g}$ and for all $(a,b)\in {\cal M}\oplus {\cal N},$ we define $
g^*(a,b):=(g^*a+ C^{st} {\cal L}(b), g^*b).$ For the space of tensor densities of weight $\lambda,$ ${\cal F}_\lambda,$ viewed as a module over the Lie algebra of smooth vector fields ${\mathrm{Vect}}(\mathbb{RP}^1),$ the classification of nontrivial extensions $$0\rightarrow {\cal F}_{\mu}\rightarrow \cdot \rightarrow {\cal
F}_{\lambda}\rightarrow 0,$$ leads Feigin and Fuks in [@gf] to compute the cohomology group ${\mathrm H}^1({\mathrm{Vect}}(\mathbb{RP}^1); {\mathrm{Hom}}({\cal
F}_{\lambda},{\cal F}_{\mu})).$ Later, Ovsienko and the author in [@bo1] have computed the corresponding relative cohomology group with respect to ${\mathrm{sl}}(2,\mathbb{R}),$ namely $$\label{chmes} {\mathrm H}^1({\mathrm{Vect}}(\mathbb{RP}^1),
{\mathrm{sl}}(2,\mathbb{R}); {\mathrm{Hom}}_{\mathrm {diff} } ({\cal
F}_{\lambda},{\cal F}_{\mu})).$$ In fact, the study of the cohomology group (\[chmes\]) has arisen from the study of the equivariant quantization procedure introduced in [@cmz; @lol]. It has been proved that there exists an ${\mathrm{sl}}(2,\mathbb{R})$-equivariant quantization map from the space of symbols and to the space of differential operators, but it is not ${\mathrm{Vect}}(\mathbb{RP}^1)$-equivariant. The obstruction here is given by the 1-cocycles that span the cohomology group (\[chmes\]) (see [@bo1; @gar]). The computation is based on an old result of Gordan [@g] on the classification of ${\mathrm{sl}}(2,\mathbb{R})$-invariant bilinear differential operators that act on tensor densities. Moreover, the case of a higher-dimensional manifold has been studied in [@b; @lo2], and the case of a Riemann surface has been studied in [@bg]. In this paper, we will compute the first cohomology group $${\mathrm H}^1({\mathrm{Vect}}(\mathbb{RP}^1), {\mathrm{sl}}(2,\mathbb{R});
{\mathrm{Hom}}_{\mathrm{diff}} ({\cal F}_{\lambda}\otimes {\cal
F}_{\nu},{\cal F}_{\mu})).$$ It turns out that the dimension of the cohomology group above can reach three, for some particular values of $\lambda$ and $\nu,$ which is a contradistinction with that of the cohomology group (\[chmes\]) in which the dimension is almost one.
Moreover, we compute the cohomology group $${\mathrm H}^1({\mathrm{sl}}(2,\mathbb{R}); {\mathrm{Hom}}_{\mathrm{diff}} ({\cal
F}_{\lambda}\otimes {\cal F}_{\nu},{\cal F}_{\mu})).$$ For linear differential operators, the analogue of the cohomology group above has been studied by Lecomte in [@l].
${\mathrm{Vect}}(\mathbb{R})$-module structures on the space of bilinear differential operators
===============================================================================================
Consider the standard action of ${\mathrm{SL}}(2,\mathbb{R})$ on $\mathbb{RP}^1$ by projective transformations. It is given in homogenous coordinates by $$x\mapsto \frac{ax+b}{cx+d}, \quad \mbox{ where } \left(
\begin{array}{cc}
a & b\\
c & d
\end{array}
\right )\in {\mathrm{SL}}(2,\mathbb{R}).$$ This action generates global vector fields $$\frac{d}{dx},\quad x\frac{d}{dx},\quad x^2\frac{d}{dx},$$ that form a Lie subalgebra of ${\mathrm{Vect}}(\mathbb{RP}^1),$ isomorphic to ${\mathrm{sl}}(2,\mathbb{R})$ (see e.g. [@olv]). Throughout this paper, ${\mathrm{sl}}(2,\mathbb{R})$ will be refered to this subalgebra.
The space of tensor densities on $\mathbb{RP}^1$
------------------------------------------------
The space of tensor densities of weight $\lambda$ on $\mathbb{RP}^1$, denoted by ${\cal F}_{\l}$, is the space of sections of the line bundle $ (T^*\mathbb{RP}^1)^{\otimes \l}.$ This space coincides with the space of functions and differential forms for $\lambda=0$ and for $\lambda=1,$ respectively. The Lie algebra ${\mathrm{Vect}}(\mathbb{RP}^1)$ acts on $\cal F_\l$ by the Lie derivative. For all $X\in {\mathrm{Vect}}(\mathbb{RP}^1)$ and for all $\phi
\in \cal F_\l:$ $$\label{dens} L_{X}^{\l}(\phi)=X\phi' +\l \phi \,X',$$ where the superscript $'$ stands for $d/dx.$
The space of bilinear differential operators as a ${\mathrm{Vect}}(\mathbb{RP}^1)$-module
-----------------------------------------------------------------------------------------
We are interested in defining a three-parameter family of ${\mathrm{Vect}}(\mathbb{RP}^1)$-modules on the space of bilinear differential operators. The counterpart ${\mathrm{Vect}}(\mathbb{RP}^1)$-modules of the space of linear differential operators is a classical object (see e.g. [@w]).
Consider bilinear differential operators that act on tensor densities: $$\label{Op} A:{\cal F}_{\lambda}\otimes {\cal F}_{\nu}\to {\cal
F}_{\mu}.$$ The Lie algebra ${\mathrm{Vect}}(\mathbb{RP}^1)$ acts on the space of bilinear differential operators as follows. For all $\phi \in
{{\mathcal{F}}}_\lambda$ and for all $\psi \in {{\mathcal{F}}}_\nu:$ $$\label{act} L_X^{\l,\nu;\mu}(A)(\phi, \psi)=L_X^\mu \circ A(\phi,
\psi) -A (L_X^\l\,\phi,\psi)-A (\phi,L_X^\nu\,\psi).$$ where $L_X^\l$ is the action (\[dens\]). We denote by ${\cal
D}_{\lambda,\mu;\nu}$ the space of bilinear differential operators (\[Op\]) endowed with the defined ${\mathrm{Vect}}(\mathbb{RP}^1)$-module structure (\[act\]).
Cohomology of ${\mathrm{sl}}(2,{\mathbb{R}})$ acting on ${\cal D}_{\lambda,\nu;\mu}$
====================================================================================
In this section, we will compute the “differentiable” cohomology of the Lie algebra ${\mathrm{sl}}(2,{\mathbb{R}})$ with coefficients in the space of bilinear differential operators ${\cal D}_{\lambda,\nu;\mu}.$ Namely, we consider only cochains that are given by differentiable maps.
\[leco\] (i) If $\mu=\lambda+\nu$ then $$\nonumber \mathrm
H^1({\mathrm{sl}}(2,{\mathbb{R}}); {{\mathcal{D}}}_{\lambda,\nu;\mu})=\mathbb{R}.$$ (ii) If $\mu-\lambda-\nu=k,$ where $k$ is a positive integer, then $$\nonumber \mathrm
H^1({\mathrm{sl}}(2,{\mathbb{R}}); {{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{ll}
{\mathbb{R}}^3,& \mbox{if }\, (\lambda,\nu)=(-\frac{s}{2},-\frac{t}{2}),
\mbox{where } 0\leq s,t\leq k-1, t>k-s-2\\
{\mathbb{R}},& \mbox{otherwise}
\end{array}
\right.$$ (iii) If $\mu-\lambda-\nu$ is not a positive integer, then $$\nonumber \mathrm
H^1({\mathrm{sl}}(2,{\mathbb{R}}); {{\mathcal{D}}}_{\lambda,\nu;\mu})=0.$$
To proof the theorem above, we are required to proof the following two Lemmas.
\[lem1\] Let $Y$ be a vector fields in ${\mathrm{sl}}(2,{\mathbb{R}})$ and let $c:{{\mathcal{F}}}_{\lambda}\otimes {{\mathcal{F}}}_{\nu}\rightarrow {{\mathcal{F}}}_{\mu}$ be a bilinear differential operator defined as follows. For all $\phi\in{{\mathcal{F}}}_{\lambda}$ and for all $\psi\in{{\mathcal{F}}}_{\nu}:$ $$c(\phi,\psi)=\sum_{i+j=k} \alpha_{i,j}\,Y'\,\phi^{(i)}\,\psi^{(j)}
+\sum_{i+j=k-1}\beta_{i,j}\,Y''\,\phi^{(i)}\,\psi^{(j)},$$ where $\alpha_{i,j}$ and $\beta_{i,j}$ are constants. Then, for all $X\in {\mathrm{sl}}(2,{\mathbb{R}}),$ we have $$\begin{array}{ccl}
L_X^{\lambda,\nu;\mu} c(\phi,\psi)&\!\!\!\!=\!\!\!\!&\displaystyle
-\frac{1}{2}\,Y'X'' \sum_{i+j=k-1}\left((i+1)\left (i+2\lambda
\right )\alpha_{i+1,j}+(j+1)\left (j+2\nu
\right) \alpha_{i,j+1}\right )\,\phi^{(i)}\psi^{(j)}\\[2mm]
&& + \displaystyle X'Y''\sum_{i+j=k-1}(\mu-\lambda-\nu -i-j)\,
\beta_{i,j}\,\phi^{(i)}\,\psi^{(j)}\displaystyle + XY''
\sum_{i+j=k} \alpha_{i,j}\,\phi^{(i)}\psi^{(j)}
\end{array}$$
[**Proof.**]{} Straightforward computation using the definition (\[act\]).
\[lem2\] Let $c:{{\mathcal{F}}}_{\lambda}\otimes {{\mathcal{F}}}_{\nu}\rightarrow
{{\mathcal{F}}}_{\mu}$ be a bilinear differential operator defined as follows. For all $\phi\in{{\mathcal{F}}}_{\lambda}$ and for all $\psi\in{{\mathcal{F}}}_{\nu}:$ $$c(\phi,\psi)=\sum_{i+j=k} c_{i,j}\,\phi^{(i)}\,\psi^{(j)},$$ where $c_{i,j}$ are constants. Then, for all $X\in {\mathrm{sl}}(2,{\mathbb{R}}),$ we have $$L_X^{\lambda,\nu;\mu} c(\phi, \psi)=\frac{1}{2}\sum_{i+j=k-1}\left
( (i+1)(2i+\lambda)\,c_{i+1,j}+ (j+1)(2j+\nu)\,c_{i,j+1}\right
)X'' \phi^{(i)}\,\psi^{(j)}$$
[**Proof.**]{} Straightforward computation using the definition (\[act\]).\
Now we are in position to prove Theorem (\[leco\]). Any 1-cocycle on ${\mathrm{sl}}(2,{\mathbb{R}})$ should retains the following general form: $$\label{sam}
c(X,\phi,\psi)=\sum_{j+j=k}\alpha_{i,j}\,X'\phi^{(i)}\psi^{(j)}+
\sum_{j+j=k-1}\beta_{i,j}\,X''\phi^{(i)}\psi^{(j)},$$ where $\alpha_{i,j}$ and $\beta_{i,j}$ are constants. The higher degree terms on $X$ are absent from the formula above, as they vanish on the Lie algebra ${\mathrm{sl}}(2,{\mathbb{R}}).$
The 1-cocycle condition reads as follows: for all $\phi \in
{{\mathcal{F}}}_{\lambda},$ for all $\psi\in {{\mathcal{F}}}_{\nu}$ and for all $X\in{\mathrm{sl}}(2,{\mathbb{R}}),$ we have $$c([X,Y],\phi,\psi)
-L_X^{\lambda,\nu;\mu}\,c(Y,\phi,\psi)+L_Y^{\lambda,\nu;\mu}\,
c(X,\phi,\psi)=0.$$ A direct computation, and by using Lemma \[lem1\], proves that the coefficient of the component $\phi^{(m)}\,\psi^{(n)}$ in the 1-cocycle condition above is equal to $$\label{sl} \frac{1}{2}\,(Y'X''-X'Y'') \left ((m+1)\left(m
+2\lambda \right )\alpha_{m+1,n}+(n+1)\left(n+2\nu \right
)\alpha_{m,n+1}\right ).$$ The formula (\[sl\]) turns into zero once restricted to the affine Lie algebra $\mbox{Span}\{\frac{d}{dx},x\frac{d}{dx}\}.$ We are required, therefore, to study the annihilation of the formula (\[sl\]) for the two vector fields $X=x\frac{d}{dx}$ and $Y=x^2\frac{d}{dx}.$ For these vector fields, the 1-cocycle property will be equivalent to the system $$\label{lam} (m+1)\left(m+2\lambda \right
)\alpha_{m+1,n}+(n+1)\left(n+2\nu \right )\alpha_{m,n+1}=0,$$ where $m+n=k-1.$
Now we are going to deal with trivial 1-cocycles, and show how the general 1-cocycles (\[sam\]) can be eventually trivial. Any trivial 1-cocycle should be of the form $$L_X^{\lambda,\nu;\mu}c,$$ where $c$ is an operator $c:{{\mathcal{F}}}_{\lambda}\otimes
{{\mathcal{F}}}_{\nu}\rightarrow {{\mathcal{F}}}_{\mu}$ defined as $c(\phi,
\psi)=\sum_{i+j=k}c_{i,j}\phi^{(i)}\psi^{(j)}.$ By using Lemma \[lem2\], we have $$\label{tya} L_x^{\lambda,\nu;\mu}c=\frac{1}{2}\,X''\sum_{m+n=k-1}
\left (-(m+1)\left(m +2\lambda \right )c_{m+1,n}-(n+1)\left(n
+2\nu \right )c_{m,n+1}\right )\phi^{(m)}\,\psi^{(n)}.$$ We emphasize on the fact that the component $X'$ is absent from the formula above. To complete the proof we distinguish many cases:
\(i) If $\lambda\not=-\frac{s}{2}$ and $\nu\not=-\frac{t}{2},$ where $s,t\in \{0,\ldots,k-1\},$ then the space of solutions of the system (\[lam\]) is one-dimensional; it is generated by $\alpha_{0,k}.$ Now, we will explain how the constant $\beta_{i,j}$ can be eliminated from our initial cocycle (\[sam\]). We add the coboundary $L_X^{\lambda,\nu;\mu}c $ of equation (\[tya\]) to our 1-cocycle (\[sam\]). The constants $c_{i,j}$ are chosen such that $$\beta_{m,n}=-2((m+1)\left(m+2\lambda \right
)c_{m+1,n}+(n+1)\left(n +2\nu) \right )c_{m,n+1}.$$ This requirement is always possible, as $\lambda\not=-\frac{s}{2}$ and $\nu\not=-\frac{t}{2}.$ Therefore, our 1-cocycle (\[sam\]) should only contain components in $X'.$ Now, by using (\[lam\]) we can see that the cohomology group in question is one-dimensional, generated by the following 1-cocycle. $$\begin{aligned}
{\mathfrak a}(X,\phi,\psi)&=&X'\phi\,\psi^{(k)}\\
&&+\sum_{u+v=k-1}
(-1)^{k-v}\binom{k}{v}\frac{(v+2\nu)(v+1+2\nu)\cdots (k-1+2\nu)
}{(u+2\lambda)(u-1+2\lambda)\cdots
(2\lambda)}\,X'\phi^{(u+1)}\psi^{(v)}. \nonumber\end{aligned}$$ (ii) If $\nu=-\frac{t}{2}$ and $\lambda\not=-\frac{s}{2},$ then the constants $\alpha_{k-t,t},\alpha_{k-t+1,t-1},\ldots,\alpha_{k,0}$ are zero and the space of solutions of the system (\[lam\]) is one-dimensional, generated by $\alpha_{0,k}.$ The constant $\beta_{i,j}$ can be eliminated by the same method as in Part (i). We have just proved that the cohomology group in question is one-dimensional, generated by the 1-cocycle: $${\mathfrak
b}(X,\phi,\psi)=X'\phi\,\psi^{k}+\sum_{u+1+v=k}\alpha_{u+1,v}
\,X'\phi^{(u+1)}\psi^{(v)},$$ where $$\alpha_{u+1,v}=\left\{
\begin{array}{ll}
0,& \mbox{if} \quad v\leq t\\
(-1)^{k-v}\binom{k}{v}\frac{(v+2\nu)(v+1+2\nu)\cdots (k-1+2\nu)
}{(u+2\lambda)(u-1+2\lambda)\cdots (2\lambda)},&\mbox{otherwise}
\end{array}
\right.$$ (iii) If $\lambda=-\frac{s}{2}$ and $\nu\not=-\frac{t}{2},$ then - and as in Part (ii) - the cohomology group in question is one-dimensional, generated by the 1-cocycle: $${\mathfrak
c}(X,\phi,\psi)=X'\phi^{k}\,\psi+\sum_{u+1+v=k}\alpha_{u,v+1}
\,X'\phi^{(u)}\psi^{(v+1)},$$ where $$\alpha_{u,v+1}=\left\{
\begin{array}{ll}
0,& \mbox{if} \quad u\leq s\\
(-1)^{k-u}\binom{k}{u}\frac{(u+2\lambda)(u+1+2\lambda)\cdots
(k-1+2\lambda) }{(v+2\nu)(v-1+2\nu)\cdots
(2\nu)},&\mbox{otherwise}
\end{array}
\right.$$ (iv) If $\lambda=-\frac{s}{2}$ and $\nu=-\frac{k-s-1}{2},$ where $s\in \{0,\ldots,k-1\},$ then the space of solutions of the system (\[lam\]) is two dimensional; it is generated by $\alpha_{s+1,k-s-1}$ and $\alpha_{s,k-s}.$ Now, we will explain how the constant $\beta_{i,j}$ (but not $\beta_{s,k-s-1}$) can be eliminated. We add the coboundary $L_X^{\lambda,\nu}c $ of (\[tya\]) to our 1-cocycle (\[sam\]). The constant $c_{i,j}$ are chosen such that $$\beta_{m,n}=-2((m+1)\left(m+2\lambda \right
)c_{m+1,n}+(n+1)\left(n+2\nu )\right )c_{m,n+1}.$$ This requirement is always satisfied, except for $\beta_{s,k-s-1}$ because the component $\phi^{s}\,\psi^{k-s-1}$ of our trivial 1-cocycle (\[tya\]) has a trivial coefficient. Finally, we have just proved that the cohomology group in question is three-dimensional, generated by the 1-cocycles: $$\begin{array}{ccl}
{\mathfrak d}(X,\phi,\psi)&=&\beta_{s,k-s-1}\,
X''\phi^{s}\,\psi^{k-s-1}+\alpha_{0,k}\,X'\phi\,\psi^{k}+
\alpha_{k,0}\,X'\phi^{k}\,\psi\\[2mm]
&&\displaystyle +\sum_{\substack{u+v=k\\u,v\not= 0}}\alpha_{u,v}
\,X'\phi^{(u)}\psi^{(v)},
\end{array}$$ where $$\alpha_{u,v}=\left\{
\begin{array}{ll}
\displaystyle(-1)^{k-v}\binom{k}{v}\frac{(v+2\nu)(v+1+2\nu)\cdots
(k-1+2\nu) }{(u-1+2\lambda)(u-2+2\lambda)\cdots
(2\lambda)}\alpha_{0,k},& \mbox{if}
\quad u\leq s\\[3mm]
\displaystyle
(-1)^{k-u}\binom{k}{u}\frac{(u+2\lambda)(u+1+2\lambda)\cdots
(k-1+2\lambda) }{(v-1+2\nu)(v-2+2\nu)\cdots
(2\nu)}\alpha_{k,0},&\mbox{if}\quad u\geq s+1
\end{array}
\right.$$ (v) If $\lambda=-\frac{s}{2}$ and $\nu=-\frac{t}{2},$ where $s,t\in \{0,\ldots,k-1\}$ but $ t\leq k-s-2,$ then the space of solutions of the system (\[lam\]) is one-dimensional, generated by $\alpha_{s+1,k-s-1}.$ The constant $\beta_{i,j}$ can be eliminated as explained before. Thus, the cohomology group in question is one-dimensional, generated by the 1-cocycle: $$\begin{array}{ccl}
{\mathfrak e}
(X,\phi,\psi)=X'\,\phi^{(s+1)}\,\psi^{(k-s-1)}\displaystyle
+\sum_{\substack{u+v=k\\u\not= s+1}}\alpha_{u,v}
\,X'\phi^{(u)}\psi^{(v)},
\end{array}$$ where $$\alpha_{u,v}=\left\{
\begin{array}{ll}
0,&\mbox{if}\quad u\leq s\\
0,&\mbox{if}\quad v\leq t\\
(-1)^{k-s-1-v}\frac{(v+1)(v+2)\cdots n }{u(u-1)\cdots s
}\frac{(v+2\nu)(v+1+2\nu)\cdots (k-s-2+2\nu)
}{(u-1+2\lambda)(u-2+2\lambda)\cdots (s+1+2\lambda)},&
\mbox{otherwise}
\end{array}
\right.$$ (vi) If $\lambda=-\frac{s}{2}$ and $\nu=-\frac{t}{2},$ where $s,t\in \{0,\ldots,k-1\}$ but $ t>k-s-2,$ then the space of solutions of the system (\[lam\]) is two-dimensional, generated by $\alpha_{s+1,k-s-1}$ and $\alpha_{k-t-1,t+1}.$ The constant $\beta_{i,j}$ can be eliminated as explained before, except for $\beta_{k-t-1,t}.$ Thus, the cohomology group in question is three-dimensional, generated by the 1-cocycle: $$\begin{array}{ccl}
{\mathfrak f}(X,\phi,\psi)&=&\beta_{k-t-1,t}\,
X''\phi^{k-t-1}\,\psi^{t}+\alpha_{0,k}\,X'\phi\,\psi^{k}+
\alpha_{k,0}\,X'\phi^{k}\,\psi\\[2mm]
&&\displaystyle +\sum_{\substack{u+v=k\\u,v\not= 0}}\alpha_{u,v}
\,X'\phi^{(u)}\psi^{(v)},
\end{array}$$ where $$\alpha_{u,v}=\left\{
\begin{array}{ll}
\displaystyle(-1)^{k-v}\binom{k}{v}\frac{(v+2\nu)(v+1+2\nu)\cdots
(k-1+2\nu) }{(u-1+2\lambda)(u-2+2\lambda)\cdots
(2\lambda)}\alpha_{0,k},& \mbox{if}
\quad v\geq t+1\\[3mm]
\displaystyle
(-1)^{k-u}\binom{k}{u}\frac{(u+2\lambda)(u+1+2\lambda)\cdots
(k-1+2\lambda) }{(v-1+2\nu)(v-2+2\nu)\cdots
(2\nu)}\alpha_{k,0},&\mbox{if}\quad u\geq s+1
\end{array}
\right.$$
[The first cohomology group of the Lie algebra ${\mathrm{sl}}(2,\mathbb{R})$ with coefficients in the space of linear differential operators has been computed in [@l]. The explicit 1-cocycles that span this cohomology group has first arisen in [@gar].]{}
${\mathrm{sl}}(2,{\mathbb{R}})$-invariant differential operators
================================================================
In this section we will investigate differential operators on tensor densities that are ${\mathrm{sl}}(2,\mathbb{R})$-invariant. These results will be useful for the computation of cohomology.
[@g] \[gor\] There exist unique (up to constants) ${\mathrm{sl}}(2,\mathbb{R})$-invariant bilinear differential operators $J_k^{\l,\mu}: {{\mathcal{F}}}_\l\otimes {{\mathcal{F}}}_\nu \rightarrow {{\mathcal{F}}}_{\l+\nu+k}$ given by $$\label{hmida} J_k^{\lambda,\nu}(\phi, \psi ) =\sum_{i+j=k}
c_{i,j}\,\phi^{(i)}\,\psi^{(j)},$$ where the constants $c_{i,j}$ are characterized as follows:
\(i) If $\lambda, \nu \not\in \{0,-1/2,\ldots,-s/2,\ldots\},$ the coefficients $c_{i,j}$ are given by $$c_{i,j}=\displaystyle (-1)^{i} {k \choose
i}\frac{(2\l-i)(2\lambda -i-1)\ldots (2\lambda
-k+1)}{(2\nu-j)(2\nu-j+1)\ldots (2\nu -k+1)}.$$
\(ii) If $\lambda$ or $\nu \in \{0,-1,-1/2,\ldots, -s/2,\ldots\},$ the coefficients $c_{i,j}$ satisfy the recurrence relation $$\label{ak}
(i+1)(i+2\lambda)\,c_{i+1,j}+(j+1)(j+2\nu)\,c_{j,i+1}=0.$$ Moreover, the space of solutions of the system (\[ak\]) is two-dimensional if $\lambda=-\frac{s}{2}$ and $\nu=-\frac{t}{2}$ but $t>k-s-2,$ and one-dimensional otherwise.
\[inv\] There exist ${\mathrm{sl}}(2,\mathbb{R})$-invariant trilinear differential operators $K_k^{\l,\nu,\tau}: {{\mathcal{F}}}_\l\otimes
{{\mathcal{F}}}_\nu\otimes {{\mathcal{F}}}_{\tau} \rightarrow {{\mathcal{F}}}_{\l+\nu+\tau+k}$ given by $$\label{hmida2} K_k^{\lambda,\nu, \tau}(\phi, \varphi, \psi )
=\sum_{i+j+l=k} c_{i,j,l}\,\phi^{(i)}\,\varphi^{(j)}\,\psi^{(l)} ,$$ where the constants $c_{i,j,l}$ are characterized by the recurrence formula $$\label{ars}
i(i-1+2\lambda)\,c_{i,j,l}+(j+1)(j+2\nu)\,c_{i-1,j+1,l}+
(l+1)(l+2\tau)\,c_{i-1,j,l+1}=0,$$ where $i+j+l=k.$
If $\lambda, \nu$ and $\tau$ are generic, then the space of solutions is $(k+1)$-dimensional.
\[inv2\] There exist ${\mathrm{sl}}(2,\mathbb{R})$-invariant trilinear differential operators\
$K_k^{\nu,\tau}: {\mathrm{Vect}}(\mathbb{RP}^1)
\otimes {{\mathcal{F}}}_\nu \otimes {{\mathcal{F}}}_{\tau} \rightarrow {{\mathcal{F}}}_{\nu+\tau+k-1}$ that vanishe on ${\mathrm{sl}}(2,\mathbb{R})$ given by $$\label{hmida2} K_k^{\nu, \tau}(X, \varphi, \psi ) =\sum_{i+j+l=k}
c_{i,j,l}\,X^{(i)}\,\varphi^{(j)}\,\psi^{(l)} ,$$ where the constants $c_{i,j,l}$ are as in (\[ars\]) but $c_{0,j,k-j}=c_{1,j,k-j-1}=c_{2,j,k-j-2}=0.$ Moreover, the space of solutions is $(k-2)$-dimensional, for all $\nu$ and $\tau.$
[**Proof of Proposition (\[inv\]) and (\[inv2\]).**]{} We are going to prove Proposition (\[inv\]) and (\[inv2\]) simultaneously. Any differential operator $K_k^{\l,\nu,\tau}:
{{\mathcal{F}}}_\l\otimes {{\mathcal{F}}}_\nu\otimes {{\mathcal{F}}}_{\tau} \rightarrow {{\mathcal{F}}}_{\mu}$ is of the form $$\nonumber K_k^{\lambda,\nu, \tau}(\phi, \varphi, \psi )
=\sum_{i+j+l=k} c_{i,j,l}\,\phi^{(i)}\,\varphi^{(j)}\,\psi^{(l)} ,$$ where $c_{i,j,l}$ are functions.
The ${\mathrm{sl}}(2,\mathbb{R})$-invariant property of the operators $K_k^{\l,\nu,\tau}$ reads as follows. $$L_X^\nu K_k^{\lambda,\nu, \tau}(\phi, \varphi, \psi
)=K_k^{\lambda,\nu, \tau}(L_X^\lambda\phi, \varphi, \psi )+
K_k^{\lambda,\nu, \tau}(\phi, L_X^\nu\varphi, \psi
)+K_k^{\lambda,\nu, \tau}(\phi, \varphi, L_X^\tau\psi ).$$ The invariant property with respect to the affine Lie algebra $\mathrm{Span}\{\frac{d}{dx}, x\frac{d}{dx}\}$ implies that $c'_{i,j,l}=0$ and $\mu=\l+\nu+\tau+k.$ On the other hand, the invariant property with respect to the vector fields $x^2\frac{d}{dx}$ is equivalent to the system (\[ars\]). If $\lambda,$ $\nu$ and $\tau$ are generic, then the space of solutions is $(k+2)$-dimensional, generated by $c_{k-1,1,0},
c_{k-1,0,1}, c_{k-2,2,0}, c_{k-3,3,0},\ldots, c_{0,k,0}.$
Now, the proof of Proposition (\[inv2\]) follows as above by putting $\lambda=-1.$ In this case, the space of solutions is $(k-2)$-dimensional, spanned by $c_{3,k-3,0}, c_{3,k-4,1},
c_{3,k-5,2},\ldots, c_{3,0,k-3}.$
Cohomology of ${\mathrm{Vect}}(\mathbb{RP}^1)$ acting on ${\cal D}_{\lambda,\nu;\mu}$
=====================================================================================
In this section, we will compute the first cohomology group of ${\mathrm{Vect}}(\mathbb{RP}^1)$ with values in ${{\mathcal{D}}}_{\lambda,\nu;\mu},$ vanishing on ${\mathrm{sl}}(2,\mathbb{R}).$
\[main\] (i) If $\mu-\lambda-\nu=2,$ then $$\nonumber
\mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{ll}
{\mathbb{R}},& \mbox{if} \quad \lambda+\nu+1=0\\
{\mathbb{R}},& \mbox{if} \quad \lambda=0 \mbox{ together with }
\nu\not =-\frac{1}{2}, \mbox{ and vice versa}\\
0,& \mbox{otherwise}
\end{array}
\right.$$ (ii) If $\mu-\lambda-\nu=3,$ then $$\nonumber \mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{lll}
{\mathbb{R}}^2,& \mbox{ if }\quad (\lambda,\nu) = &(0,0), (-2,0), (0, -2),
(-\frac{2}{3}, -\frac{2}{3})\\[2mm]
{\mathbb{R}},&\mbox{ if }\quad (\lambda,\nu)\not =
&(-\frac{1}{2},-1),(-1,-\frac{1}{2}), (-1,-1),\\[2mm]
& \mbox{ }\quad & (0,-1), (-1,0),
(-\frac{1}{2},-\frac{1}{2})\\
0,& \mbox{otherwise}
\end{array}
\right.$$ (iii) If $\mu-\lambda-\nu=4,$ then $$\nonumber \mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{ll}
{\mathbb{R}}^2,& \mbox{ if }\quad (\lambda,\nu) \not= (0,-\frac{3}{2}),
(-\frac{3}{2},0),(-1,-1)\\[2mm]
& (-\frac{1}{2}, -\frac{3}{2}), (-\frac{3}{2}, -\frac{1}{2})
,(-\frac{3}{2},-1),(-1, -\frac{3}{2}),\\[2mm]
&(-\frac{3}{2}, -\frac{3}{2}), (-\frac{1}{2},-1), (-1,
-\frac{1}{2})\\[2mm]
{\mathbb{R}},& \mbox{otherwise}
\end{array}
\right.$$ (iv) If $\mu-\lambda-\nu=5,$ then $$\nonumber
\mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{ll}
{\mathbb{R}}^3,& \mbox{if} \quad (\lambda,\nu) = (0,0),
(-4,0),(0,-4)\\[2mm]
{\mathbb{R}}^2,& \mbox{if} \quad (\lambda,\nu)\not =
(-\frac{1}{2},-\frac{3}{2}),(-\frac{3}{2},-\frac{1}{2}),
(-\frac{1}{2},-2),\\[2mm]& (-2,-\frac{1}{2}),
(-1,-2) (-2,-1),(-1,-\frac{3}{2}),\\[2mm]& (-\frac{3}{2},
-1),(-1,-1)(-\frac{3}{2},-2), (-2,-\frac{3}{2}),\\[2mm] &
(-\frac{3}{2},-\frac{3}{2})\\[2mm]
{\mathbb{R}},& \mbox{otherwise}
\end{array}
\right.$$ (v) If $\mu-\lambda-\nu=6,$ then $$\nonumber
\mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{ll}
{\mathbb{R}}^3,& \mbox{if} \quad (\lambda,\nu) = (
\frac{-5\pm\sqrt{19}}{2},0), (0,
\frac{-5\pm\sqrt{19}}{2}),\\[2mm]
&(\frac{\sqrt{19}-5}{2}, \frac{-\sqrt{19}-5}{2}),
(\frac{-\sqrt{19}-5}{2},\frac{\sqrt{19}-5}{2})\\[2mm]
{\mathbb{R}}^2,& \mbox{if} \quad (\lambda,\nu) \not= (-2,-\frac{1}{2}),
(-\frac{1}{2},-2),
(-\frac{5}{2},-\frac{1}{2}),\\[2mm]
&(-\frac{1}{2},-\frac{5}{2}),(-1,-\frac{3}{2}),
(-\frac{3}{2},-1),(-1,-2),\\[2mm]
& (-2,-1),(-1,-\frac{5}{2}),(-\frac{5}{2},-1),
(-\frac{3}{2},-\frac{3}{2}),\\[2mm]
& (-\frac{3}{2},-2),(-2,-\frac{3}{2}),(-\frac{3}{2},-\frac{5}{2}),
(-\frac{5}{2},\frac{3}{2}),\\[2mm]
& (-2,-2),
(-2, -\frac{5}{2}), (-\frac{5}{2}, -2),\\[2mm]
{\mathbb{R}},& \mbox{otherwise}
\end{array}
\right.$$ (vi) If $\mu-\lambda-\nu=7,$ then $$\nonumber
\mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})= \left\{
\begin{array}{ll}
{\mathbb{R}}^2,& \mbox{if} \quad (\lambda,\nu) =
(\frac{-5-\sqrt{19}}{2},\frac{-5-\sqrt{19}}{2}),
(\frac{\sqrt{19}-5}{2},\frac{\sqrt{19}-5}{2}),\\[2mm]
&(0,\nu), (\lambda,0), (\lambda, -6-\lambda),
(\sqrt{19}-1,\frac{-5-\sqrt{19}}{2}),\\[2mm]
&(\frac{-5-\sqrt{19}}{2},\sqrt{19}-1),
(-\sqrt{19}-1,\frac{-5+\sqrt{19}}{2}),\\[2mm]
&(\frac{-5+\sqrt{19}}{2},-\sqrt{19}-1)\\[2mm]
{\mathbb{R}},& \mbox{if} \quad (\lambda,\nu) \not = (-\frac{5}{2},
-\frac{1}{2}),
(-\frac{1}{2}, -\frac{5}{2}), (-\frac{1}{2}, -3),
\\[2mm]
& (-3,-\frac{1}{2}),(-1,-2), (-2,-1),(-1,-3),\\[2mm]
& (-3,-1), (-\frac{3}{2}, -2),(-2, -\frac{3}{2}),
(-2,-1),\\[2mm]
&(-1,-2),(-2,-\frac{3}{2}), (-\frac{3}{2}, -2),
(-\frac{3}{2},-\frac{5}{2}), \\[2mm]
&(-2,-3),(-3,-2),(-\frac{5}{2},-\frac{3}{2}),
(-2,-\frac{5}{2}),\\[2mm]
& (-\frac{5}{2}, -2), (-3,-\frac{5}{2}),
(-\frac{5}{2}, -3) \\[2mm]
0, & \mbox{otherwise}
\end{array}
\right.$$ (vii) If $\mu-\lambda-\nu$ is not like above but $\lambda$ and $\nu$ are generic then $$\nonumber
\mathrm H^1({\mathrm{Vect}}(\mathbb{RP}^1),{\mathrm{sl}}(2,{\mathbb{R}});
{{\mathcal{D}}}_{\lambda,\nu;\mu})=0.$$
Proof of Theorem (\[main\])
===========================
To proof Theorem (\[main\]) we proceed bye following the three steps:
1. We will investigate the dimension of the space of operators that satisfy the 1-cocycle condition. By Proposition (\[inv2\]), its dimension is at most $k-2,$ where $k=\mu-\lambda-\nu,$ since any 1-cocycle that vanishes on ${\mathrm{sl}}(2,\mathbb{R})$ is certainly ${\mathrm{sl}}(2,\mathbb{R})$-invariant (cf. [@bo1; @lo2]).
2. We will study all trivial 1-cocycles, namely, operators of the form $$L_XB,$$ where $B$ is a bilinear operator. As our 1-cocycles vanish on the Lie algebra ${\mathrm{sl}}(2,\mathbb{R}),$ it follows that the operator $B$ coincides with the transvectant $J_k^{\lambda,\nu}.$ By using Proposition (\[gor\]), we will determine different values of $\lambda$ and $\nu$ for which the space of operators of the form $L_X J_k^{\lambda,\nu}$ is zero, one or two-dimensional.
3. By taking into account Part 1 and part 2 - and depending on $\lambda$ and $\nu$ - the dimension of the cohomology group $\mathrm{H}^1({\mathrm{Vect}}(\mathbb{RP}), {\mathrm{sl}}(2,\mathbb{R});
{{\mathcal{D}}}_{\lambda,\nu;\mu})$ will be equal to $$\mbox{dim(operators that are 1-cocycles)}-\mbox{dim(operators of
the form } L_XJ_{k}^{\lambda,\nu})$$
We need also the following Lemma.
\[fun\] Every 1-cocycle on $\mathrm{Vect}(\mathbb{RP}^1)$ with values in ${{\mathcal{D}}}_{\lambda,\nu;\mu}$ is differentiable.
[**Proof.**]{} See [@lo2].
Now we are in position to prove Theorem (\[main\]). By Lemma (\[fun\]), any 1-cocycle on ${\mathrm{Vect}}(\mathbb{RP}^1)$ should retains the following general form: $$\label{sam2}
c(X,\phi,\psi)=\sum_{j+j+l=k}\alpha_{l,i,j}\,X^{(l)}\phi^{(i)}\psi^{(j)},$$ where $\alpha_{l,i,j}$ are constants. The fact that this 1-cocycle vanishes on ${\mathrm{sl}}(2,\mathbb{R})$ implies that $$c_{0,i,j}=c_{1,i,j}=c_{2,i,j}=0.$$ The 1-cocycle condition reads as follows: for all $\phi \in {{\mathcal{F}}}_{\lambda},$ for all $\psi\in
{{\mathcal{F}}}_{\nu}$ and for all $X\in{\mathrm{Vect}}(\mathbb{RP}^1),$ one has $$c([X,Y],\phi,\psi)
-L_X^{\lambda,\nu;\mu}\,c(Y,\phi,\psi)+L_Y^{\lambda,\nu;\mu}\,
c(X,\phi,\psi)=0.$$
The case when $k=2$
-------------------
If $\mu-\lambda-\nu=2,$ equation (\[sam2\]) shows that only one 1-cocycle spans the cohomology group of Theorem (\[main\]); it is given by $$\label{co3}
{\mathfrak L}(X,\phi,\psi):=X'''\,\phi\,\psi.$$ Let us study the triviality of this 1-cocycle. A direct computation proves that $$L_XJ_2^{\lambda,\nu}(\phi,\psi)=(-\lambda c_{2,0}-\nu
c_{0,2})X'''\,\phi\,\psi,$$
1. If $\lambda=0$ and $\nu\not= -\frac{1}{2},$ then by Part (ii) of Proposition (\[gor\]), $c_{0,2}=0$ and the 1-cocycle is not trivial. The result holds when $\nu=0$ and $\lambda\not=
-\frac{1}{2}$ as well.
2. If $\lambda=0$ and $\nu=-\frac{1}{2},$ then $c_{0,2}=-\frac{1}{\nu}$ and therefore $L_XJ_2^{0,-\frac{1}{2}}(\phi, \psi)={\mathfrak L}(X,\phi, \psi).$ Hence, the 1-cocycle is trivial. The result holds true when $\nu=0$ and $\lambda= -\frac{1}{2}.$
3. If $\lambda$ and $\nu$ are not like above, then Part (i) of Proposition (\[gor\]) implies that $$-\lambda \,c_{2,0}-\nu\,
c_{0,2}=-2\nu\frac{\lambda+\nu+1}{1+2\lambda}c_{1,1},$$ where $c_{1,1}\not =0.$ Thus, for $\lambda+\nu+1=0$ the 1-cocycle (\[co3\]) is not trivial; otherwise, the 1-cocycle is trivial.
The case when $k=3$
-------------------
If $\mu-\lambda-\nu=3,$ equation (\[sam2\]) shows that the 1-cocycles that span the cohomology group of Theorem (\[main\]) are of the form $$\label{co3} {\mathfrak
M}(X,\phi,\psi):=\gamma\,X^{(4)}\,\phi\,\psi+(\alpha_{1}\,\phi'\,\psi
+\alpha_{2}\,\phi\,\psi')\,X''',$$ where $\gamma, \alpha_1$ and $\alpha_2$ are constants. The 1-cocycle condition implies that $$\gamma=-\frac{\lambda\,\alpha_1+\nu\,\alpha_2}{2}.$$ Let us study the triviality of this 1-cocycle. A direct computation proves that $$\begin{array}{lcl}
L_XJ_3^{\lambda,\nu}(\phi,\psi)&=&(-\lambda \,c_{3,0}-\nu\,
c_{0,3})X^{(4)}\,\phi\,\psi-((1+3\lambda)\,c_{3,0}+\nu\,c_{1,2})
X'''\phi'\,\psi\\[2mm]
&&-((1+3\nu)\,c_{0,3}+\lambda\,c_{2,1}) X'''\phi\,\psi'
\end{array}$$
1. Here we will characterize all values of $\lambda$ and $\nu$ for which $L_X J_3^{\lambda,\nu}=0.$ An easy computation using Proposition (\[gor\]) proves that these values are $(0,0),
(-2,0), (0,-2),$ and $(-\frac{2}{3},-\frac{2}{3}).$ If $(\lambda,\nu)=(0,0),$ then $c_{0,3}=c_{3,0}=0$ and $c_{1,2}=-c_{2,1}.$ It follows that the cohomology group is two-dimensional, spanned by two 1-cocycles given as in (\[co3\]) for $(\alpha_1,\alpha_2)=(1,0)$ and $(0,1).$ The results holds true for $(\lambda,\nu)=(-2,0), (0,-2),$ and $(-\frac{2}{3},-\frac{2}{3}).$
2. Here we will characterize all values of $\lambda$ and $\nu$ for which $L_X J_3^{\lambda,\nu}$ is generated by two parameters. An easy computation using Proposition (\[gor\]) proves that these values are $(-\frac{1}{2},-1), (-1,-\frac{1}{2}),
(0,-1),(-1,0),(-\frac{1}{2},-\frac{1}{2}),$ and $(-1,-1).$ If $(\lambda,\nu)=(-\frac{1}{2},-1),$ then $c_{1,2}=0,$ and $c_{3,0}=\frac{2}{3}c_{2,1};$ however, $c_{0,3}$ and $c_{2,1}$ are arbitrary. Therefore, the constants $c_{0,3}$ and $c_{2,1}$ can be chosen such that $$L_X J^{\lambda,\nu}_3(\phi,\psi)= {\mathfrak M}(X,\phi,\psi).$$ It follows that the cohomology group is trivial. The result holds true for $(0,-1),(-1,0),$ $(-\frac{1}{2},-\frac{1}{2}),(-1,-\frac{1}{2}),$ and $(-1,-1).$
3. If $\lambda$ and $\nu$ are not like above, then the transvectant $J_3$ is unique by Proposition (\[gor\]). Whatever the weights $\lambda$ and $\nu$ can take, the trivial 1-cocycle $
L_X J^{\lambda,\nu}_3$ is never identically zero. Therefore, one of the constants $\gamma,$ $\alpha_1$ or $\alpha_2$ can be eliminated by just adding the trivial 1-cocycle $ L_X
J^{\lambda,\nu}_3.$ Hence, the cohomology group is one-dimensional.
The case when $k=4$
-------------------
If $\mu-\lambda-\nu=4,$ equation (\[sam2\]) shows that the 1-cocycles that span the cohomology group of Theorem (\[main\]) are of the form $$\label{co4} {\mathfrak
N}(X,\phi,\psi):=\gamma\,X^{(5)}\,\phi\,\psi+X^{(4)}\,\left (
\alpha_{1}\,\phi'\,\psi +\alpha_{2}\,\phi\,\psi'\right )+X'''\left
(\beta_1\,\phi''\,\psi+\beta_2\,\phi\,\psi''+\beta_3\,\phi'\,
\psi' \right ),$$ where $\gamma, \alpha_1, \alpha_2, \beta_1, \beta_2$ and $\beta_3$ are constants. The 1-cocycle condition is equivalent to the following system $$\begin{array}{ll}
5\gamma=-\lambda\,\alpha_1-\nu\,\alpha_2,&
2\alpha_1=-(1+2\lambda)\,\beta_1-\nu \,\beta_3,\\
2\alpha_2=-(1+2\nu)\,\beta_2-\lambda\, \beta_3.
\end{array}$$ The space of solutions of the system above is three-dimensional. Let us study the triviality of the 1-cocycle (\[co4\]). A direct computation proves that $$\begin{array}{lcl}
L_XJ_4^{\lambda,\nu}(\phi,\psi)&\!\!\!\!\!\!=\!\!\!\!\!\!&(-\lambda
\,c_{4,0}-\nu\,
c_{0,4})X^{(5)}\,\phi\,\psi-((1+4\lambda)\,c_{4,0}+\nu\,c_{1,3})
X^{(4)}\phi'\,\psi\\[2mm]
&&-((1+4\nu)\,c_{0,4}+\lambda\,c_{3,1}) X^{(4)}\,\phi\,\psi'-
((4+6\lambda)\,c_{4,0}+\nu\, c_{2,2})\,X'''\,\phi''\,\psi\\[2mm]
&&-(((1+3\lambda)\,c_{3,1}+(1+3\nu)\, c_{1,3})\,\phi'\,\psi'-
((4+6\nu)\,c_{0,4}+\lambda\, c_{2,2})\,\phi\,\psi'')\,X'''
\end{array}$$
1. For all values of $\lambda$ and $\nu$ one can easily prove that the equation $L_XJ_4^{\lambda, \nu}=0$ has no solutions.
2. Now we will characterize all values of $\lambda$ and $\nu$ for which $L_X J_4^{\lambda,\nu}$ is generated by two parameters. An easy computation using Proposition (\[gor\]) proves that these values are $(0,-\frac{3}{2}), (-\frac{3}{2},0),
(-\frac{1}{2},-\frac{3}{2}), (-\frac{3}{2},-\frac{1}{2}),
(-\frac{1}{2},-1),(-1,-\frac{1}{2}), (-1,-1), (-1,-\frac{3}{2}),$ $(-\frac{3}{2}, -1),$ and $(-\frac{3}{2}, -\frac{3}{2}).$ If $(\lambda,\nu)=(0,-\frac{3}{2}),$ then the constants $c_{1,3}$ and $c_{0,4}$ are arbitrary. On the other hand, $$c_{2,2}=\frac{3}{2}\,c_{1,3}, \quad c_{3,1}=c_{3,1},\quad
c_{4,0}=\frac{1}{4}\,c_{1,3}.$$ The constant $c_{1,3}$ and $c_{0,4}$ can be chosen in such a way that once adding the trivial 1-cocycle $L_XJ_4^{\lambda,\nu}$ above to our 1-cocycle (\[co4\]), the constants $\beta_1$ and $\beta_2$ disappear completely. Hence, the cohomology group is one-dimensional. The result holds true for the other values of $(\lambda,\nu).$
3. If $\lambda$ and $\nu$ are not like above, then the transvectant $J_4$ is unique by Proposition (\[gor\]). Whatever the weights $\lambda$ and $\nu$ can take, the trivial 1-cocycle $ L_X
J^{\lambda,\nu}_4$ is never identically zero. Therefore, one of the constants $\gamma,$ $\alpha_1,$ $\alpha_2,$ $\beta_1,$ $\beta_2$ or $\beta_3$ can be eliminated by just adding the trivial 1-cocycle $ L_X J^{\lambda,\nu}_4.$ Hence, the cohomology group is two-dimensional.
The case when $k=5$
-------------------
If $\mu-\lambda-\nu=5,$ equation (\[sam2\]) shows that the 1-cocycles that span the cohomology group of Theorem (\[main\]) are of the form $$\begin{aligned}
{\mathfrak
O}(X,\phi,\psi)&:=&\gamma\,X^{(6)}\,\phi\,\psi+X^{(5)}\,\left (
\alpha_{1}\,\phi'\,\psi +\alpha_{2}\,\phi\,\psi'\right
)+X^{(4)}\left
(\beta_1\,\phi''\,\psi+\beta_2\,\phi\,\psi''+\beta_3\,\phi'\,
\psi' \right )\nonumber\\
&&X'''\left
(\eta_1\,\phi^{(3)}\,\psi+\eta_2\,\phi^{(2)}\,\psi'+\eta_3\,\phi'\,
\psi^{(2)}+ \eta_4\,\phi\,\psi^{(3)}\right )\label{co5}\end{aligned}$$ The 1-cocycle condition is equivalent to the following system $$\begin{array}{ll}
5\gamma=-\lambda\,\beta_1-\nu\,\beta_2+\lambda \,\eta_1+\nu\,
\eta_4,
&9\gamma=-\lambda\,\alpha_1-\nu\,\alpha_2,\\
5\alpha_1=-(1+2\lambda)\,\beta_1-\nu \,\beta_3,&
5\alpha_2=-(1+2\nu)\,\beta_2-\lambda\, \beta_3,\\
2\beta_1=-(3+3\lambda)\,\eta_1-\nu\,\eta_2,&
2\beta_2=-\lambda\,\eta_3-(3+3\nu)\,\eta_4\\
2\beta_3=-(1+2\lambda)\,\eta_2-(1+2\nu)\,\eta_3.
\end{array}$$ The space of solutions of the system above is four-dimensional for $(\lambda,\nu)=(0,0),$ $(0,-2),$ $(-2,0), $ $(0,-4),$$
(-4,0),$ and $(-2,-2),$ and three-dimensional otherwise. According to these values, let us study the triviality of the 1-cocycle (\[co5\]). A direct computation proves that $$\begin{array}{lcl}
L_XJ_5^{\lambda,\nu}(\phi,\psi)&\!\!\!\!\!\!\!=\!\!\!\!\!\!\!
&(-\lambda
\,c_{5,0}-\nu\,
c_{0,5})X^{(6)}\,\phi\,\psi-((1+5\lambda)\,c_{5,0}+\nu\,c_{1,4})
X^{(5)}\phi'\,\psi\\[2mm]
&&-((1+5\nu)\,c_{0,5}+\lambda\,c_{4,1}) X^{(5)}\,\phi\,\psi'-
X^{(4)}(((5+10\lambda)\,c_{5,0}-\nu\, c_{2,3})\,\phi''\,\psi\\[2mm]
&&-((1+3\lambda)\,c_{4,1}+(1+3\nu)\, c_{1,4})\,\phi'\,\psi'-
((5+10\nu)\,c_{0,5}+\lambda\,
c_{2,3})\,\phi\,\psi'')\\[2mm]
&&-X'''((\nu\,c_{3,2}+(10+12\lambda)\,c_{5,0})\,\phi^{(3)}\,\psi
-(\nu\, c_{2,3}+(10+12\nu)\,c_{0,5})\,\phi\,\psi^{(3)}\\[2mm]
&&-((1+3\lambda)\, c_{3,2}+(4+6\nu)\,c_{1,4})\,\phi'\,\psi''
-((1+3\nu)\, c_{2,3}+(4+6\lambda)\,c_{4,1})\,\phi''\,\psi')
\end{array}$$
1. If $(\lambda,\nu)=(0,0),$ then the constant $c_{1,4}$ is arbitrary. On the other hand, $$c_{0,5}=c_{5,0}, \quad c_{4,1}=-c_{1,4},\quad
c_{2,3}=6\,c_{4,1}, \quad c_{3,2}=-6\,c_{4,1}.$$ The constant $c_{1,4}$ can be chosen in such a way that once adding the trivial 1-cocycle $L_XJ_5^{\lambda,\nu}$ above to our 1-cocycle (\[co5\]), the constant $\eta_2$ disappears completely. Therefore, our 1-cocycle is generated by $\beta_1, \beta_2$ and $\eta_3.$ Hence, the cohomology group is three-dimensional. The result holds true for $(\lambda,\nu)=(0,-4), (-4,0).$
2. If $(\lambda,\nu)=(0,-2),$ then the constant $c_{4,1}$ and $c_{0,5}$ are arbitrary. On the other hand, $$c_{1, 4} = c_{4, 1}, \quad c_{2, 3} = 2\,c_{4, 1}, \quad
c_{3, 2}= 2\, c_{4, 1}, \quad c_{5, 0}= \frac{1}{5}\,c_{4, 1}.$$ The constant $c_{1,4}$ and $c_{0,5}$ can be chosen in such a way that once adding the trivial 1-cocycle $L_XJ_5^{\lambda,\nu}$ above to our 1-cocycle (\[co5\]), the constant $\eta_1$ and $\eta_4$ disappear completely. Therefore, our 1-cocycle is generated by $\beta_3$ and $\eta_2.$ Hence, the cohomology group is two-dimensional. The result holds true for $(\lambda,\nu)=(-2,0)$ and $(-2,-2).$
3. If $(\lambda,
\nu)=(-\frac{3}{2}, -\frac{1}{2})$ then by Proposition (\[gor\]), the transvectant $J_5$ is not unique. A direct computation proves that the constants $c_{5, 0}$ and $c_{2, 3}$ are arbitrary. On the other hand, $$c_{0, 5} = \frac{c_{2, 3}}{10}, \quad
c_{1, 4}=\frac{c_{2, 3}}{2}, \quad c_{3, 2}= c_{2, 3}, \quad c_{4,
1}=5\,c_{5, 0}.$$ The constant $c_{2,3}$ and $c_{5,0}$ can be chosen in such a way such that once adding the trivial 1-cocycle $L_XJ_5^{\lambda,\nu}$ above to our 1-cocycle (\[co5\]), the constant $\eta_1$ and $\eta_4$ disappear completely. Therefore, our 1-cocycle is generated by $\eta_2.$ Hence, the cohomology group is one-dimensional. The result holds true for $(\lambda,
\nu)=(-\frac{1}{2}, -2), (-1,-1), (-1, -\frac{3}{2}),$ $(-1, -2),$ $(-\frac{1}{2}, -\frac{3}{2}),$ $(-\frac{3}{2}, -1),$ $(-\frac{3}{2}, -\frac{3}{2}), (-\frac{3}{2}, -2), (-2,
-\frac{1}{2}), (-2, -1), (-2, -\frac{3}{2}).$
4. If $\lambda$ and $\nu$ are not like above, then the transvectant $J_5$ is unique by Proposition (\[gor\]). Whatever the weights $\lambda$ and $\nu$ can take, the trivial 1-cocycle $ L_X J^{\lambda,\nu}_5$ is never identically zero. Therefore, one of the constants $\gamma,$ $\alpha_1,$ $\alpha_2,$ $\beta_1,$ $\beta_2,$ $\beta_3,$ $\eta_1, \eta_2,\eta_3$ or $\eta_4$ can be eliminated by just adding the trivial 1-cocycle $ L_X J^{\lambda,\nu}_5.$ Hence, the cohomology group is two-dimensional.
The case when $k=6$
-------------------
If $\mu-\lambda-\nu=6,$ equation (\[sam2\]) shows that the 1-cocycles that span the cohomology group of Theorem (\[main\]) are of the form $$\begin{aligned}
{\mathfrak
P}(X,\phi,\psi)&:=&\gamma\,X^{(7)}\,\phi\,\psi+X^{(6)}\,\left (
\alpha_{1}\,\phi'\,\psi +\alpha_{2}\,\phi\,\psi'\right
)+X^{(5)}\left
(\beta_1\,\phi''\,\psi+\beta_2\,\phi'\,\psi'+\beta_3\,\phi\,
\psi'' \right )\nonumber\\
&&+X^{(4)}\left
(\eta_1\,\phi^{(3)}\,\psi+\eta_2\,\phi''\,\psi'+\eta_3\,\phi'\,
\psi''+ \eta_4\,\phi\,\psi^{(3)}\right )\nonumber \\
&&+X'''\left
(\xi_1\,\phi^{(4)}\,\psi+\xi_2\,\phi^{(3)}\,\psi'+\xi_3\,\phi''\,
\psi''+ \xi_4\,\phi'\,\psi^{(3)} + \xi_5\,\phi\,\psi^{(4)} \right
) \label{co6}\end{aligned}$$ The 1-cocycle condition is equivalent to the following system $$\begin{array}{ll}
14\gamma=-\lambda\,\beta_1-\nu\,\beta_3+\lambda
\,\xi_1+\nu\,\xi_5,
&14\gamma=-\lambda\,\alpha_1-\nu\,\alpha_2,\\
5\alpha_1=-(1+3\lambda)\,\eta_1+(1+4\lambda)\,\xi_1+\nu
(\xi_4-\eta_3),& 9\alpha_1=-(1+2\lambda)\,\beta_1-\nu \beta_2,
\\
5\alpha_2=-(1+3\nu)\,\eta_4+(1+4\nu)\,\xi_5+\lambda
(\xi_2-\eta_2), &
9\alpha_2=-(1+2\nu)\,\beta_3-\lambda\, \beta_2,\\
5\beta_1=-(3+3\lambda)\,\eta_1-\nu\,\eta_2,&
5\beta_3=-\lambda\,\eta_3-(3+3\nu)\,\eta_4\\
5\beta_2=-(1+2\lambda)\,\eta_2-(1+2\nu)\,\eta_3.&
2\eta_1=-(6+4\lambda)\, \xi_1 -\nu \, \xi_2,\\
2\eta_2=-(3+3\lambda)\, \xi_2 -(1+2\nu ) \xi_3,&
2\eta_3=-(1+2\lambda)\, \xi_3 -(3+3\nu)\xi_4,\\
2\eta_4=-(6+4\nu)\, \xi_5 -\lambda \, \xi_4,
\end{array}$$ The space of solutions of the system above is four-dimensional for $(\lambda, \nu)=(-\frac{5}{2}, -\frac{5}{2}),$ $(-\frac{5}{2},0),$ $(0,-\frac{5}{2}),$ $(0, \frac{-5\pm\sqrt{19}}{2}),$ $(
\frac{-5\pm\sqrt{19}}{2},0),$ $(\frac{-5-\sqrt{19}}{2},
\frac{-5+\sqrt{19}}{2}),$ $(\frac{-5+\sqrt{19}}{2},
\frac{-5-\sqrt{19}}{2}),$ and three-dimensional otherwise. According to these values, let us study the triviality of the 1-cocycle (\[co6\]). A direct computation proves that $$\begin{gathered}
\nonumber L_XJ_5^{\lambda,\nu}(\phi,\psi)=(-\lambda
\,c_{6,0}-\nu\,
c_{0,6})X^{(7)}\,\phi\,\psi-((1+6\lambda)\,c_{6,0}+\nu\,c_{1,5})
X^{(6)}\phi'\,\psi\\\nonumber
\phantom{L_XJ_5}{-((1+6\nu)\,c_{0,6}+\lambda\,c_{5,1})
X^{(6)}\,\phi\,\psi'-
((6+15\lambda)\,c_{6,0}+\nu\, c_{2,4})\,X^{(5)}\,\phi''\,\psi}
\qquad\quad\\
\phantom{L_XJ_5}{-(((1+5\lambda)\,c_{5,1}+(1+5\nu)\,
c_{1,5})\,\phi'\,\psi'- ((6+15\nu)\,c_{0,6}+\lambda\,
c_{4,2})\,\phi\,\psi'')\,X^{(5)}}\quad\\
\phantom{L_XJ_5}{-((\nu\,c_{3,3}+(15+20\lambda)\,c_{6,0})\,\phi^{(3)}\,\psi
-((1+4\nu)\, c_{2,4}+(6+9\nu)\,c_{5,1})\,\phi''\,\psi'}\qquad\quad
\\
\phantom{L_XJ_5}{-((1+4\,\lambda)c_{4,2}+(6+9\nu)\,c_{1,5})\,\phi'\,\psi''
-(\lambda\,
c_{3,3}+(15+20\nu)\,c_{0,6})\,\phi\,\psi^{(3)}))\,X^{(4)}}
\\
\phantom{L_XJ_5}{-(((20+15\lambda)\,c_{6,0}+\nu\,c_{4,2})\,\phi^{(4)}\,\psi
-((10+10\lambda)\,
c_{5,1}+(1+3\nu)\,c_{3,3})\,\phi^{(3)}\,\psi'}\quad
\\
\phantom{L_XJ_5}{-((4+6\lambda)\,c_{4,2}+(4+6\nu)\,c_{2,4})\,\phi''\,\psi''
-((1+3\lambda)\, c_{3,3}+(10+10\nu)\,c_{1,5}) \,\phi'\,\psi^{(3)}}\\
\phantom{L_XJ_5}
{-(\lambda\,c_{2,4}+(20+15\nu)\,c_{0,6})\,\phi\,\psi^{(4)})\,X'''}\qquad\qquad
\qquad\qquad\qquad\qquad\qquad\qquad\quad\end{gathered}$$
1. If $(\lambda,\nu)=(0,\frac{-5-\sqrt{19}}{2}),$ then the constant $c_{5,1}$ is arbitrary. On the other hand, $$c_{0, 6}= 0, \quad
c_{1, 5}= -\frac{4}{45}(-35+8\sqrt{19})\,c_{5, 1},
\quad c_{2, 4}= 2(-13+ 3\sqrt{19})\,c_{5, 1},$$ $$c_{3, 3}= -\frac{4}{3}(-31+7\sqrt{19})\,c_{5, 1},
\quad c_{4, 2}= \frac{10}{3}(-4+ \sqrt{19})\,c_{5, 1},
\quad c_{6, 0}= \frac{1}{30}(5+ \sqrt{19})\, c_{5, 1}$$ The constant $c_{5,1}$ can be chosen in such a way such that once adding the trivial 1-cocycle $L_XJ_6^{\lambda,\nu}$ above to our 1-cocycle (\[co6\]), the constant $\xi_4$ disappears completely. Hence, the cohomology group is three-dimensional. The result holds true for $(\lambda,\nu)=(\frac{-5-\sqrt{19}}{2},0),
(\frac{-5+\sqrt{19}}{2},0), (0,\frac{-5+\sqrt{19}}{2}),
(\frac{-5-\sqrt{19}}{2},\frac{-5+\sqrt{19}}{2}),
(\frac{-5+\sqrt{19}}{2},\frac{-5-\sqrt{19}}{2}).$
2. If $(\lambda,\nu)=(-\frac{5}{2},-\frac{5}{2}),$ then the constant $c_{4,2}$ and $c_{1,5}$ are arbitrary. On the other hand, $$c_{0, 6}= 0, \quad c_{1, 5}=-\frac{c_{4, 2}}{10}, \quad
c_{2, 4}=c_{4, 2}, \quad c_{3, 3}= -2\, c_{4, 2}, \quad
c_{5, 1}=-\frac{c_{4, 2}}{10}, \quad c_{6, 0}= 0.$$ The constant $c_{4,2}$ and $c_{1,5}$ can be chosen in such a way that once adding the trivial 1-cocycle $L_XJ_6^{\lambda,\nu}$ above to our 1-cocycle (\[co6\]), two of the constants disappear completely. Hence, the cohomology group is two-dimensional. The result holds true for $(\lambda,\nu)=(-\frac{5}{2},0)$ , $(0,-\frac{5}{2}).$
3. If $(\lambda,\nu)=(-\frac{1}{2}, -2),$ then by Proposition (\[gor\]), the transvectant $J_6$ is not unique. A direct computation proves that $$c_{1, 5}=6\,c_{0, 6}, \quad c_{2, 4}=\frac{5}{2}\,c_{5, 1}, \quad
c_{3, 3}=\frac{10}{3}\, c_{5, 1}, \quad c_{4, 2}= \frac{5}{2}
c_{5, 1}, \quad c_{6, 0}=\frac{1}{6}\,c_{5, 1}.$$ The constant $c_{0,6}$ and $c_{5,1}$ can be chosen in such a way that once adding the trivial 1-cocycle $L_XJ_6^{\lambda,\nu}$ above to our 1-cocycle (\[co6\]), the constant $\xi_2$ and $\xi_5$ disappear completely. Hence, the cohomology group is one-dimensional. The result holds true for $(\lambda,\nu)=(-2,-\frac{1}{2}),$ $(-\frac{1}{2},-\frac{5}{2}),$ $ (-\frac{5}{2}, -\frac{1}{2}),$ $(-1, -\frac{3}{2}),$ $(-\frac{3}{2}, -1),$ $(-1, -2),$ $(-2, -1),$ $(-\frac{5}{2},
-1),$ $ (-1,-\frac{5}{2}),$ $(-\frac{3}{2},-\frac{3}{2}),$ $(-2,
-\frac{3}{2}),$ $(-\frac{3}{2},-2),$ $(-\frac{3}{2},
-\frac{5}{2}),$ $(-\frac{5}{2}, -\frac{3}{2}), (-2,-2),$ $(-2,
-\frac{5}{2}),$ and $(-\frac{5}{2}, -2).$
4. If $\lambda$ and $\nu$ are not like above, then the transvectant $J_6$ is unique by Proposition (\[gor\]). Whatever the weights $\lambda$ and $\nu$ can take, the trivial 1-cocycle $ L_X J^{\lambda,\nu}_6$ is never identically zero. Therefore, one of the constants $\gamma,
\alpha_1,\ldots$ can be eliminated by just adding the trivial 1-cocycle $ L_X J^{\lambda,\nu}_6.$ Hence, the cohomology group is two-dimensional.
The case when $k=7$
-------------------
The proof here is the same as in the previous section. We just point out that the space of solutions of the 1-cocycle property (\[sam2\]) is four-dimensional for $(\lambda, \nu)=(0,-3), (-3,
0),$ and $(-3,-3);$ three-dimensional for $(\lambda,
\nu)=(-2,-2),$ $ (\lambda,0),$ $ (0, \nu),$ $
(\lambda,-6-\lambda),$ $ (-\frac{3}{2}, -3),$ $
(-3,-\frac{3}{2}),$ $ (-\frac{3}{2}, -\frac{3}{2}), $ $(-\frac{5}{2}, -1), (-1, -\frac{5}{2}),$ $ (-\frac{5}{2},
-\frac{5}{2}),$ $(\sqrt{19}-1,\frac{-5-\sqrt{19}}{2}),$$
(\frac{-5-\sqrt{19}}{2},\sqrt{19}-1),$\
$
(-\sqrt{19}-1,\frac{-5+\sqrt{19}}{2}),$$
(\frac{-5+\sqrt{19}}{2},-\sqrt{19}-1),
(\frac{-5-\sqrt{19}}{2},\frac{-5-\sqrt{19}}{2}),
(\frac{-5+\sqrt{19}}{2},\frac{-5+\sqrt{19}}{2}),$ and two-dimensional otherwise.
The case when $k\geq 8$
-----------------------
For $k\geq 8,$ the number of variables generating any 1-cocycle is much smaller than the number of equations coming out from the 1-cocycle condition - for instance, for $k=8$ the $\sharp
\mbox{(Variables)}=28,$ while $\sharp \mbox{(Equations)}=33.$ For generic $\lambda$ and $\nu,$ the number of equations will generates a one-dimensional space, which give a unique cohomology class. This cohomology class is indeed trivial because the expression $L_XJ_k^{\lambda, \nu}$ is also a 1-cocycle.
[For $k\geq 8$ and for particular values of $\lambda$ and $\nu,$ the cohomology group $\mathrm{H}^1({\mathrm{Vect}}(\mathbb{RP}^1),
{\mathrm{sl}}(2,\mathbb{R}); {{\mathcal{D}}}_{\lambda,\nu;\mu})$ may not be trivial. For instance, for $k=8$ we have $$\mathrm{H}^1({\mathrm{Vect}}(\mathbb{RP}^1), {\mathrm{sl}}(2,\mathbb{R});
{{\mathcal{D}}}_{0,-\frac{7}{2};\frac{9}{2}})\simeq \mathbb{R}.$$]{}
Acknowledgments {#acknowledgments .unnumbered}
---------------
I am grateful to V. Ovsienko for his constant supports.
[99]{} S. Bouarroudj, [Projective and conformal Schwarzian derivatives and cohomology of Lie algebras vector fields related to differential operators,]{} Preprint.
S. Bouarroudj & H. Gargoubi, Projectively invariant cocycles for holomorphic vector fields on an open Riemann surface. [*Tokyo J. of Maths.*]{} 2002, vol. 25, no. 1, 33–40.
S. Bouarroudj & V. Ovsienko, [Three cocycles on ${\mathrm{Diff}}(S^1)$ generalizing the Schwarzian derivative]{}, [*Internat. Math. Res. Notices*]{} [**1998**]{}, No.1, 25–39.
E. Cartan, [*Leçons sur la théorie des espaces à connexion projective,*]{} Gauthier -Villars, Paris - 1937.
Cohen P, Manin Yu and Zagier D, [Automorphic pseudodifferential operators,]{} [*Algebraic aspects of integrable systems,*]{} 17–47, Progr. Nonlinear Differential Equations Appl., 26, Birkhäuser Boston, Boston, MA, 1997.
C. Duval & V. Ovsienko, [Conformally equivariant quantum Hamiltonians.]{} [*Selecta Math. (N.S.)*]{} [**7**]{} (2001), no. 3, 291–320.
B. L. Feigin & D. B. Fuchs, [Homology of Lie algebras on vector fields on the line,]{} [*Funkts. Anal. Prilozhen.,*]{} 16, No. 2, 47-63 (1982).
D. B. Fuks, [*Cohomology of infinite-dimensional Lie algebras*]{}, Contemp. Soviet. Math., Consultants Bureau, New-York, 1986.
H. Gargoubi, [Sur la géométrie de l’espace des opérateurs différentiels linéaires sur $\mathbb{R}.$]{} [*Bull. Soc. Roy. Sci. Liège.*]{} Vol. 69, 1, 2000, 21–47.
P. Gordan, [*Invariantentheorie,*]{} teubner, Leipzig, 1887.
P. B. A. Lecomte & V. Ovsienko, [Cohomology of the vector fields Lie algebra and modules of differential operators on a smooth manifold,]{} [*Compositio Mathematica.*]{} [**124:**]{} No.1, 2000, 95–110.
P. B. A. Lecomte & V. Ovsienko, [Projectively invariant symbol calculus]{}, [*Lett. Math. Phy.*]{} [**49**]{} (3) (1999), 173–196.
P. B. A. Lecomte, [On the cohomology of ${\mathrm{sl}}(n+1,\mathbb{R})$ acting on differential operators and ${\mathrm{sl}}(n+1,\mathbb{R})$-equivariant symbols,]{} [*Indag. Math. NS.*]{} 11 (1), 2000, 95-114.
P. Olver, [*Applications of Lie groups to differential equations.*]{} [Springer Verlag.]{}
E. J. Wilczynski, [*Projective differential geometry of curves and ruled surfaces,*]{} Leipzig - Teubner - 1906.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Lyonell Boulton$^1$\
[`[email protected]`]{}
- |
Maria Pilar Garcia del Moral$^2$\
[`[email protected]`]{}
- |
Alvaro Restuccia$^{2,3}$\
[`[email protected]`]{}
date: '17/3/15'
title: 'Massless ground state for a compact $SU(2)$ matrix model in 4D'
---
Introduction
============
Determining the ground state of the $(0+1)$ matrix models is a longstanding open problem [@halpern; @hoppe; @dwhn]. This is a subject of interest in different areas, including: matrix models [@bfss], Yang-Mills theories [@yi; @sethi-stern; @porrati] and M-theory [@hoppe; @fh; @bgmr]. In this paper we show the existence of the ground state wavefunction of a matrix model which takes values on a compact space and is subject to the constraints associated to gauge symmetry. This model will serve as a benchmark in order to illustrate a new method to demonstrate the existence and uniqueness of massless supersymmetric ground state wavefunctions of supersymmetric matrix models at finite temperature.
Boundary conditions can change dramatically the spectral properties of field theories and at the same time they may imply that certain symmetries may be partially or even totally broken. Indeed, for supersymmetric theories ($N=2$) only periodic, Dirichlet and Neumann boundary conditions preserve partial supersymmetry $N=1$, [@manolo]. These type of models can be of interest to test certain aspects of AdS/CFT dualities, such as characterizing black holes at finite temperature. Yang-Mills or QED fields on a box have been largely considered, for example in bag-models or compact QED models [@abls], to study confinement properties of hadrons, phase transitions and chiral symmetry breaking. In a similar fashion as QCD, compact QED in a strong-coupling region exhibits charge confinement as well as spontaneous chiral symmetry breaking. Both these aspects are thought to be related to the existence of monopoles. Many authors have studied this problem in the past, see for example [@Takahashi] and the references there in. The nature of the phase transitions in compact QED has in fact been under debate for a long time.
The deconfinement temperature of finite-temperature compact electrodynamics in $2+1$ dimensions has been shown to be insensitive to external fields [@ics]. Matrix models of the compact QED are likely to be relevant in the analysis of these theories, subject to a slow-mode regime. Effectively, they serve as toy models of such a phenomenon. From the M-theory point of view, $SU(2)$ matrix models correspond to a supermembrane (regularized via $SU(2)$) propagating in a 4-dimensional Minkowski space-time [@dwhn; @dwln]. The $SU(2)$ ground state wave function has been examined also in [@hlt2; @su2; @hlt]. The question of whether a unique ground state with zero energy exits for a Minkowski spacetime was not completely settled in [@dwhn].
In [@octonions] we addressed the existence and uniqueness of the ground state wavefunction for unconstrained models restricted to Dirichlet boundary conditions. Determining the ground state wave function from the point of view of the supermembrane is an open question which was posed when the model was originally formulated. This is expected to corresponds to an 11D wavefunction constructed in terms of the 11D supermultiplet of supergravity. The hamiltonian of the supermembrane has two independent contributions one associated to the movement of the center of mass in 11D Minkowski the space-time and a second one associated to the supermembrane excitations. The existence and uniqueness of the ground state wave function of the 11D supermembrane in the case when the massive excitations are forced to lie in a compact space with Dirichlet boundary conditions will be analyzed elsewhere [@bgmr6].
In this paper we examine a particular model, the $SU(2)$ gauge supermembrane. Our goal is to draft the relevant step towards finding the ground state wave function of the regularized supermembrane on a 11D flat background[^1], assuming that the center of mass of the supermembrane propagates freely in the 4D spacetime but its membrane excitations are confined to a compact space of arbitrary large radius $R$. We will show that, given a Dirichlet boundary condition, there exists a unique massless ground state wave function for the mass operator of the model.
The paper is organized as follows. In Section 2 we present the $SU(2)$ matrix model, the mass operator, the constraint associated to the local $SU(2)$ symmetry and the supersymmetric algebra. In Section 3 we formulate the Dirichlet problem. In Section 4 we prove the existence and uniqueness of the ground state. A final section is devoted to highlighting our main conclusions.
The $SU(2)$ regularized matrix model
====================================
The model we will consider was introduced in [@dwhn] and it arises from the 11D mass operator by taking all the fields to zero $X^{mA}=\lambda_{\alpha'}^A=0$ for $m=1,\dots,7$ and $\alpha'=1,\dots,7$, except those transforming under $U(1)$ via $(Z^A,\overline{Z}^A, \lambda^A_8=\lambda^A)$. In conjunction with the light cone gauge fields $X^+$ and $X^-$, those fields describe a regularized supermembrane propagating in a 4D space-time. We will also restrict to the simplest nonabelian gauge symmetry, given by $G=SU(2)$. The ground state wave function expressed as a superfield admits and expansion in the superfields that does not admit a factorization. The solution may have an even or odd number of odd grassmanian coordinates. The even ones have the expression $$\label{wf3}
\Psi=\phi_0(Z,\overline{Z})+\epsilon^{ABC}\phi^A(Z,\overline{Z})\lambda^B\lambda^C.$$ This is constructed in terms of four functions $(\phi_0,\phi_A)$ with $A=1,2,3$.
The associated hamiltonian is $$H=-\nabla^2+V_B+V_F= -\frac{\partial^2}{\partial Z_A\partial\overline{Z}^A}+ V(Z, \overline{Z},\lambda)$$ with $$\begin{aligned}V_B &=\frac{1}{4}\epsilon^E_{AB}\epsilon_{CDE}\left(2Z^A\overline{Z}^B\overline{Z}^CZ^D\right) \qquad \text{and}\\
V_F&=\frac{1}{\sqrt{2}}\epsilon_{ABC}\left(Z^A\lambda^B\lambda^C-\overline{Z}^A\frac{\partial}{\partial\lambda_B}\frac{\partial}{\partial\lambda_C}\right),\end{aligned}$$ subject to the first class constrain $\varphi^A\vert \Psi \rangle=0$ where $$\varphi^A=\epsilon^{ABC}\left(Z^B\frac{\partial}{\partial Z^C}+\overline{Z}^B\frac{\partial}{\partial \overline{Z}^C}+\lambda_B\frac{\partial}{\partial\lambda_C}\right).$$ The associated supercharges are the following $$\begin{aligned}
Q&=\sqrt{2}\frac{\partial}{\partial Z^A}\frac{\partial}{\partial\lambda_A}-\epsilon_{ABC}Z^A\overline{Z}^B\lambda^C\\
Q^{\dag}&=-\sqrt{2}\frac{\partial}{\partial \overline{Z}^A}\lambda^A+\epsilon_{ABC}Z^A\overline{Z}^B\frac{\partial}{\partial\lambda^C}
\end{aligned}$$ and the superalgebra satisfies the conditions $$\{Q,Q\}=2\sqrt{2}\overline{Z}^A\varphi_A,\quad
\{Q^{\dagger},Q^{\dagger}\}=2\sqrt{2}Z^A\varphi_A \quad \text{and} \quad
\{Q,Q^{\dagger}\}=2H.$$
Existence and uniqueness of a solution in a compact domain
==========================================================
We now consider the existence and uniqueness of the ground state wave function for a system that is restricted by a first class constraint given by the $SU(2)$ gauge condition. We proceed in a similar fashion as in [@octonions], but the presence of a constraint implies an added difficulty to the wavefunction analysis. Let us firstly show that there is no need to solve the constraint in this type of matrix models explicitly. This simplifies the arguments significantly. We use the representation of the wave function in terms of an anticommuting grassman coordinate (Lemma \[lema1\]) or its representation in the Fock space (Lemma \[lema2\] and Theorem \[teorema1\]) when convenient.
The constraint $\varphi^A\Psi=0$ defines a closed subspace of the Sobolev Hilbert space $\mathcal{H}^1(\Omega)$. Denote by $X$ the closure of this subspace in the norm of $L_2(\Omega)$. For the $SU(2)$ regularized supermembrane in four dimensions (RSM) we are interested in the following homogeneous problem. Given $g\in \mathcal{H}^2(\overline{\Omega})\cap X$, find $\Phi\in \mathcal{H}^2(\Omega)$ such that $$\label{I}
\begin{cases}
(-\nabla^2+ V)\Phi=0 & \textrm{in } \Omega\\
\Phi=g\quad & \textrm{on } \partial\Omega\\
\Phi\in X.
\end{cases}$$ Here $V$ is the potential of the hamiltonian $H$ of the RSM. We call $\Phi$ the ground state wavefunction of the hamiltonian in $\Omega$ since it corresponds to the restriction to $\Omega$ of the ground state wavefunction of the hamiltonian in $R^{D(N^2-1)}$, with $D=2$ and $N=2$. Besides, $\Phi$ minimizes the Dirichlet form, associated to the hamiltonian of RSM among the states which satisfy the constraint and the boundary condition. Let $D(\Lambda,\Phi)$, with $\Lambda\in \mathcal{H}^1(\Omega)$ and $\Phi\in \mathcal{H}^1(\Omega)$, be the Dirichlet form associated to the operator $-\nabla^2+V$. It is defined by $$D(\Lambda,\Phi)=(\nabla \Lambda,\nabla\Phi)+ (\Lambda,V\Phi)$$ where $(\cdot,\cdot)$ denotes the internal product in $L_2(\Omega)$. In particular if $\chi\in C_0^{\infty}(\Omega)$ we have $$\label{chi}
D(\chi,\chi)=(\chi,(-\nabla^2+V)\chi)\ge 0$$ due to the supersymmetric structure of the mass operator. It then follows $$\label{varphi} D(\varphi,\varphi)\ge 0$$ for all $\varphi\in \mathcal{H}_0^1(\Omega).$
Let $\Phi\in \mathcal{H}^1(\Omega)\cap X$, $\phi=g$ on $\partial\Omega$ be a solution of the Dirichlet problem $$\label{v-p}D(\varphi,\Phi)=0$$ for all $\varphi\in \mathcal{H}_0^1(\Omega).$ Then for any $\Lambda\in \mathcal{H}^1(\Omega)\cap X$, $\Lambda=g$ on $\partial\Omega$ we obtain $$\Lambda-\Phi=\varphi,$$ where $\varphi\in \mathcal{H}_0^1(\Omega)\cap X.$ It follows that $$D(\Lambda,\Lambda)=D(\Phi,\Phi)+D(\Phi,\varphi)+D(\varphi,\Phi)+D(\varphi,\varphi),$$ we now use (\[varphi\]) and (\[v-p\]) to get $$D(\Lambda,\Lambda)\ge D(\Phi,\Phi).$$ Consequently $D(\Phi,\Phi)$ is the minimum of the values of the Dirichlet form evaluated on the states $\Lambda\in \mathcal{H}^1(\Omega)\cap X$, $\Lambda=g$ on $\partial\Omega$. This is analogous to the Dirichlet principle in Electrostatics. We are going to show that there exists a unique $\Phi$ solution to the Dirichlet problem (\[v-p\]), moreover the solution $\Phi\in \mathcal{H}^2(\Omega)\cap X$. We may then integrate by parts in (\[v-p\]) to obtain a unique solution to (\[I\]). The minimum of the Dirichlet form is then obtained at the solution of problem (\[I\]). Let $f:=(\nabla^2-V)g$. The following inhomogeneous problem is a re-formulation of . Find $\Psi\in \mathcal{H}_0^1(\Omega)\cap \mathcal{H}^2(\Omega)$ such that $$\label{II}
\begin{cases}
(-\nabla^2+ V)\Psi=f & \textrm{in } \Omega\\
\Psi=0 & \textrm{on } \partial\Omega\\
\Psi\in X.
\end{cases}$$ If $\Psi$ is a solution of , then $\Phi=\Psi+g$ is a solution of .
In the following we will take $\Omega$ to be a ball of radius $R\ne0$. If $\Psi\in \mathcal{H}_0^1(\Omega)\cap \mathcal{H}^2(\Omega)$ then $\varphi^A\Psi\in \mathcal{H}_0^1(\Omega)$. Geometrically it means that $\Omega$ remains invariant under the symmetry generated by the first class constraint of the theory. The main result of this work is given by the following theorem:
\[teorema1\] Let $g\in \mathcal{H}^2(\Omega)\cap X$. There exists a unique solution for the problem which lies in $\mathcal{H}_0^1(\Omega)\cap \mathcal{H}^2(\Omega)\cap X$. Consequently, there exists a unique solution $\Phi\in \mathcal{H}^2(\Omega)\cap X$ to the problem .
The proof of this theorem relies on two auxiliary lemmas.
\[lema1\] Let $Q$ and $Q^\dag$ be the supercharge operators associated to the RSM. If $\Psi\in \mathcal{H}_0^1(\Omega)$ satisfies the conditions $$\label{kernel}
Q\Psi=Q^{\dag}\Psi=0 \qquad \text{in} \qquad \Omega,$$ then $\Psi=0$ in $\overline{\Omega}$. \[eql\]
Assume . According to the conditions of the supersymmetric algebra, we have $$H\Psi=0 \qquad \textrm{in}\quad \Omega.$$ By elliptic regularity, it immediately follows that $\Psi\in \mathcal{H}_0^1(\Omega)\cap \mathcal{H}^2(\Omega)$. Moreover, , which originally held true in $\Omega$, can be extended smoothly to the boundary, $\partial\Omega$.
Define $\rho^2=Z^A\overline{Z}^A$ and $\Psi_{\rho^2}= \frac{\partial\Psi}{\partial\rho^2}$, the normal derivative at $$\partial\Omega=\{Z^A, \overline{Z}^A: Z^A\overline{Z}^A= R^2\}.$$ Then using (\[kernel\]) on $\partial\Omega$ we obtain $$\frac{\partial}{\partial Z^A}\frac{\partial}{\partial \lambda_8^A}\Psi=0\quad \text{and} \quad\frac{\partial}{\partial\overline{Z}^A}\lambda_8^A\Psi=0\quad\textrm{on}\quad\partial\Omega.$$ Rewriting the latter in terms of $\Psi_{\rho^2}$, and using that all the tangential derivatives at $\partial\Omega$ are zero, gives $$\overline{Z}^A\frac{\partial}{\partial\lambda_8^A}\Psi_{\rho^2}=0 \quad \text{and} \quad Z^A\lambda_8^A\Psi_{\rho^2}=0\quad\textrm{on\quad}\partial\Omega.$$ From this we conclude that $Z^A\overline{Z}^A\Psi_{\rho^2}=0$ on $\partial\Omega$. That is $R^2\Psi_{\rho^2}=0$ for any $R^2\ne 0.$ Consequently $\Psi_{\rho^2}=0$ on $\partial\Omega$.
We therefore have $\Psi=0$ and $\Psi_{\rho^2}=0$ on $\partial\Omega$. By virtue of the Cauchi-Kovalewski Theorem, it follows that $\Psi=0$ in a neighborhood of $\partial\Omega$. Since $V$ is analytic in $\Omega$, in fact $\Psi=0$ in $\overline{\Omega}$.
\[lema2\] Let $f\in L_2(\Omega)\cap X=X$, there always exists a solution $\Psi\in \mathcal{H}_0^1(\Omega)$ to the Dirichlet problem $$\label{12}
\mathcal{D}(\chi,\Psi)=(\chi,f) \quad\textrm{for all}\quad \chi\in \mathcal{H}_0^1(\Omega),$$ Remark: The regularity properties of the Dirichlet form ensure that $\Psi\in \mathcal{H}_0^1(\Omega)\cap\mathcal{H}^2(\Omega).$
For the hamiltonian of the RSM, $\mathcal{D}$ is coercive since $V$ is bounded from below in $\Omega$, see [@Folland]. We may then use the theorem (7.21) from [@Folland], which states that $\Psi$ exists, provide the subspace $$K=\{\xi\in \mathcal{H}_0^1(\Omega): \mathcal{D}(\chi,\xi)=0\quad \textrm{for all}\quad \chi\in \mathcal{H}_0^1(\Omega)\}$$ is orthogonal in $L_2(\Omega)$ to $f$.
From the regularity property of the Dirichlet form, we obtain $\xi\in \mathcal{H}_0^1(\Omega)\cap C^{\infty}(\Omega)$, hence $(-\nabla^2 +V)\xi=0$ in $\Omega$. Consequently $Q\xi=0$ and $Q^{\dag}\xi=0$ in $\Omega$. According to Lemma \[lema1\], we conclude that $\xi=0$ in $\Omega$, hence $K=\{0\}$ is orthogonal to $f$. The regularity properties of elliptic operators ensure that $\Psi\in \mathcal{H}_0^1(\Omega)\cap \mathcal{H}^2(\Omega).$
The proof of Theorem \[teorema1\] is completed as follows. According to Lemma \[lema2\], there exists a solution $\Psi\in \mathcal{H}_0^1(\Omega)\cap \mathcal{H}^2(\Omega)$ to the Dirichlet problem . By integration by parts $$(-\nabla^2+V)\Psi=(\nabla^2-V)g \quad\textrm{in}\quad \Omega.$$ Given $\chi\in C_0^{\infty}(\Omega)$ we consider $$(\varphi^A \chi, (-\nabla^2+V)\Psi)=(\varphi^A\chi, (\nabla^2-V)g)$$ then, using that $(-\nabla^2+V)$ commutes with $\varphi^A$, we get $$( (-\nabla^2+V)\chi,\varphi^A\Psi)=((\nabla^2-V)\chi,\varphi^A g)=0.$$ We thus have $$D(\chi,\varphi^A \Psi)=0$$ for all $\chi\in C^{\infty}_0(\Omega)$ and hence, taking limits, for all $\chi\in \mathcal{H}^1_0(\Omega).$ Consequently, using that $\varphi^A\Psi\in \mathcal{H}_0^1(\Omega)$, we obtain $\varphi^A\Psi=0,$ that is $\Psi\in\mathcal{H}_0^{1}(\Omega)\cap\mathcal{H}^2(\Omega)\cap X$ is a solution to the problem . It is unique, since $K\subset\{0\}$.
Discussion
==========
Let us compare our analysis, on a smooth bounded domain, with the analysis presented in [@su2] and [@queiroz] on unbounded regions. In [@su2] the ground state of the Yang-Mills quantum mechanics is considered, and upper and lower bounds for the minimum eigenvalue are obtained. The hamiltonian has a similar structure to the hamiltonian of the bosonic membrane, its potential has a quartic dependence on the configuration variables with valleys extending to infinity. The hamiltonian is bounded from below by a hamiltonian with a basin shaped potential which has a discrete spectrum with non-zero minimal eigenvalue [@simons; @luscher].
Another approach for the same problem was considered in [@inertia] using the Molchanov mean value condition for the potential [@molchanov; @mazya]. The discreteness of the spectrum is originated by the non-zero ground-state energy of the bosonic harmonic oscillator. This argument cannot be extended to supersymmetric matrix models because the supersymmetric harmonic oscillator has zero ground state energy. In the supersymmetric matrix models describing the regularized supermembrane the spectrum is continuous from $[0,\infty)$ and the main problem is to analyze whether $0$ is an eigenvalue or not.
The boundary condition in [@su2] prescribes that the wave functions should decay at infinity. We believe that this may be too restrictive for the quantum mechanics of $SU(2)$ supermembrane in an unbounded domain. As emphasizes in [@dwhn] one has to consider the whole Hilbert space $L_2(R^D)$ which includes wave functions which are square integrable functions but diverge to infinity in some directions. This problem is still open, and it requires a more delicate analysis of the boundary conditions at infinity. For that reason the problem formulated on a bounded domain with a nontrivial boundary condition is of relevance.
The authors of [@queiroz] constructed a matrix model of QCD incorporating non-trivial topological aspects of the theory. The hamiltonian matrix model has similar properties to the bosonic hamiltonian of the supermembrane with non-trivial central charges or non-trivial winding [@bgmr0; @bgmr]. In the latter paper, the fermionic potential is a relatively bounded operator with respect to the bosonic hamiltonian, hence the spectrum of the supersymmetric matrix model has the same qualitative behaviour as its bosonic sector [@bgmr]. However, the main problem in this paper is to resolve the spectrum of the $D=11$ supermembrane with zero winding in which case one expects that the ground state should correspond to the $D=11$ supergravity multiplet.
The approach we have in mind for the extension of our analysis to the unbounded region $R^D$ consists in three steps. The first one is to solve the ’internal’ problem that is the existence and uniqueness of the ground state on a smooth bounded domain $\Omega$. The second step is to analyze the problem on the exterior region of $\Omega$. Finally we will consider the matching of the two solutions. So far we have solved the first step for the $SU(2)$ problem and expect to generalize these arguments for the $SU(N)$ regularization of the $D=11$ Supermembrane. The analysis of the exterior problem introduces a new aspect compared to the first step. It is the behaviour of the potential at infinity. The matching conditions, a well-developed topic in elliptic partial differential equations, will give the final answer concerning to the existence or not of the ground state wave function of the $D=11$ Supermembrane. A different approach is to consider a sequence of balls of increasing radius and to show the convergence of the sequence of solutions under a suitable a priori boundary condition function $g\in L_2(R^D)$.
Conclusions
===========
We presented an $SU(2)$ gauge supersymmetric matrix model whose center of mass propagates freely in a 4D spacetime with its transversal oscillations restricted to a compact space subject to Dirichlet boundary conditions. We establish the existence and uniqueness of the massless ground state of the theory. Our analysis relies heavily on the property that the constraint is of first class, which is the standard case of gauge theories. Extension of this analysis to matrix models subject to second class constraints are worth of further study, and are beyond of the scope of the present paper.
Our proofs simplify significantly, due to the fact that we do not require to solve explicitly the constraint. The approach we have illustrated here can be extended in order to determine the ground state wavefunction of other matrix models associated to supersymmetric gauge systems, such as AdS/CFT at finite temperatures or QED matrix models in compact spacetime.
Aknowledgements
===============
MPGM would like to thank to Manuel Asorey for interesting comments and to the Theoretical Physics Department at U. Zaragoza, Spain, for kind invitation while part of this work was done. MPGM is supported by Mecesup ANT1398, Universidad de Antofagasta, (Chile). A.R. is partially supported by Projects Fondecyt 1121103 (Chile).
[99]{}
M.B. Halpern, C. Schwartz, [*Int.J.Mod.Phys. A*]{} (1998) [**13**]{}, 4367. J. Hoppe, [*On The Construction of Zero Energy States in Supersymmetric Matrix Models III*]{}, arXiv:hep-th/9711033. B. de Wit, M. Luscher, H. Nicolai, [*Nucl. Phys. B*]{} (1989)[**B320**]{}, 135. T. Banks, W. Fischler, S.H. Shenker, L. Susskind, [*Phys.Rev. D*]{} (1997) [**55**]{}, 5112-5128. P. Yi,[*Nucl. Phys. B*]{} (1997) [**505**]{}, 307. S. Sethi, M. Stern, [*Commun. Math. Phys.*]{} (1998) [**194**]{}, 675. M. Porrati, A. Rozenberg, [*Nucl. Phys. B*]{} (1998) [**515**]{}, 184-202. J. Froehlich, J. Hoppe, [*On Zero-Mass Ground States in Super-Membrane Matrix Models*]{} arXiv:hep-th/9701119. L. Boulton, M.P. Garcia del Moral, A. Restuccia, [*Nucl.Phys. B*]{} (2012) [**856**]{}, 716-747. N. Acharyya, M. Asorey, A.P. Balachandran and S. Vaidya, [*Supersymmetry: Boundary Conditions and Edge States*]{} arXiv:1501.00634 \[hep-th\]. G. Arnold, B. Bunk, Th. Lippert, K. Schilling, [*Nucl.Phys.Proc.Suppl.*]{}(2003) [**119**]{} 864-866. Takahashi, [*Low-lying Dirac eigenmodes and monopoles in 4D compact QED*]{}, arXiv:hep-lat/0703023v3 M.N. Chernodub, E.M. Ilgenfritz, A. Schiller, [*Monopoles, confinement and deconfinement in lattice compact QED in (2+1)D with external fields*]{} arXiv:hep-lat/0110038v1. B. de Wit, J. Hoppe, H. Nicolai, [*Nucl. Phys. B*]{} (1988) [**305**]{}, 545. J. Hoppe, D. Lundholm, M. Trzetrzelewski, [*Nucl.Phys. B*]{} (2009) [**817**]{}, 155-166. A.M. Khvedelidze and H.P. Pavel, [*Phys.Lett.A*]{} (2000) [**267**]{}, 96-100. J. Hoppe, D. Lundholm, M. Trzetrzelewski, [*Annales Henri Poincare*]{} (2009) [**10**]{} 339. L. Boulton, M.P. Garcia del Moral, A. Restuccia, [*Phys.Lett.B*]{} (2015) [**744**]{} 260-262. L. Boulton, M.P. Garcia del Moral, A. Restuccia, [*The ground state of the D=11 supermembrane and matrix models on compact regions*]{} arXiv:1504.04071 G. Folland, Introduction to Partial Differential Equations. Second Edition. Princeton University Press, Princeton, 1995. A.P. Balachandra, S. Vaidya, A. R. de Queiroz [*A Matrix Model for QCD*]{} arXiv:1412.7900 M. Reed, B. Simon, Methods of modern mathematical physics. Vol 4: Analysis of Operators, Academic Press, New York, 1978; B. Simon, Ann. Phys 146 (1983)209 M. Luscher, [*Nucl.Phys.B*]{} 219:233-261,1983. M.P. Garcia del Moral, L. Navarro, A.J. Perez, A. Restuccia, [*Nucl.Phys. B*]{} (2007) [**765** ]{} 287-298. A.M. Molchanov On the discreteness of the spectrum conditions for selfadjoint differential equations of the second order, Proc. Moskow Math. Society 2 (1953) 169-199 (Russian). V. Mazy’a, M. Shubin, Annals of Mathematics, 162 (2005), 919-942. L. Boulton, M.P. Garcia del Moral, A. Restuccia, [*Nucl.Phys. B*]{} (2003) [**671**]{} 343-358.
[^1]: The mass operator of the regularized supermembrane in a 11D Minkowski spacetime corresponds to the $N=16$ supersymmetric matrix model.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce the sequence $(a_n) \subset (0,1]$ and prove that the asymptotic behaviour of $\sum_{k=1}^n a_k$ is the same than $\pi(n)$, the prime-counting function. We also obtain that $\pi(n) \sim n a_n$ and we estimate $\frac{1}{a_n}-\frac{n}{\pi(n)}$ showing that $\lim_{n \rightarrow \infty} \frac{1}{a_n}-\frac{n}{\pi(n)}$ is convergent.'
address:
- 'Alejandro Miralles. Instituto Universitario de Matemáticas y Aplicaciones de Castellón (IMAC), Universitat Jaume I de Castelló (UJI), Castelló, Spain. *e*.mail: [email protected]'
- 'Damià Torres. Universitat Politècnica de Catalunya - BarcelonaTech (UPC), Barcelona, Spain. *e*.mail: [email protected]'
author:
- Alejandro Miralles$^1$
- Damià Torres
title: Counting primes by sums of frequencies
---
Introduction and background
===========================
Introduction
------------
In this work we introduce the numbers $a_n \in \left( 0,1\right]$ for any $n \in \mathbb{N}$ in two equivalent ways: first we use a sieve method which results in a subset $A_n \subset \{1,2,\cdots n\}$ and then we consider the frequency $a_n=\frac{|A_n|}{n}$. We prove that this is equivalent to consider the recurrence sequence $a_n$ given by $a_1=1$ and $$a_{n}=a_{n-1}\left(1-\frac{a_{n-1}}{n}\right)$$ as in Definition \[def pr\]. We will prove that the asymptotic behaviour of the partial sums $c(n)=\sum_{k=1}^n a_n$ is the same than the prime counting function given by $$\pi(n)=\#\{p \in \mathbb{N} : p \mbox{ is prime and } p \leq n \}$$
for any $n \in \mathbb{N}$. We will also conclude other related results related to $\pi(n)$ as now that $\pi(n) \sim n a_n$ and we will estimate $\frac{1}{a_n}-\frac{n}{\pi(n)}$ showing that $\lim_{n \rightarrow \infty} \frac{1}{a_n}-\frac{n}{\pi(n)}$ is convergent to a real number between $2$ and $3$.
Classical results.
------------------
How prime numbers are distributed among positive integers plays an important role in number theory. Let $\pi(x)$ be the prime-counting function that gives the number of primes less than or equal to $x$ for any real number x.
The Prime Number Theorem gives the asymptotic behaviour of distribution of prime numbers. It was proved by Hadamard and de la Vallée-Poussin in 1896 using complex theory (see [@H] and [@V]). A more elementary proof was given by Erdös and Selberg in 1948 (see [@E] and [@S]). This classical result states that $$\pi(x) \sim \frac{x}{\log x}$$
and also that $$\pi(x) \sim Li(x),$$ where the asymptotic notation $f(x) \sim g(x)$ means $$\lim_{x \rightarrow \infty} \frac{f(x)}{g(x)}=1,$$
where $x$ could denote either a real or a positive integer and the Logarithmic Integral Function $Li(x)$ is given by $$Li(x)= \int_{2}^x \frac{dt}{\log t} dt,$$ for any $x \geq 2$. We will consider the Harmonic numbers $H_n$ given by $$H_n=1+\frac{1}{2}+\cdots+\frac{1}{n}$$
for any $n \in \mathbb{N}$. It is well-known that $$\begin{aligned}
\label{desig2}
0< H_n -\log n - \gamma < \frac{1}{n},\end{aligned}$$ where $\gamma \approx 0'577$ denotes the Euler-Mascheroni constant.
Approaching the frequency $a_n$: sieving the positive integers
--------------------------------------------------------------
Fix $n$ in ${\mathbb{N}}$, consider a *good integer* $M$ (to be determined later) and let $A_0=\{1,2,\cdots,M\}$. For any $1 \leq k \leq n$, we will sieve some numbers up to $M$ and we will obtain a sequence of decreasing subsets $A_0 \supseteq A_1 \supseteq A_2 \supseteq \cdots \supseteq A_n$ and a decreasing sequence of real numbers $a_1 \geq a_2 \geq \cdots \geq a_n$ given by the frequency $a_k=\frac{|A_k|}{M}$ for any $k$.
Fix $n=4$. We will study the frequencies $a_k$ for $k=1,2,3,4$. Consider $M=576$ and $A_0=\{1,2, \cdots, 576 \}$.
**Step 1.** For $k=1$ we start from the set $A_0$ and choose the subset $A_1$ of all the multiples of $1$, that is, the whole $A_0$. The frequency $a_1$ is given by $a_1=\frac{|A_1|}{|A_0|}=1$.
**Step 2.** For $k=2$, we start from the previous set $A_1$ of $|A_1|=576$ integers. We consider the subset $B_2$ by keeping every number in $A_1$ from the first one by counting up in increments of $2$ and crossing out the remaining ones, that is, $B_2=\{1,3,5, \cdots,575\}$, so $|B_2|=288$. Now we apply the previous step (step 1) to $B_2$, that is, we consider all the elements of $B_2$, which gives the set $C_2=\{1,3,5,\cdots,575\}$. The set $A_2$ is given by $A_1 \setminus C_2$ and, since $C_2 \subset A_1$, we have that $|A_2|=288$ and the frequency $a_2$ is given by $\frac{|A_2|}{M}=288/576=0.5$.
**Step 3.** For $k=3$, we start from the previous set $A_2$ of $|A_2|=288$ integers, that is, $A_2=\{2, 4, 6, \cdots, 576\}$. We consider the subset $B_3$ by keeping every number in $A_2$ from the first one by counting up in increments of $3$ and crossing out the remaining ones, that is, $B_3=\{2,8,14,20,\cdots, 566, 572\}$, so $|B_3|=288/3=96$. We apply the previous step (step 2) to $B_3$, that is, we keep numbers in $B_3$ from the first one by counting up in increments of $2$ and crossing up the remaining ones, that is, $C_3=\{2,14,26, \cdots, 566\}$ satisfying $|C_3|=\frac{|A_2|}{M}|B_3|=48$ numbers. The set $A_3$ is given by $A_2 \setminus C_3$ and, since $C_3 \subset A_2$, we have that $|A_3|=288-48=240$ and the frequency $a_3$ is given by $\frac{|A_3|}{M}=240/512=5/12$.
**Step 4.** For $n = 4$, we start from the previous set $A_3$ of $|A_3|=240$ integers. We consider the subset $B_4$ by keeping every number in $A_3$ from the first one by counting up in increments of $4$ and crossing out the remaining ones, so $|B_4|=60$. We apply the previous step (step 3) to $B_4$ which gives the set $C_4$ satisfying $|C_4|=\frac{|A_3|}{M} |B_4|=5/12\cdot60=25$ numbers. The set $A_4$ is given by $A_3 \setminus C_4$ and since $C_4 \subset A_3$, we have that $|A_4|=240-25=215$ and the frequency $a_4$ is given by $\frac{|A_4|}{M}=215/576$.
To define the frequency $a_k$ for any $k \geq 4$, fix $n \geq k$ and consider a good integer $M$ for $n$ (to be determined later). Suppose $A_{k-1}$ and $a_{k-1}=\frac{|A_{k-1}|}{M}$ are defined.
**Step $k$.** We start from the previous set $A_{k-1}$ of $|A_{k-1}|=a_{k-1} M$ integers. We consider the subset $B_k$ by keeping every number in $A_{k-1}$ from the first one by counting up in increments of $k$ and crossing out the remaining ones. We apply the previous step (step k-1) to $B_k$ which gives the set $C_k$ satisfying $|C_k|=\frac{|A_{k-1}|}{M}|B_k|=\frac{|A_{k-1}|^2}{M k}=a_{k-1} |B_k|$. The set $A_k$ is given by $A_{k-1} \setminus C_k$ and since $C_k \subset A_{k-1}$, we have that $|A_k|=|A_{k-1}|-|C_k|$ and the frequency $a_k$ is given by $a_k=\frac{|A_k|}{M}$.
Let $n \in {\mathbb{N}}$ and $1 \leq k \leq n$. In order to define the frequencies $a_k$, we can interchange the action to construct sets $B_k$ and $C_k$ and frequencies $a_k$ do not change. This means that for any $k$, we can start from the previous set $A_{k-1}$ of $|A_{k-1}|=a_{k-1} M$ integers. We can apply the previous step (step k-1) to $A_{k-1}$ which gives the set $B_k$ satisfying $|B_k|=|A_{k-1}|a_{k-1}$. We can consider the subset $C_k$ of $B_k$ by keeping every number in $B_k$ from the first one by counting up in increments of $k$ and crossing out the remaining ones. The set $A_k$ is given by $A_{k-1} \setminus C_k$, we have that $|A_k|=|A_{k-1}|-|C_k|$ and $a_k=\frac{|A_k|}{M}$.
A good integer $M$.
-------------------
Given $n \in {\mathbb{N}}$ and $1 \leq k \leq n$ it is necessary to determine good integers $M$ to make the construction above.
A good integer $M$ is given by a positive integer such that in any step $k$ we can sieve the set $A_{k-1}$ correctly. That is, $|A_{k-1}|$ must be multiple of $k$ in order to construct $B_k$ and since we need to apply the previous step in $B_k$, we need that $|A_{k-1}||B_k|$ is multiple of $M$, so $\frac{|A_{k-1}|^2}{k M}$ must be an integer.
The following lemma is clear from definition of $A_k$.
\[lema exp\] Let $n \in {\mathbb{N}}$ and a good integer $M$ for $n$. For any $1 \leq k \leq n$ we have that $$\begin{aligned}
\label{expr}
|A_k|=|A_{k-1}|-\frac{|A_{k-1}|^2}{k M}.\end{aligned}$$
Fix $n \in {\mathbb{N}}$. We look for a good integer $M$ for $n$. Notice that the set $A_1$ satisfies that $|A_1|=1M=M$ so to assure that $|A_1|$ is an integer, it is sufficient to consider $M=1$.
Now, $|A_2|=|A_1|-|A_1|^2/2M=M-M^2/M=M/2$, so it is sufficient to consider $M=2$ to assure that $|A_2|$ is an integer. Fix $2 < k \leq n$. If $|A_{k-1}|=\frac{\alpha}{\beta} M$, we have that $M=\beta$ is sufficient. Since $$|A_k|=|A_{k-1}|-\frac{|A_{k-1}|^2}{k M}=\left( \frac{\alpha}{\beta} -\frac{\alpha^2}{\beta^2 k}\right)M = \frac{\alpha \beta k- \alpha^2}{\beta^2 k}M,$$ it is clear that from step $k-1$ to step $k$ we change $\beta$ by $\beta^2 k$. Hence, we can give a sequence of good integers $M$ for all $n \in {\mathbb{N}}$.
A good integer for $n=1$ is given by $M_1=1$ and the recurrence $M_n=n \beta^2 =n M_{n-1}^2$ gives good integers $M_n$ for any $n \in {\mathbb{N}}$.
Notice that for any $n \in {\mathbb{N}}$, any multiple $M$ of $M_n$ is also a good integer for $n$ since we can still sieve correctly the set $\{1,2,\cdots,M\}$.
From the recurrence, it is easy by induction that
Let $n \in {\mathbb{N}}$. Then, $$M_n=2^{2^{n-2}} 3^{2^{n-3}} 4^{2^{n-4}} \cdots (n-1)^{2^{1}} n^{2^0}.$$
The frequencies $a_n$.
----------------------
Notice that the frequency $a_n$ is given by $\frac{|A_n|}{M}$. It is clear that this expression does not depends on $M$ since it is homogeneous of degree $0$ on $M$. We have used good integers $M$ in order to sieve the sets $A_n$ correctly but we can avoid this by considering upper limits when $M \rightarrow \infty$:
For any $n \in {\mathbb{N}}$ we have $$a_n=\limsup_{M \rightarrow \infty} \frac{|A_n|}{M}.$$
Dividing by $M$ in expression (\[expr\]) in Lemma \[lema exp\], we can express $a_n$ by $a_1=1$ and $$a_{n}=a_{n-1} \left(1-\frac{a_{n-1}}{n} \right).$$ Hence,
\[def pr\] For any $n \in {\mathbb{N}}$, we define the frequency $a_n$ by $a_1=1$ and $$a_{n}=a_{n-1} \left(1-\frac{a_{n-1}}{n}\right)$$ for any $n \geq 2$. We will denote by $c(n)$ the sum $\sum_{k=1}^n a_k$.
Notice that $$a_{n}=a_{n-1}-\frac{a_{n-1}^2}{n},$$ so $a_{n}-a_{n-1}=-\frac{a_{n-1}^2}{n}$. Roughly speaking, this means that the “derivative" of $``a_{n-1}"$ equals $``-a_{n-1}^2/n"$. Notice that the function $y=\frac{1}{\log x}$ also satisfies $y'=\frac{-1}{x}\frac{1}{\log^2 x}$ so, as we will show, the behaviour of $a_n$ will be similar to $\frac{1}{\log n}$.
Results
=======
The following lemma is an easy calculation.
\[lema\] For any $n \geq 2$, we have that $$\frac{1}{a_{n}}=\frac{1}{a_{n-1}}+\frac{1}{n-a_{n-1}}.$$
\[lema2\] The series $$\sum_{k=2}^{\infty} \frac{a_{k-1}}{k(k-a_{k-1})}$$ is convergent to $S \approx 0.662834$.
[[ ]{}]{}It is clear that the series is convergent since $\sum_{k=1}^\infty \frac{1}{k^2}$ is convergent and $0 \leq a_k \leq 1$. An easy computation estimates the value of the sum.
The following proposition states that $1/a_n$ is very close to $\log n$. Indeed,
\[prop1\] Let $b_n=\frac{1}{a_n}-\log n$ for any $n \in {\mathbb{N}}$. Then,
- The sequence $(b_n)$ is bounded: $$\frac{1}{2}+\gamma < b_n < 1+\gamma.$$
- The sequence $b_n$ is increasing.
- The sequence $b_n$ is convergent to $\gamma + S \approx 1.24005$.
[[ ]{}]{}a) Using Lemma \[lema\], we have that for any $k \geq 2$, $$\begin{aligned}
\label{equ}
b_{k}-b_{k-1}=\frac{1}{k-a_{k-1}}+\log (k-1) - \log k.\end{aligned}$$ Since $b_1=1$ and considering equality (\[equ\]) for indexes from $2$ to $n$, we do a telescoping sum and obtain $$\begin{aligned}
\label{equ2}
b_n=1+\sum_{k=2}^{n} \frac{1}{k-a_{k-1}}-\log n.\end{aligned}$$ Bearing in mind that $a_1=1$ and $0 \leq a_n \leq 1$ for any $n \geq 2$, we have making easy calculations that $$\frac{1}{2}+H_{n}-\log n < b_n < 1-\frac{1}{n}+H_{n}-\log n,$$ so bearing in mind inequality (\[desig2\]), we have $$\frac{1}{2} +\gamma < b_n < 1+\gamma$$ and we are done.
b\) We will show that $b_{n} > b_{n-1}$ for any $n \geq 2$. This is true if and only if $$\frac{1}{a_{n}}-\log n-\left( \frac{1}{a_{n-1}}-\log (n-1) \right) >0,$$ if and only if $$\frac{1}{a_{n}}-\frac{1}{a_{n-1}} -\log \left(\frac{n}{n-1} \right) >0.$$ Using definition of the sequence $a_n$, this inequality is equivalent to $$\frac{a_{n-1}}{n a_{n}}-\log \left(\frac{n}{n-1} \right) >0 \ \mbox{ if and only if }
\ a_{n-1} > \log \left(\frac{n}{n-1} \right)^n a_{n}.$$ Notice that $(a_n)$ is decreasing since $a_{n}-a_{n-1}=-\frac{a_{n-1}^2}{n} <0$, so $$a_{n-1} > a_{n} > \log \left(\frac{n}{n-1} \right)^n a_{n},$$ where last inequality is true since $\left(\frac{n}{n-1} \right)^n$ is an increasing sequence which tends to $e$ and $\log x$ is an increasing function on its domain.
c\) Since $(b_n)$ is increasing and bounded, it is convergent to some limit $\ell$. It is clear that $$b_n=H_{n} - \log n + \sum_{k=2}^{n} \left( \frac{1}{k-a_{k-1}} - \frac{1}{k} \right)=H_{n} - \log n + \sum_{k=2}^{n} \frac{a_{k-1}}{k(k-a_{k-1})}.$$ Last sum is a convergent series to $S$ by Lemma \[lema2\]. Then, $$\lim_{n \rightarrow \infty} b_n=\lim_{n \rightarrow \infty} H_{n}- \log n+\sum_{k=2}^{n} \frac{a_k}{k(k-a_k)}=\gamma+S$$ and we are done.
We have that $\frac{1}{a_n} \sim \log n$.
Notice that from equality (\[equ2\]) in Proposition \[prop1\], we have that
\[corol\] For any $n \geq 2$ we have $$\frac{1}{a_n}=1+\sum_{k=2}^{n} \frac{1}{k-a_{k-1}}.$$
The following lemma is an easy consequence of $\pi(x) \sim Li(x)$ and integral calculus.
Consider the function $$q(n)=\sum_{k=2}^{n} \frac{1}{\log k}$$ for any $n \in {\mathbb{N}}$. Then, $\pi(n) \sim q(n)$.
\[series\] For any real number $C$ and an integer number $m \geq 2$, we consider the sequence $$x_{m,C}=\sum_{n=2}^{m} \frac{1}{\log(n)+C}.$$ Then, $x_{m,C}$ is divergent and $x_{m,C_1} \sim x_{m,C_2}$ for any $C_1, C_2 \geq 0$.
[[ ]{}]{}It is well-known that $x_{m,1}$ is a divergent series and by the limit criterium, $x_{m,C}$ is also divergent for any $C \geq 0$. Hence, the limit $$\ell = \lim_{m \rightarrow \infty} \frac{x_{m,C_1}}{x_{m,C_2}}$$ can be calculated using the Stolz criterium, so we obtain that $$\ell= \lim_{m \rightarrow \infty} \frac{\frac{1}{\log(m+1)+C_1}}{\frac{1}{\log(m+1)+C_2}}=\lim_{m \rightarrow \infty} \frac{\log(m+1)+C_2}{\log(m+1)+C_1}=1.$$
The sum $c(n)=\sum_{k=1}^n a_k$ satisfies $c(n) \sim \pi(n)$.
[[ ]{}]{}By Corollary \[corol\], $$\frac{1}{2} + H_{k} < \frac{1}{a_k} < 1+ H_k-\frac{1}{k}.$$ By inequality \[desig2\], we have that $\log k + \gamma < H_k < \log k + \gamma + \frac{1}{k},$ so $$1+ \log k < \frac{1}{2}+ \log k + \gamma \leq \frac{1}{a_k} \leq 1+\log k + \gamma < 2+\log k$$ for any $k \geq 2$. Hence, we have that $$\frac{1}{\log k+2} \leq a_k \leq \frac{1}{\log k+1},$$ so bearing in mind that $a_1=1$, we have $$\begin{aligned}
\label{desig}
\sum_{k=1}^{n} \frac{1}{\log k+2} \leq \sum_{k=1}^n a_k \leq \sum_{k=1}^{n} \frac{1}{\log k+1}.\end{aligned}$$
Using Lemma \[series\], dividing by $x_{n,0}$ in inequality (\[desig\]), we obtain that $$1 \leq \lim_{n \rightarrow \infty} \frac{c(n)}{\sum_{k=2}^n \frac{1}{\log k}} \leq 1,$$ and by the Sandwich criterium, we are done.
We have $\frac{n}{\pi(n)} \sim \frac{1}{a_n}$ and $\pi(n) \sim n a_n$.
[[ ]{}]{}By Proposition \[prop1\], we have that $\frac{1}{a_n} \sim \log n$. Since $\frac{n}{\log n} \sim \pi(n)$, we have that $\frac{n}{\pi(n)} \sim \log n \sim \frac{1}{a_n}$. The other statement is also clear.
Evaluating $n/\pi(n)$.
----------------------
It was first Chebyshev who proved (see [@C]) that there exists $x_0 \in {\mathbb{N}}$, $c_1 \approx 0.92$ and $c_2 \approx 1.1$ such that $$c_1 \frac{x}{\log x} \leq \pi (x) \leq c_2 \frac{x}{\log x}$$ for any $x \geq x_0$. The following result was obtained in [@RS]: $$\frac{x}{\log x-1/2} < \pi(x) < \frac{x}{\log x-3/2},$$ where first inequality is true for $x \geq 67$ and the second one is true for $x \geq e^{3/2}$. L. Panaitopol (see [@P]) improved this result: $$\begin{aligned}
\label{pan}
\frac{x}{\log x-1+\frac{1}{\sqrt{\log x}}} < \pi(x) < \frac{x}{\log x-1-\frac{1}{\sqrt{\log x}}},\end{aligned}$$ where first inequality is true for $x \geq 59$ and the second one for $x \geq 6$.
Hence, we obtain the following result:
We have the following estimates for $n \geq 59$:
- We have that $$\left|\frac{n}{\pi(n)}-\log n +1 \right| \leq \frac{1}{\sqrt{\log n}},$$ so $$\lim_{n \rightarrow \infty} \left| \frac{n}{\pi(n)}-\log n+1 \right| =0.$$
- We have that $$\left|\frac{1}{a_n}-\frac{n}{\pi(n)} \right| \leq 2+ \gamma +\frac{1}{\sqrt{\log n}} < 3.$$
- We have that $$\lim_{n \rightarrow \infty} \frac{1}{a_n}-\frac{n}{\pi(n)}=\gamma+S+1 \approx 2.24005.$$
[[ ]{}]{}The first one is true by inequality \[pan\] from the Panaitopol result. The second result is an easy consecuence from the first one and Proposition \[prop1\]. To prove the third one, notice that from the first result we have that $\log n-\frac{n}{\pi(n)}-1 \rightarrow 0$ when $n \rightarrow \infty$ so $$\frac{1}{a_n}-\frac{n}{\pi(n)} = \frac{1}{a_n}-\log n + 1+\log n -\frac{n}{\pi(n)}-1$$ which tends to $\gamma+S+1$ when $n \rightarrow \infty$ by Proposition \[prop1\].
[10]{} L. Chebyshev, *Mémoire sur les nombres premiers*, Journal de Math. Pures et Appl. **17** (1852), 366–390.
P. Erdös, *On a new method in elementary number theory which leads to an elementary proof of the prime number theorem*, Proc. Nat. Acad. Scis. U.S.A. **35** (1949), 374–384. J. Hadamard, *Étude sur les proprietés des fonctions entiéres et en particulier d’une fonction considerée par Riemann*, J. de Math. Pures Appl. (4) **9**, 171–215 (1893); reprinted in *Oeuvres de Jacques Hadamard* **1**, C.N.R.S., Paris (1968), 103–147. L. Panaitopol, *Inequalities concerning the function $\pi(x)$: applications*, Acta Arith. 94 (2000), no. 4, 373–381.
J. B. Rosser and L. Schoenfeld, *Approximate formulas for some functions of prime numbers*, Illinois J. Math. **6** (1962), 64–94.
A- Selberg, *An elementary proof of the prime number theorem*, Ann. of Math. (2) **50** (1949), 305–313; reprinted in Atle Selberg Collected Papers **1**, Springer-Verlag, Berlin Heidelberg New York (1989), 379–387.
C.J. de la Vallée-Poussin, *Recherches analytiques sur la théorie des nombres premiers*, Ann. Soc. Sci. Bruxelles **20** (1896), 183–256.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate a construction of a pseudo BL-algebra out of an $\ell$-group called a kite. We show that many well-known examples of algebras related to fuzzy logics can be obtained in that way. We describe subdirectly irreducible kites. As another application, we exhibit a new countably infinite family of varieties of pseudo BL-algebras covering the variety of Boolean algebras.'
address:
- '$^1$Mathematical Institute, Slovak Academy of Sciences, Štefánikova 49, SK–814 73 Bratislava, Slovakia. Department of Algebra and Geometry, Faculty of Sciences, Palacký University, tř. 17. listopadu 1192/12, CZ-771 46 Olomouc, Czech Republic'
- '$^2$The University of Melbourne, Department of Mathematics and Statistics, Parkville, VIC 3010, Australia'
author:
- Anatolij Dvurečenskij$^1$
- Tomasz Kowalski$^2$
title: 'Kites and Pseudo BL-algebras'
---
Introduction
============
Lattice-ordered groups ($\ell$-groups) are intimately connected with certain algebras related to fuzzy logics. Indeed, the discovery of one such connection predates fuzzy logics, as it was made by Chang [@Cha] in his algebraic proof of completeness of infinitely-valued Ł[u]{}kasiewicz’s logic. Much later Mundici [@Mun] proved a categorical equivalence between the variety of MV-algebras and the class of Abelian $\ell$-groups with strong unit. This was extended by Dvurečenskij [@Dvu1] to an equivalence between the variety of *pseudo MV-algebras* (a noncommutative generalization of MV-algebras) and the class of all $\ell$-groups with strong unit. For *BL-algebras*, which constitute an algebraic semantics of classical fuzzy logic, the $\ell$-group connection was investigated by Aglianò and Montagna [@AgMo] who proved that all linearly ordered BL-algebras consist of building blocks that are either negative cones of $\ell$-groups, or negative intervals in unital $\ell$-groups. A similar result was proved by Dvurečenskij for representable (i.e., such that subdirectly irreducibles are linearly ordered) pseudo BL-algebras (cf. [@Dvu2]). Jipsen and Montagna [@JiMo] constructed a subdirectly irreducible pseudo BL-algebra, which is not linearly ordered, yet it is made out of an $\ell$-group in a rather special way, which may be visualized as resembling a kite: a two-dimensional head joined to a one-dimensional tail. In this paper we generalize their construction, and show that many well-known examples of the algebras mentioned above can be seen as particular cases of the generalized construction. As our generalization consists essentially in allowing more dimensions for the head and tail, we call our algebras *kites*.
The paper is organized as follows. Basic notions and notations are in Section 2. Section 3 defines kites, they are always pseudo BL-algebras. It gathers the main properties of kites. In particular, some kites are pseudo MV-algebras. Section 4 presents a list of important kites. Subdirectly irreducible kites and their classification are described in Section 5. Very important kites are finite-dimensional ones which are studied in Section 6, where it is shown that the variety of pseudo BL-algebras generated by all kites is generated by all finite-dimensional kites. Finally, Section 7 shows some applications of theory of kites, in particular, countably many covers generated by kites of the variety of Boolean algebras in the variety of pseudo BL-algebras are presented.
Basic notions and notation
==========================
In terminology and notation we will follow [@GJKO], which is also our standard reference for undefined notions and details. All classes (varieties) of algebras we deal with in the paper can be viewed as subclasses (subvarieties) of *FL-algebras*, that is, algebras $\mathbf{A} = \langle
A;\wedge,\vee,\ld,\rd,\cdot,0,1\rangle$ of type $(2,2,2,2,2,0,0)$ such that
- $\langle A;\wedge,\vee\rangle$ is a lattice,
- $\langle A;\cdot,\ld,\rd, 1\rangle$ is a residuated monoid,
- $0$ is an element of $A$.
The operations $\ld$ and $\rd$ are called, respectively, *left division* (or *right residuation*) and *right division* (or *left residuation*). Two unary operations: *left negation* ${\mathord{\sim}}x = x\ld 0$, and *right negation* ${\mathord{-}}x = 0\rd x$, are commonly used, and will play a role in the paper. Negations bind stronger than multiplication, which binds stronger than divisions, which in turn bind stronger than the lattice connectives. The following identities will be of some importance later.
1. \[integr\] $1\geq x$,
2. \[0-bound\] $0\leq x$,
3. \[rl\] $1 = 0$,
4. \[divis\] $x(x\ld (x\wedge y)) = x\wedge y = ((x\wedge y)\rd x)x$,
5. \[div-int\] $x(x\ld y) = x\wedge y = (y\rd x)x$,
6. \[prelin\] $x\ld y \vee y\ld x = 1 = y\rd x \vee x\rd y$,
7. \[MV-int\] $x\rd(y\ld x) = x\vee y = (x\rd y)\ld x$,
8. \[MV-gen\] $x\rd((x\vee y)\ld x) = x\vee y = (x\rd(x\vee y))\ld x$,
9. \[dbl-neg\] ${\mathord{\sim}}{\mathord{-}}x = x = {\mathord{-}}{\mathord{\sim}}x$,
10. \[good\] ${\mathord{\sim}}{\mathord{-}}x = {\mathord{-}}{\mathord{\sim}}x$,
11. \[lg\] $1 = x(x\ld 1)$,
12. \[comm\] $xy = yx$.
We write $\mathsf{FL}$ for the variety of all FL-algebras. In general, if X-algebras are defined and happen to form a variety, we will use sans-serif $\mathsf{X}$ for that variety. Thus, $\mathsf{FL_i}$ is the variety of $\text{FL}_i$-algebras: these are FL-algebras satisfying (\[integr\]); they are also known as *integral*. FL-algebras satisfying (\[0-bound\]) are called $\text{FL}_o$-algebras, or *zero-bounded*. FL-algebras that are both integral and zero-bounded are known as $\text{FL}_w$-algebras. FL-algebras satisfying (\[rl\]) are *residuated lattices*; this trick allows us to make the constant $0$ disappear; it is especially useful for viewing lattice-ordered groups ($\ell$-groups) as a variety of FL-algebras.
The identities (\[divis\]) and (\[div-int\]) are both known as *divisibility*; in varieties of integral FL-algebras (\[divis\]) is equivalent to (\[div-int\]), but not so in general. FL-algebras satisfying divisibility in the form (\[divis\]) are *GBL-algebras*; integral GBL-algebras are typically defined using (\[div-int\]) instead. This has the advantage of making lattice meet definable by means of (either of) the divisions. As integrality is derivable from (\[div-int\]), it is a natural and economical choice of an equational base. For more on $\mathsf{GBL}$ we refer the reader to [@GaTs; @JiMo]. The identities in (\[prelin\]) are together known as *prelinearity*. $\text{FL}_w$-algebras satisfying prelinearity and divisibility are *pseudo BL-algebras* (cf. e.g., [@DGI1; @DGI2]). This variety is of prime importance in this paper, since the kites from the title are certain pseudo BL-algebras constructed out of $\ell$-groups in a special way. We use $\mathsf{psBL}$ as a name for the variety of pseudo BL-algebras.
An important subvariety of $\mathsf{psBL}$ is the variety $\mathsf{psMV}$ of *pseudo MV-algebras*: these are pseudo BL-algebras satisfying (\[MV-int\]), or, equivalently[^1], pseudo BL-algebras satisfying (\[dbl-neg\]). Pseudo MV-algebras are categorically equivalent to the class of $\ell$-groups with a strong unit, as shown in [@Dvu1]. The role of (\[MV-gen\]) with respect to (\[MV-int\]) is analogous to that of (\[divis\]) with respect to (\[div-int\]), namely, (\[MV-gen\]) it is a non-integral version of (\[MV-int\]). In particular, (\[MV-gen\]) holds in $\ell$-groups, while (\[MV-int\]) does not. For this to make sense, we need to interpret $\ell$-groups as residuated lattices, and this is done with the help of (\[lg\]). Namely, the subvariety of residuated lattices satisfying (\[lg\]) is term equivalent to the variety of $\ell$-groups, upon defining $x^{-1} = x\ld 1$ one way, and $x\ld y = x^{-1}y$, $x\rd y = xy^{-1}$ the other.
Pseudo BL-algebras satisfying (\[good\]), a natural weakening of (\[dbl-neg\]), are known by an unassuming name of *good*. For example, every pseudo MV-algebra is good. It was an open question for some time whether every pseudo BL-algebra was good (cf. [@DGI2 Problem 3.21]). It was resolved in the negative in [@DGK], and the counterexample was in fact a special type of what we now call a kite, defined in [@JiMo].
Finally, commutative pseudo BL-algebras are just *BL-algebras* and commutative pseudo MV-algebras are *MV-algebras*. This reflects the order of discovery: “pseudo” varieties[^2] were discovered as noncommutative generalizations of BL- and MV-algebras, respectively.
Let $\mathbf{A}$ be an FL-algebra. Given $a \in A,$ we define $a^1:=a$ and $a^{n+1}:=a^na,$ for $n \ge 1.$ An element $a\in A$ is said to be (i) *idempotent* if $a^2 = a$, (ii) *Boolean* if it is idempotent and ${\mathord{-}}{\mathord{\sim}}a=a={\mathord{\sim}}{\mathord{-}}a$. Let $B(\mathbf A)$ be the set of Boolean elements of $\mathbf A.$ It is the greatest Boolean subalgebra of $\mathbf A.$ If $\mathbf A$ is a BL-algebra, by [@DGI2 Prop 2.10], an element $a\in A$ is Boolean iff $a \vee {\mathord{\sim}}a =1$ iff $a\vee {\mathord{-}}a$. Then ${\mathord{-}}a= {\mathord{\sim}}a$.
We now recall some basic facts about arithmetical and structural properties of FL-algebras, residuated lattices and pseudo BL-algebras that we wish to use freely later.
\[arith\] Let $\mathbf{A}$ be an FL-algebra. Then the equivalences $$x\leq z\rd y \text{\quad if{}f\quad} xy\leq z \text{\quad
if{}f\quad} y\leq x\ld z$$ hold for any $x,y,z\in A$. If $\mathbf{A}$ is integral, then $xy
\leq x\wedge y$ always holds, so in particular $x^2\leq x$; moreover $x\ld y = 1 = y\rd x$ hold iff $x\leq y$. If $\mathbf{A}$ is divisible and $x\geq y$, there exist elements $z_1$ and $z_2$ such that $xz_1 = y = z_2x$; in fact $z_1 = x\ld y$ and $z_2 = y\rd x$. If $\mathbf{A}$ is divisible and $z$ is idempotent, then $xz =
x\wedge z = zx$ holds for any $x\in A$.
Let $\mathbf{A}$ be an FL-algebra. A *left conjugate* of an element $x\in A$ by an element $y\in A$ is the element $y\ld
xy\wedge 1$, and its *right conjugate* is $yx\rd y\wedge 1$. An element $x\in A$ is *central* if $yx = xy$ holds for all $y\in
A$. By the last statement of Lemma \[arith\] all idempotent elements in pseudo BL-algebras are central. A *filter* of $\mathbf{A}$ is a set $F\subseteq A$ such that (i) $F$ is a subalgebra of $0$-free reduct of $\mathbf{A}$ and (ii) $F$ is convex as an ordered set. A filter $F$ is normal if it is closed under conjugates, i.e., for all $x\in F$ and all $y\in A$ both $y\ld xy$ and $yx\rd y$ belong to $F$. Normal filters are also called *convex normal subalgebras* (e.g. in [@GJKO]), which makes good sense for residuated lattices but can be confusing for FL-algebras: convex normal subalgebras may not be subalgebras in the proper sense because they do not need to contain $0$. If $\mathbf{A}$ is integral, filters can be alternatively defined as subsets of $A$ that are upward closed and closed under multiplication. For a set $S\subseteq A$ we denote its upward closure by $\up S$. If $S = \{s\}$ we write $\up s$ instead of $\up\{s\}$.
\[cong\] Let $\mathbf{A}$ be an FL-algebra. For each congruence $\theta$ on $\mathbf{A}$, its class $1/\theta$ is a normal filter. Conversely, each normal filter $F$ corresponds to a unique congruence $\theta_F$ on $\mathbf{A}$. This correspondence establishes a lattice isomorphism between the lattices of normal filters of $\mathbf{A}$ and congruences of $\mathbf{A}$. If $\mathbf{A}$ is integral and $z\in A$ is idempotent and central, then $\up z$ is a normal filter. If $\mathbf{A}$ is, moreover, divisible and $y\in A$ is idempotent, then $\up y$ is normal.
Kites
=====
Let $\mathbf{G}$ be an $\ell$-group, and $I$, $J$ be sets with $|J|\leq |I|$. Since only the cardinalities of $I$ and $J$ matter for the construction, it is harmless to think of these sets as ordinals. We will do so explicitly in the next section. Let further $\lambda,\rho\colon J\to I$ be injections. Now we define an algebra with the universe $(G^+)^J \uplus (G^-)^I$. We order its universe by keeping the original coordinatewise ordering within $(G^+)^J$ and $(G^-)^I$, and setting $x\leq y$ for all $x\in(G^+)^J$, $y\in(G^-)^I$. It is easy to verify that this is a (bounded) lattice ordering of $(G^+)^J \uplus (G^-)^I$. Notice also that the case $I=J$ is not excluded, so the element $e^I$ may appear twice: at the bottom of $(G^+)^J$ and at the top of $(G^-)^I$. To avoid confusion in the definitions below, we adopt a convention of writing $a_i^{-1},b_i^{-1}, \dots$ for elements of $(G^-)^I$ and $f_j,g_j,\dots$ for elements of $(G^+)^J$. In particular, we will write $e^{-1}$ for $e$ as an element of $G^-$. We also put $1$ for the constant sequence $(e^{-1})^I$ and $0$ for the constant sequence $e^J$. With these conventions in place we are ready to define multiplication, putting: $$\begin{aligned}
\langle a_i^{-1}\colon i\in I\rangle\cdot\langle b_i^{-1}\colon i\in I\rangle &=
\langle(b_ia_i)^{-1}\colon i\in I\rangle\\
\langle a_i^{-1}\colon i\in I\rangle\cdot\langle f_j\colon j\in J\rangle &=
\langle a_{\lambda(j)}^{-1}f_j\vee e\colon j\in J\rangle\\
\langle f_j\colon j\in J\rangle\cdot\langle a_i^{-1}\colon i\in I\rangle &=
\langle f_ja_{\rho(j)}^{-1}\vee e\colon j\in J\rangle\\
\langle f_j\colon j\in J\rangle\cdot\langle g_j\colon j\in J\rangle &=
\langle e\colon j\in J\rangle = 0.\\\end{aligned}$$
\[resid\] Divisions, $\rd$ and $\ld$, corresponding to multiplication defined as above on $(G^+)^J \uplus (G^-)^I$ are defined by: $$\begin{aligned}
\langle a_i^{-1}\colon i\in I\rangle\ld\langle b_i^{-1}\colon i\in I\rangle &=
\langle a_ib_i^{-1}\wedge e^{-1}\colon i\in I\rangle\\
\langle b_i^{-1}\colon i\in I\rangle\rd\langle a_i^{-1}\colon i\in I\rangle &=
\langle b_i^{-1}a_i\wedge e^{-1}\colon i\in I\rangle\\
\langle a_i^{-1}\colon i\in I\rangle\ld\langle f_j\colon j\in J\rangle &=
\langle a_{\lambda(j)}f_j\colon j\in J\rangle\\
\langle f_j\colon j\in J\rangle\rd\langle a_i^{-1}\colon i\in I\rangle &=
\langle f_ja_{\rho(j)}\colon j\in J\rangle\\
\langle f_j\colon j\in J\rangle\ld\langle g_j\colon j\in J\rangle &= \langle a_i^{-1}\colon i\in I\rangle,\\
\text{ where } a_i^{-1} &=\begin{cases}
f_{\rho^{-1}(i)}^{-1}g_{\rho^{-1}(i)}\wedge e^{-1} &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise}
\end{cases}\\
\langle g_j\colon j\in J\rangle\rd\langle f_j\colon j\in J\rangle &= \langle b_i^{-1}\colon i\in I\rangle,\\
\text{ where } b_i^{-1} &=\begin{cases}
g_{\lambda^{-1}(i)}f_{\lambda^{-1}(i)}^{-1}\wedge e^{-1} &
\text{ if } \lambda^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise},
\end{cases}\\
\langle a_i^{-1}\colon i\in I\rangle\rd \langle f_j\colon j\in
J\rangle &=(e^{-1})^I= \langle f_j\colon j\in J\rangle \ld \langle
a_i^{-1}\colon i\in I\rangle.\end{aligned}$$
We will call the algebra we have just defined a *kite* of $\mathbf{G}$, and write $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ for it. Observe that if we take $I=J$, then $\lambda$ and $\rho$ become permutations of the set of coordinates and so the kite construction is reminiscent of wreath product. This analogy is not mistaken, as we will see later. For the moment, let us focus on the algebra $K_{I,J}^{\lambda,\rho}(\mathbf{G})$. The next lemma shows that the algebra $K_{I,J}^{\lambda,\rho}(\mathbf{G}) =
\langle (G^+)^J \uplus (G^-)^I;\vee,\wedge,\cdot,\ld,\rd, 1,0\rangle$ is a pseudo BL-algebra.
\[hyper-psBL\] For any $\ell$-group $\mathbf{G}$ and any choice of appropriate sets $I,J$ and maps $\lambda,\rho$, the algebra $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is a pseudo BL-algebra.
It is clear that $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is a lattice-ordered groupoid with unit. To show that multiplication is associative, first observe that triples from $(G^-)^I$ associate because $(G^-)^I$ is just the negative cone of $\mathbf{G}^I$. Next, triples involving at least two elements from $(G^+)^J$ associate because both the products equal $0$. The remaining cases all involve one element from $(G^-)^I$ and two from $(G^+)^J$. One such case is: $$\begin{aligned}
(\langle a_i^{-1}\colon i\in I\rangle\cdot\langle f_j\colon j\in J\rangle)\cdot\langle b_i^{-1}\colon i\in I\rangle &=
\langle a_{\lambda(j)}^{-1}f_j\vee e\colon j\in J\rangle\cdot\langle b_i^{-1}\colon i\in I\rangle\\
&= \langle( a_{\lambda(j)}^{-1}f_j\vee e)b_{\rho(j)}^{-1}\vee e\colon j\in J\rangle\\
&= \langle a_{\lambda(j)}^{-1}f_jb_{\rho(j)}^{-1}\vee b_{\rho(j)}^{-1}\vee e\colon j\in J\rangle\\
&= \langle a_{\lambda(j)}^{-1}f_jb_{\rho(j)}^{-1}\vee e\colon j\in J\rangle\\
&= \langle a_{\lambda(j)}^{-1}f_jb_{\rho(j)}^{-1}\vee a_{\lambda(j)}^{-1}\vee e\colon j\in J\rangle\\
&= \langle a_{\lambda(j)}^{-1}(f_jb_{\rho(j)}^{-1}\vee e)\colon j\in J\rangle\\
&= \langle a_i\colon i\in I\rangle\cdot\langle f_jb_{\rho(j)}^{-1}\vee e\colon j\in J\rangle\\
&= \langle a_i^{-1}\colon i\in I\rangle\cdot(\langle f_j\colon j\in J\rangle\cdot\langle b_i^{-1}\colon i\in I\rangle).\end{aligned}$$ Other cases follow by similar calculations. Now, to show that divisibility holds, we also proceed case by case. Let us deal with two cases here. The first is: $$\begin{aligned}
\langle f_j\colon j\in J\rangle\cdot(\langle f_j\colon j\in J\rangle\ld\langle g_j\colon j\in J\rangle)
&= \langle f_j\colon j\in J\rangle\cdot\langle a_i^{-1}\colon i\in I\rangle \\
&= \langle f_ja_{\rho(j)}^{-1}\vee e\colon j\in J\rangle\end{aligned}$$ where $$a_i^{-1} =\begin{cases}
f_{\rho^{-1}(i)}^{-1}g_{\rho^{-1}(i)}\wedge e^{-1} &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise}
\end{cases}$$ but, observe that $\rho^{-1}(\rho(j))$ is always defined and equals $j$, so calculating further we obtain $f_ja_{\rho(j)}^{-1} = g_j\wedge f_j$ for every $j\in J$, and therefore $$\langle f_ja_{\rho(j)}^{-1}\vee e\colon j\in J\rangle =
\langle(f_j\wedge g_j)\vee e\colon j\in J\rangle =
\langle f_j\wedge g_j\colon j\in J\rangle$$ as required. For the second, take: $$\begin{aligned}
\langle a_i^{-1}\colon i\in I\rangle\cdot(\langle a_i^{-1}\colon i\in I\rangle\ld\langle f_j\colon j\in J\rangle) &=
\langle a_i^{-1}\colon i\in I\rangle\cdot\langle a_{\lambda(j)}f_j\colon j\in J\rangle\\
&= \langle a_{\lambda(j)}^{-1}a_{\lambda(j)}f_j\vee e\colon j\in J\rangle\\
&= \langle f_j\colon j\in J\rangle.\end{aligned}$$ All other cases are straightforward. It remains to show prelinearity. Since multiplication and divisions in $(G^-)^I$ are defined coordinatewise, prelinearity for $x,y\in (G^-)^I$ is inherited from $G^-$. If $x\in (G^-)^I$ and $y\in (G^+)^J$, or *vice versa* prelinearity holds trivially. For the only remaining case, calculating $$\langle f_j\colon j\in J\rangle\ld\langle g_j\colon j\in J\rangle\vee
\langle g_j\colon j\in J\rangle\ld\langle f_j\colon j\in J\rangle$$ yields two cases: (1) if $\rho^{-1}(i)$ is defined, we have $$\begin{aligned}
& f_{\rho^{-1}(i)}^{-1}g_{\rho^{-1}(i)}\wedge e^{-1}
\vee g_{\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)}\wedge e^{-1}\\
&= (f_{\rho^{-1}(i)}^{-1}g_{\rho^{-1}(i)}\vee
g_{\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)})\wedge e^{-1})\\
&= e^{-1}\end{aligned}$$ and (2) if $\rho^{-1}(i)$ is not defined, we have $$a^{-1}_i \vee a^{-1}_i = e^{-1}\vee e^{-1}= e^{-1}$$ as well. Thus, prelinearity holds and that finishes the proof of all the claims in the lemma.
Somewhat surprisingly, many kites turn out to be pseudo MV-algebras.
\[hyper-psMV\] Let $\mathbf{G}$ be an $\ell$-group, and suppose $|I| = |J|$ and $\lambda,\rho$ are bijections. Then $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is a pseudo MV-algebra.
By Lemma \[hyper-psBL\] we only need to show that under the conditions of the lemma, the identity $x\rd(y\ld x) = x\vee y = (x\rd y)\ld x$ holds. We have two nontrivial cases to consider:
*Case 1*. $x\in(G^+)^J$ and $y\in(G^-)^I$. Then, $x = \langle f_j\colon j\in J\rangle \leq \langle a_i^{-1}\colon i\in I\rangle = y$, and so we calculate: $$\begin{aligned}
\langle f_j\colon j\in J\rangle\rd(\langle a_i^{-1}\colon i\in I\rangle\ld\langle f_j\colon j\in J\rangle) &=
\langle f_j\colon j\in J\rangle\rd\langle a_{\lambda(j)}f_j\colon j\in J\rangle\\
&= \langle f_{\lambda^{-1}(i)}(a_{\lambda(j)}f_j)^{-1}_{\lambda^{-1}(i)}\wedge e^{-1}\colon i\in I\rangle\\
&= \langle f_{\lambda^{-1}(i)}f_{\lambda^{-1}(i)}^{-1}a_i^{-1}\colon i\in I\rangle\\
&= \langle a_i^{-1}\colon i\in I\rangle\end{aligned}$$ which shows that $x\rd(y\ld x) = y = x\vee y$ holds. Notice that we used bijectiveness of $\lambda$ to pass from the first to the second equality above.
*Case 2*. $x,y\in(G^+)^J$. Then, $x = \langle f_j\colon j\in J\rangle$ and $y = \langle g_j\colon j\in J\rangle$, and so we calculate: $$\begin{aligned}
\langle f_j\colon j\in J\rangle\rd(\langle g_j\colon j\in J\rangle\ld\langle f_j\colon j\in J\rangle) &=
\langle f_j\colon j\in J\rangle\rd\langle g_{\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)}\wedge e^{-1}\colon i\in I\rangle\\
&= \langle f_j(g_{\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)}\wedge e^{-1})^{-1}_{\rho(j)}\colon j\in J\rangle\\
&= \langle f_j(f_j^{-1}g_j\vee e)\colon j\in J\rangle\\
&= \langle g_j\vee f_j\colon j\in J\rangle\end{aligned}$$ which again shows that $x\rd(y\ld x) = x\vee y$ holds. The proofs for $x\vee y = (x\rd y)\ld x$ are symmetric.
As we already mentioned, it was an open problem ([@DGI2 Problem 3.21]) for a while whether every pseudo BL-algebra is good. We found a negative solution in [@DGK] using special types of kites from [@JiMo]. Below we characterize good kites, and in the last section we will exhibit a countably infinite family of varieties of kites with the property that their all and only good members are Boolean algebras.
\[good-kites\] Let $\mathbf{G}$ be an $\ell$-group, and $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ a kite.
1. $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is good if and only if $\lambda(J)=\rho(J)$.
2. $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is a pseudo MV-algebra if and only if $\lambda(J)=I=\rho(J)$.
\(1) Let $x= \langle a_i\colon i\in I\rangle.$ Then ${\mathord{-}}x = \langle
a_{\rho(j)}\colon j\in J\rangle$ and ${\mathord{\sim}}x= \langle
a_{\lambda(j)}\colon j\in J\rangle.$ In addition, ${\mathord{\sim}}{\mathord{-}}x=
\langle z_i\colon i\in I\rangle, $ where $$z_i^{-1}=\begin{cases} a_i^{-1} &
\text{ if } \lambda^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise},
\end{cases}$$ and ${\mathord{-}}{\mathord{\sim}}x= \langle y_i\colon i\in I\rangle$, where $$y_i^{-1}=\begin{cases} a_i^{-1} &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise}.
\end{cases}$$ Now, if $x = \langle f_j\colon j\in J\rangle$, we have ${\mathord{\sim}}{\mathord{-}}x = \langle g_j\colon j\in J\rangle$, where $$g_j =
\begin{cases} f_j & \text{ if } \rho^{-1}(i) \text{ is defined}\\
e & \text{ otherwise},
\end{cases}$$ and ${\mathord{-}}{\mathord{\sim}}x = \langle h_j\colon j\in J\rangle$, where $$h_j =
\begin{cases} f_j & \text{ if } \lambda^{-1}(i) \text{ is defined}\\
e & \text{ otherwise}.
\end{cases}$$ Hence, if $\lambda(J)=\rho(J),$ the kite $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is good.
Conversely, assume that the kite $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is good, and let $\lambda(J)\ne
\rho(J).$ For our aims we can assume that each $a_i^{-1} \ne
e^{-1}.$ There is an $i\in I$ such that either $\lambda^{-1}(i)$ or $\rho^{-1}(i)$ is not defined. Equivalently, $x_i^{-1}= e^{-1}$ and $x_i^{-1}=a_i^{-1}$ or $y_i^{-1}=e^{-1}$ and $y_i^{-1}=a_i^{-1}.$ Hence, $\lambda(J)=\rho(J).$
\(2) If $\lambda(J)=I=\rho(J)$, then $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is a pseudo MV-algebra by Lemma \[hyper-psMV\]. Conversely, let the kite be a pseudo MV-algebra. Since every pseudo MV-algebra is good, by the first part of the present proof, we have $\lambda(J)=\rho(J)$. Now assume by absurd that there is an $i\in I\setminus \lambda(J)$. Then both $\lambda^{-1}(i)$ and $\rho^{-1}(i)$ are not defined and whence $x_i^{-1}=e^{-1} = y_i^{-1} \ne a_i$ which contradicts the property ${\mathord{\sim}}{\mathord{-}}x=x={\mathord{-}}{\mathord{\sim}}x$. Whence $\lambda(J)=I=\rho(J)$.
One-dimensional elements
------------------------
One property of kites that is frequently used in calculations is that double negations amount to certain shifts (often, rotations) of coordinate system. Below, we will state this observation in the form that will later help characterize finite-dimensional kites.
Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a kite. An element $a\in (G^-)^I$ will be called $\alpha$-dimensional, for some cardinal $\alpha$, if $|\{i\in I\colon a(i) \neq e\}| = \alpha$. Similarly, this notion applies to elements of $(G^+)^J$. One-dimensional elements are particularly easy to work with, and, moreover, it is immediately seen that every element of a kite is a join or a meet of one-dimensional elements.
\[rotations\] Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a kite and $a\in
(G^-)^I$ be one-dimensional, with $a(i)\neq e^{-1}$. Then ${\mathord{\sim}}{\mathord{\sim}}a$, ${\mathord{\sim}}{\mathord{-}}a$, ${\mathord{-}}{\mathord{\sim}}a$, and ${\mathord{-}}{\mathord{-}}a$ are at most one-dimensional. If $\lambda^{-1}(i)$ and $\rho^{-1}(i)$ are defined, these elements are exactly one-dimensional. More precisely, the following hold:
1. ${\mathord{\sim}}{\mathord{\sim}}a< 1$ and ${\mathord{-}}{\mathord{\sim}}a<1$ iff $\lambda^{-1}(i)$ is defined,
2. ${\mathord{-}}{\mathord{-}}a< 1$ and ${\mathord{\sim}}{\mathord{-}}a<1$ iff $\rho^{-1}(i)$ is defined,
3. ${\mathord{-}}{\mathord{\sim}}a = a$ iff $\lambda^{-1}(i)$ is defined,
4. ${\mathord{\sim}}{\mathord{-}}a = a$ iff $\rho^{-1}(i)$ is defined,
5. ${\mathord{\sim}}{\mathord{\sim}}a\vee a = 1$ iff $\rho(\lambda^{-1}(i))\neq i$,
6. ${\mathord{-}}{\mathord{-}}a\vee a = 1$ iff $\lambda(\rho^{-1}(i))\neq i$.
To begin with, since ${\mathord{\sim}}{\mathord{-}}a \geq a$ and ${\mathord{-}}{\mathord{\sim}}a\geq a$ hold in any residuated lattice, ${\mathord{\sim}}{\mathord{-}}a$ and ${\mathord{-}}{\mathord{\sim}}a$ are at most one-dimensional. Let us calculate ${\mathord{\sim}}{\mathord{\sim}}a$. To make the notation less cumbersome, we first regard $a$ as a sequence $\langle
e^{-1},\dots, a^{-1}(i),e^{-1},\dots\rangle$, and then write $\langle a^{-1}(i)\rangle$ for $\langle
e^{-1},\dots,a^{-1}(i),e^{-1},\dots\rangle$ and $\langle
a(i)\rangle$ for $\langle e,\dots, a(i), e,\dots\rangle$. Similarly, we put $\langle e\rangle$ for $e^J = 0$ and $\langle e^{-1}\rangle$ for $(e^{-1})^I = 1$. Calculating $\langle
a^{-1}(i)\rangle\ld\langle e\rangle$ we see that it is different from $e$ at coordinate $\lambda(j)$ if and only if $\lambda(j) = i$. Therefore $$\langle a^{-1}(i)\rangle\ld\langle e\rangle =\begin{cases}
\langle a(\lambda^{-1}(i))\rangle &
\text{ if } \lambda^{-1}(i) \text{ is defined}\\
\langle e\rangle & \text{ otherwise}.
\end{cases}$$ Then, calculating $\langle a(\lambda^{-1}(i))\rangle\ld\langle e\rangle$ in turn, we get that it is different from $e^{-1}$ at coordinate $k$ if and only if $\rho^{-1}(k)$ is defined and equal to $\lambda^{-1}(i)$, that is if and only if $k = \rho(\lambda^{-1}(i))$. Altogether, we have $$\bigl(\langle a^{-1}(i)\rangle\ld\langle e\rangle\bigr)\ld\langle e\rangle =\begin{cases}
\langle a^{-1}(\rho\circ\lambda^{-1}(i))\rangle &
\text{ if } \lambda^{-1}(i) \text{ is defined}\\
\langle e^{-1}\rangle & \text{ otherwise}.
\end{cases}$$ By similar calculations we also obtain $$\begin{aligned}
\langle e\rangle\rd\bigl(\langle a^{-1}(i)\rangle\ld\langle e\rangle\bigr) &=\begin{cases}
\langle a^{-1}(i)\rangle &
\text{ if } \lambda^{-1}(i) \text{ is defined}\\
\langle e^{-1}\rangle & \text{ otherwise}
\end{cases}\\
\bigl(\langle e\rangle\rd\langle a^{-1}(i)\rangle\bigr)\ld\langle e\rangle &=\begin{cases}
\langle a^{-1}(i)\rangle &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
\langle e^{-1}\rangle & \text{ otherwise}
\end{cases}\\
\langle e\rangle\rd\bigl(\langle e\rangle\rd\langle a^{-1}(i)\rangle\bigr) &=\begin{cases}
\langle a^{-1}(\lambda\circ\rho^{-1}(i))\rangle &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
\langle e^{-1}\rangle & \text{ otherwise}.
\end{cases}\end{aligned}$$ Then all the claims follow from these calculations and coordinatewise ordering of $(G^-)^I$. To give just one example, for (5) we have ${\mathord{\sim}}{\mathord{\sim}}a\vee a = 1$ if and only if $\langle
a^{-1}(\rho\circ\lambda^{-1}(i))\rangle\vee \langle a^{-1}(i)\rangle
= \langle e^{-1}\rangle$ if and only if $\rho\circ\lambda^{-1}(i)\neq i$.
Examples of kites
=================
As we have already remarked, in the kite construction we can think of the index sets $I$ and $J$ as ordinals. Doing so systematically also makes classification of kites easier, so throughout this section we assume that $I$ and $J$ are ordinals. Below we give five examples of rather familiar algebras that can be rendered as kites.
Boolean algebras
----------------
Let $I = 0 = J$. Then, $(G^-)^I$ and $(G^+)^J$ are both singletons, and $\lambda = \rho$ can only be the empty function (hence, an injection). Thus, $K_{0,0}^{\emptyset,\emptyset}(\mathbf{G})$ is the two-element Boolean algebra for any $\ell$-group $\mathbf{G}$. We also get the two-element Boolean algebra in another way. Namely, let $\mathbf{O}$ be the trivial $\ell$-group. Then $\lambda = id =
\rho$, and $K_{I,J}^{id,id}(\mathbf{O})$ is the two-element Boolean algebra for any choice of $I$ and $J$.
Product logic algebras
----------------------
Let $I = 1$ and $J=0$. As the only function from $J$ to $I$ is the empty function (which is an injection), the kite $K_{1,0}^{\emptyset,\emptyset}(\mathbf{G})$ is well-defined for any $\ell$-group $\mathbf{G}$. Let $\mathbb{R}^+$ stand for the $\ell$-group of positive reals under multiplication. Then $K_{1,0}^{\emptyset,\emptyset}(\mathbb{R}^+)$ is isomorphic to the *standard product logic algebra*, i.e., the real interval $[0,1]$ with usual multiplication, and divisions given by $x\rd y =
\frac{x}{y} = y\ld x$ for $y\neq 0$ and $x\rd 0 = 1 = 0\ld x$. Taking $\mathbb{Z}$ for $\mathbf{G}$, we obtain as $K_{1,0}^{\emptyset,\emptyset}(\mathbb{Z})$ the algebra $\mathbb{Z}^-_\bot =
\langle\mathbb{Z}^-\cup\{\bot\};\mathrm{max},\mathrm{min},
+,-,\bot,0\rangle$, where $\bot = -\infty$. This is also a product logic algebra. Both $K_{1,0}^{\emptyset,\emptyset}(\mathbb{R}^+)$ and $K_{1,0}^{\emptyset,\emptyset}(\mathbb{Z})$ generate the whole variety of product logic algebras. This variety covers the variety of Boolean algebras.
Jipsen-Montagna algebras
------------------------
Taking $I = 2$, $J =1$, $\lambda(0) = 0$, and $\rho(0) = 1$, we get that $K_{2,1}^{\lambda,\rho}(\mathbf{G})$ is a Jipsen-Montagna algebra, for any $\ell$-group $\mathbf{G}$. In particular, the algebra $K_{2,1}^{\lambda,\rho}(\mathbb{Z})$ is a pseudo BL-algebra which is not good [@DGK] and it generates another cover of the variety of Boolean algebras. In Section \[appl\] we will show that the same holds for $K_{n+1,n}^{\lambda,\rho}(\mathbb{Z})$ with $\lambda(i) = i$ and $\rho(i) = i+1$, for an arbitrary $n\in
\omega$.
Chang chain and its subdirect powers
------------------------------------
Now, let $I = 1 = J$. Then $K_{1,1}^{id,id}(\mathbb{Z})$ is the Chang chain, denoted $\mathbf{S}^\omega_1$ in Komori’s paper, and $\mathbf{C}_\infty$ in [@GJKO]. The variety generated by $K_{1,1}^{id,id}(\mathbb{Z})$ is also a cover of the variety of Boolean algebras. For $I = J > 1$, the kite $K_{I,I}^{id,id}(\mathbb{Z})$ is subdirectly embeddable in $(\mathbf{C}_\infty)^I$ and thus $\mathsf V(K_{1,1}^{id,id}(\mathbb{Z})) =
\mathsf V(K_{I,I}^{id,id}(\mathbb{Z}))$.
Intervals in Scrimger groups
----------------------------
Taking $I = J = n$ for $n\geq 2$, and putting $\lambda(i) = i$ and $\rho(i) = i+1$ ($\mathrm{mod}\ n$), we get that $K_{I,I}^{\lambda,\rho}(\mathbb{Z})$ is isomorphic to $\Gamma(\mathbf{G}_n, (\langle 0\rangle,1))$, where $\mathbf{G}_n$ is the subgroup of $\mathbb{Z}\wr\mathbb{Z}$ (antilexicographically ordered), consisting of the elements $\langle\langle a_i\colon
i\in\mathbb{Z}\rangle, b\rangle$, such that $i=j$ ($\mathrm{mod}\
n$) implies $a_i = a_j$.
Subdirectly irreducible kites
=============================
We will characterise subdirectly irreducible kites and show they fall into five broad classes. To begin with, we state a few facts on normal filters in kites.
\[up-max-norm\] Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a kite. Then $(G^-)^I$ is a maximal normal filter of $K_{I,J}^{\lambda,\rho}(\mathbf{G})$.
It is clear that $(G^-)^I$ is a maximal filter, so we only need to show that $(G^-)^I$ is closed under conjugation by elements outside $(G^-)^I$. Let $x\in (G^-)^I$ and $y\in(G^+)^J$. Taking $y\ld xy$ we observe that $xy\in(G^+)^J$. By definition of divisions, we then get that $y\ld xy\in(G^-)^I$. By symmetry the same holds for right conjugates.
\[hyper-filters\] Let $\mathbf{G}$ be an $\ell$-group, and $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ a kite. Let $F$ be a convex normal subgroup of $\mathbf{G}$, and $N = F|_{G^-}$. Then $N^I =
\{\langle a^{-1}_i\colon i\in I\rangle\colon a^{-1}_i\in N\}$ is a normal filter of $K_{I,J}^{\lambda,\rho}(\mathbf{G})$.
Since multiplication and order are defined coordinatewise on $(G^-)^I$, it is obvious that $N^I$ is a filter. We need to show that $N^I$ is closed under conjugates. Now, conjugation by an element of $(G^-)^I$ also proceeds coordinatewise, so $N^I$ is closed under all such conjugates. Take an element $\langle f_j\colon j\in J\rangle$ and consider $\langle f_j\colon j\in J\rangle\ld
\langle a_i^{-1}\colon i\in I\rangle\cdot\langle f_j\colon j\in J\rangle$, for some $\langle a_i^{-1}\colon i\in I\rangle\in N^I$. This is equal to $\langle f_j\colon j\in J\rangle\ld
\langle a_{\lambda(j)}^{-1}f_j\vee e\colon j\in J\rangle$, which in turn equals $\langle b_i^{-1}\colon i\in I\rangle$, where $$b_i^{-1} =\begin{cases}
\bigl(f_{\rho^{-1}(i)}^{-1}a_{\lambda\circ\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)}\wedge
e^{-1}\bigr) \vee f_{\rho^{-1}(i)}^{-1} &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise}
\end{cases}\eqno(0)\\$$ as one can verify by a series of simple calculations. Notice that $f_{\rho^{-1}(i)}^{-1}a_{\lambda\circ\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)}$ is a conjugate of a member of $N$ and thus belongs to $F$ by normality. Therefore $f_{\rho^{-1}(i)}^{-1}a_{\lambda\circ\rho^{-1}(i)}^{-1}f_{\rho^{-1}(i)}\wedge
e^{-1}$ belongs to $N$. Thus, by upward closedness of $N$ we get that $b_i\in N$, for each $i\in I$, and so $\langle b_i^{-1}\colon
i\in I\rangle\in N^I$. This shows that $N^I$ is normal, as required.
The converse of Lemma \[hyper-filters\] is also true, but it will be useful to define a technical notion before we show it. Namely, for a proper normal filter $N$ of a kite $K_{I,J}^{\lambda,\rho}(\mathbf{G})$, and a $k\in I$, we define the restriction of $N$ to a coordinate $k$ as the set of elements $a\in
N$ with $a(i) \neq e^{-1}$ iff $i = k$.
\[restrict\] Let $\mathbf{G}$ be an $\ell$-group, $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ a kite, and $N$ a normal filter of $K_{I,J}^{\lambda,\rho}(\mathbf{G})$. For any $k\in I,$ the restriction of $N$ to $k$ is a normal filter of $\mathbf{G}^-$, and therefore the negative part of a convex normal subgroup of $\mathbf{G}$.
Since $N$ is proper, it is contained in $(G^-)^I$, which is a direct power of $\mathbf{G}^-$. Then, the restriction of $N$ to $k$ is just the $k$-th projection of $N$, and the claim follows.
The next lemma shows that subdirectly irreducible kites can only arise out of subdirectly irreducible $\ell$-groups, and a first characterization of subdirectly irreducible kites follows.
\[si-only-if-si\] Let $\mathbf{G}$ be an $\ell$-group, and $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ a kite. If $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is subdirectly irreducible, so is $\mathbf{G}$.
We will prove the contrapositive. Suppose $\mathbf{G}$ is not subdirectly irreducible. Then, there exists a set $\{\mathbf{M}_s\colon s\in S\}$ of nontrivial convex normal subgroups of $\mathbf{G}$, such that $\bigcap_{s\in S}M_s = \{e\}$. Let $M^-_s = M_s|_{G^-}$. Then, by Lemma \[hyper-filters\], $(M^-_s)^I$ is a normal filter of $K_{I,J}^{\lambda,\rho}(\mathbf{G})$, for each $s\in S$. Suppose $\langle a_i\colon i\in I\rangle$ belongs to $(M^-_s)^I$ for each $s\in S$. Then, for any coordinate $k\in I$ we have that $a_k\in
M^-_s$ for all $s\in S$, and thus $a_k = e$. Therefore, $\bigcap_{s\in S}(M^-_s)^I = \{e^I\}$, showing that the set $\{(M^-_s)^I\colon s\in S\}$ of nontrivial normal filters of $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ intersects trivially. Thus, $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is not subdirectly irreducible, proving the claim.
\[si-kites\] Let $\mathbf{G}$ be an $\ell$-group, and $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ a kite. The following are equivalent:
1. $\mathbf{G}$ is subdirectly irreducible and for all $i,j\in I$ there exists $m\in\omega$ such that $(\rho\circ\lambda^{-1})^m(i) = j$ or $(\lambda\circ\rho^{-1})^m(i) = j$.
2. $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is subdirectly irreducible.
To prove that (1) implies (2), let $\mathbf{M}$ be the smallest nontrivial convex normal subgroup of $\mathbf{G}$, and $N$ be the restriction of $M$ to the negative cone of $\mathbf{G}$. Then, $N$ is the smallest nontrivial normal filter of $\mathbf{G}^-$. By Theorem \[hyper-filters\] we have that $N^I$ is a normal filter of $K_{I,J}^{\lambda,\rho}(\mathbf{G})$, so it remains to prove that $N^I$ is the smallest nontrivial such. Since for any $a\in
N^I\setminus\{1\}$ there is a one-dimensional element $a'$ with $a\leq a'<1$, it suffices to prove that any one-dimensional element $b\in N^I\setminus\{1\}$ generates $N^I$. Without loss of generality assume $b = \langle b^{-1}_0,e^{-1},\dots\rangle$; this is always achievable by a suitable re-ordering of $I$, regardless of its cardinality. Observe that $b^{-1}_0$ generates $N$, since $N$ is the smallest nontrivial normal filter of $\mathbf{G}^-$. It follows that $b$ generates all members of $N^I$ of the form $\langle a^{-1},
e^{-1},\dots\rangle$, using only conjugates of the same form. Consider an arbitrary $i\in I$. By assumption, there is $m\in\omega$ with $\rho(\lambda^{-1})^m(0) = i$ or $\lambda(\rho^{-1})^m(0) =
i$. Now, repeating $m$ times the calculation from the proof of Lemma \[rotations\], we obtain that for an element $u = \langle
a^{-1}, e^{-1},\dots\rangle$, one of the following must be the case:
- if $\rho(\lambda^{-1})^m(0) = i$, then $\underbrace{{\mathord{\sim}}\dots{\mathord{\sim}}}_{\text{$2m$-times}}u= \langle
e^{-1},\dots,e^{-1},a^{-1},e^{-1},\dots\rangle$,
- if $\lambda(\rho^{-1})^m(0) = i$, then $\underbrace{{\mathord{-}}\dots{\mathord{-}}}_{\text{$2m$-times}}u = \langle
e^{-1},\dots,e^{-1},a^{-1},e^{-1},\dots\rangle$,
where $a^{-1}$ occurs at a coordinate $i$; again by a suitable renumbering of $I$ we can assume it to be the $m$-th coordinate. By taking appropriate meets it then follows that every element of $N^I$ can be generated, which proves the claim.
For the the converse, by Lemma \[si-only-if-si\] we can assume $\mathbf{G}$ is subdirectly irreducible. Then, suppose there are $i,j\in I$ such that for all $m\in \omega$ we have $\rho(\lambda^{-1})^m(i) \neq j$ and $\lambda(\rho^{-1})^m(i) \neq
j$. We will call such $i$ and $j$ *disconnected*; otherwise, $i$ and $j$ will be called *connected*. If all distinct members of a $K\subseteq I$ are connected, we will call $K$ a *connected component* of $I$. Now, let $I_0$ and $I_1$ be connected components of $I$ such that $i\in I_0$ and $j\in I_1$. Clearly, $I_0$ and $I_1$ are disconnected, that is no member of $I_0$ is connected to any member of $I_1$. We will prove that $N^{I_0}\cap N^{I_1} = \{1\}$, from which it follows immediately that $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is not subdirectly irreducible. In fact, it suffices to show that for an element $u =
\langle a^{-1}_i\colon i\in I\rangle$ such that $a^{-1}_i = e^{-1}$ for all $i\notin I_0$, and for any element $b$, the conjugate $b\ld
ub$ has $b\ld ub(i) = e^{-1}$ if $i\notin I_0$, and the same holds for $bu\rd b$. Take $b = \langle b_j\colon j\in J\rangle$. We have $$\begin{aligned}
b\ld ub &=
\langle b_j\colon j\in J\rangle\ld
\langle a^{-1}_i\colon i\in I\rangle\cdot
\langle b_j\colon j\in J\rangle\\
&= \langle b_j\colon j\in J\rangle\ld
\langle a_{\lambda(j)}^{-1}b_j\vee e\colon j\in J\rangle\\
&= \langle c_i^{-1}\colon i\in I\rangle\end{aligned}$$ where $c_i^{-1} =\begin{cases}
b_{\rho^{-1}(i)}^{-1}
(a_{\lambda(\rho^{-1}(i))}b_{\rho^{-1}(i)}\vee e)
\wedge e^{-1} &
\text{ if } \rho^{-1}(i) \text{ is defined}\\
e^{-1} & \text{ otherwise.}
\end{cases}$\
Now, by assumption $a^{-1}_i = e^{-1}$ for $i\notin I_0$, and by connectedness, $\lambda(\rho^{-1}(i))\notin I_0$ if $i\notin I_0$. Therefore, $c_i^{-1}$ can be different from $e^{-1}$ only if $i\in I_0$, and thus $b\ld ub$ is of the required form. The claim for the other conjugate follows by symmetry.
If $I$ and $J$ are finite, the form of subdirectly irreducible kites is even more restricted than Theorem \[si-kites\] explicitly states. In such a case, $I$ can only be the same size as $J$ or bigger by one, and $\lambda$ and $\rho$ are essentially determined by the sizes of $I$ and $J$.
\[si-sizes\] If $\mathbf{G}$ is an $\ell$-group, $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ a subdirectly irreducible kite, and $I$ and $J$ are finite, then $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is isomorphic to one of:
1. $K_{n,n}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j$ and $\rho(j) = j+1\ (\mathrm{mod}\ n)$,
2. $K_{n+1,n}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j$ and $\rho(j) = j+1$.
Assume first that $|I| = n = |J|$. Then, since $\lambda$ and $\rho$ are injections, we can number the elements of $I$ and $J$ so that $\lambda(j) = j$. If $\rho(k) = k$ for some $k$, then $\{k\}$ is a connected component of $I$ and thus, by Theorem \[si-kites\], $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is not subdirectly irreducible, contradicting the assumption. So, $\rho(j) \neq j$, for every $j$. We can then renumber $I$ and $J$ so that $\rho(j) = j+1\
(\mathrm{mod}\ n)$.
For the second part, assume $|J|<|I|$. Then, there is $i\in I$ not in the range of $\rho$. We start numbering $I$ by putting $i = 0$. By Theorem \[si-kites\] we get that $I$ is connected, and so $0\in I$ is in the range of $\lambda$. We start numbering $J$ by putting $\lambda(0) = 0$. Now, $\rho(0)\neq 0$, so we put $\rho(0) = 1$. Then, there are two cases to consider.
*Case 1.* $\lambda^{-1}(1)$ is not defined. Then, if $J$ contains a nonempty subset $J'$ disconnected with $\{0,1\}$, Theorem \[si-kites\] yields a contradiction. Thus, $J = \{0,1\}$ and, since $\lambda$ and $\rho$ are injections, $I = \{0\}$. The claim holds in this case.
*Case 2.* $\lambda^{-1}(1)$ is defined. Then, we extend the numbering of $J$ putting $\lambda^{-1}(1) = 1$. Since $J|<|I|$, there is $j\in J\setminus\{0,1\}$. Now, if $\rho(1) = 0$, then $\{0,1\}$ and $j$ are disconnected, contradicting subdirect irreducibility. Therefore, $\rho(1)\neq 0$, and by injectiveness $\rho(1)\neq 1$. We then put $\rho(1) = 2$ extending the numbering of $I$.
Then, we repeat the procedure recursively, and by finiteness it must terminate. By inspection of the two cases, it is clear that it terminates in Case 1, for $j\in J$ such that $\lambda^{-1}(j)$ is not defined. Observe that $j = n+1 = \rho(n)$, where $n$ was numbered at the immediately preceding stage employing Case 2. This results in the required numbering of $I$ and $J$ and at the same time shows that $|I| = |J|+1$.
Kites of the form $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ with $I$ and $J$ finite will be called *finite-dimensional*. We will deal with these in more detail in the next section. Next, we focus on subdirectly irreducible kites with $I$ or $J$ infinite.
\[countable-dim\] Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a subdirectly irreducible kite. Then, $I$ and $J$ are at most countably infinite.
Suppose $I$ is uncountable, but $J$ is countable. Then, $\lambda(J)$ and $\rho(J)$ are also countable, and therefore $I\setminus \bigl(\lambda(J)\cup\rho(J)\bigr)$ is nonempty. Thus for any $k$ belonging to that set, we have that for $i = k = j$ neither of the conditions stated in the second part of Theorem \[si-kites\](2) can hold.
Suppose $J$ is uncountable, and thus so is $I$ because $\lambda$ and $\rho$ are injections. Take $i\in \lambda(J)$. The set $\bigcup_{m\in\omega}(\rho\circ\lambda^{-1})^m(i)$ is also countable and therefore $I\setminus \bigcup_{m\in\omega}(\rho\circ\lambda^{-1})^m(i)$ is uncountable. If $i\notin\rho(J)$, then any pair $(i, j)$ with $j\in I\setminus
\bigcup_{m\in\omega}(\rho\circ\lambda^{-1})^m(i)$ witnesses a failure of the conditions from the second part of Theorem \[si-kites\](2). If $i\in\rho(J)$, then, since $\bigcup_{m\in\omega}(\lambda\circ\rho^{-1})^m(i)$ is countable as well, we have a $k\in I\setminus\bigcup_{m\in\omega}(\lambda\circ\rho^{-1})^m(i)$. Then, the pair $(i, k)$ witnesses a failure of these conditions.
\[countable-dim-maps\] Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a subdirectly irreducible kite with countably infinite $I$ and $J$. Then, one of the following three cases must obtain:
1. $\lambda$ and $\rho$ are bijections.
2. $\lambda$ is a bijection and $|I\setminus\rho(J)| = 1$.
3. $\rho$ is a bijection and $|I\setminus\lambda(J)| = 1$.
From the conditions in Theorem \[si-kites\](2) it follows that $\rho(J)\cup\lambda(J) = I$. Suppose $i,j\in\lambda(J)\setminus\rho(J)$. Observe that any path of alternating $\lambda$ and $\rho$ between $i$ and $j$ must begin with $\lambda^{-1}$. Therefore, it must end with $\rho$, and so $i = j$ and $m = 0$. So, $|\lambda(J)\setminus\rho(J)|\leq 1$. Similarly, $|\rho(J)\setminus\lambda(J)|\leq 1$. We have three cases to consider.
*Case 1.* Suppose $|\lambda(J)\setminus\rho(J)| = 1$ and $|\rho(J)\setminus\lambda(J)| = 1$. Put $i = \lambda(J)\setminus\rho(J)$ and $j = \rho(J)\setminus\lambda(J)$. Since $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is subdirectly irreducible, there is an $m\in \omega$ such that $(\rho\circ\lambda^{-1})^m(i) = j$. For $0\leq n\leq m$, put $k_n = (\rho\circ\lambda^{-1})^n(i)$. Take $k\in J\setminus \{k_0,\dots,k_m\}$ and $k_n$ with $0\leq n \leq m$. It is not difficult to show, by case analysis, that no path of alternating $\lambda$ and $\rho$ can exist between $k$ and $k_n$, unless $i=j$. But then either $|\lambda(J)\setminus\rho(J)| \neq 1$ or $|\rho(J)\setminus\lambda(J)| \neq 1$, contradicting assumptions. This case is thus excluded.
*Case 2.* Suppose $|\lambda(J)\setminus\rho(J)| = 0$ and $|\rho(J)\setminus\lambda(J)| = 1$. Then, $\lambda(J)\subset\rho(J)$ and so $I = \lambda(J)\cup\rho(J) = \rho(J)$. Thus, $\rho$ is a bijection and $|I\setminus\lambda(J)| = 1$, i.e., (3) holds. By symmetry, if $|\rho(J)\setminus\lambda(J)| = 0$ and $|\lambda(J)\setminus\rho(J)| = 1$, we get that (2) holds.
*Case 3.* Finally, suppose $|\lambda(J)\setminus\rho(J)| = 0 =
|\rho(J)\setminus\lambda(J)|$. Then $\lambda(J) = \rho(J) = I$ and (1) holds.
Three infinite-dimensional kites
--------------------------------
It is consistent with Lemma \[countable-dim-maps\] that no infinite-dimensional subdirectly irreducible kites exist. The examples below show that cases (1), (2) and (3) in Lemma \[countable-dim-maps\] are non-void.
Case (1). Take $I = J = \mathbb{Z}$ and put $\lambda(i) = i$, $\rho(i) = i+1$. The kite $K_{\mathbb{Z},\mathbb{Z}}^{\lambda,\rho}(\mathbb{Z})$ is subdirectly irreducible; its smallest nontrivial normal filter is the set of all sequences $\langle k_i\colon i\in\mathbb{Z}\rangle$ with $k_i \neq 0$ for finitely many $i$.
Case (2). Take $I = J = \omega$ and put $\lambda(i) = i$, $\rho(i) =
i+1$. The kite $K_{\mathbb{Z},\mathbb{Z}}^{\lambda,\rho}(\mathbb{Z})$ is subdirectly irreducible; its smallest nontrivial normal filter is the set of all sequences $\langle k_i\colon i\in\omega\rangle$ with $k_i \neq 0$ for finitely many $i$.
Case (3). Take $I = J = \omega$ and put $\lambda(i) = i+1$, $\rho(i)
= i$. We obtain an example which is symmetric, but not isomorphic, to the previous one.
Classification
--------------
Gathering all results of this section together, we obtain the classification of subdirectly irreducible kites promised at the beginning.
\[classif\] Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a subdirectly irreducible kite. Then, $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is isomorphic to precisely one of:
1. $K_{n,n}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j$ and $\rho(j) = j+1\ (\mathrm{mod}\ n)$.
2. $K_{\mathbb{Z},\mathbb{Z}}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j$ and $\rho(j) = j+1$.
3. $K_{\omega,\omega}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j$ and $\rho(j) = j+1$.
4. $K_{\omega,\omega}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j+1$ and $\rho(j) = j$.
5. $K_{n+1,n}^{\lambda,\rho}(\mathbf{G})$, with $\lambda(j) = j$ and $\rho(j) = j+1$.
Moreover, types [(1)]{} and [(2)]{} consist entirely of pseudo MV-algebras, the other types contain no pseudo MV-algebras except the two-element Boolean algebra. A kite of type [(3)]{} or [(4)]{} is good if and only if it is a two-element Boolean algebra. A kite of type [(5)]{} is good if and only if $J=\emptyset$.
The cases with $I$ and $J$ finite follow from Lemma \[si-sizes\]. The infinite-dimensional cases can be derived from Lemma \[countable-dim-maps\] using arguments mimicking those of Lemma \[si-sizes\]. The ‘moreover’ statements follow from Lemma \[good-kites\].
For kites of the types from Theorem \[classif\] above it will be convenient from now on to use the following notational convention:
1. $K_{n,n}^{0,1}(\mathbf{G})$
2. $K_{\mathbb{Z},\mathbb{Z}}^{0,1}(\mathbf{G})$
3. $K_{\omega,\omega}^{0,1}(\mathbf{G})$
4. $K_{\omega,\omega}^{1,0}(\mathbf{G})$
5. $K_{n+1,n}^{0,1}(\mathbf{G})$
where the upper indices correspond to the functions $\lambda$ and $\rho$ in such a way that whenever $\lambda$ or $\rho$ is the identity function, we replace it by $0$; otherwise we replace it by $1$ (which sits well with the fact that it is then a kind of successor function).
A subdirectly irreducible kite is good if and only if it is either a pseudo MV-algebra or it is of the form $K_{1,0}^{\emptyset,\emptyset}(\mathbf{G})$, for a subdirectly irreducible $\ell$-group $\mathbf{G}$.
The last result in this section shows that subdirectly irreducible kites are building blocks from which all kites can be built. Namely, we will show that each kite is a subdirect product of subdirectly irreducible ones. This is not a trivial corollary of Birkhoff’s subdirect representation theorem, the value added is that the subdirect factors are kites as well.
\[si-kites-are-all\] Let $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ be a kite. Then, $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is subdirectly embeddable into a product of kites of the form $K_{I',J'}^{\lambda,\rho}(\mathbf{G})$, where $I'\cup J'$ is a connected component of the graph of $\lambda\circ\rho^{-1}\cup
\rho\circ\lambda^{-1}$ paths in $I\cup J$.
It is not difficult to show, using Lemmas \[hyper-filters\] and \[restrict\], that if the $\ell$-group $\mathbf{G}$ is not subdirectly irreducible, then any subdirect representation $\mathbf{G}\leq \prod_{s\in S}\mathbf{G}_s$ naturally gives rise to a subdirect representation $K_{I,J}^{\lambda,\rho}(\mathbf{G})\leq
\prod_{s\in S}K_{I,J}^{\lambda,\rho}(\mathbf{G}_s)$. We leave the details to the reader.
Now, assume $\mathbf{G}$ is subdirectly irreducible and let $P$ stand for the graph of $\lambda\circ\rho^{-1}\cup
\rho\circ\lambda^{-1}$ paths in $I\cup J$. Let further $\mathcal{C}$ be the set of all connected components of $P$. From Theorem \[si-kites\] it follows that for each connected component $C = I_C\cup J_C$ of $P$ the kite $K_{I_C,J_C}^{\lambda,\rho}(\mathbf{G})$ is subdirectly irreducible. For each $C\in\mathcal{C}$ let $N_C$ be the set of all $a = \langle
a^{-1}_i\in I\rangle\in (G^-)^I$ such that $i\notin C$ implies $a_i
= e$. It is straightforward to see that $N_C$ is a normal filter for each $C\in\mathcal{C}$, and that $K_{I,J}^{\lambda,\rho}(\mathbf{G})/\Theta_C$, where $\Theta_C$ is the congruence corresponding to $N_C$, is isomorphic to $K_{I_C,J_C}^{\lambda,\rho}(\mathbf{G})$. Since $\bigcap_{C\in\mathcal{C}} N_C = \{1\}$, this proves claim.
\[si-kites-generate\] Let $\mathsf{K}$ be the variety generated by all kites. Then, $\mathsf{K}$ is generated by all subdirectly irreducible kites.
Finite-dimensional kites
========================
A kite $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ will be called *$n$-dimensional* if $|I| = n\in\omega$, and *finite-dimensional* if it is $n$-dimensional for some $n$. We write $\mathcal{K}_n$ for the class of all $n$-dimensional kites and $\mathsf{K_n}$ for the variety generated by $\mathcal{K}_n$. In this section we will show that $\mathsf{K}$ is generated by finite-dimensional kites, and thus $\mathsf{K}$ is the varietal join of $\mathsf{K_n}$ for $n\in\omega$. Our proof proceeds by embedding infinitely dimensional kites from Theorem \[classif\] into a quotient of a subalgebra of a product of kites of finitely dimensional kites.
For any $n\in \omega$, take the kite $K_{2n+1,2n+1}^{0,1}(\mathbf{G})$. We think of the set $2n+1$ as the universe of the additive group $\mathbb{Z}/(2n+1)\mathbb{Z}$ and label its elements accordingly so that $2n+1 =
\{-n,-n+1,\dots,-1,0,1,\dots,n-1,n\}$. Consider the direct product $\prod_{n\in\omega}K_{2n+1,2n+1}^{0,1}(\mathbf{G})$. As usual, for $u\in\prod_{n\in\omega}K_{2n+1,2n+1}^{0,1}(\mathbf{G})$, we will write $u(i)$ for the $i$-th element of $u$. Then, employing our numbering convention, we will write $u(i)$ as $\langle
u_{-i},\dots,u_0,\dots,u_i\rangle$, and so we have $$u = \bigl\langle\langle u_{-n},\dots,u_0,\dots,u_n\rangle\colon n\in\omega\bigr\rangle.$$ It is easy to see that the set $S_\mathbf{G} = U_\mathbf{G}\cup L_\mathbf{G}$, where $$\begin{aligned}
U_\mathbf{G} &= \bigl\{\langle a^{-1}_{-n},\dots,a^{-1}_i,\dots,a^{-1}_{n}\rangle,
\colon a^{-1}_i\in G^-,\ n\in\omega\bigr\}\\
L_\mathbf{G} &= \bigl\{\langle f_{-n},\dots,f_i,\dots,f_{n}\rangle
\colon f_i\in G^+,\ n\in\omega\bigr\}\end{aligned}$$ is a subuniverse of $\prod_{n\in\omega}K_{2n+1,2n+1}^{0,1}(\mathbf{G})$. Let $\mathbf{S}_\mathbf{G}$ be the corresponding subalgebra of $\prod_{n\in\omega}K_{2n+1,2n+1}^{0,1}(\mathbf{G})$. Now, on $\mathbf{S}_\mathbf{G}$ we define a binary relation, putting $$u\sim w \quad\text{if{}f}\quad
\exists k\in\omega\ \forall n\geq k\colon [\![u(n)\neq w(n)]\!]\cap [-n+k,n-k] = \emptyset$$ where $[\![u(n)\neq w(n)]\!] = \{-n\leq i\leq n\colon u_i\neq w_i\}$, as usual. Intuitively, $u\sim w$ holds if, for sufficiently large $n\in\omega$, the sequences $u(n)$ and $w(n)$ differ only at a bounded number of initial and final places. It should be clear that $u\sim w$ can hold only if either $u,w\in U_\mathbf{G}$ or $u,w\in L_\mathbf{G}$.
\[sim-is-cong\] The relation $\sim$ is a congruence on $\prod_{n\in\omega}K_{2n+1,2n+1}^{0,1}(\mathbf{G})$.
Reflexivity and symmetry are obvious. For transitivity, suppose $u\sim v$ and $v\sim w$. By definition then, there are $k_1,k_2\in\omega$ such that $$\forall n\geq k_1\colon
[\![u(n)\neq v(n)]\!]\cap [-n+k_1,n-k_1] = \emptyset$$ and $$\forall n\geq k_2\colon
[\![v(n)\neq w(n)]\!]\cap [-n+k_2,n-k_2] = \emptyset.$$ Then, putting $k = \mathrm{max}\{k_1,k_2\}$ we obtain $$\forall n\geq k\colon
[\![u(n)\neq w(n)]\!]\cap [-n+k,n-k] = \emptyset$$ and so $u\sim w$ as required. It remains to show that $\sim$ preserves the operations. We will only prove it for two cases of multiplication. Let $u\sim w$ and $v\sim s$, with $u,w\in L_\mathbf{G}$ and $v,s\in U_\mathbf{G}$. Thus, $$\begin{aligned}
u(n) &= \langle f_{-n},\dots,f_i,\dots,f_{n}\rangle\\
w(n) &= \langle g_{-n},\dots,g_i,\dots,g_{n}\rangle\\
v(n) &= \langle a^{-1}_{-n},\dots,a^{-1}_i,\dots,a^{-1}_{n}\rangle\\
s(n) &= \langle b^{-1}_{-n},\dots,b^{-1}_i,\dots,b^{-1}_{n}\rangle\end{aligned}$$ with $a_i^{-1},b_i^{-1}\in G^-$ and $f_i,g_i\in G^+$, for any $n\in\omega$. Further, by definition of $\sim$, there are $k_1,k_2\in\omega$ such that $$\begin{aligned}
\forall n\geq k_1\colon
[\![u(n)\neq w(n)]\!]\cap [-n+k_1,n-k_1] &= \emptyset\\
\forall n\geq k_2\colon
[\![v(n)\neq s(n)]\!]\cap [-n+k_2,n-k_2] &= \emptyset.\end{aligned}$$ Consider $uv$ and $ws$. $$\begin{aligned}
uv &=
\langle u(n)\colon n\in\omega\rangle\cdot\langle v(n)\colon n\in\omega\rangle\\
&= \langle u(n)\cdot v(n)\colon n\in\omega\rangle\\
&= \bigl\langle\langle f_{-n},\dots,f_n\rangle\cdot
\langle a^{-1}_{-n},\dots,a^{-1}_n\rangle\colon n\in\omega\bigr\rangle\\
&= \bigl\langle\langle f_{-n}a^{-1}_{-n+1}\vee e,\dots,
f_{n-1}a^{-1}_n\vee e,f_na^{-1}_{{n+1}(\mathrm{mod}\ 2n+1)}\vee e\rangle\colon n\in\omega\bigr\rangle\\
&= \bigl\langle\langle f_{-n}a^{-1}_{-n+1}\vee e,\dots,
f_{n-1}a^{-1}_n\vee e,f_na^{-1}_{-n}\vee e\rangle\colon n\in\omega\bigr\rangle\end{aligned}$$ and similarly, $$\begin{aligned}
ws &= \langle w(n)\cdot s(n)\colon n\in\omega\rangle\\
&= \bigl\langle\langle g_{-n}b^{-1}_{-n+1}\vee e,\dots,
g_{n-1}b^{-1}_n\vee e,g_nb^{-1}_{-n}\vee e\rangle\colon n\in\omega\bigr\rangle.\end{aligned}$$ Now, for a given $n\geq k = \mathrm{max}\{k_1,k_2\}$, let us compare $$\begin{aligned}
u(n)\cdot v(n) &=
\langle f_{-n}a^{-1}_{-n+1}\vee e,\dots,f_na^{-1}_{-n}\vee e\rangle\\
w(n)\cdot s(n) &=
\langle g_{-n}b^{-1}_{-n+1}\vee e,\dots,g_nb^{-1}_{-n}\vee e\rangle.\end{aligned}$$ From the definition of multiplication on $K_{2n+1,2n+1}^{0,1}(\mathbf{G})$ it follows immediately that the two can differ only at initial places up to $f_{-n+k}a^{-1}_{-n+k+1}$, $g_{-n+k}b^{-1}_{-n+k+1}$, respectively, and at final places from $f_{n-k-1}a^{-1}_{-n-k}$, $g_{n-k-1}b^{-1}_{-n-k}$, respectively. It follows that $$\forall n\geq k+1\colon
[\![u(n)\cdot v(n)\neq w(n)\cdot s(n)]\!]\cap [-n+k+1,n-k-1] =\emptyset$$ and thus $uv\sim ws$ as required. Other cases of multiplication are similar, and analogous arguments prove that $\sim$ preserves all the operations, completing the proof.
We are now ready to show that $K_{\mathbb{Z},\mathbb{Z}}^{0,1}(\mathbf{G})$ embeds into $\mathbf{S}_\mathbf{G}/\kern-3pt\sim$. To this end, first define a map $\mu^-\colon K_{\mathbb{Z},\mathbb{Z}}^{0,1}(\mathbf{G})
\to \mathbf{S}_\mathbf{G}$ putting $$\langle u_i\colon i\in\mathbb{Z}\rangle \mapsto
\bigl\langle\langle u_{-n},\dots, u_0,\dots, u_{n}\rangle, n\in\omega\bigr\rangle$$ where $\langle u_i\colon i\in\mathbb{Z}\rangle$ is either $\langle a_i^{-1}\colon i\in\mathbb{Z}\rangle$ with $a_i\in G^{-}$ for each $i$, or $\langle f_i\colon i\in\mathbb{Z}\rangle$ with $f_i\in G^{+}$ for each $i$. The map $\mu^-$ extends naturally to a map $\mu\colon K_{\mathbb{Z},\mathbb{Z}}^{0,1}(\mathbf{G})
\to \mathbf{S}_{\mathbf{G}}/\kern-3pt\sim$, namely $\mu = \mu^-/\kern-3pt\sim$.
\[embed-Z\] The map $\mu$ above is an embedding of $K_{\mathbb{Z},\mathbb{Z}}^{0,1}(\mathbf{G})$ into $\mathbf{S}_{\mathbf{G}}/\kern-3pt\sim$.
To show that $\mu$ is one-one, let $u = \langle u_i\colon
i\in\mathbb{Z}\rangle$ and $w = \langle w_i\colon
i\in\mathbb{Z}\rangle$ be such that $u_i= w_i$ for all $i\neq \ell$ and $u_\ell \neq w_\ell$, where $\ell\in\mathbb{Z}$ is arbitrary, but fixed. By definition, $\mu(u)\neq\mu(w)$ iff $\mu^-(u)\not\sim\mu^-(w)$. Now, observe that for any $k\in \omega$, if we take $n\geq k+|\ell|+1$, then $-n+k<\ell<n-k$, and therefore $\ell\in [\![u(n)\neq w(n)[\!]$. It follows that $$\forall k\ \exists n\geq k\colon
[\![u(n)\neq w(n)[\!]\cap[-n+k,n-k]\neq\emptyset$$ and so $\mu^-(u)\not\sim\mu^-(w)$ as needed.
To prove that $\mu$ is a homomorphism we proceed case by case. We only present two cases of multiplication here. Let $u = \langle a_i^{-1}\colon i\in \mathbb{Z}\rangle$ and $w = \langle f_i\colon i\in \mathbb{Z}\rangle$. Then, $uw = \langle a_i^{-1}f_i\vee e\colon i\in \mathbb{Z}\rangle$ and $wu = \langle f_ia_{i+1}^{-1}\vee e\colon i\in \mathbb{Z}\rangle$. Thus, $$\mu^-(uw) =
\bigl\langle\langle a_{-n}^{-1}f_{-n}\vee e,\dots,
a_0^{-1}f_0\vee e,\dots, a_{n}^{-1}f_{n}\vee e\rangle, n\in\omega\bigr\rangle$$ and $$\mu^-(wu) =
\bigl\langle\langle f_{-n}a_{-n+1}^{-1}\vee e,\dots,
f_0a_1^{-1}\vee e,\dots, f_{n}a_{n+1}^{-1}\vee e\rangle, n\in\omega\bigr\rangle$$ On the other hand, we have $$\begin{aligned}
\mu^-(u)\mu^-(w) &=
\bigl\langle\langle a_{-n}^{-1},\dots, a_0^{-1},\dots, a_{n}^{-1}\rangle,
n\in\omega\bigr\rangle\cdot
\bigl\langle\langle f_{-n},\dots, f_0,\dots, f_{n}\rangle,
n\in\omega\bigr\rangle\\
&= \bigl\langle\langle a_{-n}^{-1}f_{-n}\vee e,\dots, a_0^{-1}f_0\vee e,
\dots, a_{n}^{-1}f_n\vee e\rangle, n\in\omega\bigr\rangle\end{aligned}$$ and also, $$\begin{aligned}
\mu^-(w)\mu^-(u) &= \bigl\langle\langle f_{-n},\dots, f_0,\dots,
f_{n}\rangle, n\in\omega\bigr\rangle\cdot \bigl\langle\langle
a_{-n}^{-1},\dots, a_0^{-1},\dots, a_{n}^{-1}\rangle,
n\in\omega\bigr\rangle\\
&= \bigl\langle\langle f_{-n}a_{-n+1}^{-1}\vee e,\dots, f_0a_1^{-1}\vee e,
\dots, f_na_{{n+1}(\mathrm{mod}\ 2n+1)}^{-1}\vee e\rangle, n\in\omega\bigr\rangle\\
&= \bigl\langle\langle f_{-n}a_{-n+1}^{-1}\vee e,\dots, f_0a_1^{-1}\vee e,
\dots, f_na_{-n}^{-1}\vee e\rangle, n\in\omega\bigr\rangle.\end{aligned}$$ Thus, for every $n\in\omega$ we have that $\bigl(\mu^-(uw)\bigr)(n)
= \bigl(\mu^-(u)\mu^-(w)\bigr)(n)$ and the sequences $\bigl(\mu^-(wu)\bigr)(n)$ and $\bigl(\mu^-(w)\mu^-(u)\bigr)(n)$ differ at most at the final place. It follows that $$[\![\bigl(\mu^-(uw)\bigr)(n)\neq
\bigl(\mu^-(u)\mu^-(w)\bigr)(n)]\!]\cap[-n+1,n-1] = \emptyset$$ and $$[\![\bigl(\mu^-(wu)\bigr)(n)\neq
\bigl(\mu^-(w)\mu^-(u)\bigr)(n)]\!]\cap[-n+1,n-1] = \emptyset$$ proving $\mu^-(uw)\sim \mu^-(u)\mu^-(w)$ and $\mu^-(wu)\sim
\mu^-(w)\mu^-(u)$, as required.
Next we turn to $K_{\omega,\omega}^{0,1}(\mathbf{G})$. The argument is by and large analogous to the one just used, but simpler. Take the direct product $\prod_{n\in\omega}K_{n+1,n}^{0,1}(\mathbf{G})$. We have that $S'_\mathbf{G} = U'_\mathbf{G}\cup L'_\mathbf{G}$, where $$\begin{aligned}
U'_\mathbf{G} &= \bigl\{\langle
a^{-1}_0,\dots,a^{-1}_i,\dots,a^{-1}_{n+1}\rangle,
\colon a^{-1}_i\in G^-,\ n\in\omega\bigr\}\\
L'_\mathbf{G} &= \bigl\{\langle f_0,\dots,f_i,\dots,f_{n}\rangle
\colon f_i\in G^+,\ n\in\omega\bigr\}\end{aligned}$$ is a subuniverse of $\prod_{n\in\omega}K_{n+1,n}^{0,1}(\mathbf{G})$. Let $\mathbf{S}'_\mathbf{G}$ be the corresponding subalgebra of $\prod_{n\in\omega}K_{n+1,n}^{0,1}(\mathbf{G})$. Here we can embed $K_{\omega,\omega}^{0,1}(\mathbf{G})$ directly into $\mathbf{S}'_\mathbf{G}$. Define a map $\nu\colon K_{\omega,\omega}^{0,1}(\mathbf{G})
\to \mathbf{S}'_{\mathbf{G}}$ putting $$\langle f_i\colon i\in\omega\rangle \mapsto
\bigl\langle\langle f_0,\dots, f_n\rangle, n\in\omega\bigr\rangle$$ where $f_i\in G^+$ for all $i$, and $$\langle a_i^{-1}\colon i\in\omega\rangle \mapsto \bigl\langle\langle
a_0^{-1}, a_1^{-1}, \dots, a_{n+1}^{-1}\rangle,
n\in\omega\bigr\rangle$$ where $a_i^{-1}\in G^-$ for all $i$. It is then rather straightforward to prove the following lemma.
\[embed-omega-case1\] The map $\nu$ above is an embedding of $K_{\omega,\omega}^{0,1}(\mathbf{G})$ into $\mathbf{S}'_{\mathbf{G}}$.
Finally, we deal with $K_{\omega,\omega}^{1,0}(\mathbf{G})$. We again use the algebra $\mathbf{S}''_\mathbf{G}$, but this time define a map $\nu'\colon K_{\omega,\omega}^{1,0}(\mathbf{G}) \to
\mathbf{S}'_{\mathbf{G}}$ by reversing the ordering in $\nu$, namely, by $$\langle f_i\colon i\in\omega\rangle \mapsto
\bigl\langle\langle f_n,f_{n-1},\dots, f_0\rangle, n\in\omega\bigr\rangle$$ where $f_i\in G^+$ for all $i$, and $$\langle a_i^{-1}\colon i\in\omega\rangle \mapsto \bigl\langle\langle
a_{n+1}^{-1}, a_n^{-1}, \dots, a_0^{-1}\rangle,
n\in\omega\bigr\rangle$$ where $a_i^{-1}\in G^-$ for all $i$. We obtain the lemma below.
\[embed-omega-case2\] The map $\nu'$ above is an embedding of $K_{\omega,\omega}^{0,1}(\mathbf{G})$ into $\mathbf{S}'_{\mathbf{G}}$.
\[generation\] The variety $\mathsf{K}$ is generated by all finite-dimensional kites.
By Corollary \[si-kites-generate\], the variety $\mathsf{K}$ is generated by all subdirectly irreducible kites. Let $\mathcal{K}_{\mathrm{fin}}$ be the class of all finitely dimensional kites. By Lemmas \[embed-Z\], \[embed-omega-case1\], and \[embed-omega-case2\], every subdirectly irreducible kite belongs to $SHSP(\mathcal{K}_{\mathrm{fin}})$. Thus $\mathsf{K} =
V(\mathcal{K}_{\mathrm{fin}})$ as claimed.
The variety $\mathsf{K}$ generated by all kites is the varietal join of varieties $\mathsf{K}_n$, generated by $n$-dimensional kites. Briefly, $$\mathsf{K} = \bigvee_{n\in\omega}\mathsf{K}_n.$$
Applications {#appl}
============
Let $\mathbf{A}=(A;\ld,\rd, \cdot, 0,1)$ be a pseudo BL-algebra and let $g\in A \setminus \{1\}.$ Then there is a filter, $V,$ of $A$ not containing $g$ and is maximal with respect to this property. We call it a *value* of $g,$ and the filter $V^*$ generated by $V$ and the element $g$ is said to be a *cover* of $V$. As in $\ell$-groups, we say that $\mathbf{A}$ is *normal-valued* if every value is normal in its cover. Let $\mathsf{NVpsBL}$ be the class of normal-valued pseudo BL-algebras. In [@DGK] it was proved that $\mathsf{NVpsBL}$ is a variety, and in contrast to the variety of $\ell$-groups, it is not the greatest proper subvariety of the variety of $\ell$-groups because $\mathsf{NVpsBL}$ is a proper subvariety of the system of all pseudo BL-algebras $\mathbf{A}$ such that every maximal filter of $\mathbf A$ is normal that is also a variety, see [@DGK].
According to Wolfenstein [@Dar Thm 41.1], an $\ell$-group $G$ is normal-valued iff every $a,b \in G^-$ satisfy $a^2b^2 \le ba$, or in our language $$a^2\cdot b^2 \le b\cdot a.\eqno(1)$$ In [@BDK] it was shown a large variety of pseudo BL-algebras that are normal-valued iff they satisfy (1). We note that the same is true for kites.
\[norm-val\] Let $\mathbf{G}$ be an $\ell$-group. Then the kite $K_{I,J}^{\lambda,\rho}(\mathbf{G})$ is normal-valued if and only if [(1)]{} holds for all $a,b \in
K_{I,J}^{\lambda,\rho}(\mathbf{G}).$
It follows from the definition of kites and Wolfenstein’s criterion, [@Dar Thm 41.1].
We note that if $x \in K_{I,J}^{\lambda,\rho}(\mathbf{G}),$ then $$x^2 = 0 \quad \mbox{or}\quad ({\mathord{\sim}}x)^2 =0. \eqno(2)$$ Let $\mathsf{V}(K_{I,J}^{\lambda,\rho}(\mathbf{G}))$ denote the variety of pseudo BL-algebras generated by a kite $K_{I,J}^{\lambda,\rho}(\mathbf{G})$.
Let $n\ge 2$ be a fixed integer and we set $J_n=\{1,\ldots,n\},$ $I_n=\{1,\ldots,n,n+1\},$ $J_n'=\{2,\ldots,n\},$ and $I_n'=\{2,\ldots,n,n+1\}.$ We set $\lambda,\rho: J_n\to I_n$ by $\lambda(i)=i$ and $\rho(i)=i+1$ for each $i \in J_n.$ Take two terms $f_\sim(x):= {\mathord{\sim}}x$ and $f_-(x)={\mathord{-}}x$. Then on $K_{I_n,J_n}^{\lambda_n,\rho_n}(\mathbf{G})$ we have $$f^{2n+1}_\sim(x) \in \{0,1\} =
B(K_{I_n,J_n}^{\lambda_n,\rho_n}(\mathbf{G}))\eqno(3)$$ for every $x\in K_{I_n,J_n}^{\lambda_n,\rho_n}(\mathbf{G}).$
For abbreviation, given an integer $n\ge 1,$ we set $\mathbb
Z_n^\dag = K_{I_n,J_n}^{\lambda_n,\rho_n}(\mathbf{Z}).$ In addition, we define $\mathbb Z_0^\dag:=
K_{I_1,J_1}^{\lambda_1,\rho_1}(\mathbf{O}),$ where $\mathbf O$ is the trivial $\ell$-group consisting only from the identity. Then $\mathbb Z_0^\dag$ is the two-element Boolean algebra, therefore, $\mathbb Z_0^\dag$ generates the variety of Boolean algebras, $\mathsf{BA}.$
\[covers\] For any integer $n \ge 1$, ${\mathsf{V}}(\mathbb Z_{n}^\dag)$ is a cover of the variety $\mathsf{BA},$ and $n\ne m$ implies ${\mathsf{V}}(\mathbb Z_{n}^\dag) \ne {\mathsf{V}}(\mathbb Z_{m}^\dag).$
For $n = 0$, the variety ${\mathsf{V}}(\mathbb Z_{n}^\dag)$ is precisely the variety of *product logic algebras* and it is well-known that this variety covers $\mathsf{BA}$. For $n = 1$, it was proved in [@JiMo Thm 11] that the variety $\mathsf V(\mathbb Z^\dag_1)$ covers $\mathsf{BA}$. Thus, we can assume that $n \ge 1$.
It follows from Theorem \[classif\] that $\mathbb Z_n^\dag$ is subdirectly irreducible.
*Claim 1. Any nontrivial element of $\mathbb
Z^\dag_n$ generates a subalgebra of $\mathbb Z^\dag_n$ that is an isomorphic copy of $\mathbb Z^\dag_n.$*
Assume $a=\langle a_1^{-1},\ldots,a_{n+1}^{-1}\rangle$ is an element from $(G^-)^I$ and let $A$ be the subalgebra of $\mathbb Z^\dag_n$ generated by $a$. Let $i_0$ be the first index such that is different of $e^{-1}$. There is an integer $n_0$ such that $f_\sim^{n_0}(a)=\langle e^{-1},\ldots, e^{-1},a_{i_0}^{-1} \rangle
\in A$. There is also an integer $m$ such that $f_-^m(\langle
e^{-1},\ldots, e^{-1},a_{i_0}^{-1} \rangle) =\langle
a_{i_0},e^{-1},\ldots,e^{-1}\rangle \in A.$ In addition, for every $i=1,\ldots, n+1,$ the element $x_i= \langle
b_1^{-1},\ldots,b_{n+1}^{-1}\rangle$ belongs to $A,$ where $b_i^{-1}
= a_{i_0}^{-1}$ and $b_{j}^{-1}=e^{-1}$ for $j\ne i.$
Since ${\mathord{\sim}}\langle a_1^{-1},\ldots,a^{-1}_{n+1}\rangle =
{\mathord{\sim}}\langle a_1^{-1},\ldots,a_n^{-1}\rangle, $ $\langle f_1,\ldots,f_n\rangle
= \langle e^{-1},f_1^{-1},\ldots,f_n^{-1}\rangle$, we see that the algebra $A$ can be generated equivalently by some appropriate element from $(G^{-})^I$ or from $(G^+)^J$.
Hence, if $a=\mbox{g.c.d.}\{a_1,\ldots,a_{n+1}\},$ we see that $A$ is generated e.g. by the element $\langle a^{-1},e^{-1},\ldots,
e^{-1}\rangle,$ and it is an isomorphic copy of $\mathbb Z^\dag_n.$
*Claim 2. If $\mathbf{A}$ is an algebra from $\mathsf{V}(\mathbb Z^\dag_n)$, then for any $x\in A$, we have $$f_\sim^{2n+1}(x) \in B(A).\eqno(4)$$*
Indeed, let $C$ be a subdirectly irreducible algebra from $\mathsf{V}(\mathbb Z^\dag_n).$ By the Jónsson Lemma, $C$ is a homomorphic image of a subalgebra $D$ of an ultrapower $\mathcal
U$ of $\mathbb Z^\dag_n$ on the index set $U.$ By (3), if $x \in
\mathbb Z^\dag_n,$ then $f_\sim^{2n+1}(x)\in \{0,1\}= B(\mathbb
Z^\dag_n).$ If $x \in \langle t_u\colon u \in U\rangle, $ then $f_\sim^{2n+1}(x) \in B((\mathbb Z^\dag_n)^U).$ Similarly, if $x \in
\langle t_u \colon u\in U\rangle/\mathcal U,$ $f^{2n+1}(x) \in
B(\mathbb (\mathbb Z^\dag_n)^U/\mathcal U).$ In the same way, we can prove that if $x \in C,$ then $f^{2n+1}_\sim(x) \in B(C).$ The general case follows from the statement follows from the fact that $A$ is isomorphic to a subdirect product of algebras from $\mathit
{HSP}_U(\mathbb Z^\dag_n).$
*Claim 3. Any subdirectly irreducible algebra of $\mathsf{V}(\mathbb Z^\dag_n)$ is either the two-element Boolean algebra or has a subalgebra isomorphic to $\mathbb Z^\dag_n.$*
Let $C$ be a subdirectly irreducible algebra of $\mathsf{V}(\mathbb Z^\dag_n)$ and assume that it has a nontrivial element $0<c<1.$ Now we use again the Jónsson Lemma, and we assume that $C$ is a homomorphic image of a subalgebra $D$ of an ultrapower $\mathcal U$ of $\mathbb Z^\dag_n$ on the index set $U.$ Let $d\in D$ be a preimage of $c$ under the homomorphism. We show that $a=d^\sim$ generates a subalgebra of $D$ that is isomorphic to $\mathbb Z^\dag_n.$ Then $a = \langle a_u: u \in
U\rangle/\mathcal U.$ Consider the set $V =\{u\in U: a^2_u =0\}$. If $V \in \mathcal U,$ then $a$ generates in the same way as $\langle
1,\ldots, 1\rangle \in \mathbb (\mathbb Z^+)^n.$ If $V \notin \mathcal
U,$ we can use the generator like $\langle 0,\ldots,0,-1\rangle
\in \mathbb (\mathbb Z^-)^{n+1}.$
*Claim 4. Every $\mathsf{V}(\mathbb Z^\dag_{n})$ is a cover variety of the variety of Boolean algebras.*
Let $\mathcal V$ be a subvariety of $\mathsf{V}(\mathbb Z^\dag_{n})$ containing properly the variety of Boolean algebras, and let $A$ be a subdirectly irreducible algebra from $\mathcal V.$ By Claim 3, $A$ is either a two-element Boolean algebra or contains an isomorphic copy of $\mathbb Z^\dag_{n}.$ Therefore, in the later case, $A
\notin \mathsf{BA}$ and we have $\mathsf V(A) = \mathsf{V}(\mathbb
Z^\dag_{n})\subseteq \mathcal V$, which proves the statement of the Theorem.
*Claim 5. If $n$ and $m$ are two different positive integers, then $\mathsf{V}(\mathbb Z^\dag_{n}) \ne \mathsf{V}(\mathbb Z^\dag_{m}).$*
Assume $n<m$. Verifying (4), we see that the element $x=\langle-1,\ldots,-1 \rangle$ from $\mathbb Z^\dag_{m}$ does not belong to $\mathsf{V}(\mathbb Z^\dag_{n}).$
The next result follows immediately from Lemma \[good-kites\], Theorem \[covers\], and the subdirect representation theorem.
An algebra $\mathbf{A}\in \mathsf V(\mathbb{Z}^\dag_n)$, for any $n\in\omega$, is good if and only if $\mathbf{A}$ is a Boolean algebra.
[DvRa2]{}
P. Aglianò and F. Montagna, *Varieties of BL-algebras I: general properties*, J. Pure Appl. Algebra [**181**]{} (2003), 105–129.
M. Botur, A. Dvurečenskij, T. Kowalski, *On normal-valued basic pseudo hoops*, Soft Computing [**16**]{} (2012), 635–644. DOI: 10.1007/s00500-011-0763-7,
C.C. Chang, *A new proof of the completeness of Łukasiewicz axioms*, Trans. Amer. Math. Soc. **93** (1959), 74–80.
M.R. Darnel, *Theory of Lattice-Ordered Groups*, Marcel Dekker, Inc., New York, 1995.
A. Di Nola, G. Georgescu, and A. Iorgulescu, *Pseudo-BL algebras I*, Multiple Val. Logic [**8**]{} (2002), 673–714.
A. Di Nola, G. Georgescu, and A. Iorgulescu, *Pseudo-BL algebras II*, Multiple Val. Logic [**8**]{} (2002), 715–750.
A. Dvurečenskij, [*Pseudo MV-algebras are intervals in $\ell$-groups*]{}, J. Austral. Math. Soc. [**72**]{} (2002), 427–445.
A. Dvurečenskij, *Aglianò–Montagna type decomposition of linear pseudo hoops and its applications*, J. Pure Appl. Algebra [**211**]{} (2007), 851–861.
A. Dvurečenskij, R. Giuntini, and T. Kowalski, *On the structure of pseudo BL-algebras and pseudo hoops in quantum logics*, Found. Phys. [**40**]{} (2010), 1519–1542. DOI:10.1007/s10701-009-9342-5
N. Galatos, C. Tsinakis, *Generalized MV-algebras*, J. Algebra [**283**]{} (2005), 254–291.
N. Galatos, P. Jipsen, T. Kowalski, H. Ono, *Residuated Lattices. An Algebraic Glimpse at Substructural Logic*, Elsevier, 2007.
P. Jipsen and F. Montagna, *On the structure of generalized BL-algebras*, Algebra Universalis [**55**]{} (2006), 226–237.
D. Mundici, [*Interpretations of $AF$ $C^\star$-algebras in Łukasiewicz sentential calculus*]{}, J. Funct. Analysis [**65**]{} (1986), 15–63.
[^1]: Over pseudo BL-algebras; the equivalence does not hold in general.
[^2]: Not pseudovarieties, which are classes closed under finite direct products, subalgebras and homomorphic images.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present formal expressions for the optical scalars in terms of the curvature scalars in the weak gravitational lensing regime at second order in perturbations of a flat background without mentioning the extension of the lens or their shape. Also, by considering the thin lens approximation for static and axially symmetric configurations we obtain an expression for the second-order deflection angle which generalizes our previous result presented in [@PhysRevD.83.083007]. As applications of these formulas we compute the optical scalars for some known family of metrics and we recover expressions for the deflection angle. In contrast to other works in the subject, our formalism allows a straightforward identification of how the different components of the curvature tensor contribute to the optical scalars and deflection angle. We also discuss in what sense the Schwarzschild solution can be thought as a true thin lens at second order.'
author:
- 'Gabriel Crisnejo$^1$ and Emanuel Gallo$^1,^2$'
title: Expressions for optical scalars and deflection angle at second order in terms of curvature scalars
---
Introduction {#section1}
============
The phenomenon of gravitational lensing has become an active area of research at least from the beginning of the 1980’s when the observation of the first astrophysical gravitational lens was announced[@1979Natur.279..381W]. Since then, a plethora of different astrophysical and cosmological gravitational lenses have been discovered and studied in depth. For these reasons, even today, this subject is still being theoretical considered due to the numerous applications in both astrophysical and cosmological settings, such as the study of CMB power spectrum[@Lewis:2006fu; @Nguyen:2017zqu; @Marozzi:2016und; @Peloton:2016kbw; @Fabbian:2017wfp; @Pratten:2016dsm], the distance-redshift relationship[@Kaiser:2015iia; @Fleury:2016fda], or the estimation of mass content of groups or clusters of galaxies[@Hoekstra:2013via; @2015IAUS..311...86M; @Giocoli:2013tga], to cite some examples.
In [@PhysRevD.83.083007], we presented a new approach to the study of gravitational lensing in the weak field regime, which proved to be very useful since gives explicit gauge invariant expressions not only for the optical scalars but also to the deflection angle. These expressions can be easily implemented in any gauge. In particular, in that reference we also presented expressions for these optical quantities in terms of the components of the energy-momentum tensor instead of the usual presentation in terms of metric potentials. Applications of the formalism can be found in [@Bozza:2015haa; @Gallo:2011hi], and its results have been recently extended to the cosmological framework[@Boero:2016nrd].
In the case of an axially symmetric lens configuration, it was also shown that at first order the deflection angle $\alpha$ in a weak field regime of an asymptotically flat spacetime can be written in terms of the impact parameter $J$, and the projected Ricci and Weyl scalars $\hat\Phi_{00}$ and $\hat\Psi_0=-\hat\psi_0e^{2i\theta}$ (We refer to [@PhysRevD.83.083007] for more details) in a very compact form which we reproduce here $$\label{angulito}
\alpha(J)=J\left(\hat\Phi_{00}(J)+\hat\psi_0(J)\right).$$
On the other hand, the second-order gravitational lens theory, that is, the study of the optical scalars at second order in perturbation of a given background metric and its applications has been extensively covered. For example, in [@Vanderveld:2011sj] the weak field lens equation at second order in $\tilde\varepsilon=G/c^{2}$ has been worked out in the post-Newtonian formalism. Fritelli [*et al.*]{} [@Frittelli:1999yf] studied the exact lens equation in Schwarzschild spacetime and, in particular, the lens equation at second order in $\tilde\varepsilon$. The light deflection at second order in $\tilde\varepsilon$ has been studied in [@Brugmann:2005ft] for a system of two bounded point masses. In the cosmological context we found several works concerning the second order effect of weak lensing in the CMB power spectrum[@Hagstotz:2014qea; @1475-7516-2015-06-050; @Kaiser:2015iia; @Marozzi:2016uob; @Petri:2016qya; @Schaefer:2005up; @Cooray:2002mj; @Marozzi:2016qxl]. A general expression for the convergence at second order in the Poisson gauge can be obtained using the area distance [@BenDayan:2012wi; @Fanizza:2013doa; @Marozzi:2014kua] or alternatively using the so-called galaxy number counts [@DiDio:2014lka; @DiDio:2015bua]. In the same gauge a general expression for the shear at second order can be found in [@Bernardeau-2010].
However, in all these works the different expressions for the optical scalars or the deflection angle are generally written in a particular gauge or a given family of gauges. Although all these studies are very useful and relevant in their respective range of validity, it is our intention to extend some of the results presented in [@PhysRevD.83.083007]. In particular, we will obtain expressions for the optical scalars in terms of curvature scalars which take into account second order perturbations of a flat metric, and which are independent of the extending or shape of the lens. It is also our intention to generalize at second order. .
One of the advantages of the present approach is that allows us to recognize how the different aspects of the curvature of the spacetime contribute to the deflection angle and optical scalars and on the other hand they are ready for use in any gauge.
The organization of the paper is as follows. In Sec. \[section2\] we present a short review of the main equations of weak gravitational lensing, including a brief discussion of the geodesic deviation equation and we solve the geodesic deviation equation up to second-order in perturbation of a flat metric and we obtain, via the amplification matrix, general expressions for the optical scalars. In Sec. \[section4\], we focus on the thin lens approximation and give an explicit expression for the deflection angle for static and axially symmetric lenses which generalizes . In Sec. \[Applications\], we provide some examples, applying the formalism to the description of the optical scalars for some known metrics and recovering expressions for the weak lensing quantities in those spacetimes. We conclude with general remarks and comments in the last section. Some auxiliary but relevant relations have been included in the four Appendixes.
The geodesic deviation equation: second order solution {#section2}
======================================================
Preliminaries
-------------
As is well-known, the weak lens equation relates the angular position $\beta^a$ of the source (which should be the observed angular position if there were no gravitational lens) and the actual observed angular position $\theta^a$ due to the presence of the lens, $$\label{eq: 0.1}
\beta^{a} = \theta^{a} - \frac{\lambda_{ls}}{\lambda_{s}} \alpha^{a};$$ with $\alpha$ the deflection angle [@Ehlers:1992dau]. The differential of the Eq. can be written as: $$\label{eq:0.00}
\delta \beta^{a} = A^{a}{}_{b} \ \delta \theta^{b},$$ where $A^{a}{}_{b}$ is called the *amplification matrix* and is given by: $$\label{Aij}
A^a{}_b=\begin{pmatrix}
1 - \kappa - \gamma_1 & -\gamma_2 - \hat{\omega} \\
-\gamma_2 + \hat{\omega} & 1 - \kappa + \gamma_1 \\
\end{pmatrix};$$ where $\kappa$, $\gamma \equiv \gamma_1 + i \, \gamma_2$ and $\hat{\omega}$ are called *optical scalars*: convergence, shear and rotation, respectively.
A powerful way to study weak gravitational lensing is through the use of the geodesic deviation equation. A complete discussion of this equation and its use in the description of gravitational lensing can be found in [@Bartelmann:2010fz; @Seitz:1994xf].
Let us consider a null geodesic congruence starting at the source position $S$ and ending at the observer position $O$. That is, a null congruence belonging to the past null cone of $O$. The tangent vector to a fiducial null geodesic of this congurence is given by $\ell = \frac{\partial}{\partial \lambda}$. At the position of the observer, we can construct a null tetrad $\{\ell^a,n^a,m^a,\bar{m}^a\}$ satisfying the standard normalization conditions, and with $m^a, \bar{m}^{a}$ complex null vectors which are orthogonal to the four-velocity $u^a$ of $O$. The deviation vector connecting two neighboring null geodesics in the congruence can be expressed by $$\label{zeta_forma}
\zeta^{a} = \zeta \bar{m}^{a} + \bar{\zeta} m^{a} + \zeta_{\ell} \, \ell^{a}.$$
As it is well discussed in the literature, the geodesic deviation equation can be written as [@Bartelmann:2010fz; @Seitz:1994xf; @Frittelli:2000bc]: $$\label{dg:so}
\ell(\ell(\mathcal{X})) = - Q \mathcal{X};$$ where $$\mathcal{X} =
\begin{pmatrix}
\zeta \\
\bar{\zeta}
\end{pmatrix},$$ and $$Q = \begin{pmatrix}
\Phi_{00} & \Psi_{0} \\
\bar{\Psi}_{0} & \Phi_{00}
\end{pmatrix},$$ with $$\Phi_{00} = -\frac{1}{2} R_{ab} \ell^{a} \ell^{b} , \ \ \ \Psi_{0} = C_{abcd} \ell^{a} m^{b} \ell^{c} m^{d}.$$
Second-order solution {#subsection3}
---------------------
Now we will solve the Eq. at second order in a perturbation of a flat spacetime following the iterative method used in [@PhysRevD.83.083007].
We define $$X =
\begin{pmatrix}
\mathcal{X} \\
\mathcal{V}
\end{pmatrix}, \ \ \
\mathcal{V} = \frac{d \mathcal{X}}{d \lambda}.$$ Therefore, Eq.(\[dg:so\]) is reduced to a first order differential equation given by: $$\label{dg:fo}
\ell(X) = \mathbb{A} \, X,$$ where $$\mathbb{A} =
\begin{pmatrix}
0 & \mathds{1} \\
-Q & 0
\end{pmatrix}.$$ Following [@PhysRevD.83.083007] we start from a seed $X_{0}$, and we will perform the integrations from the observer position $\lambda_{o}$ (which can be taken without a loss of generality to be zero ) to the source position in $\lambda_{s}$. The beam has initially a vanishing departure, and therefore the seed is taken as (see [@PhysRevD.83.083007] for details) $$X_{0} = \left(
\begin{array}{c}
0 \\
\mathcal{V}(0)
\end{array}\right).$$ Then we construct the following sequence of approximate solutions: $$\label{eq:2.1}
X_{1}(\lambda_s) = X_{0} + \int^{\lambda_s}_{0} \mathbb{A}(\lambda{'}) X_{0} d\lambda{'},$$ $$\label{eq:2.2}
X_{2}(\lambda_s) = X_{0} + \int^{\lambda_s}_{0} \mathbb{A}(\lambda{'}) X_{1} d\lambda{'},$$ and so on. By replacing into we get $$X_{2}(\lambda_s) = X_{1} + \int^{\lambda_s}_{0}\int^{\lambda{'}}_{0} \mathbb{A}(\lambda{'})\mathbb{A}(\lambda{''}) X_{0} d\lambda{''} d\lambda{'};$$ in this way we can obtain a solution of to the desired order. We see that in the $X_{\emph{n}}$ step there will be *n* products of matrices $\mathbb{A}$. As we are interested in second-order solutions, we only need to compute up to $X_{5}$ because in the next step only appear cubic quantities in Q. Accordingly, by considering up to quadratic terms in Q we get $$\label{eq:2.3}
\begin{aligned}
&\mathcal{X}(\lambda_s) = \bigg[ \mathds{1}\lambda_{s} - \int^{\lambda_s}_{0} \int^{\lambda{'}}_{0}\lambda{''} Q(\lambda{''})
d\lambda{''} d\lambda{'} \\
&+ \int_{0}^{\lambda_s}\int_{0}^{\lambda{'}}\int_{0}^{\lambda{''}}\int_{\lambda_ {0}}^{\lambda{'''}}\lambda{''''}
Q(\lambda{''}) Q(\lambda{''''}) d\lambda{''''} d\lambda{'''}d\lambda{''}d\lambda{'} \bigg] \\
&\times \mathcal{V}(0),
\end{aligned}$$ On the other hand, if the metric were flat $(Q = 0)$ then we should have $$\mathcal{X}_{s} \equiv \mathcal{X}(\lambda_{s}) = \lambda_{s} \mathcal{V}(0) \ \ \Rightarrow \ \ \mathcal{V}(0)
= \frac{\mathcal{X}(\lambda_{s})}{\lambda_{s}}.$$ But in the presence of a gravitational lens, if an observer sees an image of size $\mathcal{X}_{o}$, which means $\mathcal{X}_{o} \equiv\lambda_{s} \mathcal{V}_{0}$ then it should be produced by a source of size $\mathcal{X}_{s} = \mathcal{X}(\lambda_{s})$, as described by Eq..
Using the following two relations obtained by integration by parts \[see Appendix \]: $$\label{identity1}
\int_{0}^{\lambda_{s}}\int_{0}^{\lambda{'}}\lambda{''} Q(\lambda{''}) d\lambda{''}d\lambda{'}
= \int^{\lambda_{s}}_{0} \lambda(\lambda_{s} - \lambda) Q(\lambda) d\lambda,$$ $$\label{identity2}
\begin{aligned}
&\int^{\lambda_{s}}_{0} \int^{\lambda{'}}_{0} \int^{\lambda{''}}_{0} \lambda{'''} (\lambda{''}-\lambda{'''})
Q(\lambda{''}) Q(\lambda{'''}) d\lambda{'''} d\lambda{''} d\lambda{'} \\
&= \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})Q(\lambda)Q(\lambda{'})d\lambda{'}d\lambda,
\end{aligned}$$ we finally obtain $$\label{eq:2.44}
\begin{aligned}
\mathcal{X}_{s} &= \bigg[\mathds{1} - \frac{1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) Q(\lambda) d\lambda \\
&+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})Q(\lambda)Q(\lambda{'})d\lambda{'}d\lambda \bigg] \mathcal{X}_{o};
\end{aligned}$$ where the deviation vector at the source position $\mathcal{X}_{s}$ and at the observer position $\mathcal{X}_{0}$ are given by $$\label{eqxs}
\mathcal{X}_{s}=\left(
\begin{array}{c}
\zeta_{s} \\
\bar{\zeta}_{s}
\end{array}\right), \ \ \ \ \ \ \ \ \ \ \mathcal{X}_{0}=\left(
\begin{array}{c}
\zeta_{o} \\
\bar{\zeta}_{o}
\end{array}\right).$$
By replacing into and using the explicit expression for the $Q$ matrix we have: $$\label{eq:2.41}
\begin{aligned}
\zeta_{s} =& \bigg[ 1 - \frac{1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Phi_{00}(\lambda) d\lambda
+ \frac{1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \int^{\lambda}_{0} \lambda{'} \\
&(\lambda_{s} - \lambda)(\lambda-\lambda{'})
\bigg(\Phi_{00}(\lambda) \Phi_{00}(\lambda{'})
+ \Psi_{0}(\lambda)\bar{\Psi}_{0} (\lambda{'})\bigg) \\
&d\lambda{'}d\lambda \bigg] \zeta_{o}
+ \bigg[- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}} \lambda
(\lambda_{s} - \lambda) \Psi_{0}(\lambda) d\lambda \\
&+ \frac{1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \int^{\lambda}_{0} \lambda{'} (\lambda_{s}
- \lambda)(\lambda-\lambda{'}) \bigg(\Phi_{00}(\lambda) \Psi_{0}(\lambda{'}) \\
&+ \Psi_{0}(\lambda) \Phi_{00}(\lambda{'})\bigg)
d\lambda{'}d\lambda\bigg]
\bar{\zeta}_{o}.
\end{aligned}$$ This is the general solution of the geodesic deviation equation at second order. We proceed now to connect this equation with the lens equation . In order to do that, we decompose $\Psi_{0}$, $\zeta_{s}$ and $\zeta_{o}$ into their real and imaginary parts, $$\begin{aligned}
\Psi_{0} = \Psi_{0R} + i \; \Psi_{0I}, \ \
\zeta_{s} = \zeta_{sR} + i \;\zeta_{sI}, \ \
\zeta_{o} = \zeta_{oR} + i \;\zeta_{oI},
\end{aligned}$$ which after replacing in gives the following equations: $$\label{eq:2777}
\begin{aligned}
\zeta_{sR} &= A \; \zeta_{oR} + B \; \zeta_{oI}, \\
\zeta_{sI} &= A{'} \; \zeta_{oR} + B{'} \; \zeta_{oI},
\end{aligned}$$ where the explicit expressions for the coefficients $\{A, B, A{'}, B{'}\}$ can be found in the Appendix \[abcd\].
From the linearity of the lens equation, we know that the deviation vectors must also be related by the same lens mapping matrix $A^i{}_j$ [@Uzan:2000xv], $$\label{eq: 2.77}
\zeta^{i}_{s} = A^{i}{}_{j} \, \zeta^{j}_{o},$$ and where $\{\zeta_{s}^{i},\zeta_{o}^{i}\}$ are the spatial vector associated with $\{\zeta_{s},\zeta_{o}\}$.
Since we are not interested in the component $\zeta_\ell$ of $\zeta^{a}$ along $\ell^{a}$, we will consider only the projection of the deviation vector in the two-space spanned by $\{ m^{a}, \bar{m}^{a} \}$, $$\zeta_{\perp}^{a} = \zeta \bar{m}^{a} + \bar{\zeta} m^{a}.$$ Introducing an orthonormal spatial Sachs basis $\{a^{a},b^{a}\}$ at the observer position and by parallel transport of this basis to the other points in the past null cone we can always express $m^{a}$ by $$m^{a} = \frac{1}{\sqrt{2}} (a^{a} + i \, b^{a}).$$ Hence, we obtain $$\zeta_{\perp}^{a} = {\sqrt{2}} (\zeta_{R} a^{a} + \zeta_{I} b^{a}).$$ where $\{ \zeta_{R}, \zeta_{I} \}$ are the real and imaginary part of the component $\zeta$, respectively. Therefore, from we see that: $$\begin{pmatrix}
\zeta_{sR} \\ \zeta_{sI}
\end{pmatrix}
=
\begin{pmatrix}
1 - \kappa - \gamma_1 & -\gamma_2 - \hat{\omega} \\
-\gamma_2 + \hat{\omega} & 1 - \kappa + \gamma_1 \\
\end{pmatrix}
\begin{pmatrix}
\zeta_{oR} \\ \zeta_{oI}
\end{pmatrix}.$$ Finally, by comparison with we arrive to the following expressions for the optical scalars:
$$\label{convergencia:gral2}
\begin{aligned}
\kappa =& \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Phi_{00}(\lambda) d\lambda
- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda) \\
&(\lambda-\lambda{'})
\bigg(\Phi_{00}(\lambda)\Phi_{00}(\lambda{'})
+ \Re \{\Psi_{0}(\lambda) \bar{\Psi}_{0}(\lambda{{'}}) \} \bigg)
d\lambda{'} d\lambda,
\end{aligned}$$
$$\label{shear:gral2}
\begin{aligned}
\gamma &= \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Psi_{0}(\lambda) d\lambda
- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda) \\
&(\lambda-\lambda{'})
\bigg( \Phi_{00}(\lambda) \Psi_{0}(\lambda{'})
+ \Psi_{0}(\lambda) \Phi_{00}(\lambda{'}) \bigg)
d\lambda{'}d\lambda,
\end{aligned}$$
$$\label{rotacion:gral2}
\hat{\omega} = \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \Im \{ \Psi_{0}(\lambda) \bar{\Psi}_{0}(\lambda{{'}}) \} d\lambda{'}d\lambda;$$
where $\Re\{\cdot\}$ and $\Im\{\cdot\}$ indicate real and imaginary part, respectively.
The way in which the different curvature scalars appear in the expression for the convergence $\kappa$ and the shear $\gamma$ is expected taking into account that they are the integrated version of the well-known relations between the local quantities $\rho$ and $\sigma$ [@Newman-1962] $$\label{rho_rho}
\ell(\rho) = (\rho^{2} + \sigma \bar{\sigma}) + \Phi_{00},$$ $$\label{sigma_sigma}
\ell(\sigma) = (\rho + \bar{\rho})\sigma + \Psi_{0}.$$
On the other hand even when the congruence is twist free, $$\omega=\frac{1}{2}\left(\nabla_{[a}l_{b]}\nabla^a l^b\right)^{1/2}=\frac{1}{2}(\rho-\bar\rho)=0;$$ the rotation of the image described by the scalar $\hat\omega$ could be different from zero due to a cumulative effect of shearing in different directions when the light beams pass different regions of lensing [@Holz-1998].
Now, as we are interested in expressions up to second-order in the formalism of weak lenses, we need to address two issues: first, we only need the curvature components $\Phi_{00}$ and $\Psi_{0}$ up to second-order in the flat metric perturbation; that is $$\begin{aligned}
\Phi_{00} &= \Phi_{00}^{(1)} + \Phi_{00}^{(2)} + \mathcal{O}(\varepsilon^{3}), \\
\Psi_{0} &= \Psi_{0}^{(1)} + \Psi_{0}^{(2)} + \mathcal{O}(\varepsilon^{3}),
\end{aligned}$$ where the superscript in parenthesis indicates the order in $\varepsilon$ of the respective quantity, with $\varepsilon$ the parameter which measure the perturbation of a flat metric. Consequently, we need to compute the Ricci and Weyl tensor up to second-order in $\varepsilon$ and perform the parallel transport of the null tetrad at first order also in $\varepsilon$. Second, as we are working in the weak lensing regime, we can approximate the actual path of the beam as follows: $$x^{a}_{actual}(\lambda) = x^{(0)a}(\lambda) + \delta x^{(1)a}(\lambda) + \mathcal{O}(\varepsilon^{2});$$ that is, we can express the actual beam path as its path in the background plus higher order corrections. Here, we only need to consider corrections at first order. This method of approximation which goes beyond of the Born approximation is frequently used in the literature of second order lensing [@1475-7516-2015-06-050; @Cooray:2002mj; @Bernardeau-2010; @Hagstotz:2014qea; @Marozzi:2016uob; @Petri:2016qya; @Schaefer:2005up].
Therefore, we expand $\Phi_{00}$ and $\Psi_{0}$ along the background geodesic at second-order as, $$\label{aproxxx}
\begin{aligned}
\Phi_{00}(x^{a}_{actual}(\lambda)) =& \Phi_{00}^{(1)}(x^{(0)a}(\lambda)) + \delta x^{(1)a}(\lambda) \frac{\partial \Phi_{00}^{(1)}}{\partial x^{a}}\bigg |_{x^{(0)a}(\lambda)} \\
& + \Phi_{00}^{(2)}(x^{(0)a}(\lambda)), \\
\Psi_{0}(x^{a}_{actual}(\lambda)) =& \Psi_{0}^{(1)}(x^{(0)a}(\lambda)) + \delta x^{(1)a}(\lambda) \frac{\partial \Psi_{0}^{(1)}}{\partial x^{a}}\bigg |_{x^{(0)a}(\lambda)}\\
&+ \Psi_{0}^{(2)}(x^{(0)a}(\lambda)).
\end{aligned}$$ In conclusion, the final expressions for the optical scalars are formally written in terms of the curvature scalars as $$\label{convergencia:gral}
\begin{aligned}
\kappa =& \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \bigg( \Phi^{(1)}_{00}(\lambda) + \Phi^{(2)}_{00}(\lambda)
+ \delta x^{(1)a}(\lambda) \\
&\frac{\partial \Phi^{(1)}_{00}}{\partial x^{a}}\bigg |_{\lambda} \bigg) d\lambda
- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \\
&\bigg(\Phi^{(1)}_{00}(\lambda)\Phi^{(1)}_{00}(\lambda{'})
+ \Re \{\Psi_{0}^{(1)}(\lambda)
\bar{\Psi}_{0}^{(1)}(\lambda{{'}}) \} \bigg) d\lambda{'} d\lambda,
\end{aligned}$$ $$\label{shear:gral}
\begin{aligned}
\gamma =& \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \bigg( \Psi^{(1)}_{0}(\lambda) + \Psi^{(2)}_{0}(\lambda)
+ \delta x^{(1)a}(\lambda) \\
&\frac{\partial \Psi^{(1)}_{0} }{\partial x^{a}}\bigg |_{\lambda} \bigg) d\lambda
- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \\
&\bigg( \Phi^{(1)}_{00}(\lambda) \Psi^{(1)}_{0}(\lambda{'}) + \Psi^{(1)}_{0}(\lambda) \Phi^{(1)}_{00}(\lambda{'}) \bigg)
d\lambda{'}d\lambda,
\end{aligned}$$ $$\label{rotacion:gral}
\hat{\omega} = \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \Im \{ \Psi_{0}^{(1)}(\lambda) \bar{\Psi}_{0}^{(1)}(\lambda{'}) \} d\lambda{'}d\lambda.$$ Equations (\[convergencia:gral\]), (\[shear:gral\]) and (\[rotacion:gral\]) are the most general formal expressions for the optical scalars which can be obtained without mentioning neither the extension of the lens nor their shape. We want to emphasize that unlike the expressions (\[convergencia:gral2\]), (\[shear:gral2\]), (\[rotacion:gral2\]), in these last expressions the integrals are made over the background geodesic and only contain quantities up to second-order. These formulas for the optical scalars generalize to second order the relations found in [@PhysRevD.83.083007] (see also[@0264-9381-22-14-005]).
The thin lens approximation {#section4}
===========================
Preliminary considerations
--------------------------
As is well-known, the thin lens approximation is based on the assumption that the *lens-observer* and *lens-source* distances are significantly larger that the lens size [@Ehlers:1992dau]. In this situation, the different optical scalars can be found by projection of the curvature scalars in the line of sight [@0264-9381-22-14-005; @Seitz:1994xf].
Let $C$ or $D$ be any of the $\{ \Phi_{00}, \Psi_{0} \}$ scalars so, following the approach used in [@PhysRevD.83.083007] the first order thin lens approximation is completely contained in the following expression $$\label{aprox-delgada}
\widetilde{C}(\lambda{}) \equiv \int^{\lambda{}}_{0} C(\lambda{'}) d\lambda{'} \cong \left\{
\begin{array}{ll}
0 & \forall \ \lambda{'} < \lambda_{l} - \delta \\
\hat{C} & \forall \ \lambda{'} \geq \lambda_{l} + \delta \\
\end{array}
\right.$$ where $\delta \ll \lambda_{l}, \delta \ll \lambda_{ls}, \delta \ll \lambda_{s}$ ($\lambda_{ls} \equiv \lambda_{s}-\lambda_{l}$).
On the other hand, even when $\delta \ll \lambda_{l}$, $\delta \ll \lambda_{ls}$ implies that is a good approximation in the computation of first order quantities, it does not mean that it remains sufficiently precise in order to compute the second order quadratic terms. We prove in the Appendix \[app-thin-lens\], that if we only consider the approximation in the computation of quantities like $$\label{terminos_cuadraticos_tipo}
\int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
C(\lambda{}) D(\lambda{'}) d\lambda{'} d\lambda{};$$ then these quantities should be zero. However, even for a point mass lens, as we will show in Sec.\[Applications\] terms like make a non-negligible contribution. It should not be seen as a surprising result, for example even in a toy model where quantities like $C(\lambda)$ are represented by step functions centered in $\lambda_l$ and with a width $\delta$, integrals like make nontrivial contributions proportional to the width $\delta$. Unfortunately, even when in the thin lens situation we can find more simple expressions for the integrals which depend linearly from the curvature scalars (they can be written in terms of projected $\hat{C}$ quantities), we can not do the same with the quadratic ones. Therefore, in the following we only express the linear integrals in $C$ in terms if $\hat{C}$ and rewrite in a more compact way the integrals which are quadratic in the curvature scalars.
In order to implement the thin lens approximation to the optical scalars , , , we need to manipulate these expressions as follows. Using (\[aprox-delgada\]) in (\[convergencia:gral2\]) and (\[shear:gral2\]) we have terms of the form $$\label{a100}
\begin{aligned}
\int_{0}^{\lambda_s} \lambda (\lambda_{s}-\lambda) C(\lambda) d\lambda %&= \lambda^{'} (\lambda_{s}-\lambda^{'}) \widetilde{C}(\lambda^{'})
%\bigg|^{\lambda_s}_{0} - \int_{0}^{\lambda_{s}} (\lambda_{s}-2\lambda^{'}) \widetilde{C}(\lambda^{'}) d\lambda^{'}
%= - \int_{0}^{\lambda_{s}} (\lambda_{s}-2\lambda^{'}) \widetilde{C}(\lambda^{'}) d\lambda^{'} \\
%&\cong - \hat{C} \int_{\lambda_{l}}^{\lambda_{s}} (\lambda_{s} - 2\lambda^{'}) d\lambda^{'}
%= -\hat{C} \bigg[ \lambda_{s} (\lambda_{s} - \lambda_{l}) - \lambda_{s}^{2} + \lambda_{l}^{2} \bigg]
= \hat{C} \lambda_{l} \lambda_{ls}.
\end{aligned}$$ That is, the integrals which are linear in the curvature scalars $C$ can be written in a more compact way in terms of the projected quantities $\hat{C}$ [@PhysRevD.83.083007; @0264-9381-22-14-005].
On the other hand, even when we can not do the same with the terms which are quadratic in $C$ and $D$, we use the following notation which will be useful later. The contribution of these terms to the optical scalars will be denoted as $$\begin{aligned}
\left[\cdot\right]^{(2)}_{CD}=&-\frac{1}{\lambda_s}\int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \\
& \times C(\lambda{}) D(\lambda{'}) d\lambda{'}d\lambda{};
\end{aligned}$$ where into the squares brackets $\left[\cdot\right]$ should be placed the associated quantity to the respective optical scalar. Therefore, we have quadratic contributions to the second order optical scalars with terms like ${\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}$, ${\kappa}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\bar\Psi$}}$, ${\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}$ and ${\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}$.
Multiplying each of them by the prefactor $\frac{\lambda_s}{\lambda_l\lambda_{ls}}$, we define the associated scalars $\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}$, $\tilde{\kappa}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\bar\Psi$}}$, $\tilde{\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}$ and $\tilde{\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}$ by: $$\label{tildeka}
\begin{aligned}
\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}=\frac{\lambda_s}{\lambda_l\lambda_{ls}}{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}=
-\frac{1}{\lambda_l\lambda_{ls}}\int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \\
\times \; \Phi_{00}(\lambda{'}) \Phi_{00}(\lambda{})d\lambda{'} d\lambda;&
\end{aligned}$$ with similar relations for the rest of the quadratic terms. Finally, using the approximation we obtain for the convergence and shear $$\label{eq:2.34}
\begin{aligned}
\kappa &= \frac{\lambda_{l} \lambda_{ls}}{\lambda_{s}} \bigg( \hat{\Phi}_{00}^{(1)} + \delta \hat{\Phi}_{00}^{(2)} +\hat{\Phi}_{00}^{(2)}+\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}+\Re [\tilde{\kappa}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\bar\Psi$}}] \bigg),
\end{aligned}$$ $$\label{eq:2.33}
\begin{aligned}
&\gamma = \frac{\lambda_{l} \lambda_{ls}}{\lambda_{s}} \bigg( \hat{\Psi}_{0}^{(1)} + \delta \hat{\Psi}_{0}^{(2)} +\hat{\Psi}_{0}^{(2)} + \tilde{\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}} +\tilde{\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}\bigg).
\end{aligned}$$
As $\hat\omega$ does not have linear terms in the curvature, it remains the same as in . Note that in the previous expressions, the projected hat quantities are given by $$\begin{aligned}
\hat{\Phi}_{00}^{(j)} =& \int^{\lambda_{s}}_{0} \Phi_{00}^{(j)} d\lambda, \ \ \text{$j = 1, 2$},\\
\hat{\Psi}_{0}^{(j)}=& \int^{\lambda_{s}}_{0} \Psi_{0}^{(j)} d\lambda, \ \ \text{$j = 1, 2$},\\
\delta \hat{\Phi}_{00}^{(2)} =& \int^{\lambda_{s}}_{0} \delta x^{(1)a}(\lambda) \frac{\partial \Phi_{00}^{(1)}}{\partial x^{a}}\bigg|_{x^{(0)}(\lambda)} d\lambda, \\
\delta \hat{\Psi}_{0}^{(2)} =& \int^{\lambda_{s}}_{0} \delta x^{(1)a}(\lambda) \frac{\partial \Psi_{0}^{(1)}}{\partial x^{a}}\bigg|_{x^{(0)}(\lambda)} d\lambda.
\end{aligned}$$
Axisymmetric lenses
-------------------
Let us consider a static and axisymmetric gravitational lens in the thin lens approximation where the axis of symmetry corresponds to the line of sight which crosses the central region of the matter distribution from the observer. We select a Cartesian coordinate system with the origin in the lens plane and the line of sight in the negative $y$ direction. As in [@PhysRevD.83.083007], in the lens plane we identify the first component with the $z$ coordinate and the second component with the $x$ coordinate. In this plane, it is convenient to work with new coordinates $(J,\vartheta)$ given by $$\label{eq:a.01}
\begin{aligned}
z &= J \cos(\vartheta), \\
x &= J \sin(\vartheta),
\end{aligned}$$ where $J$ is the impact parameter and $\vartheta$ is the polar angle measured from $z$.
Since $\Phi_{00}$ and $\Psi_{0}$ are quantities with spin weight 0 and 2 respectively, we assume that in the case of static and axial symmetries they have the following functional dependence[@0264-9381-22-14-005]: $$\label{psipsi}
\begin{aligned}
\Phi_{00} &= \Phi_{00}(y,J), \\
\Psi_{0} &= -\psi_{0}(y,J) e^{2i\vartheta},
\end{aligned}$$ for some arbitrary phase in the choice of $m^{a}$. Then, the projected curvature scalars are given by $$\label{hatspherically}
\begin{aligned}
\hat{\Phi}_{00} &= \hat{\Phi}_{00}(J), \\
\hat{\Psi}_{0} &= -\hat{\psi}_{0}(J) e^{2i\vartheta} ,
\end{aligned}$$ where $$\hat{\psi}_{0}(J) = -e^{-2i\vartheta} \int^{\lambda_{s}}_{0} \Psi_{0}(\lambda) d\lambda.$$ According to the expressions and we can write the optical scalars , as follows: $$\label{kappathin}
\begin{aligned}
\kappa =& \frac{\lambda_{l} \lambda_{ls}}{\lambda_{s}} \bigg( \hat{\Phi}_{00}^{(1)}(J) + \delta \hat{\Phi}_{00}^{(2)}(J) + \hat{\Phi}_{00}^{(2)}(J)\\ &+\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}(J,\lambda_l,\lambda_{ls})+\Re [\tilde{\kappa}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\bar\Psi$}}(J,\lambda_l,\lambda_{ls})]\bigg) ,
\end{aligned}$$ $$\label{gamma1thin}
\begin{aligned}
\gamma_{1} =& - \frac{\lambda_{l} \lambda_{ls}}{\lambda_{s}} \bigg( \hat{\psi}_{0}^{(1)}(J) + \delta \hat{\psi}_{0}^{(2)}(J) + \hat{\psi}_{0}^{(2)}(J)\bigg.\\
&+\bigg.\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}}(J,\lambda_l,\lambda_{ls}) +\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}}(J,\lambda_l,\lambda_{ls})\bigg)
\cos(2\vartheta),
\end{aligned}$$ $$\label{gamma2thin}
\begin{aligned}
\gamma_{2} =& - \frac{\lambda_{l} \lambda_{ls}}{\lambda_{s}} \bigg( \hat{\psi}_{0}^{(1)}(J) + \delta \hat{\psi}_{0}^{(2)}(J) + \hat{\psi}_{0}^{(2)}(J)\bigg.\\
&+\bigg.\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}}(J,\lambda_l,\lambda_{ls}) +\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}}(J,\lambda_l,\lambda_{ls}) \bigg)
\sin(2\vartheta),
\end{aligned}$$ where $$\hat{\psi}_{0}^{(j)}= \int^{\lambda_{s}}_{0} \psi_{0}^{(j)} d\lambda, \ \ \text{$j = 1, 2$},$$ $$\delta \hat{\psi}_{0}^{(2)} = \int^{\lambda_{s}}_{0} \delta x^{(1)a}(\lambda) \frac{\partial \psi_{0}^{(1)}}{\partial x^{a}}\bigg|_{x^{(0)}(\lambda)} d\lambda,$$ and $$\begin{aligned}
\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}}=& -\frac{1}{\lambda_{l}\lambda_{ls}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\psi_{0}(\lambda,J) \\ &\Phi_{00}(\lambda{'},J) d\lambda{'} d\lambda;
\end{aligned}$$ with a similar definition for $\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}}$.
Deflection angle
----------------
Now, we want to obtain an expression for the deflection angle in the thin lens approximation which generalizes .
From the lens equation we see that in this approximation the amplification matrix $A^i{}_j$ can be expressed as [@Bartelmann:2010fz; @Narayan:1996ba] $$\label{eq:thin}
A^{i}{}_{j} = \frac{d\beta^{i}}{d\theta^{j}} %= \delta^{i}{}_{j} - \frac{\lambda_{ls}}{\lambda_{s}} \frac{d\alpha^{i}}{d\theta^{j}}
= \delta^{i}{}_{j} - \frac{\lambda_{ls}\lambda_{l}}{\lambda_{s}} \frac{d\alpha^{i}}{dx^{j}},$$ where we have used that in the thin lens approximation $\frac{d}{d\theta^{i}} \approx \lambda_{l}\frac{d}{d x^{i}}$. We define the components of $\alpha^{i} = (\alpha^{1},\alpha^{2})$ as $$\label{eq:alphai}
\alpha^{i} = \alpha(J,\lambda_l,\lambda_{ls})\bigg(\frac{z}{J},\frac{x}{J} \bigg).$$
Hence, from (\[eq:thin\]), and follow that $$\label{eq:2.66}
\kappa - \gamma_{1} \cos(2\vartheta) - \gamma_{2} \sin(2\vartheta) = \frac{\lambda_{ls}\lambda_{l}}{\lambda_{s}} \frac{\alpha}{J}.$$
Therefore, since we are interested in the deflection angle in the asymptotic region, that is where $\lambda_l$ and $\lambda_{ls}$ go to infinity, from (\[eq:2.66\]) and the expressions , , for the optical scalars, we finally obtain
$$\label{eqf02}
\begin{aligned}
\alpha(J) &= J \bigg( \hat{\Phi}^{(1)}_{00}(J) + \hat{\psi}^{(1)}_{0}(J) + \delta \hat{\Phi}^{(2)}_{00}(J) + \delta \hat{\psi}^{(2)}_{0}(J)
+ \hat{\Phi}^{(2)}_{00}(J) + \hat{\psi}^{(2)}_{0}(J)+\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}(J)+\Re [\tilde{\kappa}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\bar\Psi$}}(J)]+\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}}(J) +\underline{\tilde{\gamma}^{(2)}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}}(J) \bigg);
\end{aligned}$$
where $$\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}(J):=\lim_{\lambda_l,\lambda_{ls}\to\infty}\tilde{\kappa}^{(2)}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}(J,\lambda_l,\lambda_{ls}),$$ and with similar definitions for the other quadratic terms.
This expression generalizes our previous formula to second order. At the difference of similar relations which are written in terms of metric components, this quantity is explicitly gauge invariant, due to it only depends on well-defined quantities as the impact parameter $J$ and curvature scalars. We will see in the examples of the next section how using this formula to compute the deflection angle at second order in two different coordinate systems yields the same result.
Applications {#Applications}
============
Optical scalars for a Schwarzschild metric at second order
----------------------------------------------------------
As a first example we will compute explicitly the optical scalars and deflection angle for a Schwarzschild point mass lens in two different coordinate systems.
For this, we proceed as follows. First, we choose a null tetrad such that far away from the point mass, in the asymptotically flat region it reduces to $$\begin{aligned}
\ell^{a} &= (-1,0,1,0), \ \ \ \ \ m^{a} = \frac{1}{\sqrt{2}} (0,i,0,1), \\
n^{a} &= \frac{1}{2} (-1,0,-1,0), \ \bar{m}^{a} = \frac{1}{\sqrt{2}} (0,-i,0,1).
\end{aligned}$$
On the other hand, we choose the origin of the coordinate system in the lens’s position and parametrize the geodesic by $$( x(\lambda), y(\lambda), z(\lambda) ) = ( x, \lambda - \lambda_{l}, z),$$ that is, $\lambda = 0$ indicates the observer’s position and $\lambda = \lambda_{s}$ the source’s position at $\lambda_{ls}$. Without loss of generality we can take $x=J$ and $z=0$ (and therefore $\vartheta=\frac{\pi}{2}$). In order to analyze how the different aspects of the curvature contribute to the optical scalars and taking into account the Eq. we define $$\begin{aligned}
\label{eq:redef}
\tilde\kappa&:=&\frac{\lambda_s}{\lambda_{ls}\lambda_l}\kappa,\\
\tilde\gamma&:=&\frac{\lambda_s}{\lambda_{ls}\lambda_l}\gamma.\end{aligned}$$
### Isotropic coordinates
Let us consider a static, spherically symmetric body acting as a gravitational lens in such a way that its external gravitational field can be described by the Schwarzschild metric. This metric can be expressed in isotropic coordinates by $$\label{isotropic}
ds^{2} = \frac{\bigg(1 - \frac{\varepsilon}{4 r}\bigg)^{2}}{\bigg(1 + \frac{\varepsilon}{4 r}\bigg)^{2}} dt^{2} -
\bigg(1 + \frac{\varepsilon}{4 r}\bigg)^{4} (dx^{2} + dy^{2} + dz^{2}),$$ where $\varepsilon = 2 M $ and $r = \sqrt{x^{2}+y^{2}+z^{2}}$.
The corresponding null tetrad, curvature scalars and the first order correction to the geodesic are shown in the Appendix \[A:sch\]. The integrals needed in the computation of the optical scalars were made with the help of MAPLE and the GRTensor package.
The resulting leading order behavior for the optical scalars is given by $$\label{schw-gamma}
\begin{aligned}
\tilde{\gamma}(J) =& \bigg[\frac{2}{J^2} + \mathcal{O}(\frac{1}{\lambda_{ls}},\frac{1}{\lambda_{l}}) \bigg] \varepsilon
+ \bigg[ \frac{45 \pi}{32 J^{3}} + \mathcal{O}(\frac{1}{\lambda_{ls}},\frac{1}{\lambda_{l}}) \bigg] \varepsilon^{2}\\& + \mathcal{O}(\varepsilon^{3});
\end{aligned}$$ $$\label{schw-kappa}
\begin{aligned}
\tilde{\kappa}(J) &= \bigg[ - \frac{15 \pi}{32 J^{3}} + \mathcal{O}(\frac{1}{\lambda_{ls}},\frac{1}{\lambda_{l}}) \bigg] \varepsilon^{2}
+ \mathcal{O}(\varepsilon^{3}),
\end{aligned}$$ while $\hat{\omega} = 0$.
If we take the limits $\lambda_{l} \to \infty $ and $ \lambda_{ls} \to \infty$ of these relations and using \[or equivalently \], we obtain an expression for the deflection angle which agrees with the familiar result for the bending angle at second-order for the Schwarzschild metric $$\label{eq:Salpha}
\alpha(J) = \frac{4\,M}{J} + \frac{15\,\pi}{4} \frac{M^{2}}{J^{2}}.$$
### Quasi-Minkowskian coordinates
It is important to note that, the expressions , and are coordinate independent because they only depend on the impact parameter $J$ and the total mass $M$. We could make similar computations for the same metric in a different gauge. For example, using the so-called quasi-Minkowskian coordinate system where the Schwarzschild metric reads[@Weinberg-1973]: $$\label{cartesian-coord}
\begin{aligned}
ds^{2} = &\bigg(1 - \frac{\varepsilon}{r} \bigg) dt^{2} - (dx^{2} + dy^{2} + dz^{2}) \\
&- \bigg[ \bigg( 1 - \frac{\varepsilon}{r} \bigg)^{-1} - 1 \bigg] r^{-2} ( x dx + y dy + z dz )^{2},
\end{aligned}$$ where again $\varepsilon = 2M$ and $r=\sqrt{x^{2}+y^{2}+z^{2}}$, we obtain the same results for the optical quantities. (For a complete discussion and more details, see table \[T1\] and Appendix \[A:sch\]).
Parametrized-Post-Newtonian point mass lens
-------------------------------------------
We will consider now a more general metric known as the *Parametrized-post-Newtonian (PPN) point mass metric* whose line element is given by $$ds^{2} = \bigg( 1 - \frac{\varepsilon}{r} + \frac{\beta \varepsilon^{2}}{2 r^{2}} \bigg) dt^{2} - \bigg( 1 + \frac{\mu \varepsilon}{r} + \frac{3 \nu \varepsilon^{2}}{8 r^{2}} \bigg)
(dx^{2} + dy^{2} + dz^{2}),$$ where $\varepsilon = 2 M $ and $r = \sqrt{x^{2}+y^{2}+z^{2}}$. For the particular choice $\beta = \mu = \nu = 1$, this metric reduces to the second order approximation of the Schwarzschild metric in isotropic coordinates. In the Appendix \[A:sch\] are described the corresponding null tetrad, first order correction to the geodesic and curvature scalars.
The leading order behavior for the optical scalars in the PPN metric is given by $$\tilde{\kappa} = \bigg[\frac{\pi(-8+4\beta-3\nu-8\mu)}{32J^{3}} + \mathcal{O}(\frac{1}{\lambda_{ls}},\frac{1}{\lambda_{l}})\bigg]\varepsilon^{2}+ \mathcal{O}(\varepsilon^{3}),$$ $$\begin{aligned}
\tilde{\gamma} =& \bigg[\frac{(1+\mu)}{J^2}+\mathcal{O}(\frac{1}{\lambda_{ls}},\frac{1}{\lambda_{l}})\bigg]\varepsilon - \bigg[\frac{3\pi(-8+4\beta-3\nu-8\mu)}{32J^{3}} \\ &+\mathcal{O}(\frac{1}{\lambda_{ls}},\frac{1}{\lambda_{l}})\bigg]\varepsilon^{2}+ \mathcal{O}(\varepsilon^{3}).
\end{aligned}$$ Again, if we take the limits $\lambda_{l} \to \infty $ and $ \lambda_{ls} \to \infty$ and using we recover the well-known result for the deflection angle[@Epstein-1980] $$\alpha(J)=2(1+\mu)\frac{M}{J} + \pi(2 - \beta + 2 \mu + \frac{3}{4}\nu)\frac{M^{2}}{J^{2}}.$$
[lccc]{} Optical scalars &Isotropic[^1] & q-Minkowskian[^2] & PPN[^3]\
$\tilde{\kappa}_{\text{\tiny $\Phi^{(1)}$}}$ & 0 & 0 & 0\
$\tilde{\kappa}_{\text{\tiny $\Phi^{(2)}$}}$ & 0 & 0 & $\frac{\pi(-13+8\beta-6\nu+2\mu+9\mu^{2})}{16} \frac{M^{2}}{J^{3}}$\
$\tilde{\kappa}_{\text{\tiny $\delta$}\text{\tiny $\Phi$}}$ & 0 & 0 & $-\frac{\pi(-2+\mu+\mu^{2})}{8} \frac{M^{2}}{J^{3}}$\
$\tilde{\kappa}_{\text{\tiny $\Phi$}\text{\tiny $\Phi$}}$ & 0 & 0 & $\frac{\pi(1-2\mu+\mu^{2})}{32} \frac{M^{2}}{J^{3}}$\
$\tilde{\kappa}_{\text{\tiny $\Psi$}\text{\tiny $\Psi$}}$ & $-\frac{15\pi}{8} \frac{M^{2}}{J^{3}}$ & $-\frac{15\pi}{8} \frac{M^{2}}{J^{3}}$ & $-\frac{15\pi(1+2\mu+\mu^{2})}{32} \frac{M^{2}}{J^{3}}$\
\
$\tilde{\kappa}$ (Total contribution at second order) & $-\frac{15\pi}{8} \frac{M^{2}}{J^{3}}$ & $-\frac{15\pi}{8} \frac{M^{2}}{J^{3}}$ & $\frac{\pi(-8+4\beta-3\nu-8\mu)}{8}\frac{M^2}{J^{3}}$\
\
\
$\tilde{\gamma}_{\text{\tiny $\Psi^{(1)}$}}$ & $\frac{4M}{J^2}$ & $\frac{4M}{J^2}$ & $2(1+\mu) \frac{M}{J^2}$\
$\tilde{\gamma}_{\text{\tiny $\Psi^{(2)}$}}$ & $\frac{3\pi}{8} \frac{M^{2}}{J^{3}}$ & $\frac{15\pi}{4} \frac{M^{2}}{J^{3}}$ & $\frac{\pi(45-24\beta+18\nu+6\mu-39\mu^{2})}{16} \frac{M^{2}}{J^{3}}$\
$\tilde{\gamma}_{\text{\tiny $\delta$}\text{\tiny $\Psi$}}$ & $\frac{21\pi}{4} \frac{M^{2}}{J^{3}}$ & $\frac{15\pi}{8} \frac{M^{2}}{J^{3}}$ & $\frac{3\pi(2+7\mu+5\mu^{2})}{8} \frac{M^{2}}{J^{3}}$\
$\tilde{\gamma}_{\text{\tiny $\Phi$}\text{\tiny $\Psi$}}$ & 0 & 0 & $\frac{9\pi(\mu^{2}-1)}{32} \frac{M^{2}}{J^{3}}$\
$\tilde{\gamma}_{\text{\tiny $\Psi$}\text{\tiny $\Phi$}}$ & 0 & 0 & $\frac{9\pi(\mu^{2}-1)}{32} \frac{M^{2}}{J^{3}}$\
\
$\tilde{\gamma}$ (Total contribution at second order) & $\frac{4M}{J^2}
+ \frac{45\pi}{8}\frac{M^{2}}{J^{3}}$ & $\frac{4M}{J^2}
+ \frac{45\pi}{8}\frac{M^{2}}{J^{3}}$ & $2{(1+\mu)}\frac{M}{J^2} + \frac{3\pi(8-4\beta+3\nu+8\mu)}{8}\frac{M^2}{J^{3}}$\
\
Term by term contribution
-------------------------
Using the definitions , we denote the different terms in the expressions of the optical scalars as follows:
$$\begin{aligned}
\tilde{\kappa} =& \frac{1}{\lambda_{ls} \lambda_{l}} \bigg[ \underbrace{ \int_{0}^{\lambda_{s}} \lambda (\lambda_{s}-\lambda) \Phi_{00}^{(1)}(\lambda) d\lambda}_{\text{\large ${\kappa}^\dagger_{\text{\tiny $\Phi^{(1)}$}}$}}
+ \underbrace{ \int_{0}^{\lambda_{s}} \lambda (\lambda_{s}-\lambda) \Phi_{00}^{(2)}(\lambda) d\lambda}_{\text{\large ${\kappa}^\dagger_{\text{\tiny $\Phi^{(2)}$}}$}}
+ \underbrace{ \int_{0}^{\lambda_{s}} \lambda (\lambda_{s}-\lambda) \delta x^{(1)a}(\lambda)
\frac{\partial\Phi_{00}^{(1)}}{\partial x^{a}}\bigg |_{\lambda} d\lambda}_{\text{\large ${\kappa}^\dagger_{\text{\tiny $\delta\Phi$}}$}} \\
&\underbrace{- \int_{0}^{\lambda_{s}} \int_{0}^{\lambda^{}} \lambda^{'}(\lambda_{s}-\lambda)(\lambda-\lambda^{'})
\Phi_{00}^{(1)}(\lambda) \Phi_{00}^{(1)}(\lambda^{'}) d\lambda^{'} d\lambda}_{\text{\large ${\kappa}^\dagger_{\text{\tiny $\Phi\Phi$}}$}}
\underbrace{- \int_{0}^{\lambda_{s}} \int_{0}^{\lambda^{}} \lambda^{'}(\lambda_{s}-\lambda)(\lambda-\lambda^{'})
\Re \{\Psi_{0}^{(1)}(\lambda) \Psi_{0}^{(1)}(\lambda^{'})\} d\lambda^{'} d\lambda}_{\text{\large ${\kappa}^\dagger_{\text{\tiny $\Psi\Psi$}}$}} \bigg];
\end{aligned}$$
$$\begin{aligned}
\tilde{\gamma} =& \frac{1}{\lambda_{ls} \lambda_{l}} \bigg[ \underbrace{ \int_{0}^{\lambda_{s}} \lambda (\lambda_{s}-\lambda) \Psi_{0}^{(1)}(\lambda) d\lambda}_{\text{\large ${\gamma}^\dagger_{\text{\tiny $\Psi^{(1)}$}}$}}
+ \underbrace{ \int_{0}^{\lambda_{s}} \lambda (\lambda_{s}-\lambda) \Psi_{0}^{(2)}(\lambda) d\lambda}_{\text{\large ${\gamma}^\dagger_{\text{\tiny $\Psi^{(2)}$}}$}}
+ \underbrace{ \int_{0}^{\lambda_{s}} \lambda (\lambda_{s}-\lambda) \delta x^{(1)a}(\lambda)
\frac{\partial\Psi_{0}^{(1)}}{\partial x^{a}}\bigg |_{\lambda} d\lambda}_{\text{\large ${\gamma}^\dagger_{\text{\tiny $\delta\Psi$}}$}} \\
&\underbrace{- \int_{0}^{\lambda_{s}} \int_{0}^{\lambda^{}} \lambda^{'}(\lambda_{s}-\lambda)(\lambda-\lambda^{'})
\Phi_{00}^{(1)}(\lambda) \Psi_{0}^{(1)}(\lambda^{'}) d\lambda^{'} d\lambda}_{\text{\large ${\gamma}^\dagger_{\text{\tiny $\Phi\Psi$}}$}}
\underbrace{- \int_{0}^{\lambda_{s}} \int_{0}^{\lambda^{}} \lambda^{'}(\lambda_{s}-\lambda)(\lambda-\lambda^{'})
\Psi_{0}^{(1)}(\lambda) \Phi_{00}^{(1)}(\lambda^{'}) d\lambda^{'} d\lambda}_{\text{\large ${\gamma}^\dagger_{\text{\tiny $\Psi\Phi$}}$}} \bigg].
\end{aligned}$$
Using these definitions, in Table \[T1\] we show how each of these terms contribute to the optical scalars. In the case of the Schwarzschild metric we obtain unequal expressions for $\tilde{\gamma}_{\text{\tiny $\Psi^{(2)}$}}$ and $\tilde{\gamma}_{\text{\tiny $\delta$}\text{\tiny $\Psi$}}$ when the metric is expressed in two different coordinate systems. It is not unexpected because we are considering the same geodesic but in two different coordinate systems. However, their addition contribute in the same way to the total shear. Note also that at second order, the convergence is different from zero as a consequence of the Weyl-Weyl interaction. It means, that even when the Schwarzschild solution can be though as a thin lens, there exist nontrivial contributions to the convergence which comes from interactions between the curvature components in closely but different regions.
In the case of PPN metrics the contribution to the convergence from the term $\tilde{\kappa}_{\text{\tiny $\Phi^{(2)}$}}$ is different from zero, due to the fact that the Ricci scalar $\Phi_{00}$ for this general family of metrics is not zero. Despite that, in the thin lens approximation $\tilde{\kappa}_{\text{\tiny $\Phi^{(1)}$}}=0$.
Final remarks
=============
In recent years, the theoretical study of gravitational lenses has been fundamental for the description and analysis of different astrophysical and cosmological phenomena of our Universe. In this work, we have shown how to express different optical scalars and the deflection angle at second order in terms of curvature scalars. These formulas are general, and allow us to find explicit expressions for the optical quantities once the gauge is fixed.
As an example of the formalism, we have shown that by expressing the Schwarzschild’s solution in two different coordinate systems, one obtains at second order the same final expressions for the (asymptotic) optical scalars. This is not surprising, because the evaluation of such scalars in the asymptotic region must only depend from the well-defined quantities as the total ADM mass of the spacetime and the involved impact parameter of the considered null geodesics [@2003AmJPh].
For a thin lens situation, we also have shown that even when at first order the lens can be thought as if the whole distribution of matter of the lens is placed in a single plane, when one consider second-order corrections, one must be more careful and preserve quadratic terms that at a first sight could be thought that they do not contribute.
\[grafico1\] ![Top: a comparative plot of the first and second order contributions to the shear for a spherical mass distribution with $M=10^{12}M_\odot$ situated at a distance of 1500 Mpc and with $\lambda_{s}/\lambda_{l}=2$. Bottom: the quotient between the leading and second order contribution to the shear $\gamma$.[]{data-label="grafico1"}](shear.eps "fig:"){width="95mm"}
Finally, one would like to make some comments about how important second-order contributions to the optical scalars could be. In order to give an example of the estimated order of magnitude involved, we consider a model of a galaxy as a spherical mass distribution of the order of $10^{12}$ solar masses located at a distance $\lambda_l=1500$Mpc from us. The sources are assumed to be at a distance $\lambda_s=2\lambda_l$. Of course, even when a point mass model is a rough model and we would also take into account cosmological corrections, it nevertheless serves to have an idea of the order of magnitude involved. In fact, compared to more realistic matter distribution models such as those coming from an NFW density profile, it remains a reasonable estimator [@Lasky:2009]. In Figure \[grafico1\] we have plotted the contribution that $\gamma^{(2)}$ and $\gamma^{(1)}$ make to the total shear and also its quotient. As it can be seen in the figure, second-order contribution to the shear for this kind of astrophysical system is of the order $10^{-6}$ of the main contribution. A similar crude analysis follows for groups and cluster of galaxies. For example, for the Coma cluster which is placed at 100 Mpc from us, with an estimated mass of $M_{\text{Coma}}=7\times 10^{14}M_\odot$ at $J\approx 2$ Mpc[@Gallo:2011hi], $\gamma^{(2)}/\gamma^{(1)}\approx 8\times 10^{-5}$. In general for a point mass, or in the exterior region of a spherical mass the quotient $\gamma^{(2)}/\gamma^{(1)}$ is proportional to the ratio $ {r_{H}/ J}$, with $r_{H}$ the Schwarzschild radius.
We would also like to mention that in the literature, it can also be found the discussion of gravitational lensing for exotic objects [@Bozza:2015haa; @Asada:2017vxl; @Kitamura:2013tya; @Izumi:2013tya; @Tsukamoto:2014dta]. For some of these kind of objects, the first order effect in the optical scalars or deflection angle are of the same order of magnitude that the second order effects coming from a point mass in a Schwarzschild metric. For example, it follows from [@Tsukamoto:2014dta] that for a four-dimensional projection of a five-dimensional Tangherlini spacetime with mass $M$, the deflection angle is proportional to $M^2/J^2$, and therefore the convergence and shear of this metric in the weak field regime have a similar functional dependence as the second order contribution to the optical scalars for a four-dimensional Schwarzschild metric with the same mass.
On the other hand, the contribution of second order weak lensing to the CMB power spectrum or in a general cosmological context is extensively discussed in [@Hagstotz:2014qea; @1475-7516-2015-06-050; @Kaiser:2015iia; @BenDayan:2012wi; @Fanizza:2013doa; @Marozzi:2014kua; @Marozzi:2016uob; @Petri:2016qya; @DiDio:2014lka; @DiDio:2015bua; @Schaefer:2005up; @Cooray:2002mj; @Marozzi:2016qxl; @Bernardeau-2010]. We mention that recently in [@Boero:2016nrd], Boero and Moreschi extended the analysis of Ref.[@PhysRevD.83.083007] to the cosmological setting. Due to the relevance of second order effects in the cosmological framework, it would be desirable to generalize some of the results of [@Boero:2016nrd] to second order.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank an anonymous referee whose comments and suggestions helped to improve this manuscript. We acknowledge support from CONICET and SeCyT-UNC.
{#app:int}
Proof of the identity
----------------------
In order to prove the identity we only have to integrate by parts the integral in $\lambda{'}$ in the left-hand side taking a function $$u=\int_{0}^{\lambda{'}}\lambda{''}Q(\lambda{''})d\lambda{''},$$ and the other function $v$ such that $$\frac{d v}{d \lambda{'}}=1;$$ then, using the well know expression $$\label{intbyparts}
\int u(s) \frac{d v}{ds} ds= u v - \int \frac{d u}{ds} v(s) ds$$ and making the identification $\lambda{'}$$\to$$\lambda$ the identity is proven.
Proof of the identity
----------------------
The identity is a particular case of the following relation: $$\label{identity2a}
\begin{aligned}
&\int^{\lambda_{s}}_{0} \int^{\lambda{'}}_{0} \int^{\lambda{''}}_{0} \lambda{'''} (\lambda{''}-\lambda{'''})
f(\lambda{'''},\lambda{''}) d\lambda{'''} d\lambda{''} d\lambda{'} \\
&= \int_{0}^{\lambda_{s}} \int_{0}^{\lambda{}}\lambda{'}(\lambda_{s}-\lambda{})(\lambda-\lambda{'})f(\lambda{'},\lambda) d\lambda{'}d\lambda{}.
\end{aligned}$$ In order to prove the identity we have to integrate by parts the integral in $\lambda{'}$ in the left-hand side taking a function $$u=\int_{0}^{\lambda{'}} \int_{0}^{\lambda{''}}\lambda{'''}(\lambda{''}-\lambda{'''})f(\lambda{'''},\lambda{''}) d\lambda{'''}d\lambda{''},$$ and the other function such that $$\frac{d v}{d \lambda{'}}=1;$$ then, using and making the identifications $\lambda{''}$$\to$$\lambda{'}$, $\lambda{'}$$\to$$\lambda$, the identity is proven.
{#abcd}
Explicit expressions for the coefficients $A$, $B$, $A'$ and $B'$ that appear in the solution of the geodesic deviation equation at second order
------------------------------------------------------------------------------------------------------------------------------------------------
$$\begin{aligned}
A=& 1 - \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Phi_{00}(\lambda) d\lambda
+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\bigg(\Phi_{00}(\lambda)\Phi_{00}(\lambda{'}) + \Psi_{0R}(\lambda) \Psi_{0R}(\lambda{'}) \\
&+ \Psi_{0I}(\lambda) \Psi_{0I}(\lambda{'}) \bigg) d\lambda{'} d\lambda
- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}} \lambda (\lambda_{s} - \lambda) \Psi_{0R}(\lambda) d\lambda
+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\bigg(\Phi_{00}(\lambda) \Psi_{0R}(\lambda{'}) \\
&+ \Psi_{0R}(\lambda) \Phi_{00}(\lambda{'}) \bigg) d\lambda{'}d\lambda,
\end{aligned}$$
$$\begin{aligned}
B=& - \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Psi_{0I}(\lambda) d\lambda
+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\bigg( \Phi_{00}(\lambda) \Psi_{0I}(\lambda{'}) + \Psi_{0I}(\lambda) \Phi_{00}(\lambda{'}) \bigg)
d\lambda{'}d\lambda \\
&- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \bigg( \Psi_{0I}(\lambda) \Psi_{0R}(\lambda{'}) - \Psi_{0R}(\lambda) \Psi_{0I}(\lambda{'}) \bigg)
d\lambda{'}d\lambda,
\end{aligned}$$
$$\begin{aligned}
A{'}=& - \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Psi_{0I}(\lambda) d\lambda
+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\bigg( \Phi_{00}(\lambda) \Psi_{0I}(\lambda{'}) + \Psi_{0I}(\lambda) \Phi_{00}(\lambda{'}) \bigg) d\lambda{'}d\lambda \\
&+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'}) \bigg( \Psi_{0I}(\lambda) \Psi_{0R}(\lambda{'}) - \Psi_{0R}(\lambda) \Psi_{0I}(\lambda{'}) \bigg)
d\lambda{'}d\lambda,
\end{aligned}$$
$$\begin{aligned}
B{'}=& 1 - \frac {1}{\lambda_{s}} \int^{\lambda_{s}}_{0} \lambda (\lambda_{s} - \lambda) \Phi_{00}(\lambda) d\lambda
+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\bigg(\Phi_{00}(\lambda)\Phi_{00}(\lambda{'}) + \Psi_{0R}(\lambda) \Psi_{0R}(\lambda{'}) \\
&+ \Psi_{0I}(\lambda) \Psi_{0I}(\lambda{'}) \bigg) d\lambda{'} d\lambda
+ \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}} \lambda (\lambda_{s} - \lambda) \Psi_{0R}(\lambda)d\lambda
- \frac{1}{\lambda_{s}} \int_{0}^{\lambda_{s}}\int_{0}^{\lambda}\lambda{'}(\lambda_{s}-\lambda)(\lambda-\lambda{'})
\bigg(\Phi_{00}(\lambda) \Psi_{0R}(\lambda{'}) \\
&+ \Psi_{0R}(\lambda) \Phi_{00}(\lambda{'}) \bigg) d\lambda{'}d\lambda.
\end{aligned}$$
{#A:integral}
About the contribution of the quadratic terms in the thin lens approximation {#app-thin-lens}
----------------------------------------------------------------------------
In this Appendix, we show that if we take the naive approximation to be sufficient in the computation of the quadratic terms, then they should be vanishing. In (\[convergencia:gral2\]), (\[shear:gral2\]), (\[rotacion:gral2\]) we have terms of the form $$\label{terminos_cuadratios_tipo}
\int^{\lambda_{s}}_{0} \int^{\lambda}_{0} \lambda{'} (\lambda_{s}-\lambda) (\lambda-\lambda{'})
C(\lambda) D(\lambda{'}) d\lambda{'} d\lambda.$$ In order to find the contribution of this expression in the thin lens approximation we proceed as follows. First, we integrate by parts the integral in $\lambda{'}$ taking one function as $u(\lambda{'})=\lambda{'}(\lambda-\lambda{'})$ and the other as $\frac{d v}{d\lambda{'}}=D(\lambda{'})$. The expression is now reduced to $$\label{a:1}
\int_{0}^{\lambda_{s}}(\lambda-\lambda_{s})G(\lambda)C(\lambda) d\lambda$$ where $$\label{a:1aa}
G(\lambda) = \int_{0}^{\lambda} (\lambda-2\lambda{'}) \widetilde{D}(\lambda^{'}) d\lambda{'}.$$ Second, we integrate by parts taking $\tilde{u}(\lambda)=(\lambda-\lambda_{s})G(\lambda)$ and $\frac{d \tilde{v}}{d\lambda}=C(\lambda)$: $$\label{rell}
\begin{aligned}
&\int^{\lambda_{s}}_{0} \widetilde{C}(\lambda) \bigg[ \int_{0}^{\lambda}(2\lambda{'}+\lambda_{s})\widetilde{D}(\lambda^{'}) d\lambda{'} - \int_{0}^{\lambda}2\lambda\widetilde{D}(\lambda^{'})d\lambda{'} \\
&- (\lambda_{s}-\lambda)\lambda\widetilde{D}(\lambda) \bigg] d\lambda
\end{aligned}$$ Finally, implementing the thin lens approximation (\[aprox-delgada\]) to we obtain the anticipated result $$\label{a101}
\int^{\lambda_{s}}_{0} \int^{\lambda}_{0} \lambda{'} (\lambda_{s}-\lambda) (\lambda-\lambda{'})
C(\lambda) D(\lambda{'}) d\lambda{'} d\lambda = 0.$$
{#A:sch}
In this Appendix we consider the second order approximation of the Scharwzschild metric in isotropic and quasi-Minkowskian coordinate systems and also the PPN metric.
We give the expressions of the parallel propagated vectors $\{ \ell^{a}, m^{a}\}$ at first order, the correction to the background geodesics and the required curvature scalars. All these computations were done with the help of the GRTensor package and MAPLE.
These ingredients are necessary to compute the optical scalars and the deflection angle at second order.
As in the main text, according to , the spherical symmetry allows us to set without loss of generality $$x=J, \ \ \ z=0.$$
Schwarzschild lens: Isotropic coordinates
-----------------------------------------
In order to compute the optical scalars and the deflection angle for the second order approximation of the Schwarzschild exterior metric written in isotropic coordinates we need to compute the parallel transport of the vectors $\{ \ell^{a}, m^{a} \}$ at first order in $\varepsilon$, $$\begin{aligned}
\ell^{t} =& -1 + \bigg( -\frac{1}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}} + \frac{1}{2\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{x} =& \frac{1}{J} \bigg( \frac{\lambda_{l}-\lambda}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}}
- \frac{\lambda_{l}}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
\ell^{y} =& 1 - \frac{\varepsilon}{2\sqrt{J^{2}+\lambda_{l}^{2}}}
+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{z} =& \mathcal{O}(\varepsilon^{2});
\end{aligned}$$ $$\begin{aligned}
m^{t} &= \frac{1}{2\sqrt{2}} \frac{i}{J} \bigg( \frac{\lambda-\lambda_{l} }{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} }
+ \frac{\lambda_{l} }{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
m^{x} &= \frac{i}{\sqrt{2}} - \frac{i}{2\sqrt{2}} \frac{\varepsilon}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}} + \mathcal{O}(\varepsilon^{2}), \\
m^{y} &= \frac{i}{2\sqrt{2}J} \bigg( \frac{\lambda-\lambda_{l} }{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} }
+ \frac{\lambda_{l} }{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
m^{z} &= \frac{1}{\sqrt{2}} - \frac{1}{2\sqrt{2}} \frac{\varepsilon}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2} }} + \mathcal{O}(\varepsilon^{2}).
\end{aligned}$$ The correction to the background null geodesic, which follows from the integration of the $\ell^a$ components at first order is $$\begin{aligned}
\delta x^{t} =& \bigg( \frac{\lambda}{2\sqrt{J^{2}+\lambda_{l}^{2}}} - \text{arcsinh}(\frac{\lambda-\lambda_{l}}{J}) - \text{arcsinh}(\frac{\lambda_{l}}{J}) \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{x} =& \frac{1}{J} \bigg( \sqrt{J^{2}+\lambda_{l}^{2}}-\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}
- \frac{\lambda\lambda_{l}}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon \\
&+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{y} =& -\frac{1}{2}\frac{\lambda}{\sqrt{J^{2}+\lambda_{l}^{2}}} \varepsilon
+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{z} =& \mathcal{O}(\varepsilon^{2}).
\end{aligned}$$ The only non-vanishing curvature scalar at second order is $\Psi_{0}$,
$$\label{psi-iso-coord}
\begin{aligned}
\Psi_{0}(J,\lambda) =& \frac{3}{2} \frac{J^2}{(J^{2}+(\lambda - \lambda_{l})^{2})^{5/2}} \varepsilon
- \frac{3}{4}\frac{1}{\sqrt{J^{2}+\lambda_{l}^{2}}(J^{2}+(\lambda-\lambda_{l})^{2})^{7/2}} \bigg[ \sqrt{J^{2}+\lambda_{l}^{2}} \sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} \\
&\times ( J^{2}-4\lambda^{2} + 8\lambda\lambda_{l} -4\lambda_{l}^{2} ) + 2J^{2} ( J^{2}+3\lambda_{l}^{2} + \lambda^{2} -4\lambda\lambda_{l} ) + 4\lambda_{l} (3\lambda^{2}\lambda_{l} - 3\lambda\lambda_{l}^{2}+ \lambda_{l}^{3} - \lambda^{3})
\bigg] \varepsilon^{2}
+ \mathcal{O}(\varepsilon^{3}).
\end{aligned}$$
Schwarzschild lens: Quasi-Minkoskian coordinates {#apendiceA}
------------------------------------------------
In order to compute the optical scalars and the deflection angle for the Schwarzschild point mass lens written in quasi-Minkowskian coordinates we need to compute the parallel transport of the vectors $\{ \ell^{a}, m^{a} \}$ at first order in $\varepsilon$: $$\begin{aligned}
\ell^{t} =& -1 + \bigg( -\frac{1}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}} + \frac{1}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon \\&+ \mathcal{O}(\varepsilon^{2}),\\
\ell^{x} =& \frac{1}{2J} \bigg( \frac{3 J^{2}(\lambda_{l}-\lambda) + 2 (\lambda_{l}-\lambda)^{3}}{(J^{2}+(\lambda-\lambda_{l})^{2})^{3/2}}
- \frac{3\lambda_{l} J^{2} + 2 \lambda_{l}^{3}}{(J^{2}+\lambda_{l}^{2})^{3/2}} \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{y} =& 1 + \frac{ J^{2}}{2} \bigg( \frac{1}{(J^{2}+(\lambda-\lambda_{l})^{2})^{3/2}} - \frac{1}{(J^{2}+\lambda_{l}^{2})^{3/2}} \bigg) \varepsilon\\
&+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{z} &= \mathcal{O}(\varepsilon^{2}); \\
\end{aligned}$$ and $$\begin{aligned}
m^{t} =& \frac{1}{2\sqrt{2}} \frac{i}{J} \bigg( - \frac{(\lambda_{l}- \lambda) }{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} }
+ \frac{\lambda_{l} }{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
m^{x} =& \frac{i}{\sqrt{2}} + \frac{i \, J^{2} }{2\sqrt{2}} \bigg( -\frac{1}{(J^{2}+(\lambda-\lambda_{l})^{2})^{3/2}} \\
&+ \frac{1}{(J^{2}+\lambda_{l}^{2})^{3/2}} \bigg) \varepsilon+ \mathcal{O}(\varepsilon^{2}), \\
m^{y} =& \frac{1}{2\sqrt{2}} \frac{i}{J} \bigg( \frac{(\lambda - \lambda_{l})^{3} }{(J^{2}+(\lambda-\lambda_{l})^{2})^{3/2} }
+ \frac{\lambda_{l}^{3} }{(J^{2}+\lambda_{l}^{2})^{3/2}} \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
m^{z} =& \frac{i}{\sqrt{2}} + \mathcal{O}(\varepsilon^{2}). \\
\end{aligned}$$ The first order contribution to the actual path is $$\begin{aligned}
\delta x^{t} =& \bigg( \frac{\lambda}{\sqrt{J^{2}+\lambda_{l}^{2}}} - \text{arcsinh}(\frac{\lambda-\lambda_{l}}{J}) - \text{arcsinh}(\frac{\lambda_{l}}{J}) \bigg) \varepsilon+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{x} =& \frac{1}{2J\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}(J^{2}+\lambda_{l}^{2})^{3/2}} \bigg(\sqrt{J^{2}+\lambda_{l}^{2}}
(4J^{2}\lambda\lambda_{l}-J^{4}-3J^{2}\lambda_{l}^{2}-2J^{2}\lambda^{2}
-2\lambda^{2}\lambda_{l}^{2}
+4\lambda\lambda_{l}^{3}-2\lambda_{l}^{4})\\
&+\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}
(J^{4}
+3J^{2}\lambda_{l}^{2}-3J^{2}\lambda\lambda_{l}+2\lambda_{l}^{4}-2\lambda\lambda_{l}^{3}) \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
\delta x^{y} =& \frac{1}{2\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}(J^{2}+\lambda_{l}^{2})^{3/2}} \bigg(\sqrt{J^{2}+\lambda_{l}^{2}} (J^{2}\lambda+\lambda\lambda_{l}^{2}-J^{2}\lambda_{l}-\lambda_{l}^{3})+\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}\bigg)(J^{2}\lambda_{l}+\lambda_{l}^{3}-J^{2}\lambda) \varepsilon \\
&+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{z} =& \mathcal{O}(\varepsilon^{2}).
\end{aligned}$$ While for the curvature quantity we get $$\begin{aligned}
\Psi_{0}(J,\lambda) =& \frac{3}{2} \frac{J^{2}}{(J^{2}+(\lambda - \lambda_{l})^{2})^{5/2}} \varepsilon
+ \frac{3}{2} \frac{1}{\sqrt{J^{2}+\lambda_{l}^{2}}(J^{2}+(\lambda - \lambda_{l})^{2})^{7/2}}
\bigg( 2 (\lambda - \lambda_{l})^{2} \sqrt{J^{2}+\lambda_{l}^{2}} \sqrt{J^{2}+(\lambda - \lambda_{l})^{2}} \\
&+ J^{2} \sqrt{J^{2}+\lambda_{l}^{2}} \sqrt{J^{2}+(\lambda - \lambda_{l})^{2}}
+ 4 \lambda_{l} \lambda J^{2} - 3 \lambda_{l}^{2} J^{2} + 2 \lambda_{l} (\lambda - \lambda_{l})^{3}
- J^{4} - \lambda^{2} J^{2} \bigg) \varepsilon^{2} + \mathcal{O}(\varepsilon^{3}).
\end{aligned}$$ Despite the above expression for $\Psi_{0}$ is different of which is calculated in isotropic coordinates, the expressions for the convergence, shear and deflection angle are the same in the limits $\lambda_{l} \to \infty $, $ \lambda_{ls} \to \infty$.
Parametrized-post-Newtonian point mass lens
-------------------------------------------
For the PPN point mass metric the components of $\{\ell^{a},m^{a}\}$ and the contribution to the actual geodesic $\delta x^{a}$ at first order are given by: $$\begin{aligned}
\ell^{t} =& -1 + \bigg( -\frac{1}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}} + \frac{1}{2\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{x} =& \frac{1}{2} \frac{(\mu + 1)}{J} \bigg( \frac{\lambda_{l}-\lambda}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}}
- \frac{\lambda_{l}}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{y} =& 1 - \frac{1}{2}\bigg( \frac{(\mu-1)}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}} + \frac{1}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon
+ \mathcal{O}(\varepsilon^{2}), \\
\ell^{z} =& \mathcal{O}(\varepsilon^{2});
\end{aligned}$$ $$\begin{aligned}
m^{t} &= \frac{1}{2\sqrt{2}} \frac{i}{J} \bigg( \frac{\lambda-\lambda_{l} }{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} }
+ \frac{\lambda_{l} }{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
m^{x} &= \frac{i}{\sqrt{2}} - \frac{i \, \mu}{2\sqrt{2}} \frac{\varepsilon}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}}} + \mathcal{O}(\varepsilon^{2}), \\
m^{y} &= \frac{\mu}{2\sqrt{2}} \frac{i}{ J} \bigg( \frac{\lambda-\lambda_{l} }{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} }
+ \frac{\lambda_{l} }{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
m^{z} &= \frac{1}{\sqrt{2}} - \frac{\mu}{2\sqrt{2}} \frac{\varepsilon}{\sqrt{J^{2}+(\lambda-\lambda_{l})^{2} }} + \mathcal{O}(\varepsilon^{2});
\end{aligned}$$ $$\begin{aligned}
\delta x^{t} =& \bigg( \frac{\lambda}{2\sqrt{J^{2}+\lambda_{l}^{2}}} - \text{arcsinh}(\frac{\lambda-\lambda_{l}}{J}) - \text{arcsinh}(\frac{\lambda_{l}}{J}) \bigg) \varepsilon\\ &+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{x} =& \frac{1}{2} \frac{(\mu + 1)}{J} \bigg( \sqrt{J^{2}+\lambda_{l}^{2}}-\sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} \\
&- \frac{\lambda\lambda_{l}}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) \varepsilon + \mathcal{O}(\varepsilon^{2}), \\
\delta x^{y} =& \frac{1}{2}\bigg[ (1-\mu) \bigg(\text{arcsinh}(\frac{\lambda-\lambda_{l}}{J}) + \text{arcsinh}(\frac{\lambda_{l}}{J}) \\
&- \frac{\lambda}{\sqrt{J^{2}+\lambda_{l}^{2}}} \bigg) - \frac{\mu\lambda}{\sqrt{J^{2}+\lambda_{l}^{2}}}\bigg] \varepsilon
+ \mathcal{O}(\varepsilon^{2}), \\
\delta x^{z} =& \mathcal{O}(\varepsilon^{2});
\end{aligned}$$ and the curvature quantities $\{\Psi_{0},\Phi_{00}\}$ at second order are given by: $$\begin{aligned}
\Psi_{0}(J,\lambda) =& \frac{3}{4} \frac{(\mu + 1) J^2}{(J^{2}+(\lambda - \lambda_{l})^{2})^{5/2}} \varepsilon
+ \frac{1}{8}\frac{1}{\sqrt{J^{2}+\lambda_{l}^{2}}(J^{2}+(\lambda-\lambda_{l})^{2})^{7/2}} \bigg[ \sqrt{J^{2}+\lambda_{l}^{2}} \sqrt{J^{2}+(\lambda-\lambda_{l})^{2}} \\
&\times \bigg( 6 (\lambda - \lambda_{l})^{2} + J^{2} (13 - 8\beta - 2\mu -15\mu^{2} + 6\nu) + 6\mu ( \mu \lambda_{l}^{2} + 2 \lambda^{2}
+ \lambda^{2} \mu + 2 \lambda_{l}^{2} - 2 \lambda \lambda_{l} \mu - 4 \lambda \lambda_{l}) \bigg) \\
&- 6 J^{2} ( \lambda_{l}^{2}\mu^{2}-\lambda \lambda_{l}\mu^{2}-4\lambda_{l}\lambda\mu+\mu\lambda^{2}+3\lambda_{l}^{2}\mu-3\lambda_{l}\lambda+\lambda^{2} ) + 6\mu^{2} ( \lambda_{l}\lambda^{3}+3\lambda_{l}^{3}\lambda-3\lambda_{l}^{2}\lambda^{2}-\lambda_{l}^{4} ) \\
&+ 12\mu ( 3 \lambda \lambda_{l}^{3} - 3 \lambda^{2}\lambda_{l}^{2} + \lambda^{3} \lambda_{l} - \lambda_{l}^{4} -\frac{1}{2}J^{4})
- 6 ( \lambda_{l}^{4} + J^{4} -3\lambda_{l}^{3}\lambda+3\lambda_{l}^{2}\lambda^{2}-\lambda_{l}\lambda^{3} ) \bigg] \varepsilon^{2}
+ \mathcal{O}(\varepsilon^{3});
\end{aligned}$$ $$\begin{aligned}
\Phi_{00}(J,\lambda) =& \frac{1}{4} \frac{(\mu - 1) (J^{2} - 2 (\lambda - \lambda_{l})^{2})}{(J^{2}+(\lambda- \lambda_{l})^{2})^{5/2}} \varepsilon
+ \frac{1}{8} \frac{1}{\sqrt{J^{2}+\lambda_{l}^{2}} (J^{2}+(\lambda- \lambda_{l})^{2})^{7/2}} \bigg[ \sqrt{J^{2} + \lambda_{l}^{2}}
\sqrt{J^{2} + (\lambda - \lambda_{l})^{2}} \\
&\times \bigg( 2(1 - 2\beta - 3\nu) (\lambda-\lambda_{l})^{2} + J^{2} (4\mu - 3\mu^{2} + 4\beta - 5) + 2\mu (9\mu \lambda_{l}^{2}
- 18\mu \lambda \lambda_{l} + 9\mu \lambda^{2} - 5 \lambda_{l}^{2} + 10 \lambda \lambda_{l} \\
&- 5 \lambda^{2} ) \bigg)
+ 2 J^{2} ( \lambda_{l}^{2}\mu- \lambda_{l}\lambda + 3\mu^{2}\lambda_{l}\lambda + \mu\lambda^{2} - \lambda^{2} + 2\lambda_{l}^{2} - 2\mu\lambda_{l}\lambda - 3\mu^{2}\lambda_{l}^{2} ) - 6\mu^{2} ( 3\lambda_{l}^{2}\lambda^{2}-3\lambda_{l}^{3}\lambda - \lambda_{l} \lambda^{3} \\
&+ \lambda_{l}^{4} )
- 4\mu ( 4\lambda_{l}\lambda^{3} + 4\lambda_{l}^{3}\lambda - 6\lambda_{l}^{2}\lambda^{2} + \frac{1}{2}J^{4}- \lambda_{l}^{4} - \lambda^{4} )
- 2 ( 3\lambda_{l}^{2}\lambda^{2} - J^{4} - \lambda_{l}^{4} + \lambda_{l}^{3}\lambda + 2\lambda^{4} - 5\lambda_{l}\lambda^{3} ) \bigg] \varepsilon^{2} \\
&+ \mathcal{O}(\varepsilon^{3}).
\end{aligned}$$
[10]{}
Emanuel Gallo and Osvaldo M. Moreschi. Gravitational lens optical scalars in terms of energy-momentum distributions. 83:083007, ()2011).
D. [Walsh]{}, R. F. [Carswell]{}, and R. J. [Weymann]{}. . , 279:381–384, (1979).
Antony Lewis and Anthony Challinor. . , 429:1–65, (2006).
H. N. Nguyen, N. Sehgal and M. Madhavacheril, “Measuring the small-scale matter power spectrum with high-resolution CMB Lensing,” arXiv:1710.03747. Giovanni Marozzi, Giuseppe Fanizza, Enea Di Dio, and Ruth Durrer. . , 118:211301, 2017.
Julien Peloton, Marcel Schmittfull, Antony Lewis, Julien Carron, and Oliver Zahn. . , D95:043508, 2017.
G. Fabbian, M. Calabrese and C. Carbone, “CMB weak-lensing beyond the Born approximation: A numerical approach,” JCAP, no. 02, 050 (2018). Geraint Pratten and Antony Lewis. . , 1608(08):047, 2016.
Nick Kaiser and John A. Peacock. . , 455(4):4518–4547, 2016.
Pierre Fleury, Chris Clarkson, and Roy Maartens. , 1703(03):062, 2017.
Henk Hoekstra, Matthias Bartelmann, Haakon Dahle, Holger Israel, Marceau Limousin, and Massimo Meneghetti. . , 177:75–118, 2013.
R. [Mandelbaum]{}. . In M. [Cappellari]{} and S. [Courteau]{}, editors, [*Galaxy Masses as Constraints of Formation Models*]{}, volume 311 of [*IAU Symposium*]{}, pages 86–95, April 2015. Cambridge University Press. Cambridge, UK.
Carlo Giocoli, Massimo Meneghetti, R. Benton Metcalf, Stefano Ettori, and Lauro Moscardini. . , 440(2):1899–1915, 2014.
V. Bozza and A. Postiglione. . , 1506(06):036, 2015.
Emanuel Gallo and Osvaldo Moreschi. . , A27:1250044, 2012.
Ezequiel F Boero, Osvaldo M Moreschi; , , 475 (4), 4683-4703, 2018.
R. Ali Vanderveld, Robert R. Caldwell, and Jason Rhodes. . , D84:123510, 2011.
Simonetta Frittelli, Thomas P. Kling, and Ezra T. Newman. . , D61:064021, 2000.
Michael H. Brugmann. . , D72:024012, 2005.
Steffen Hagstotz, Björn Malte Schäfer, and Philipp M. Merkel. . , 454(1):831–838, 2015.
Camille Bonvin, Chris Clarkson, Ruth Durrer, Roy Maartens, and Obinna Umeh. Do we care about the distance to the [CMB]{}? [C]{}larifying the impact of second-order lensing. , 2015(06):050, 2015.
I. Ben-Dayan, G. Marozzi, F. Nugier and G. Veneziano, “The second-order luminosity-redshift relation in a generic inhomogeneous cosmology,” JCAP [**1211**]{}, 045 (2012) G. Fanizza, M. Gasperini, G. Marozzi and G. Veneziano, “An exact Jacobi map in the geodesic light-cone gauge,” JCAP [**1311**]{}, 019 (2013) G. Marozzi, “The luminosity distance–redshift relation up to second order in the Poisson gauge with anisotropic stress,” Class. Quant. Grav. [**32**]{}, no. 4, 045004 (2015) Erratum: \[Class. Quant. Grav. [**32**]{}, 179501 (2015)\] E. Di Dio, R. Durrer, G. Marozzi and F. Montanari, “Galaxy number counts to second order and their bispectrum,” JCAP [**1412**]{}, 017 (2014) Erratum: \[JCAP [**1506**]{}, no. 06, E01 (2015)\] E. Di Dio, R. Durrer, G. Marozzi and F. Montanari, “The bispectrum of relativistic galaxy number counts,” JCAP [**1601**]{}, 016 (2016) Giovanni Marozzi, Giuseppe Fanizza, Enea Di Dio, and Ruth Durrer. . , 1609(09):028, 2016.
A. Petri, Z. Haiman and M. May, “Validity of the Born approximation for beyond Gaussian weak lensing observables,” Phys. Rev. D [**95**]{}, no. 12, 123503 (2017). Bjoern Malte Schaefer and Matthias Bartelmann. . , 369:425–440, 2006.
Asantha Cooray and Wayne Hu. . , 574:19, 2002.
G. Marozzi, G. Fanizza, E. Di Dio and R. Durrer, “CMB-lensing beyond the leading order: temperature and polarization anisotropies,” arXiv:1612.07263 \[astro-ph.CO\]. Ezra Newman; Roger Penrose. An approach to gravitational radiation by a method of spin coefficients. , 3, 1962.
Daniel E. Holz and Robert M. Wald. New method for determining cumulative gravitational lensing effects in inhomogeneous universes. , 58, 8 1998.
Thomas P Kling and Brian Keith. The [B]{}ianchi identity and weak gravitational lensing. , 22(14):2921, 2005.
Robert V. Weinberg, Gravitation and cosmology: Principles and applications of the general theory of relativity. Wiley, New York. 1972.
Irwin I. Epstein, Reuben; Shapiro. Post-post-[N]{}ewtonian deflection of light by the [S]{}un. , 22, 12 1980.
Camille; Vernizzi Filippo Bernardeau, Francis; Bonvin. Full-sky lensing shear at second order. , 81, 4 2010.
J. [Bodenner]{} and C. M. [Will]{}. . , 71:770–773, August 2003.
M. Bartelmann, “Gravitational Lensing,” Class. Quant. Grav. [**27**]{}, 233001 (2010) S. Seitz, P. Schneider and J. Ehlers, “Light propagation in arbitrary space-times and the gravitational lens approximation,” Class. Quant. Grav. [**11**]{}, 2345 (1994) S. Frittelli, T. P. Kling and E. T. Newman, “Image distortion from optical scalars in nonperturbative gravitational lensing,” Phys. Rev. D [**63**]{}, 023007 (2000) P. Schneider, J. Ehlers, and E.E. Falco, “Gravitational lenses,” Springer-Verlag, 1992. R. Narayan and M. Bartelmann, “Lectures on gravitational lensing,” arXiv:astro-ph/9606001v2. J. P. Uzan and F. Bernardeau, “Cosmic strings lens phenomenology: General properties of distortion fields,” Phys. Rev. D [**63**]{}, 023004 (2000) N. Tsukamoto, T. Kitamura, K. Nakajima and H. Asada, “Gravitational lensing in Tangherlini spacetime in the weak gravitational field and the strong gravitational field,” Phys. Rev. D [**90**]{}, no. 6, 064043 (2014). H. Asada, “Gravitational lensing by exotic objects,” Mod. Phys. Lett. A [**32**]{}, no. 34, 1730031 (2017). T. Kitamura, K. Izumi, K. Nakajima, C. Hagiwara and H. Asada, “Microlensed image centroid motions by an exotic lens object with negative convergence or negative mass,” Phys. Rev. D [**89**]{}, no. 8, 084020 (2014). K. Izumi, C. Hagiwara, K. Nakajima, T. Kitamura and H. Asada, “Gravitational lensing shear by an exotic lens object with negative convergence or negative mass,” Phys. Rev. D [**88**]{}, 024049 (2013). P. D. Lasky, C. J. Fluke; , Monthly Notices of the Royal Astronomical Society, Volume 396, Issue 4, 11 July 2009, .
[^1]: Schwarzschild in isotropic coordinates
[^2]: Schwarzschild in quasi-Minkoskian coordinates
[^3]: Parametrized-post-Newtonian point mass metric
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
**Abstract**\
This paper demonstrates the application of Bayesian Artificial Neural Networks to Ordinary Differential Equation (ODE) inverse problems. We consider the case of estimating an unknown chaotic dynamical system transition model from state observation data. Inverse problems for chaotic systems are numerically challenging as small perturbations in model parameters can cause very large changes in estimated forward trajectories. Bayesian Artificial Neural Networks can be used to simultaneously fit a model and estimate model parameter uncertainty. Knowledge of model parameter uncertainty can then be incorporated into the probabilistic estimates of the inferred system’s forward time evolution. The method is demonstrated numerically by analysing the chaotic Sprott B system. Observations of the system are used to estimate a posterior predictive distribution over the weights of a parametric polynomial kernel Artificial Neural Network. It is shown that the proposed method is able to perform accurate time predictions. Further, the proposed method is able to correctly account for model uncertainties and provide useful prediction uncertainty bounds.
author:
- |
D. K. E. Green$^{1,2}$ [^1] $\quad$ Filip Rindler$^{1,2}$ [^2]\
\
$^1$ *The Alan Turing Institute. London, United Kingdom.*\
$^2$ *Mathematics Institute, University of Warwick. Coventry, United Kingdom.*
---
\
**Keywords:** Dynamical systems, probabilistic numerics, inverse problems, Artificial Neural Networks, Bayesian analysis, Ordinary Differential Equations, parametric polynomial kernels
Introduction {#sec:Introduction}
============
This paper explores Bayesian Machine Learning-type methodologies for the inference of chaotic dynamical system models from trajectory observations. Problems of this form are also known as inverse problems for dynamical systems. The forward time behaviour of a dynamical system can be predicted given the solution of an inverse problem by, for example, direct simulation [@iserles2009first] or as a part of a filter which incorporates observational data into time evolution estimates (see [@murphy2012machine], §17.4). The ability to predict the forward time behaviour of dynamical systems has applications across virtually all areas of science and engineering [@meiss2017differential; @strogatz2018nonlinear]. If the equations are chaotic, then the single maximum a posteriori (MAP) forward estimate from a inverse problem solver is likely to be very inaccurate after only a short time. Inverse problems for dynamical systems are almost always ill-posed in the sense that many different models could be used to generate the observed data. Thus, probabilistic techniques are required to make usable inverse model estimates.
Artificial Neural Networks (ANNs) [@Goodfellowetal2016], augmented with Bayesian uncertainty quantification techniques from [@Gal2015], are used in this paper to solve dynamical system inverse problems. Bayesian updating provides the fundamental method for solving inverse problems [@Stuart_2010]. Following [@Green2019], the terms ANN and ‘compute graph’ will be used interchangeably. Compute graphs will be used in this paper to parametrically represent systems of Ordinary Differential Equations.
Dynamical system trajectory observations can be used to estimate the parameters of a compute graph representation of an ODE by minimising an objective functional. This minimisation procedure effectively performs a search over a parameterised space of possible functions. Nonlinearities in the composed functions allow for the parametric space to represent a very large number of possible functions. This is advantageous when solving inverse problems as it allows for the amount of a priori knowledge about the unknown functional form of the dynamical system in question to be minimised.
By utilising Bayesian methods, forward prediction errors can be properly quantified. A probabilistic formulation also allows for errors in the measurement data to be incorporated in a clear manner. Techniques for doing so are described within. Until recently, Bayesian inference over compute graph parameters has been a significant challenge. Fortunately, the method in [@Gal2015] provides a computationally efficient method. The ‘dropout’ technique [@dropout2014] (originally developed as a regulariser) is used to build up probabilistic output samples from a compute graph by randomly disabling certain parameters. This forces the encoding of the solution to spread out across the entire compute graph, thereby potentially avoiding overfitting of outlier data. Bayesian parameter uncertainty can be estimated by repeatedly sampling from the compute graph while dropout is enabled. This gives a computationally tractable approximation to a Gaussian process over the graph parameters.
In [@Gal2015], forward projections of time series are made without learning an explicit ODE model. We detail a method for combining an ODE model and time discretisation errors into the solution of ODE inverse problems. Directly incorporating time discretisation into the inverse problem solution allows the learnt compute graph to be used with standard ODE solver algorithms (as described in [@boyce2017elementary]) for forward time estimation. Further, errors due to time discretisation can be quantified and more carefully controlled. Thus, the interpretability of the proposed method is improved relative to the technique employed in [@Gal2015].
The Bayesian inverse dynamical system methodology described in this paper is tested by numerical analysis of the Sprott B system [@Sprott1994]. The effect of various hyperparameters, such as derivative discretisation error and dropout rate, was examined. The proposed method was able to provide forward time estimates, solving the inverse problem in a probabilistic sense. Accurate predictions were able to be made over time intervals up to several orders of magnitude longer than the observation data sampling rate, demonstrating that the useful predictions can be made. Further, the proposed method was able to provide confidence intervals that correctly bounded test case data. The methodology presented within could be improved by further testing on more complex cases.
Bayesian Artificial Neural Networks {#sec:bayesianNetworks}
===================================
Neural networks are highly effective as nonlinear regression models. On the other hand, modern deep neural networks typically rely on using a massive number of parameters to ensure that gradient based optimisation will not get stuck in local minima. This is problematic for chaotic ODE inverse problems. The right balance between model size and learning capacity must be found. Bayesian modelling of network parameters can help by quantifying the true range of predictions the trained network is capable of producing for a given input. Full Bayesian modelling of neural network parameter uncertainty is computationally intractable. In [@Gal2015], it is shown that approximate Bayesian inference of network parameters can be carried out by introducing a probabilistic compute graph architecture. This section describes this approximation technique, which is then applied to ODE inverse problems in subsequent sections.
Compute Graphs and Artificial Neural Networks
---------------------------------------------
Artificial Neural Networks (ANNs) (composed of compute graphs, see [@bishop1995neural; @Green2019]) are able to represent nonlinear functions by weighted composition of simpler functions. Roughly, a neural network architecture is defined by a directed graph, consisting of a set of nodes and edges. A complete definition is provided in [@Green2019]. The power of compute graph representation of functions is that a large number of possible alternative function choices can be searched efficiently.
Compute graphs can be used for nonlinear regression problems. Given a compute graph architecture, the parameters defining how to weigh the composed functions can be adjusted until some error functional is minimised over the regression data points. Under certain circumstances, this optimisation can be achieved efficiently by combining Automatic Differentiation [@Rall1981j], the backpropagation method and Stochastic Gradient Descent [@bishop1995neural; @Schmidhuber2015].
For the purposes of the regression problems considered in this paper, a subset of suitable compute graphs is described. A real valued feedforward, layerwise compute graph computes a function $g_\theta(x)$ of the form $$\begin{aligned}
g_\theta\colon \mathbb{R}^m \longrightarrow \mathbb{R}^n\end{aligned}$$ as follows. Let the network have ‘layers’, each labelled by a natural number $i$ from $1$ to $L$. The input to each layer $i$ is a vector $a_{i-1} \in \mathbb{R}^{n_{i-1}}$. Each layer has a ‘parameter matrix’ (or ‘weight matrix’), $\theta_i \in \mathbb{R}^{n_{i}\times n_{i-1}}$. Each layer computes a linear transformation of its inputs, $z_i \in \mathbb{R}^{n_i}$, computed by left-multiplying $a_{i-1}$ by $\theta_i$. Finally, each layer possess a nonlinearity function, $\sigma_i$, which is applied elementwise to $z_i$. In summary, each layer computes $$\begin{aligned}
z_i &:= \theta_i a_{i-1}; \\
a_i &:= \sigma_i(z_i)\end{aligned}$$ with the additional conditions $$\begin{aligned}
a_0 &:= x; \\
g_\theta(x) &:= a_L,\end{aligned}$$ where $a_0$ is the input to the compute graph and $g_\theta(x)$ is the graph output function.
Compute graphs of the form defined above are termed ‘nonrecurrent’ (also known as ‘feedforward’) graphs. As the inputs to each layer, $i$, depend only on layers $j$ for $j<i$, the flow of information is unidirectional. Recurrent graphs, by contrast, allow for a layer $i$ to have inputs from layers $j \geq i$. Details regarding recurrent graphs can be found in [@chauvin2013backpropagation]. Recurrent graphs will not be considered further in this paper.
The search for compute graph weights, from the set of all possible weights, is an optimisation problem. Let $\theta$ be the set of all weights in the network across all layers. Then $$\begin{aligned}
\theta:= \lbrace \theta_i \rbrace_{i=1}^L.\end{aligned}$$ Further, let $\Theta$ denote the set of all possible weights such that $\theta \in \Theta$.
The output of the network can be written directly as the composition of linear combinations of inputs and the application of the nonlinearities as follows: $$\begin{aligned}
\label{eqn:basicFeedForwardNet}
g_\theta(x) = \sigma_L(\theta_L \sigma_{L-1}(\theta_{L-1}\sigma_{L-2}( \dots \sigma_1(\theta_1x) \dots))).\end{aligned}$$
If $g_\theta(x)$ should approximate some given function, the values of $\theta$ can be found by optimising some loss functional, $J(T,\theta)$, given a training data set, $T$. Define the training data set, $T$, as $$\begin{aligned}
\label{eqn:trainingData}
T := \lbrace (x_i, g_i) \rbrace_{i=1}^M,\end{aligned}$$ where $(x_i, g_i)$ are given value pairs of the function to be approximated.
Training seeks some optimal weights $\theta^\ast \in \Theta$ such that $$\begin{aligned}
\theta^\ast \in \underset{\theta \in \Theta}{\operatorname{argmin}} ~J(T,\theta).\end{aligned}$$
A common choice for the loss functional is a 2-norm over $T$: $$\begin{aligned}
\label{eqn:trainingDataJTwoNorm}
J(T,\theta) = \frac{1}{M}\sum_{i=1}^M \norm*{ g_\theta(x_i) - g_i}_2^2,\end{aligned}$$ where $\norm*{\cdot}_2$ denotes the Euclidean norm on $\mathbb{R}^n$.
Other loss functionals are also possible [@Goodfellowetal2016].
For this paper, it is assumed that $$\begin{aligned}
J(T,\theta) \geq 0\end{aligned}$$ for all $\theta$.
In the case that the nonlinearities are piecewise (or weakly) differentiable and that the graph is nonrecurrent, Stochastic Gradient Descent [@Goodfellowetal2016] can be used to find an approximation to $\theta^\ast$ by iteratively moving in the direction of decreasing $J(T,\theta)$. Let $\theta^j$ denote the $j$-th iteration of the gradient descent process. Then, for each weight, an approximation to a local minimum can be found by computing $$\begin{aligned}
\label{eqn:standardSGD}
\theta^{j+1} := \theta^j - \alpha \nabla_\theta J(T,\theta^j)\end{aligned}$$ where $\alpha$ is the learning rate (gradient descent step size) and $\nabla_\theta J(T,\theta^j)$ is the derivative of $J(T,\theta)$ with respect to all weights, computed at $\theta^j$. The gradients can be computed efficiently by the backpropagation method (an application of the chain rule [@bishop1995neural]).
The computation of these gradients is typically carried out using Automatic Differentiation methods. Many software packages exist for building and optimising compute graphs, including Tensorflow [@tensorflow2015-whitepaper]. Further, adaptive learning rates are typically used to improve the optimisation performance over the basic SGD algorithm outlined above. For instance, the Adam optimiser [@Kingma2014] works well for many problems.
Standard neural network training as a maximum likelihood estimate {#ssec:standardNetTraining}
-----------------------------------------------------------------
Standard neural network training can be viewed as obtaining a maximum likelihood estimate of the posterior, $P(\theta|T)$, where $T$ is some set of observational data that can be used to compute (or ‘train’) the weights $\theta$ as in \[eqn:trainingData\].
### Required probabilistic notation
The probabilistic notation used in this paper, summarised here, is as follows:
- $P(X)$ denotes a distribution (a measure that may be applied to events) of $X$.
- $P(X=x) = P(X)[\lbrace x \rbrace]$ denotes probability of event $X = x$.
- $P(Y|X=x)$ denotes the conditional probability of $Y$ given $X=x$, to be understood as a distribution over $Y$ that is dependent on $x$.
- Marginalisation of $Y$ from a distribution over $X$ and $Y$ is the operation $$\begin{aligned}
\label{eqn:longMarginal}
P(Y) &= \int P(Y|X=x)dP(X=x).
\end{aligned}$$ Marginalisation is also denoted by the shorthand $$\begin{aligned}
\label{eqn:shortMarginal}
P(Y) &= \int P(Y|X)dP(X)
\end{aligned}$$ in this paper.
### Gibbs measure definition {#sssec:gibbsMeasureDefn}
A definition of the Gibbs measure is also required. The Gibbs measure (defined rigorously in [@georgii2011gibbs] and roughly here) over some space consisting of $x \in X$ is given by $$\begin{aligned}
\label{eqn:gibbsMeasureDefinition}
P(X = x) = \frac{\exp \left(-\beta E(x) \right)}{\int \exp \left( -\beta E(x) \right) dx} = \frac{1}{Z(\beta)}\exp \left(-\beta E(x) \right)\end{aligned}$$ where:
- $E(x)\colon X \to \mathbb{R}$ is a so-called ‘energy function’. Energies can be used to define the relative probabilities of each $x \in X$.
- $\beta$ is a parameter which defines how ‘spread out’ $E(x)$ is over $X$. It can be considered to be analogous to the inverse of the variance of a Gaussian distribution.
- $Z(\beta)$ is a normalising function, referred to as a ‘partition function’, which ensures $P(X)$ is a valid probability measure.
The Gibbs measure as given in \[eqn:gibbsMeasureDefinition\] is defined as long as the integral in $Z(\beta)$ converges [@georgii2011gibbs]. In the limit that $\beta$ goes to positive infinity, all probability mass over $X$ will be concentrated at the minima of $E(x)$. In other words, \[eqn:gibbsMeasureDefinition\] converges (weakly) to the Dirac measure at the minimum of $E(x)$.
### Maximum likelihood approximated from Bayes theorem {#sssec:maxLikelihoodFromBayes}
Using Bayes theorem, the posterior distribution over the weights is $$\begin{aligned}
\label{eqn:standardNetPrior}
P(\theta|T) &= \frac{P(T|\theta)P(\theta)}{P(T)} \\
&= \frac{P(T|\theta)P(\theta)}{\int_\Theta P(T|\theta)dP(\theta)}.\end{aligned}$$ This section derives a maximum likelihood estimate, so that a more general probabilistic approach can be adopted in later parts of this paper.
The output of the network, after training, can be computed by marginalising over the weight posterior to calculate the posterior predictive distribution for $g(x)$. This gives $$\begin{aligned}
\label{eqn:aPosterioriANNOutput}
P(g(x)|T) = \int_\Theta P(g_\theta(x)|\theta)dP(\theta|T)\end{aligned}$$ where $g_\theta(x)$ is the output of a compute graph as in \[eqn:basicFeedForwardNet\] with weights $\theta \in \Theta$. The posterior must be estimated using \[eqn:standardNetPrior\].
Following the techniques described in [@Stuart_2010], the likelihood ratio over weight space can (by assumption) be modelled as a Gibbs measure by letting $$\begin{aligned}
\label{eqn:gibbsMeasureForWeights}
\frac{P(T|\theta)}{P(T)} = \frac{1}{Z} \exp\left( -\beta J(T, \theta) \right)\end{aligned}$$ where the loss functional in \[eqn:gibbsMeasureForWeights\] defines the error between the data $T$ and $g_\theta(x)$ as in \[eqn:trainingDataJTwoNorm\]. In \[eqn:gibbsMeasureForWeights\], the partition function $Z$ has been modified to absorb the normalising factor, $P(T)$, so that $Z$ is given by $$\begin{aligned}
\label{eqn:gibbsMeasureForWeightsPartition}
Z &= \int_\Theta \exp\left( -\beta J(T,\theta) \right)dP(\theta) \\
&= \mathbb{E}_{P(\theta)} \left[ \exp\left( -\beta J(T,\theta) \right) \right].\end{aligned}$$ The normalising factor $Z$ ensures that the posterior $P(\theta|T)$ is a probability distribution (using \[eqn:standardNetPrior\]) so $$\begin{aligned}
\int dP(\theta|T) = \frac{1}{Z}\int \exp \left(-\beta J(T,\theta)\right) dP(\theta) = 1.\end{aligned}$$
Assuming a prior, $P(\theta)$, equal to a point mass $\delta_{\theta^j}$ at $\theta^j$, then, from \[eqn:standardNetPrior\] the likelihood can be expressed as $$\begin{aligned}
\label{eqn:standardNetPriorFixedWj}
P(\theta|T) = \frac{P(T|\theta^j)}{P(T)}.\end{aligned}$$
Taking logs of \[eqn:standardNetPriorFixedWj\] gives $$\begin{aligned}
\log P(\theta|T) &= \log \frac{P(T|\theta^j)}{P(T)} \\
&= \log \left(\frac{1}{Z}\exp\left(-\beta J(T,\theta^j) \right) \right) \\
\label{eqn:logPosteriorW}
&= -\beta J(T,\theta^j) -\log{Z}.\end{aligned}$$
$P(\theta|T)$ is bounded between $0$ and $1$ so $\log P(\theta|T) < 0$. Since $\log$ is monotonic, the maximum likelihood estimate of $P(\theta|T)$ is found when $\log P(\theta|T)$ is maximised.
The $\log$ posterior can then be maximised by gradient ascent by iteratively setting $$\begin{aligned}
\label{eqn:maxLogProb}
\theta^{j+1} := \theta^{j} + \alpha \nabla_\theta \log P(\theta^{j}|T).\end{aligned}$$
Taking gradients of \[eqn:logPosteriorW\] with respect to $\theta$ (assuming that all terms in \[eqn:logPosteriorW\] are smooth in $\theta$) gives $$\begin{aligned}
\label{eqn:almostThereStillDrvZ}
\nabla_\theta \log P(\theta|T) &= -\beta \nabla_\theta J(T,\theta^j) - \nabla_\theta \log Z.\end{aligned}$$ This is simplified by noting that $\nabla_\theta \log Z = 0$ as the $Z$ defined in \[eqn:gibbsMeasureForWeightsPartition\] is a constant.
Computing the maximum of $\log P(\theta|T)$ iteratively by gradient ascent yields $$\begin{aligned}
\label{eqn:sgdIsMaxLikelihood}
\theta^{j+1} = \theta^{j} - \alpha \nabla_\theta J(T,\theta^{j}),\end{aligned}$$ where the constant parameter $\beta$ has been absorbed into $\alpha$.
The local optimisation target in \[eqn:sgdIsMaxLikelihood\] is identical to \[eqn:standardSGD\]. That is, minimisation of $J(T,\theta)$ finds the maximum likelihood estimate of the posterior distribution of the weights, given the training data, by iterating until $$\begin{aligned}
\label{eqn:basicNetMLEstimate}
\theta^{j+1} \approx \theta^\ast \in \underset{\theta \in \Theta}{\operatorname{argmin}}~J(T,\theta).\end{aligned}$$ Then $\theta^\ast$ is an approximation of the maximum likelihood $\theta$ in the posterior $P(\theta|T)$. Assuming that the maximum likelihood estimate in \[eqn:basicNetMLEstimate\] is a reasonable approximation to the true posterior gives $$\begin{aligned}
\label{eqn:posteriorMLEstimate}
P(\theta|T) \approx \delta_{\theta^\ast}.\end{aligned}$$ That is, the posterior is assumed to be approximated by the single point $\theta^\ast$.
The posterior predictive distribution for $g(x)$ is then $$\begin{aligned}
P(g(x)|T) &= \int_\Theta P(g_\theta(x)|\theta) dP(\theta|T)\\
&\approx \int_\Theta P(g_\theta(x)|\theta) d \delta_{\theta^\ast} \\
&= P(g_{\theta^{\ast}}(x)) \\
\label{eqn:basicNetOutputG}
&= \delta_{g_{\theta^{\ast}}(x)}.\end{aligned}$$ The approximate maximum a posteriori distribution for $g(x)$ in \[eqn:basicNetOutputG\], after standard neural network training, reduces to a deterministic function $g_{\theta^{\ast}}(x)$.
Unfortunately, a MAP estimate of a function is insufficient for the needs of this paper and a Bayesian method for approximation of the full posterior predictive distribution is required.
Dropout regularisation for neural networks {#ssec:dropoutReg}
------------------------------------------
Bayesian updating of large parameter spaces is numerically intractable. In [@Gal2015] an efficient approximation technique for parametric Gaussian process regression is introduced. For a compute graph with dropout layers [@dropout2014], it can be shown that introducing dropout before weight layers is equivalent to an approximation of a probabilistic deep Gaussian process [@Damianou2013]. This section introduces the original dropout regularising prior, in preparation for \[ssec:bayesANN\], which describes a method for estimating the posterior over all weights in a trained network.
Dropout randomly disconnects weights within a network. For a single layer, following the definitions for \[eqn:basicFeedForwardNet\], dropout can be implemented as follows. Define the inverse vector Bernoulli distribution (a specific sort of Bernoulli process [@kroese2013statistical]) of dimension $n$ to be a vector in $\mathbb{R}^n$ with random variable entries, $X^i$, for $1 \leq j \leq n$, such that each $X^j$ is either $0$ or $1$ and that the probability that $X^j = 1$, $p$, is the same for all $X^j$. Denote the inverse vector Bernoulli distribution by $D(r,n)$, which is such that $$\begin{aligned}
P(X^j = 1) = 1 - r \text{ for }j=1, \dots, n.\end{aligned}$$
Returning to the definition of dropout, let $z_i$ (for the $i$-th layer in a compute graph) be a vector in $\mathbb{R}^{n_i}$. The value $r$ will be referred to as the ‘dropout rate’ and a sample $d_i \sim D(r,n_i)$ referred to as a ‘dropout mask’ for layer $i$.
Define the Hadamard product, denoted $\circ$, of two vectors in $\mathbb{R}^n$ as the entrywise product $$\begin{aligned}
\circ \colon \mathbb{R}^n \times \mathbb{R}^n \longrightarrow \mathbb{R}^n\end{aligned}$$ such that, for $C = A \circ B$ for $A, B, C \in \mathbb{R}^n$, the entries of $C$ (denoted $C^j$) are given by $C^j = A^j B^j$.
The dropout mask is applied to $z_i$ by taking the Hadamard product of $d_i$ with $z_i$. The layer output, $a_i$, is modified to $$\begin{aligned}
a_i := \sigma_i(d_i \circ z_i).\end{aligned}$$
Entries of $z_i$ multiplied with entries of $d_i$ equal to zero are thus ‘dropped out’ from the computation of $a_i$. Denote the set of all (independent) dropout distributions across the network by $$\begin{aligned}
\label{eqn:fullDrDistribution}
D(r) := \lbrace D(r,n_i)\rbrace_{i=1}^L,\end{aligned}$$ so that $d \sim D(r)$ is the set of sampled dropout masks for all layers $$\begin{aligned}
d := \lbrace d_i \sim D(r,n_i) \rbrace_{i=1}^L.\end{aligned}$$
Denote the function computed by the network, with dropout mask sample $d$ applied to $\theta$ (for all layers in the network), as $$\begin{aligned}
\label{eqn:gWithDropoutApplied}
g_\theta(x|d) := g_\theta(x) \text{ with }d \text{ applied to }\theta.\end{aligned}$$
Dropout was first designed as a regularisation method (in the sense of Tikhonov regularisation, see [@murphy2012machine]). Regularisation methods are equivalent, in a Bayesian optimisation sense, to the selection of some prior over the weight space [@murphy2012machine]. In the original implementation of dropout, randomisation was used only during training, and then disabled when using the compute graph for predictions. That is, after training was completed the dropout layer was modified to have $r = 0$.
In standard dropout training, the maximum likelihood estimate of $\theta$ is found as in \[eqn:sgdIsMaxLikelihood\], with the additional step that the posterior is calculated by marginalising over $P(d) = D(r)$. Training (minimising $J(T,\theta)$) over some data set $T$ with dropout enabled computes a posterior distribution for the weights. The model likelihood given a set of weights and a dropout mask can be computed by $$\begin{aligned}
\frac{P(T|\theta)}{P(T)} &= \int_{D(r)} \frac{P(T|\theta,d=\delta)}{P(T)} dP(d=\delta|\theta) \\
\label{eqn:dropoutLikelihoodModel}
&= \int_{D(r)} \frac{P(T|\theta,d=\delta)}{P(T)} dP(d=\delta)\end{aligned}$$ as $d$ is independent of $\theta$. Then \[eqn:dropoutLikelihoodModel\] can be expressed as an expectation over $D(r)$ as follows $$\begin{aligned}
\label{eqn:dropoutLikelihoodModelExpect}
\frac{P(T|\theta)}{P(T)} = \mathbb{E}_{D(r)}\left[\frac{P(T|\theta,d)}{P(T)}\right].\end{aligned}$$
As in \[eqn:gibbsMeasureForWeights\], following [@Stuart_2010], the likelihood ratio $\frac{P(T|\theta,d)}{P(T)}$ can be assumed to be a Gibbs measure: $$\begin{aligned}
\label{eqn:gibbsMeasureDropout}
\frac{P(T|\theta,d)}{P(T)} = \frac{1}{Z}\exp \left( -\beta J(T,\theta,d) \right)\end{aligned}$$ where $J(T,\theta,d)$ is the loss functional computed using $g_\theta(x|d)$ from \[eqn:gWithDropoutApplied\] (the network output calculated after applying the dropout mask to $\theta$). For example, using a 2-norm loss functional, as in \[eqn:trainingDataJTwoNorm\], and the definition of $T$ in \[eqn:trainingData\] yields $$\begin{aligned}
J(T,\theta,d) = \frac{1}{M} \sum_{i=1}^M \| g_\theta(x_i|d) - g_i \|_2^2.\end{aligned}$$
From this point, essentially the same procedure as that used in \[eqn:sgdIsMaxLikelihood\] can be used to find the posterior maximum likelihood estimate of $\theta$ given $T$ and $D(r)$.
Using Bayes theorem, note that the posterior for $\theta$ given $T$ and a particular dropout mask $d$ is $$\begin{aligned}
P(\theta|T,d) = \frac{P(T|\theta,d)}{P(T)}P(\theta|d).\end{aligned}$$
Assuming the prior $P(\theta|d)$ is given by a point mass at $\theta^j$ and using \[eqn:gibbsMeasureDropout\], the $\log$ probability of the weight posterior given a dropout mask, $d$, is $$\begin{aligned}
\log P(\theta|T,d) &= \log \frac{P(T|\theta^j,d)}{P(T)} \\
\label{eqn:logPostGivenDropout}
&= -\beta J(T,\theta^j,d) - \log Z.\end{aligned}$$
Following the discussing in \[sssec:maxLikelihoodFromBayes\], the posterior probability for $P(\theta|T,d)$ will be maximised by the value of $\theta$ which maximises $\log P(\theta|T,d)$. Taking derivatives of the $\log$ posterior in \[eqn:logPostGivenDropout\] with respect to $\theta$, we get $$\begin{aligned}
\nabla_\theta \log P(\theta|T,d) &= -\beta \nabla_\theta J(T,\theta^j,d) - \nabla_\theta \log Z \\
\label{eqn:gradGivenDropoutNoExpect}
&= -\beta \nabla_\theta J(T,\theta^j,d)\end{aligned}$$ where, as discussed in \[ssec:standardNetTraining\], $\nabla_\theta \log Z = 0$.
Taking expectations of \[eqn:gradGivenDropoutNoExpect\] over all dropout masks gives $$\begin{aligned}
\mathbb{E}_{D(r)}\left[\nabla_\theta \log P(\theta|T,d) \right] &= \mathbb{E}_{D(r)}\left[ -\beta \nabla_\theta J(T,\theta^j,d) \right]; \\
\nabla_\theta \mathbb{E}_{D(r)}\left[ \log P(\theta|T,d) \right] &= -\beta \nabla_\theta \mathbb{E}_{D(r)}\left[ J(T,\theta^j,d) \right].\end{aligned}$$
Maximising $\mathbb{E}_{D(r)}\left[\log P(\theta|T,d) \right]$ by gradient ascent yields $$\begin{aligned}
\theta^{j+1} &= \theta^j + \alpha \nabla_\theta \mathbb{E}_{D(r)}\left[ \log P(\theta|T,d) \right] \\
\label{eqn:sgdForDropoutMaxLogProbeq}
&= \theta^j - \alpha \nabla_\theta \mathbb{E}_{D(r)}\left[ J(T,\theta^j,d) \right]\end{aligned}$$ where the term $\beta$ has been absorbed into the constant $\alpha$ in \[eqn:sgdForDropoutMaxLogProbeq\].
Then, the maximum likelihood estimate for $P(\theta|T)$, averaged across dropout masks, can be approximated by iteratively updating $\theta^{j+1}$ until $$\begin{aligned}
\label{eqn:argminForDropout}
\theta^{j+1} \approx \theta^\ast \in \underset{\theta \in \Theta}{\operatorname{argmin}}~ \mathbb{E}_{D(r)} \left[ J(T,\theta,d)\right].\end{aligned}$$
The dropout modified gradient can be estimated by Monte Carlo sampling $K$ times from $D(r)$: $$\begin{aligned}
\label{eqn:dropoutKApprox}
\mathbb{E}_{D(r)} \left[ J(T,\theta,d)\right] \approx \frac{1}{K} \sum_{k=1}^K J(T,\theta,d_k) \quad \text{ where }d_k \sim D(r).\end{aligned}$$
Training with dropout has the effect of adding a regularising prior over $\theta$, which aims to prevent the network from overfitting. The learnt parameters, $\theta^\ast$, are the maximum likelihood estimate for the posterior $P(\theta|T)$.
The posterior predictive distribution for $g(x)$, after disabling dropout, is then $$\begin{aligned}
\label{eqn:dropoutTrainPosteriorPredNoApprox}
P(g(x)|T) = \int_\Theta P(g_\theta(x))dP(\theta|T).\end{aligned}$$
As in \[eqn:posteriorMLEstimate\], standard dropout training assumes that the posterior density for $P(\theta|T)$ in \[eqn:dropoutTrainPosteriorPredNoApprox\] is approximately a point mass at the maximum likelihood estimate from \[eqn:argminForDropout\]. Then, $$\begin{aligned}
\label{eqn:dropoutTrainPosteriorPred}
P(g(x)|T) &\approx \int_\Theta P(g_\theta(x)) d\delta_{\theta^\ast} \\
&= P(g_{\theta^{\ast}}(x)).\end{aligned}$$
As in \[eqn:basicNetOutputG\], the posterior predictive distribution over the weights in \[eqn:dropoutTrainPosteriorPred\] is a deterministic approximation $g_{\theta^{\ast}}(x)$.
Bayesian neural network approximation {#ssec:bayesANN}
-------------------------------------
While the regularising action of dropout-enabled training can be beneficial, additional steps must be taken to approximate the actual posterior predictive distribution for $g(x)$ over all weights. In [@Gal2015], an extension to the dropout method is introduced which is able to estimate the full posterior. Roughly, the technique is as follows. The starting point is to first train using dropout regularisation. Then, rather than disable dropout to find a maximum a posteriori estimate of $P(g(x))$ for prediction, dropout is left active during prediction. Repeated sampling from the dropout-enabled predictive network can be used to estimate $P(g(x))$. Denote the posterior predictive distribution (used to approximate $g(x)$) after training by $$\begin{aligned}
P(\hat{g}(x)|T).\end{aligned}$$
Crucially, [@Gal2015] demonstrates that the posterior of the network with dropout-enabled prediction is approximately a Gaussian process. This is done by showing that the Kullback-Leibler divergence between the posterior of a deep Gaussian process and the posterior predictive distribution of an ANN is minimised by the dropout training objective. The first two moments of the approximate Gaussian process representing the $P(\hat{g}(x)|T)$ are sufficient to describe the trained ANN predictions. These moments can be recovered efficiently by Monte Carlo sampling.
Training for the full posterior estimation method is the same as that shown in \[eqn:sgdForDropoutMaxLogProbeq\]. However, rather than disable the dropout layers when computing $g_{\theta^\ast}(x)$, the dropout layers remain active. Denote samples from the probabilistic network as $$\begin{aligned}
g_\theta(x|d) \sim P(g(x|d)|\theta)\end{aligned}$$ where $P(g(x|d)|\theta)$ is the probability to sample some value $g(x)$ given $\theta$ and a dropout mask, $d$. The dropout mask is assumed to be sampled from $D(r)$ as in \[eqn:fullDrDistribution\].
Assigning a single value (a MAP estimate) of $\theta^\ast$ for $\theta$ to the network simplifies $P(g(x|d)|T)$ to $$\begin{aligned}
\label{eqn:apostWithDropoutPre}
P(g(x|d)|T) &:= \int_\Theta P(g(x|d)|\theta) d\delta_{\theta^{\ast}} \\
\label{eqn:apostWithDropout}
&= P(g_{\theta^\ast}(x|d))\end{aligned}$$ where $\theta^\ast$ the set of weight parameters found after training the network by gradient descent. That is, $\theta^\ast$ are the maximum likelihood parameters for $\theta$ as in \[eqn:argminForDropout\].
The $P(g(x))$ can be approximated by the posterior predictive distribution computed by marginalising over all dropout masks in $D(r)$, so $$\begin{aligned}
\label{eqn:gHatEqn}
P(\hat{g}(x)) = \int P(g_{\theta^\ast}(x|d)) dD(r).\end{aligned}$$
It is shown in [@Gal2015] that $P(\hat{g}(x))$ can be approximate efficiently by a Gaussian process of the form $$\begin{aligned}
\label{eqn:dropoutNormalPosterior}
P(\hat{g}(x)|T) \approx \mathcal{N}\left(\mu_{\theta^\ast}(x),\sigma_{\theta^\ast}(x) \right)\end{aligned}$$ where the mean, $\mu_{\theta^\ast}$, and standard deviation, $\sigma_{\theta^\ast}$, of the process are computed empirically by repeatedly sampling from the dropout-enabled network.
Denote the $k$-th sample from the dropout-enabled network by $$\begin{aligned}
g^k(x|d) \sim P(g_{\theta^{\ast}}(x|d_k)), \quad d_k \sim D(r).\end{aligned}$$ Then the empirical mean, $\mu_{\theta^\ast}(x)$, for $P(g(x|d))$ can be computed by $$\begin{aligned}
\label{eqn:sampleMuWGauss}
\mu_{\theta^\ast}(x) \approx \frac{1}{K} \sum_{k=1}^K g^k(x|d)\end{aligned}$$ where $K$ is some finite number of samples from the dropout-enabled network. The standard deviation $\sigma_{\theta^\ast}$ can be similarly estimated by $$\begin{aligned}
\label{eqn:sampleSigWGauss}
\sigma_{\theta^\ast}(x) \approx \sqrt{ \frac{1}{K-1} \sum_{k=1}^K \left( g^k(x|d) - \mu_{\theta^\ast}(x) \right)^2 }.\end{aligned}$$
Calculating the output of a compute graph for a given input is computationally cheap, so estimation of $\mu_{\theta^\ast}(x)$ and $\sigma_{\theta^\ast}(x)$ is tractable. The repeated sampling of $g^k(x|d)$ adds only a constant overhead to the computation of a prediction from the network.
As $P(\hat{g}(x))$ in \[eqn:dropoutNormalPosterior\] is a Gaussian distribution for each input, $x$, prediction confidence intervals at each $x$ can be obtained. Upper and lower confidence bounds for a Gaussian distribution can be computed respectively by: $$\begin{aligned}
\label{eqn:gConfUp}
g^U(x) &= \mu_{\theta^\ast}(x) + c \sigma_{\theta^\ast}(x); \\
\label{eqn:gConfDown}
g^L(x) &= \mu_{\theta^\ast}(x) - c \sigma_{\theta^\ast}(x).\end{aligned}$$ The factor $c$ is the number of standard deviations away from the mean required for some confidence level. For example, $c = 1.96$ for a $95\%$ confidence interval. See [@kroese2013statistical] for further details on this subject. Confidence intervals can be used to simplify the interpretation of the quality of the predictions produced by Bayesian compute graph posterior and are used to present the numerical results in \[sec:sprottBSystem\].
Probabilistic representations of dynamical systems {#sec:probApproxDynSys}
==================================================
This section presents a probabilistic ODE representation that can be used for solving inverse problems with Bayesian ANNs.
Dynamical systems
-----------------
Following the notation in §4.1 of [@Green2019], we consider in this paper continuous-time dynamical systems that can be expressed as coupled first-order Ordinary Differential Equations (ODEs) of the form $$\begin{aligned}
\label{eqn:basicODEForm}
\frac{d}{dt}u(t) = f(t,u(t))\end{aligned}$$ where:
- $t \in [0,\infty)$ represents time;
- $u(t) \in \mathbb{R}^n$ is the vector of values representing the $n$ variables of the system at time $t$;
- $f(t,u(t)) \in \mathbb{R}^n$ represents the prescribed time derivatives of $u(t)$.
For the remainder of this paper, we assume that the unknown ODE model, $f$, is autonomous [@boyce2017elementary]. Then $$\begin{aligned}
\label{eqn:odeAutonomous}
\frac{d}{dt}u(t)) = f(u(t)).\end{aligned}$$
Inverse problems for non-autonomous systems required further assumptions and, while not explored further in this paper, would be a useful avenue for future work.
For notational convenience, define $$\begin{aligned}
u_t := u(t).\end{aligned}$$
A solution to the ODE in \[eqn:basicODEForm\] satisfies $$\begin{aligned}
\label{eqn:detFlowMap}
u_{t+s} = u_t + \int_t^{t+s} f(u_\tau) d\tau.\end{aligned}$$
The deterministic (analytic) trajectory of a dynamical system from time $t$ to $t+s$ refers to the set of all states occupied by the dynamical system, $$\begin{aligned}
\label{eqn:trajectorySet}
\lbrace (a,u_{a}) : a \in [t,t+s] \rbrace,\end{aligned}$$ ordered by time. Numerical ODE solvers, described in [@boyce2017elementary], can be used to produce approximations to the analytic trajectory. These approximations are denoted $$\begin{aligned}
\label{eqn:ODESOLVE}
\lbrace u_{t_i} \rbrace_{i=1}^N = \operatorname{ODESOLVE}(\lbrace t_i \rbrace_{i=1}^N,u_{t_0},f)\end{aligned}$$ where $u_{t_0}$ is the initial value for the trajectory, $\lbrace t_i \rbrace_{i=1}^N$ is some finite set of times at which the values of $u_{t_i}$ will be computed, and $f$ is the ODE derivative function as in \[eqn:basicODEForm\].
In this paper, only continuous-time dynamical systems are investigated, although the numerical methods presented could be applied to both continuous-time and discrete-time systems.
Markov process representation of continuous-time dynamical systems {#ssec:markovProcRepOfContTimeDynSys}
------------------------------------------------------------------
Continuous-time dynamical systems can be represented as a Markov process by making reference to the Gibbs measure. An alternative representation, not discussed in this paper, is the Stochastic Partial Differential Equation (SPDE) random variable approach (as in [@holden2013stochastic]). For rigorous definitions of continuous-time stochastic processes, see §IV of [@moral2017stochastic]. We will outline only the necessary parts here.
For a dynamical system, with time derivative model $f$, the Gibbs measure can be used to define the probability to transition from one state, $u_t$, to another state, $u_{t+s}$, in $s$ time units. First, denote the probability to be in a state $u_t$ at time $t$ by $$\begin{aligned}
P_t(u_t).\end{aligned}$$
Define the so-called ‘transition probability’ as the probability to go from some state, $u_t$, to another state, $u_{t+s}$, in $s$ time units by $$\begin{aligned}
P(u_{t+s}|u_t).\end{aligned}$$ Then $$\begin{aligned}
\label{eqn:transitionProbabilityBasic}
P_{t+s}(u_{t+s}) &= \int_{\mathbb{R}^n} P(u_{t+s}|u_t) dP_t(u_t).\end{aligned}$$
### Discretised trajectory Markov model formulation
A discrete trajectory is considered to be a set of probabilities for $u_t$ at some finite set of times, $\lbrace t_i \rbrace_{i=1}^N$. The assumed Markov property of the transition probability can be used to define trajectories in terms of transitions between states at one time, $t$, to another, $t+s$, as a series of smaller steps from $t_i$ to $t_{i+1}$. A probabilistic trajectory is then given by $$\begin{aligned}
\lbrace P_{t_i}(u_{t_i}) \rbrace_{i=1}^N\end{aligned}$$ where each $P_{t_i}(u_{t_i})$ is calculated by $$\begin{aligned}
P_{t_1}(u_{t_{1}}) &= \text{assumed a priori}; \\
P_{t_{i+1}}(u_{t_{i+1}}) &= P(u_{t_{i+1}}|u_{t_i})P_{t_i}(u_{t_i}); \\
P_{t_{i+2}}(u_{t_{i+2}}) &= P(u_{t_{i+2}}|u_{t_{i+1}})P(u_{t_{i+1}}|u_{t_i})P_{t_{i+1}}(u_{t_i}); \\
\notag \text{etc}.\end{aligned}$$
The entire trajectory can be computed by chaining together conditional probabilities: $$\begin{aligned}
\label{eqn:trajectoryProbs}
P_{t_n}(u_{t_{n}}) &= \left[\prod^{n-1}_{j=1} P\left(u_{t_{j+1}}|u_{t_{j}}\right) \right] P_{t_1}(u_{t_1}) ~\quad\text{ for } n = 1 \dots N.\end{aligned}$$
ODE model parameter uncertainty {#ssec:inverseDynamicalSys}
-------------------------------
Uncertainty regarding model parameters can be included by replacing the dependency on $f$ in the transition probability with $f_\theta(u_t)$ and marginalising over all $\theta$.
The probability of an output, $f_\theta(u_t)$, given the set of inputs $\lbrace t, u(t), \theta \rbrace$ will be denoted $$\begin{aligned}
P(f_\theta(u_t)) := P(f_\theta(u_t)|\theta).\end{aligned}$$
Let $\hat{f}(u_t)$ refer to the value of $f_\theta(u_t)$ marginalised over $\theta$. Then $$\begin{aligned}
\label{eqn:hatFDefn}
P(\hat{f}(u_t)) := \int_\Theta P(f_\theta(u_t)) dP(\theta).\end{aligned}$$
The values of $P(\theta)$ can be estimated using Bayes rule given training data, $T$, yielding a posterior distribution, $P(\theta|T)$. The posterior over the weights can be used to estimate the posterior predictive distribution $$\begin{aligned}
\label{eqn:inverseProblemPosteriorPredictive}
P\left(\hat{f}(u_t)|T \right) := \int_\Theta
P\left(f_\theta(u_t) \right) dP(\theta|T).\end{aligned}$$
The posterior predictive distribution in \[eqn:inverseProblemPosteriorPredictive\] can be combined with an ODE integral discretisation technique to recover estimates of the future states of the dynamical system.
### Bayesian networks for dynamical system inverse problems {#ssec:bayesianNetworksForInverseProblems}
To solve a dynamical system inverse problem, the Bayesian Gaussian approximation can be applied to the problem of learning the posterior predictive distribution for $f_\theta(u_t)$ given observations of some process $u_t$.
Let the model for $f_\theta(u_t)$ with dropout augmentation be denoted $$\begin{aligned}
\label{eqn:dropoutEnabledF}
f_\theta(u_t|d).\end{aligned}$$
Training with dropout activated can be used to recover the MAP estimate, $\theta^\ast$. After training, a Gaussian approximation to $\hat{f}(u_t)$ can be recovered for \[eqn:inverseProblemPosteriorPredictive\] by marginalising out the dropout layers, as in \[eqn:dropoutNormalPosterior\], so that $$\begin{aligned}
P(\hat{f}(u_t)|T) &= \int \int_\Theta P(f_\theta(u_t|d)) dP(\theta|T) dD(r) \\
&\approx \int \int_\Theta P(f_\theta(u_t|d)) d\delta_{\theta^\ast} dD(r) \\
\label{eqn:dropoutODENormalPosterior}
&\approx \mathcal{N}\left(\mu_{\theta^{\ast}}(u_t),\sigma_{\theta^{\ast}}(u_t) \right)\end{aligned}$$ where $\mu_{\theta^{\ast}}(u_t)$ and $\sigma_{\theta^{\ast}}(u_t)$ can be estimated by sampling, as in \[eqn:sampleMuWGauss,eqn:sampleSigWGauss\] respectively.
Solution of the inverse problem {#sec:inverseSolution}
===============================
Although the previous section describes the posterior predictive distribution for $\hat{f}$ given $P(\theta|T)$, the method for finding $\theta$ given observations $u_t$ has not yet been described. The particulars depend on additional assumptions. First, an explicit probabilistic representation of the transition probability is required. A Gaussian (Gibbs measure) form is utilised. Second, the form of the error induced by numerical ODE integration schemes must assumed. We take the error $\epsilon$ to be additive Gaussian noise. By making these assumptions, the transition probability can be approximated as a Gaussian process. This, in combination with the Bayesian compute graph approximation method, allows for the ODE problem to be solved in a computationally tractable manner. This section derives the form of the inverse problem approximation scheme used for the numerical analysis in \[sec:sprottBSystem\].
Finding the posterior distribution, assuming a 2-norm error distribution and Euler integration {#ssec:dropoutEuler}
----------------------------------------------------------------------------------------------
The predictive Bayesian network computing $\hat{f}$ can be found by training a Bayesian compute graph on approximations to the time derivatives of $u_t$. In [@Green2019] a method for approximating $f_\theta$ given an integral discretisation was used. In this paper a simpler method is shown, based on approximations to the time derivative of $u_t$.
Denote an approximation to $f(u_t)$, computed using $u_t$ (an observation in $T$) by $f_\gamma(u_t)$. Assuming that $f_\gamma$ has some approximation error, $\gamma$, gives $$\begin{aligned}
f(u_t) &= f_\gamma(u_t) + \gamma\end{aligned}$$ where the form of the implied distribution $P(f(u_t)|f_\gamma(u_t))$ depends on the exact choice of $\gamma$.
Then a dropout-enabled network can be trained, as in \[eqn:argminForDropout\], by finding $$\begin{aligned}
\theta^\ast &\in \underset{\theta \in \Theta}{\operatorname{argmin}}~\mathbb{E}_{D(r)}\left[J(T,\theta,d)\right].\end{aligned}$$
We assume, for mathematical convenience, that the loss term $J(T,\theta,d)$ has a 2-norm representation of the form $$\begin{aligned}
J(T,\theta,d) = \int \mathbb{E}_{P(f(u_t)|f_\gamma(u_t))}\left[\norm{ f(u_t) - f_\theta(u_t|d) }_2^2 \right] du_t.\end{aligned}$$
This error can be approximated by taking $B$ Monte Carlo samples of $f_\gamma$ from $P(f(u_t)|f_\gamma(u_t))$: $$\begin{aligned}
\label{eqn:jForPractical}
J(T,\theta,d) &\approx \int \frac{1}{B}\sum_{b=1}^B \left[\norm{ f_\gamma^b(u_t) - f_\theta(u_t|d) }_2^2 \right] du_t \\
\text{ for } f^b_\gamma(u_t) &\sim P(f(u_t)|f_\gamma(u_t)).\end{aligned}$$
### Euler integral approximation for $f_\gamma$
To actually compute \[eqn:jForPractical\], the distribution $P(f(u_t)|f_\gamma(u_t))$ must be known. In this paper the data, $T$, is assumed to consist of observations of the process, $u_t$, at $N$ discrete times. Then $$\begin{aligned}
\label{eqn:trainingDataForAnODEInverseProblem}
T := \lbrace t_i, u_{t_i} \rbrace_{i=1}^N.\end{aligned}$$
Further, the $t_i$ values are assumed to be evenly spaced and sampled at a constant rate of $\frac{1}{h}$. This assumption gives $$\begin{aligned}
h = t_{i+1}-t_i.\end{aligned}$$
The simplest method to compute the approximate time derivatives $f_\gamma$ from the training observations is to use a first-order finite difference method of the form $$\begin{aligned}
\label{eqn:derivativeApproxFirstOrder}
f_\gamma(u_{t_i}) &= \frac{u_{t_{i+1}}-u_{t_i}}{h} \\
f(u_{t_i}) &\approx f_\gamma(u_{t_i}) + \mathcal{O}(h^2).\end{aligned}$$ This approximation implies $\gamma \approx \mathcal{O}(h^2)$.
The error $\gamma$ is assumed to be additive Gaussian noise. Then $$\begin{aligned}
\label{eqn:sigGammaEulerDefnPart1}
P(f(u_t)|f_\gamma(u_t)) &= \mathcal{N}\left( \frac{u_{t_{i+1}}-u_{t_i}}{h}, \sigma_\gamma \right); \\
\label{eqn:sigGammaEulerDefn}
\sigma_\gamma &:= c h^2\end{aligned}$$ for some constant $c$. For an inverse problem, the value of $c$ is unknown and must be estimated. In this paper, $c=1$ is used since the dropout rate, $r$, must also be adjusted to match the variance of the actual training data. As such, the variance induced by $c$ can be implicitly controlled by adjusting $r$. It was found that it is still useful, for a numerical problem, to include the error due to $h^2$ in $\sigma_\gamma$.
The derivative approximation in \[eqn:derivativeApproxFirstOrder\] could be replaced by some other suitable approximation, such as a higher-order Taylor series based approximation, as described in any standard reference on numerical methods [@hamming2012numerical].
### Dropout training objective assuming an Euler integral approximation
The Euler approximation can be inserted into the loss functional in \[eqn:jForPractical\] to find a computable training objective. As the training data is assumed to be sampled at $N$ discrete points, the integral over each $u_t$ can be approximated by a finite integral at each of the $N-1$ points at which $f_\gamma$ is computed. The training objectives becomes $$\begin{aligned}
\label{eqn:jWithAllApprox}
J(T,\theta,d) &\approx \frac{1}{N-1} \sum_{i=1}^{N-1} \frac{1}{B}\sum_{b=1}^B \norm*{ f^b_\gamma(u_t) - f_\theta(u_{t_i}|d) }_2^2 \\
\text{ for } f^b_\gamma(u_t) &\sim \mathcal{N}\left( \frac{u_{t_{i+1}}-u_{t_i} }{h}, \sigma_\gamma \right).\end{aligned}$$
Then, as in \[eqn:dropoutNormalPosterior\], $\theta^\ast$ can be used as a maximum likelihood estimator for computing the posterior distribution $P(\theta|T)$. The application of the posterior distribution to forward model prediction is discussed in the next section.
The posterior distribution in \[eqn:dropoutODENormalPosterior\], with $\theta^\ast$ found using \[eqn:jWithAllApprox\] is the solution to the dynamical system inverse problem. The full training procedure to calculate the posterior over weight space is summarised in \[alg:trainingBayesNet\].
Note the training data is assumed to be sampled at evenly spaced intervals, $h$. Also note that a first-order Taylor series derivative approximation has been used. The algorithm presented could be modified to make use of irregular time discretisation or different derivative approximations methods by altering the assumptions made in this section.
$(f_\theta,r,T=\lbrace t_i, u_{t_i} \rbrace_{i=1}^N)$\
Approximate derivatives using $S: = \left\lbrace \left( t_i, \frac{u_{t_{i+1}}-u_{t_i}}{h} \right) \right\rbrace_{i=1}^N$ and compute $\sigma_\gamma$ (derivative approximation error) as defined in \[eqn:derivativeApproxFirstOrder,eqn:sigGammaEulerDefnPart1\].\
Predicting future states given the posterior predictive distribution
====================================================================
The posterior predictive model, $P(\hat{f}(u_t)|T)$ in \[eqn:dropoutODENormalPosterior\] can be used for forward prediction. That is, the learnt model can be used to compute an a posteriori estimate for trajectories, defined in \[eqn:trajectoryProbs\]. If certain assumptions about approximation errors are made, then the estimated trajectory can be assumed to be a Gaussian process.
To achieve this, first a discrete-time state transition distribution is derived using ideas from probabilistic numerics [@Hennig2015]. Then, the posterior model for $\hat{f}$ is combined with the transition probability model to compute a posterior distribution over future ODE states.
Error induced by numerical approximation of the transition probability {#ssec:errorByNumericalApproxTransition}
----------------------------------------------------------------------
Discretisation must be introduced to the integral $\int_t^{t+s} f(u_\tau) d\tau$ to allow for numerical approximation of trajectories. Examples of ODE discretisation schemes include Euler integration and Runge-Kutta methods, see [@boyce2017elementary] for a detailed overview.
Denote the numerical approximation to the analytical ODE integral in \[eqn:detFlowMap\] by $$\begin{aligned}
\label{eqn:uHatDefn}
u_{t+s} = \hat{u}_{t+s}^f(u_t) + \epsilon &:= u_t + G(t,s,f) + \epsilon\end{aligned}$$ where $G(t,s,f)$ is some numerical approximation scheme with $$\begin{aligned}
\label{eqn:gApproxDefn}
G(t,s,f) \approx \int_{t}^{t+s} f(u_\tau)d\tau\end{aligned}$$ and $\epsilon$ represents the error of the approximation.
For a standard numerical integration method, $G(t,s,f)$ can be represented as a weighted sum of a set of values, $\lbrace f(u_{t_i}) \rbrace_{i=1}^N$ with $t \leq t_i \leq t+s$. The values of $u_{t_i}$ are termed the ‘evaluation points’ of the integration scheme. The integral approximation can then be written $$\begin{aligned}
\label{eqn:gAsLinearSum}
G(t,s,f) = G\left(\lbrace f(u_{t_i}) \rbrace_{i=1}^N \right) = \alpha_0 + \sum_{i=1}^N \alpha_i f(u_{t_i})\end{aligned}$$ The factors, $\alpha_i$, and the evaluation points, $u_{t_i}$, depend on the particular numerical integration scheme used.
The transition probability can be computed, given the probability to sample some value of $\hat{u}^f_{t+s}(u_t)$, by marginalisation: $$\begin{aligned}
\label{eqn:marginalisingOutG}
P(u_{t+s}|u_t) &= \int P(u_{t+s}|\hat{u}^{f}_{t+s}(u_t))dP(\hat{u}^{f}_{t+s}(u_t)).\end{aligned}$$
If $f$ is known, \[eqn:marginalisingOutG\] is deterministic. Later, $f$ will be replaced by the random-valued posterior approximation $\hat{f}$.
Gaussian representation of the integral approximation error {#ssec:twoNormIntegralError}
-----------------------------------------------------------
Combining the techniques in [@Ramsay_2007] and [@Stuart_2010], we assume that the likelihood of a numerically approximated state transition can be represented with a Gibbs measure. The error term, $\epsilon$, in \[eqn:uHatDefn\] is assumed to be independent Gaussian noise: $$\begin{aligned}
\label{eqn:gaussianNoiseForG}
\epsilon \sim \mathcal{N}(0,\sigma_\epsilon).\end{aligned}$$ In \[eqn:gaussianNoiseForG\], the standard deviation $\sigma_\epsilon$ is estimated from the amount of error induced by the choice of $G\left(\lbrace f(u_{t_i}) \rbrace_{i=1}^N \right)$. This is discussed in \[sssec:approxErrorIn2Norm\].
Then the probability that the approximation $\hat{u}^f_{t+s}(u_t)$ is equal to the latent value of $u_{t+s}$ is $$\begin{aligned}
P(u_{t+s}|\hat{u}_{t+s}^f) = \mathcal{N}\left(u_t + G(t,s,f),\sigma_\epsilon \right).\end{aligned}$$
### Approximation error of the assumed Gaussian error representation {#sssec:approxErrorIn2Norm}
The error term in \[eqn:uHatDefn\] must, in practical cases, be assumed. Knowledge of the true errors cannot be obtained. The Gaussian assumption in \[eqn:gaussianNoiseForG\] is made for mathematical convenience. As $\sigma_\epsilon$ controls the variance of the estimated probability over outcomes, it is reasonable to assume that the scale of the variance is on the order of the error of $G(t,s,f)$.
For ODE discretisation schemes, the error is typically expressed in terms of some parameter, $h$, which represents the finest time scale resolution used to approximate $\int_t^{t+s} f(u_\tau) d\tau$. For an extended discussion see [@boyce2017elementary].
Typically, discretisation errors are polynomial in $h$ and can be written as $\mathcal{O}(h^m)$ for some $m$. In this case, $\beta$ in the numerical approximation to the transition probability of a dynamical system can be estimated by setting $$\begin{aligned}
\label{eqn:trueGivenApproxAssumption}
\sigma_\epsilon &\approx h^m.\end{aligned}$$
Smaller time intervals will result in more accurate approximations, as $\sigma_\epsilon$ will be proportional to $t_{i+1}-t_i$ for a standard ODE time integral discretisation method [@boyce2017elementary].
More rigorous error analysis could potentially be used to derive a more exact error estimate. For numerical analyses, these error estimates may be adjusted to ensure that the assumptions are reasonable in the sense that inferred models predict observational data to have high probability.
Gaussian posterior predictive distribution for the numerical integration function {#ssec:twoNormIntegralError2}
---------------------------------------------------------------------------------
Given the Gaussian posterior predictive model inverse problem solution, $P(\hat{f}(u_t)|T)$, a Gaussian process over trajectories can be derived by replacing $\hat{u}^f_{t+s}(u_t)$ with $\hat{u}^{\hat{f}}_{t+s}(u_t)$. From \[eqn:dropoutODENormalPosterior\], the distribution at each integral evaluation point is $$\begin{aligned}
P(\hat{f}(u_{t_i})|T) = \mathcal{N}\left(\mu_{\theta^{\ast}}(u_{t_i}),\sigma_{\theta^{\ast}}(u_{t_i}) \right).\end{aligned}$$
As each $\hat{f}(u_{t_i})$ is a random variable, $\hat{u}_{t+s}^{\hat{f}}(u_t)$ can be expressed as a random variable $$\begin{aligned}
\hat{u}_{t+s}^{\hat{f}}(u_t) &= u_t + G(t,s,\hat{f}) + \epsilon \\
\label{eqn:uHatIsGaussian}
&= u_t + \alpha_0 + \sum \alpha_i \hat{f}(u_{t_i}) + \epsilon.\end{aligned}$$
Each $\hat{f}(u_{t_i})$ is a Gaussian random variable and $\epsilon$ is assumed to be a Gaussian random variable. As the linear combination of Gaussian random variables is also Gaussian [@kroese2013statistical], $P(u_{t+s}|u_t)$ is also normally distributed. The explicit distribution is not required, as it will be approximated by sampling.
Gaussian trajectory prediction model {#ssec:fullGaussianTrajectory}
------------------------------------
Following \[eqn:uHatIsGaussian\], $\hat{u}_{t+s}^{\hat{f}}(u_t)$ is a Gaussian random variable. Then a full trajectory estimate, as in \[eqn:trajectoryProbs\], can be computed from times $t$ to $t+s$ at $\lbrace t_i \rbrace_{i=1}^N$ using $$\begin{aligned}
\label{eqn:inverseModelPredTrajectory}
P_{t_n}(u_{t_{n}}) &= \left[\prod^{n-1}_{j=1} P(u_{t_{j+1}}|u_{t_{j}}) \right] P_{t_1}(u_{t_1}) ~\quad\text{ for } n \in [1,N].\end{aligned}$$
As \[eqn:inverseModelPredTrajectory\] is simply the product of Gaussians, the probability distribution at each $t_i$ in the trajectory, $P_{t_{i}}(u_{t_{i}})$, will also be Gaussian (see §4 in [@murphy2012machine]). This means that the full trajectory can be described by the mean and standard deviation of $u_{t_i}$ at each $t_i$. These statistics can be collected by sampling.
### Simplified version of the full trajectory estimation algorithm {#sssec:roughAlgo}
Probabilistic trajectory estimates can be obtained roughly as follows:
1. For $M$ iterations:
1. Sample an initial condition from $P_{t_1}(u_{t_1})$.
2. Use a standard ODE solver to predict $u$ at all times in $\lbrace t_i \rbrace_{i=1}^N$. To calculate the time derivative at $u(t)$ for the ODE solver, sample $\hat{f}$ from \[eqn:dropoutODENormalPosterior\] (the dropout-enabled network for $f$ with weights $\theta^\ast$).
2. Compute the sample mean and standard deviation for each $u_{t_i}$ for each of the $M$ samples.
The first and second moment sample statistics for each $u_{t_i}$ can be collected over the $M$ trajectories so $$\begin{aligned}
\mu(u_{t_n}) &= \frac{1}{M} \sum_{j=1}^M u_j \quad \text{ for } u_j \sim P(u_{t_{n}}), \\
\sigma(u_{t_n}) &= \sqrt{\frac{1}{M-1} \left( \sum_{j=1}^M u_j - \mu(u_{t_n}) \right) } \quad \text{ for } u_j \sim P_{t_{n}}(u_{t_{n}}).\end{aligned}$$
Confidence intervals for the values of $u(t_i)$ can be found using these statistics at each $t_i$ by using the same method described for $g(x)$ in \[eqn:gConfUp,eqn:gConfDown\]: $$\begin{aligned}
u^U_{t_n} &= \mu(u_{t_n}) + c \sigma(u_{t_n}); \\
u^L_{t_n} &= \mu(u_{t_n}) - c \sigma(u_{t_n});\end{aligned}$$ where $c$ is defined using the standard Gaussian confidence levels [@kroese2013statistical].
### Finite time blow up regularisation {#sssec:finiteBlowUpTime}
The full algorithm in \[alg:samplingBayesNet\] improves on the rough outline in \[sssec:roughAlgo\] by including a method to limit the effect of the severe numerical instabilities. Samples from $\hat{f}$ can easily lead to sudden blow ups (values of $u_t = \pm \infty$) which prevent simulation of full trajectories. These trajectories can, depending on the dynamical system being modelled, be considered spurious. To eliminate the effect of these instabilities on the computed trajectory statistics for the $M$ trajectory samples, trajectories that blow up are discarded from computations. In this paper, this is referred to as ‘finite blow up time regularisation’. It is, in effect, an implicit prior introduced over the posterior predictive distribution for $\hat{f}$. This implicit prior says that values of $\hat{f}$ that produce go to infinity in finite time have probability zero. The effect of this assumption is studied numerically in \[sec:sprottBSystem\].
$(f_\theta,r,u_t,s,M,N,c)$
Set $h = \frac{|(t+s)-t|}{N}$. Generate output times $\lbrace t_i \rbrace_{i=1}^N$ where $t_i = t+ih$.\
Generate running mean storage set $\mu:= \lbrace 0 \rbrace_{i=1}^N$. Define $\mu(t_i) := \mu_i$.\
Generate running standard deviation storage set $\sigma:= \lbrace 0 \rbrace_{i=1}^N$.\
Initialise $\sigma_\epsilon = h^m$ where $m$ defined by the ODE solver used. Initialise counter: $j \leftarrow 0$.\
Set $u^U_{t_i} = \mu(t_i) + c\sigma(t_i)$ and $u^L_{t_i} = \mu(t_i) - c\sigma(t_i)$.
Numerical example: Sprott B System {#sec:sprottBSystem}
==================================
System overview
---------------
To demonstrate the techniques outlined in the earlier sections of this paper, the classic ‘Sprott B system’ [@Sprott1994] was analysed. This highly chaotic system is a map from $\mathbb{R}^3$ to $\mathbb{R}^3$ defined by $$\begin{aligned}
\frac{dx}{dt} = yz; \quad
\frac{dy}{dt} = x-y; \quad
\frac{dz}{dt} = 1-xy.\end{aligned}$$
Views of an example trajectory of the system are shown in \[fig:sprottExample3d\] and \[fig:sprottExampleTimeSeries\]. The data was generated using the initial conditions $x=y=0.1$ and $z=-0.1$. An $x,y,z$ trajectory plot is shown in \[fig:sprottExample3d\]. The same data is presented as a time series in \[fig:sprottExampleTimeSeries\]. These trajectory plots demonstrate that the system oscillates around an attractor state, with difficult to predict fluctuations away from a central point.
Note that the simulated trajectory, as well as all other simulations in this section, were generated using the SciPy solve\_ivp method [@scipy2001] with the ‘RK45’ algorithm (variable 4th-5th order Runge-Kutta, see [@boyce2017elementary]).
![Sprott B system example trajectory in $\mathbb{R}^3$. Similar colours show points closer in time.[]{data-label="fig:sprottExample3d"}](figure1){width="100.00000%"}
![Example from \[fig:sprottExample3d\] shown as a time series.[]{data-label="fig:sprottExampleTimeSeries"}](figure2){width="100.00000%"}
Inverse problem task
--------------------
The inverse problem task analysed in this section is to recover the Sprott B system from observations of the trajectory. The given training data is described in \[ssec:sprottTrainingData\]. After training the compute graph architecture described in \[ssec:computeGraphArch\], forward time predictions made using the trained model are compared to a test set.
The analysis explores several facets of the theoretical developments presented earlier in this paper. The feasibility of the techniques is tested by demonstrating that it is possible to predict the behaviour of a particular dynamical system. The prediction time periods are much longer than the training data sampling period. It is shown that the prediction uncertainty can be usefully quantified. The effect of data spacing was empirically assessed. It was expected that if derivatives are estimated as described in \[sec:inverseSolution\], the model fit should be worse if larger spacings are used. This was observed. The effect of the dropout rate was also examined and found to impact of the accuracy of the estimated trajectory confidence bounds.
Training data {#ssec:sprottTrainingData}
-------------
Training data trajectories for the inverse problem are shown in \[fig:sprottTrainingData\] and were generated using initial values: $$x = y = z = 1.$$
The training data runs from times $0$ to $10$. Derivatives were estimated from the training data using a basic finite difference scheme as per the description in \[sec:inverseSolution\]. The effect of varying the parameter $h$ was tested and is described in \[ssec:numericalResults\].
![Sprott B inverse problem training data.[]{data-label="fig:sprottTrainingData"}](figure3){width="100.00000%"}
Compute Graph architecture {#ssec:computeGraphArch}
--------------------------
For the purposes of this task, the latent system model is assumed to be a polynomial function of the observable variables. As such, an appropriate choice of architecture is a parametric polynomial kernel (see [@Green2019]). The parametric polynomial kernel allows for a polynomial structure to be assumed, even if the particular polynomial is unknown. For this demonstration case, it is assumed known a priori that the solution is a at most a second-order polynomial, although Bayesian model selection could be carried out to find a suitable polynomial order [@murphy2012machine]. Bayesian architecture search can be approximated by methods such as NEAT or others detailed in [@stanley2019designing]. This would add significant computational overhead and would cloud the main point of this paper.
A parametric polynomial kernel mapping $x \in \mathbb{R}^a$ to $f_\theta(x) \in \mathbb{R}^b$ is a function of the form $$\begin{aligned}
\label{eqn:paraPolyKernel}
f_\theta(x) = W_2 \left[ \circ^m (W_1 x + B_1) \right] + B_2\end{aligned}$$ where $W_1 \in \mathbb{R}^{k\times a}$, $B_1 \in \mathbb{R}^{k}$, $W_2 \in \mathbb{R}^{b\times k}$ and $B_2 \in \mathbb{R}^{b}$ for $b \in \mathbb{N}$.
Note that $\circ^m$ is defined here to be the Hadamard (elementwise) product of the argument repeated $m$ times, $$\begin{aligned}
\circ^m(a) &:= a \circ a \circ \cdots \circ a \\
&:= a \circ^{m-1}(a); \\
\circ^1(a) &:= a \circ a; \\
\circ^0(a) &:= a\end{aligned}$$ where the Hadamard product for matrices [@horn2012matrix] is defined by $$\begin{aligned}
(A\circ B)_{ij} := (A)_{ij}(B)_{ij}.\end{aligned}$$
The dimension $k$ in \[eqn:paraPolyKernel\] can be any natural number. Increasing the size of $k$ increases the dimension of the hidden representation space for $f_\theta(x)$. The basis of this hidden space is given by the polynomials implicitly represented in \[eqn:paraPolyKernel\].
To generate a Bayesian network of the form in [@Gal2015] for a parametric polynomial kernel, a dropout layer must be incorporated. The dropout mask, for the numerical analysis in this paper, is introduced after computing $W_1x + B_1$ but before computing $\circ^m (W_1 x + B_1)$. Dropout also acts as a regulariser [@dropout2014]: by regularising the layer directly before the polynomial kernel layer, during training the network will attempt to minimise the number of polynomial terms that are actually represented by the network.
The network architecture used for the results presented in this paper was $$\begin{aligned}
f_\theta &= W_2 \left[ \circ^1 (d \circ(W_1 x + B_1)) \right] +B_2; \\
d &\sim D(r);\end{aligned}$$ where $W_1 \in \mathbb{R}^{10\times 3}$, $B_1 \in \mathbb{R}^{10}$, $W_2 \in \mathbb{R}^{3\times 10}$ and $B_2 \in \mathbb{R}^{3}$.
The choice of $k=10$ was found by trial and error. Such a search could be automated, but this is left for future work.
The vector $d \in \mathbb{R}^{10}$ is a dropout mask, sampled from $D(r)$, as described in \[ssec:dropoutReg\]. The effect of the choice of $r$ on the solution was tested by comparing results obtained using different values of $r$. These results are shown in \[sssec:effectOfChangingR\].
Compute Graph training schedule
-------------------------------
Training for all models was conducted in accordance with the procedure outlined in \[alg:trainingBayesNet\]. Training batches, $R$, were sampled $1000$ times. For each $R$ sample, the weights $\theta$ were updated using Stochastic Gradient Descent and the Adam optimiser [@Kingma2014] as follows:
1. Train for $10$ steps with learning rate $\alpha = 0.01$.
2. Train for $100$ steps with learning rate $\alpha = 0.001$.
This training schedule allows for a weak ‘simulated annealing’ effect [@kirkpatrick1983optimization] by first training at a higher learning rate to get a rough solution, before fine tuning the result using a smaller learning rate. The training schedule was developed from a combination of trial and error, experience and intuition.
Results {#ssec:numericalResults}
-------
### Overview
The second order parametric polynomial kernel with Bayesian parameter approximation was able to estimate, and bound, the future time behaviour of the Sprott B system over time periods much longer than the training data sampling frequency, $h$. As long as the prediction time exceeds the data sample acquisition time, the inverse problem solution could be successfully utilised as a filtering transition model (for example in a particle filter). Further, the probabilistic confidence interval estimates were successfully able to bound the future behaviour of the system.
To test the performance of the predictions, a test case data set with initial condition $$\begin{aligned}
x=y=z=-1.0\end{aligned}$$ was generated. All comparisons were made by investigating $x(t)$. Since the parameters $x,y,z$ are tightly coupled, the comparison results for $x(t)$ can be expected to be similar to those for $y$ and $z$. The error for the ‘RK45’ method used to generate sample traces is between $\mathcal{O}(h^4)$ and $\mathcal{O}(h^5)$ [@boyce2017elementary]. As such, a value of $\sigma_\epsilon = h^4$ (with reference to \[alg:samplingBayesNet\]) was used.
Example prediction outputs are shown in \[fig:sprottNiceTestCase\]. Three networks were used to estimate $95\%$ confidence intervals. Each network was trained slightly differently, sampling the training data at a rate of one of: $$\begin{aligned}
\label{eqn:hDefns250}
h_{250} &:= \frac{1}{250};\\
\label{eqn:hDefns500}
h_{500} &:= \frac{1}{500};\\
\label{eqn:hDefns1000}
h_{1000} &:= \frac{1}{1000}.\end{aligned}$$ All networks had a fixed dropout rate, $r = 0.25$. Confidence intervals were estimated with $1000$ sampled traces ($M=1000$ for \[alg:samplingBayesNet\]).
![Sprott B test case, $95\%$ confidence interval predictions for variable $x(t)$ after training. The test case initial condition is different from the training data. All predictions were generated with a fixed dropout rate, $r = 0.25$. $h_{i}$ refers to a network trained with a sample rate $h = \frac{1}{i}$, as per \[eqn:hDefns250,eqn:hDefns500,eqn:hDefns1000\]. Lines for each $h_i$ indicate mean estimates and $95\%$ upper and lower confidence interval bounds. []{data-label="fig:sprottNiceTestCase"}](figure4){width="100.00000%"}
Using the proposed method, it is possible to make predictions over long time periods. This is shown \[fig:sprottLongTestCase\]. The prediction for $h_{500}$ in \[fig:sprottLongTestCase\] is the same as that in \[fig:sprottNiceTestCase\], extended from $t=10$ to $t=100$ time units. However, this long time prediction comes with a number of caveats.
Although the prediction is tightly bounded around the test data over a short time, at longer times the estimate predicts only the range (but not the specific values) of the test data. Due to chaotic mixing, the estimate can track only the size of the stable manifold of the system. If there was no stable manifold present in the system analysed, one would anticipate that the confidence intervals would become increasingly wide (in relation to the Lyapanov exponents, see [@meiss2017differential]). A more detailed study of this behaviour is left for future work.
Further, the confidence bounds become jagged. This could be partly resolved, at increasing computational cost, by increasing the number of traces used to build the confidence interval predictors ($M$ in \[alg:samplingBayesNet\]). For short time periods, a small $M$ is reasonable, but over long time periods the computational expense increases.
![Prediction in \[fig:sprottNiceTestCase\] for $h_{500}$ extended from a maximum time of $t=10$ to $t=100$. The estimated confidence interval approximately matches the size of the stable manifold of the system.[]{data-label="fig:sprottLongTestCase"}](figure5){width="100.00000%"}
The other source of the jagged confidence interval predictions is the ‘finite blow up time’ regularisation (see \[sssec:finiteBlowUpTime\]) used to exclude estimated trajectories that go to positive or negative infinity in finite time. Having to discard a large number of trajectories would suggests that the posterior over the model weights is a poor match for the actual posterior distribution as strong regularisation is required.
If the trained network produces a larger number of trajectories which must be discarded, the predictions will become increasingly jagged, as shown in \[fig:sprottLongTestCaseSpikes\]. The prediction for $h_{1000}$, with $r=0.25$ and $M=1000$, in \[fig:sprottLongTestCaseSpikes\] is different from $h_{1000}$ in \[fig:sprottNiceTestCase\] (a new random seed was used to generate the network and training data randomisation). The long term prediction is quite jagged, which could be resolved with some sort of moving window smoothing (as in [@kroese2013statistical]) or outlier removal at the cost of introducing some time lag in the predictions. The overall prediction captures the stable manifold of the Sprott B system over long time periods, but the quality of the prediction is worse than over short time periods.
![Effect of poor ‘finite blow up time regularisation’ on long term time prediction. $h_{1000}$ is different from that in \[fig:sprottNiceTestCase\] (this prediction was trained using a different random seed). Jagged confidence intervals may be predicted if the learnt model is likely to blow up.[]{data-label="fig:sprottLongTestCaseSpikes"}](figure6){width="100.00000%"}
The remainder of this section investigates other impacts of varying $h$ and $r$ on prediction.
### Effect of changing $h$ given fixed dropout rate
The effect of changing $h$ with a fixed dropout rate is shown in \[fig:sprottNiceTestCase\]. Increasing the time resolution (decreasing $h$) improves the estimate in the sense that the confidence intervals more tightly bound the test case data. This, however, represents an ideal case. In \[fig:sprottBadTestCase\], the predictions for the following $h$ values are shown: $$\begin{aligned}
\label{eqn:hBadDefns50}
h_{50} &= \frac{1}{50}; \\
\label{eqn:hBadDefns100}
h_{100} &= \frac{1}{100}; \\
\label{eqn:hBadDefns1000}
h_{1000} &= \frac{1}{1000}; \\
\label{eqn:hBadDefns5000}
h_{5000} &= \frac{1}{5000};\end{aligned}$$ where the estimate for $h_{1000}$ is the same as the estimate in \[fig:sprottNiceTestCase\]. At $h_{50}$ and $h_{100}$, the confidence intervals fail to capture the test case data after around $t = 6$. This indicates, along with \[fig:sprottNiceTestCase\], that improving the time resolution can improve the time over which accurate predictions can be made. However, there is a limit to this accuracy. The error bounds for $h_{5000}$ start to become overly broad. It is possible that the poor prediction of $h_{5000}$ is due to numerical precision errors. This suggests that the maximum time resolution possible should be used, up to some limit at which accuracy begins to decrease. The quality of the prediction must be verified with a test case, separate from the training data.
![Effect of setting $h$ too low or too high on confidence interval prediction bounds for Sprott B system. $h_{i}$ refers to a network trained with a sample rate $h = \frac{1}{i}$, as per \[eqn:hBadDefns50,eqn:hBadDefns100,eqn:hBadDefns1000,eqn:hBadDefns5000\]. The estimate for $h_{1000}$ is the same as in \[fig:sprottNiceTestCase\].[]{data-label="fig:sprottBadTestCase"}](figure7){width="100.00000%"}
### Effect of changing dropout rate given fixed $h$ {#sssec:effectOfChangingR}
The dropout rate, $r$, was varied for a fixed $h = \frac{1}{500}$ to investigate the impact on predictive performance. It was anticipated that small values of $r$ (low probability to retain a network weight) will estimate wide confidence intervals. Conversely, high $r$ should indicate higher confidence, and therefore more narrow confidence interval bands. Three values of $r$ were tested: $$\begin{aligned}
r_{50} &:= 0.5; \\
r_{25} &:= 0.25; \\
r_{5} &:= 0.05.\end{aligned}$$
The results of the numerical analysis are shown in \[fig:sprottLongTestChangingRFar\] and \[fig:sprottLongTestChangingRZoom\]. While all confidence interval predictors perform well for a short time period, the behaviour is quite different over long time periods. The result for $r_{50}$ is emphasised in \[fig:sprottLongTestChangingRFar\]. This $r$ value is too high. Although the test data is always bounded by the confidence intervals, the intervals are very wide. The predictive performance can be improved by decreasing $r$. The results for $r_{25}$ and $r_{5}$ are emphasised in \[fig:sprottLongTestChangingRZoom\]. The result for $r_{25}$ is the same as that shown in \[fig:sprottLongTestCase\] for $h_{500}$. The $r_{25}$ result bounds the data well, as discussed earlier. The result for $r_{5}$ is somewhat overconfident, missing peaks in the test case data until around $t \approx 60$. By this time, the width of the $r_{5}$ estimator is quite similar to the $r_{25}$ confidence intervals.
![Sprott B test case - effect of changing dropout rate $r$ given fixed $h = \frac{1}{500}$. Zoomed out view emphasising $r_{50}$ results. The wide confidence intervals predicted capture the test data, but have poor predictive performance.[]{data-label="fig:sprottLongTestChangingRFar"}](figure8){width="100.00000%"}
![Zoomed in view of \[fig:sprottLongTestChangingRFar\] emphasising $r_{25}$ and $r_{5}$. The $r_{25}$ confidence intervals bound the data well. The $r_{5}$ interval is overconfident and misses some peaks of the test data.[]{data-label="fig:sprottLongTestChangingRZoom"}](figure9){width="100.00000%"}
Discussion
----------
The results demonstrate that Bayesian neural network Gaussian process approximation methods can be applied to learning chaotic time series data. The quality of the model predictions have been shown to be dependent on both the model parameters (for example $r$) and the quality of the available training data (simulated by altering $h$). For real problems based on observational data, the data sampling rate may be fixed and derivative estimates with sufficiently small $h$ may not be available.
Poor data fits may occur, even when careful parameter choices have been made. It is crucial that, were this process applied to real data, a set of test data is used to verify the derived predictive model. With additional computational power, model parameters like $r$ could be found by an optimisation method, such as a grid search or other more powerful techniques [@Bergstra2012]. The choice of network architecture was fixed for this example, but could also be optimised for. However, architecture search is also a difficult problem and may strongly influence the results. Search techniques, such as [@stanley2019designing], may be useful in more complex applications.
Conclusions {#sec:conclusions}
===========
This paper demonstrated a technique for the application of Bayesian artificial neural network Gaussian process approximations to inverse problems for dynamical systems. A low dimensional, very chaotic system was analysed as a test case. Analysis of higher dimensional systems is left to future work.
In the case tested, the future behaviour of the system was able to be predicted for a time period far longer than the data sampling rate. This means that the method presented could be used to update a transition model for a filtering task. In particular, the method presented could be used as an adaptive transition model, reacting to the latest observed data. Such an approach would involve less feature engineering compared to methods based on filter banks.
The method presented aims to reduce the required amount of a priori knowledge of the ODE functional form that must be injected into the inverse problem solution. However, the compute graph architecture can be considered to be a sort of implicit prior over classes of ODEs. The network architectures used in this paper were found to work well for the numerical analyses presented, but in general some sort of architecture search must be performed. Methods for architecture search are an open area of research. The traditional method, experience-based trial and error, was used. The methods in this paper could be supplemented with symbolic regression and neuroevolution methods [@Schmidt2009; @stanley2019designing; @Real2017; @Real2018]. Although these methods can be effective, they are very computationally intensive. At the very least, the parametric polynomial kernel method demonstrated within this paper should be well suited to polynomial type dynamical systems. Further exploration of the parameterisation choices and compute graphs that work well for different use cases would be an interesting direction for future work.
Using probability theory and probabilistic numerics, the demonstrated method carefully tracks sources of noise. As such, confidence intervals that correctly bound the future time behaviour of the system in question can be predicted. Reasoning about discretisation errors from a probabilistic perspective enables the inverse problem task to be written in terms of probability theory, treating both analytical and numerical methods in a unified manner. The optimisation problem can be understood as a form of approximate Bayesian inference. A useful direction for future work in this area would be to incorporate more information theoretic reasoning. Such an analysis may yield further computational benefits over the approximate Gaussian process Bayesian updating model used in this paper.
Acknowledgements
================
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme, grant agreement No 757254 (SINGULARITY) and a Lloyds Register Foundation grant for the Data-Centric Engineering programme at the Alan Turing Institute.
[10]{}
M. Abadi, A. Agarwal, P. Barham, et al. : [L]{}arge-[S]{}cale [M]{}achine [L]{}earning on [H]{}eterogeneous [S]{}ystems, 2015. Software available from tensorflow.org.
J. Bergstra and Y. Bengio. Random search for hyper-parameter optimization. , 13:281–305, February 2012.
C. M. Bishop. . Advanced Texts in Econometrics. Clarendon Press, 1995.
W. E. Boyce, R. C. DiPrima, and D. B. Meade. . Wiley, 2017.
Y. Chauvin and D.E. Rumelhart. . Developments in Connectionist Theory Series. Taylor & Francis, 2013.
A. C. Damianou and N. D. Lawrence. Deep [G]{}aussian processes. In [*Proceedings of the 16th International Conference on Artificial Intelligence and Statistics (AISTATS)*]{}, Journal of Machine Learning Research, 2013.
Y. Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In [*Proceedings of the 33rd International Conference on Machine Learning*]{}, 2016.
H. O. Georgii. . De Gruyter studies in mathematics. De Gruyter, 2011.
I. Goodfellow, Y. Bengio, and A. Courville. . MIT Press, 2016.
D. K. E. Green and F. Rindler. Model inference for ordinary differential equations by parametric polynomial kernel regression. In G. Stefanou M. Papadrakakis, V. Papadopoulos, editor, [ *Proceedings of the 3rd ECCOMAS Thematic Conference on Uncertainty Quantification in Computational Sciences and Engineering*]{}, 2019.
R. Hamming. . Dover Books on Mathematics. Dover Publications, 2012.
P. Hennig, M. A. Osborne, and M. Girolami. Probabilistic numerics and uncertainty in computations. , 471(2179):20150142, 2015.
H. Holden, B. Oksendal, J. Uboe, and T. Zhang. . Probability and Its Applications. Birkh[ä]{}user Boston, 2013.
R. A. Horn and C. R. Johnson. . Cambridge University Press, 2012.
A. Iserles. . A First Course in the Numerical Analysis of Differential Equations. Cambridge University Press, 2009.
E. Jones, T. Oliphant, P. Peterson, et al. : [O]{}pen source scientific tools for [Python]{}, 2001. Online.
D. P. Kingma and J. Ba. dam: [A]{} [M]{}ethod for [S]{}tochastic [O]{}ptimization. In [*Proceedings of the 3rd International Conference on Learning Representations (ICLR)*]{}, 2015.
S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi. Optimization by simulated annealing. , 220(4598):671–680, 1983.
D. P. Kroese and J. C. C. Chan. . SpringerLink : B[ü]{}cher. Springer New York, 2013.
J. D. Meiss. . Mathematical Modeling and Computation. SIAM, Society for Industrial and Applied Mathematics, 2017.
P. D. Moral and S. Penev. . Chapman & Hall/CRC Texts in Statistical Science. CRC Press, 2017.
K. P. Murphy. . Adaptive Computation and Machine Learning. [MIT]{} Press, 2012.
L. B. Rall. . Springer-Verlag Berlin Heidelberg, 1981.
J. O. Ramsay, G. Hooker, D. Campbell, and J. Cao. Parameter estimation for differential equations: a generalized smoothing approach. , 69(5):741–796, oct 2007.
E. Real, A. Aggarwal, Y. Huang, and Q. V. Le. Regularized evolution for image classifier architecture search. In [*Proceedings of the 33rd AAAI Conference on Artificial Intelligence*]{}, 2019.
E. Real, S. Moore, A. Selle, S. Saxena, Y. L. Suematsu, J. Tan, Q. Le, and A. Kurakin. Large-scale evolution of image classifiers. In [*Proceedings of the 34th International Conference on Machine Learning*]{}, 2017.
J. Schmidhuber. Deep learning in neural networks: An overview. , 61:85–117, 2015.
M. Schmidt and H. Lipson. Distilling free-form natural laws from experimental data. 324:81–85, 2009.
J. C. Sprott. Some simple chaotic flows. , 50(2):R647–R650, aug 1994.
N. Srivastava, G. Hinton, A. Krizhevsky I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. , 15:1929–1958, June 2014.
K. O. Stanley, J. Clune, J. Lehman, and R. Miikkulainen. Designing neural networks through neuroevolution. , 1(1):24–35, 2019.
S. Strogatz. . CRC Press, 2018.
A. M. Stuart. Inverse problems: A [B]{}ayesian perspective. , 19:451–559, may 2010.
[^1]: `[email protected]`
[^2]: `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
**Purpose**: This research investigates controversial online marketing techniques that involve buying hundreds or even thousands of upvotes, likes, comments, etc.
**Methodology**: Observation and categorization of 7,426 “campaigns” posted on the crowdsourcing platform microworkers.com over a 365 day (i.e., yearlong) period were conducted. Hypotheses about the mechanics and effectiveness of these campaigns were established and evaluated.
**Findings**: The campaigns contained a combined 1,856,316 microtasks, 89.7% of which were connected to online promotion. Techniques for search engine manipulation, comment–generating in the scale of tens of thousands, online vote manipulation, mass account creation, methods for covering tracks were discovered. The article presents an assessment of the effectiveness of such campaigns as well as various security challenges created by these campaigns.
**Research limitations**: The observed campaigns form only a small portion of the overall activity. This is due to invite-only campaigns and the presence of alternative, unobservable platforms.
**Practical implications**: The findings of this article could be input for detecting and avoiding such online campaigns.
**Social implications**: The findings show that in some conditions tremendous levels of manipulation of an online discourse can be achieved with a limited budget.
**Originality**: While there is related work on “follower factories” and “click/troll farms”, those entities offer complete “solutions” and their techniques are rather opaque. By investigating a crowdsourcing platform, this research unveils the underlying mechanics and organization of such campaigns. The research is based on a uniquely large number of observations. Small, cheap campaigns, the manipulation of less significant platforms is also included, while the related work tends to focus on mass, politically motivated efforts. **Keywords**: Crowdsourcing, Social media, Electronic word-of-mouth, facebook, twitter, youtube, reddit
author:
- |
Mihály Héder\
MTA SZTAKI\
[email protected]\
bibliography:
- 'OpMicro.bib'
title: '**A black market for upvotes and likes**'
---
Introduction
============
This article analyses marketing campaigns that have been executed through the hiring of *freelancers* or “*microworkers*” to complete short, menial tasks called *microtasks* that usually pay less than one US dollar each, and most often only around ten cents. These tasks include watching, liking, upvoting, and “+1”-ing items on web platforms featuring social media functions, like Facebook, Twitter, Reddit, and Instagram. A job description showing what such a campaign looks is given below (an actual, observed example):
Title: Facebook Like: <REDACTED String>
Payment: USD 0.15
Number of workers accepted: 100
Job description:
WARNING: we manually review almost all of the submitted tasks. Thus, if we find that you have ignored the instructions (i.e., posting on a wrong site, using non-unique or nonsensical content), you will be permanently banned from our system. You must have 50 Facebook friends.
1. Go to <REDACTED URL1>
2. Visit the URL shown at <REDACTED URL1>
3. On that page, you will see a Facebook Like button. Click on it
4. Submit your Facebook profile URL on <REDACTED URL1> to verify that you have completed the task (make sure you have set your Facebook profile to Public View in order for your task to be verified) to get a 7 character confirmation code
(...)
Completing this task pays USD 0.15 for the microworkers. In this case, one hundred freelancers were sought to do the task.
There are specialized web platforms for the brokering of small tasks. One of them is microworkers.com, which was investigated for this article. This platform lets employers run campaigns for a fee, usually 10% of the payment made to the freelancing internet users who do the job, called microworkers. Therefore, the 100 Likes above would cost the client posting the job only USD 16.5. The platform has hundreds of thousands of microworkers [@Nguyen2014].
The platform was not specifically created for promotional or marketing purposes, and indeed data processing, survey-based research, and software testing jobs are also posted on the platform. However, as shown below, the majority of tasks can be categorized as being of a promotional nature (See Table \[ActBudg\]).
Article Structure
-----------------
The structure of the present article is as follows. After this introduction, the terminology is presented and a number of ethical questions are posed. This is followed by an outline of the related work in this field. The next chapter describes the research aims and details of the observation, as well as the limitations of the research. This is followed by the Results section, which also contains a number of separate subchapters covering various aspects of the investigation into the most important gray promotional activities, including estimates about their efficiency, security concerns, and possible counter-measures, where applicable. The Results section also describes some observed campaign-supplementing techniques. Finally, the Conclusions section offers a summary and outlines some possibilities for future work.
Why call it a black market?
===========================
In the opinion of the author of this paper buying likes, followers, votes, upvotes, retweets, etc. (the generic term for these used in this paper is social media activity) for promotional purposes is an unethical practice, therefore we use the term “black market” to describe this part of the industry. This is not to say they are necessarily breaking any rules or laws—that question is outside the scope of this paper.
This kind of activity is considered unethical for the following reasons. a) The users of a social platform are misled by a page or post having an artificially inflated number of likes, followers, etc. Normally, it is impossible for users to differentiate between paid activity and genuine activity. The conventional semantic behind a like, upvote, or follow is a statement of approval of the content in question. In other words, the microworkers are paid to lie in the sense that they are paid to pretend to like/endorse/approve content that they most likely have not truly read or watched. b) The microworker’s “employment” is arguably rather exploitative, because of the low payment and the apparent lack of any powers against the clients (see table \[Payment\]). c) Content creators and social media users who do not employ such practices are clearly disadvantaged. Finally, d) it is easy to see that if paid social media activity were more universal, it would create an unsustainable social media environment.
Besides paid social activity, there are other highly controversial campaigns based around social media. One such typical activity involves creating accounts on behalf of a client and then the microworker handing over the user name and password to the client. An example of this kind of activity is given in the following (actual, observed example):
``` {#gmailacc caption="Acquiring" gmail="" accounts="" label="gmailacc"}
Title: Gmail: Create an Account
Payment: 0.16
Number of workers accepted: 230
Job description:
Note: I will Check American Name and Profile Picture otherwise I have to decline you.
1. Go to www.fakenamegenerator.com
2. Choose proper American name
3. Go to Gmail.com
4. Create a new Gmail account using details from www.fakenamegenerator.com
5. Upload a good profile picture in Gmail
For proof:
- Give Password
- Give a Recovery Mail so that I can change it later
Note: Make sure all are perfect otherwise, I will decline your payment.
Required proof that task was finished?
1. Gmail and Username
2. Password
3. Recovery email
```
In this case, the client was able to acquire 230 Gmail accounts for a mere \$ 40.48. We might speculate that these accounts (and similarly those created for YouTube, Twitter, etc.) will be used for promotional purposes, controlled by the client, while also raising all the ethical concerns highlighted in the previous example. However, this speculation might actually be optimistic. Fake accounts like these could also be used for more sinister purposes, such as as tools for spreading fake news [@Allcott2017] for political purposes or for committing fraud.
To sum up, it seems justified to state that there is a black market for “fake” accounts and social media activity and that promotion done using this market represents an ethically gray area.
The author wants to point out that this is not to say that microworkers.com or any other platform is in itself immoral or designed to be a black market—such platforms also offer a useful venue for valid projects, like data processing at scale, acquiring subjects for survey-based or interactive online scientific research, monitoring a competitor’s public online activity, or counting objects on images - several such campaigns were observed.
Related Work
============
The marketing techniques discussed in this article are related to techniques called sock puppetry and click farming. Sock puppetry means the control of many social media accounts by one person or a small group of individuals. A report on such activity was published in The New York Times [@Caldwell2007]. The method of control is a crucial difference in these efforts. As we will see some in Section \[Sec:signup\] clients buy hundreds or thousands of accounts that they can use themselves. In this case there are technical possibilities to detect the puppetry, by noticing when a very high number of users are logging in from the same network location or use the browser client fingerprint [@Laperdrix2016].
But the method of control can also be an order from a client to a cohort of users to perform some activity (but the client itself never logs in to any of their accounts). In this case detection of the activity is much harder if the client takes some precautionary steps (see Section \[Sec:signup\]). Sometimes this is called meet puppetry [@Cook2014] referring to the fact that it involves real freelancers.
Click farming is another related term. Click farms are actual workplaces in developing countries where a large number of employees are performing short task sometimes for as low as 1000 likes for 15 USD[@Arthur2013]. In current reporting these are often called troll farms [@Smith2018].
The platform investigated by this article, microworkers.com differs greatly from a click farm as it is a completely distributed crowdsourcing tool, but some of the campaigns done here might be similar to those done by a click farm.
On the deceptiveness of such campaigns in comparison with traditional advertising (where the prospective customer is aware that it is being presented with an ad) is well described by @DelRiego2009 and @Forrest2010 in connection with then-new US Federal Trade Commission guidelines on endorsements and reviews.
This article focuses on probably way smaller market available on microworkers.com, which is, however, easily accessible for freelancers and is not specialized to gray marketing in particular. In contrast with the services that directly offer followers and social media activity (Fiverr, SeoClerks, InterTwitter, FanMeNow, LikedSocial, SocialPresence, SocializeUk, ViralMediaBoost [@DeMicheli2013]), here the client has to organize its own campaign and orchestrate the freelancers on the crowdsourcing platform. This allows for creativity and innovations in the campaign methods.
@Nguyen2014 explained the idea behind microworkers.com, founded in 2009, as well as reported its user count at the time of writing (presumably 2014). @Howe2006 also reported on microworkers.com as a crowdsourcing platform.
According to Nguyen, the platform had over 600,000 users from 190 different countries. The aim of the microworkers.com as a project was to aid brokering crowdsourcing campaigns. As the article explained,
> *In crowdsourcing platforms, there is perfect meritocracy. Especially in systems like Microworkers; age, gender, race, education, and job history do not matter, as the quality of work is all that counts; and every task is available to Users of every imaginable background. If you are capable of completing the required Microtask, you’ve got the job.*
The campaign templates on the landing page of the platform are great sources of inspiration for what could possibly be achieved through crowdsourcing: participating in market research, captioning documents and video, categorizing images, testing websites and applications, and so on. The fact that the majority of public campaigns visible on the platform are mostly employing controversial promotion techniques does not seem to be the result of the platform design or intentions.
@Hirth2011 investigated microworkers.com in order to compare it to the much better understood Amazon Mechanical Turk [@Paolacci2010; @Buhrmester2011]. They correctly identified a main difference between the portals: the payment mechanism. At the time, it was basically impossible to use MTurk without a US-based credit card, while the microworkers website allowed payments to be made with Moneybookers (called Skrill today). This helps explains why the author of this paper and possibly other non-US-based researchers first discovered microworkers. Works by @Gardlo2012 and @Crone2017 aimed to assess the usefulness of the platform for scientific purposes; and indeed, scientific projects regularly, though relatively infrequently, appeared on the platform. However, it is possible that this difference in payment methods is only one of the reasons behind the nature of the campaigns conducted on each.
Hirth et al.’s work [@Hirth2011] indicated that gray promotional campaigns were already existed as long ago as 2011: “Signup”, “Click or Search”, “Voting and Rating” were already featured as campaign categories; however, the payments offered were slightly higher than today.
The connection between social media and marketing was analyzed by @Thackeray2008 as early as 2008. Of course this work concentrated on the legitimate social media strategies firms might embrace, such as paid search results, where the brand buys a presence in the search results. As @Yang2010 explained, these are usually placed in a separate area on the results page, together with being clearly to indicate that they are paid for or they may be labeled as an ad, e.g., on Facebook (the difference between “paid” and “organic” (showing up in non-paid results) links is less emphasized in today’s search engines but remains clear). @Rutz2011 demonstrated how tying paid search results to generic search terms might increase the success of a branded paid search.
The search and engage campaigns (see section \[Sec:search\]) discussed in the present paper are different. They don’t try to increase visibility by buying paid results. These represent the dark side of search engine-based advertising, whereby they try to directly manipulate the organic links.
It was envisaged [@Zhang2013] that the identification of the key influencers on social media could be crucial for effective viral marketing—but with the gray marketing techniques presented here, influence is attempted to be created directly, albeit artificially.
It was also envisioned that customer-generated content on blogs, etc. would be crucial for promotional activities—in the present paper, campaigns seeking to manufacture legitimate-looking customer content are analyzed. In other words, these campaigns, albeit unethical, are sometimes the effective counterparts of hard-to-operate social media marketing tactics, or in other words, they are controversial shortcuts to followers and likes and brand-friendly social content.
Confessore et al.’s recent work reported on in the New York Times [@Confessore2018] covered a very similar theme to this article, but was focused mostly on Twitter and on the activities of a company called Devumi.
Aims and Methods
================
Research aims
-------------
The primary aims of the present research were to identify the different schemes employed in microworker-based gray marketing on microworkers.com and to then categorize them, attempt to uncover how they fit in a wider strategy, estimate their limits and effectiveness, and to utilize this knowledge to provide general insights into these kinds of campaigns.
Second, the scale and typical budget of these campaigns, and their relative share of the overall activity on the microworkers.com platform were also measured and reported herein.
Observation of campaigns
------------------------
This research project attempted to observe all campaigns posted on microworkers.com from 22 February 2016 to 21 February 2017. The site was checked several times a day during this period. In total, 7,426 campaigns were observed during the period. Each campaign was manually categorized and the aggregate numbers of categories (payments, number of tasks) were updated.
Categorization
--------------
The campaigns were categorized in terms of two dimensions: the related target platform (Facebook, Google Plus, SoundCloud, Twitter, etc.), and the specific activity (search and engage, like, comment, sign up, etc.). The categorization was based on the campaign title and text and proved to be quite straightforward as the names of the platforms were clearly stated in the title and represented unambiguous brand names, and as the activity to be performed was almost always explained in an itemized list in the posting. Obviously certain activities were further linked to certain platforms (like retweets can only be done through Twitter), but others, like search and engage can be done on multiple platforms.
The platform labels utilized are summarized in Table \[Platforms\]. while the activity labels are summarized in Table \[Activities\].
[| p[6.5cm]{} | p[6.5cm]{} |]{}\
**Ali** (Alibaba, AliExpresS) & **Instagram**\
**Amazon** & **Bing**\
**MBS**: (microblog instant share): blogger.com, Pinterest, Digg, Tumblr, 9gag or other blogs or quick sharing platforms & **Mix**: (SoundCloud, Mixtape, datpiff)\
**Browser add-on** (e.g., Chrome extensions) & **RDT** (a traffic generator site, the name of which is redacted from this article)\
**Other**: (500px, Wordpress.com, Snapchat, Skillshare, Hotmail, LinkedIn, Coursera, Bitbucket, Snapchat, Steam, Yandex, other uncategorized) & **Question** sites: (Yahoo Answers, Quora)\
**eBay** & **Reddit**\
**Forum** (Disqus, Warrior Forum, other forums) & **Redacted**: (not visible in description because of the rotator technique, explained below)\
**Facebook** & **Smartphone** (iOS and Android)\
**Gmail** & **Twitter**\
**Google** (search) & **Yahoo** (search)\
**Google+** &\
Also, each campaign could belong to multiple activity categories, by using multiple tags.
Limitations
-----------
On microworkers.com, there are invite-only campaigns as well. Unfortunately, there is no information freely available on these campaigns or their share of the total number of campaigns. The nature of invite-only campaigns, for which clients apparently hire tested and trusted microworkers, could be a subject for further research.
Some campaigns might have slipped trough between two observations, meaning that their full life cycle very short (only a couple of hours). This does not appear to be typical but cannot be ruled out. Therefore, it can be said that in reality there were possibly more than 7,426 public campaigns and an additional, unknown number of invity-only ones.
As explained above, the campaigns were manually categorized by the author. This categorization, because it relied on objectively observable features of the campaigns, did not require significant subjective judgment. Therefore, in the context of intersubjectivity, it should not be a serious limitation that there was no multiple-person cross-checking performed for interpretation of the category labels.
Finally, there are obviously other brokering platforms for such microtasks, but these are outside the scope of this investigation (in fact, some of those platforms seem to be using microworkers for recruitment purposes).
The author is confident that these limitations do not prevent the work from meeting its stated aims, as it is an explorative rather than exhaustive description of techniques and strategies.
[| p[6.5cm]{} | p[6.5cm]{} |]{}\
**A** Answer (a question on an answer site—see under Platforms—of a forum) & **B** Bookmark or Pin (quick share on MBS platform—see above)\
**C** Comment (Facebook, YouTube, forums, etc. comments) & **D** Data processing (counting, summarizing information)\
**E** Engage (usually unspecified web activity, e.g., “use the website for a while” or social media activity that cannot be categorized in the other labels) & **F** Follow (various social media, mostly Twitter)\
**H** Share (using the share function on various social media) & **I** Install (install applications to a smartphone or computer)\
**K** Link (add a given link to a comment, question answer, share) & **L** Like/Upvote (depending on platform: YouTube thumbs up, Facebook like, Reddit upvote, Google+ +1)\
**M** Upload or Download (Upload: YouTube videos, Download: various files) & **N** Connect/Friend (Depends on social media platform, e.g., on Facebook, connects to become friends)\
**O** Other (anything that could not be categorized otherwise) & **P** Write/Post/Blogpost (create and post written content, similar to comment but usually longer)\
**Q** Ask a question (on an answer site—see under Platforms) & **R** Research participation/Survey\
**S** Search/Search and Click/Search and Visit (use the search function: Google, Yahoo, Bing, Facebook search, others) & **T** Tweet or Retweet\
**U** Signup (YouTube channel, mailing list, portal, etc.). Sometimes this involves handing over the login credentials & **V** Vote (vote on a given entrant, various voting platforms)\
**W** Watch (usually YouTube videos) & **Y** Captcha (solve capthas)\
**Z** Test (software or website testing) &\
Anonymization
-------------
Generally, all data presented in this paper (mostly microtask descriptions) is anonymized. Most of the actual URLs, person and company names, and other named entities are replaced with the string $<REDACTED(...)>$. When there are several URLs or names within one example though, they are replaced with their own unique label so that they are not mixed up. Other than this modification, the job descriptions are copied herein verbatim.
Obviously, some basic URLs like Facebook.com or fakenamegenerator.com, for example, are kept because they are reported only for uncovering the campaign method but not its content, and also because the job descriptions would not be as understandable without them.
The reason for the redaction is that the persons, websites, and Facebook accounts mentioned in these task descriptions may be unwilling targets of a campaign. It is also probable that the customers of promotion campaigns are often not aware or may even have been misled about the methods employed on their behalf.
Results
=======
Based on the observations, the following summaries were created.
Table \[ActBudg\] contains budget summaries by activity. The columns represent the activity code, number of campaigns, number of tasks, and net budget (without the 10% fee). The table is ordered by the descending number of campaigns.
**Activity** **\# campaigns** **\# tasks** **total budget**
-------------- ------------------ -------------- ------------------
L 1,303 207,811 \$ 22,757.32
P 1,293 116,682 \$ 26,796.95
S 1,229 577,444 \$ 46,802.91
U 1,210 354,180 \$ 40,344.71
C 733 227,756 \$ 27,926.15
E 495 318,387 \$ 23,310.10
I 361 16,934 \$ 8,375.01
H 357 26,305 \$ 7,547.54
Z 352 122,974 \$ 9,681.65
N 203 26,808 \$ 3,419.92
B 196 40,045 \$ 4,422.94
W 147 38,756 \$ 4,307.35
D 138 50,886 \$ 6,307.99
F 79 12,649 \$ 1,520.29
V 68 32,832 \$ 4,026.03
T 31 2,071 \$ 409.87
R 29 5,191 \$ 2,037.10
O 24 7,219 \$ 1,240.66
A 14 519 \$ 78.90
M 12 847 \$ 203.45
K 10 474 \$ 292.02
Q 3 150 \$ 24.60
Y 1 1,026 \$ 61.56
: Budget Summary by Activities[]{data-label="ActBudg"}
If we take out the category data processing, research participation, captcha solving, testing, installing and “other”, the remaining activities are purely for promotional purposes. This leaves 1,665,138 tasks, or 89.7% of the whole. It should be added that many of the install tasks appear to be promotional (see section \[Smartphone\]). Counting these in would make the figure even higher. However, since for many such campaigns this aspect is impossible to tell, they are left out.
A similar summary for the platforms is given in table \[PlatBudg\]. The first column is the platform name, the rest is the same as before. The table is ordered by the descending number of campaigns.
**Platform** **\# campaigns** **\# tasks** **\# total budget**
---------------- ------------------ -------------- ---------------------
Smartphone 1102 231,892 \$ 27,702.07
Google+ 823 57,203 \$ 18,960.50
Other 797 349,937 \$ 31,333.06
Twitter 747 83,731 \$ 14,047.21
Facebook 666 115,967 \$ 15,657.86
YouTube 616 248,914 \$ 29,322.28
Reddit 539 77,104 \$ 5,434.15
Redacted 501 132,401 \$ 13,397.94
Google 330 261,923 \$ 18,672.37
RDT 296 80,844 \$ 8,034.94
MBS 218 38,859 \$ 5,497.36
Instagram 172 19,421 \$ 2,072.35
Amazon 156 55,760 \$ 6,934.85
Mix 144 11,803 \$ 1,458.55
Question site 126 16,190 \$ 2,447.59
Forum 84 7,558 \$ 964.15
Gmail 50 10,275 \$ 2,908.35
eBay 22 6,014 \$ 570.52
Yahoo 21 45,686 \$ 2,355.13
Ali 9 3,945 \$ 551.91
Bing 4 675 \$ 69.00
Browser add-on 3 214 \$ 73.96
: Budget Summary by Platforms[]{data-label="PlatBudg"}
From these values, the average payment for tasks related to certain activities and platforms could be calculated. Here is the distribution of the payments (not equal ranges):
**task payment** **total number of tasks**
------------------ ---------------------------
\$0 99,999
\$0.05–\$0.1 1,059,172
\$0.11–\$0.2 581,146
\$0.21–\$0.3 63,904
\$0.31–\$0.5 35,893
\$0.51–\$1.0 15,671
\$1.1–\$3.0 531
: Number of tasks by payment range[]{data-label="Payment"}
The lowest paid wage was \$0, and this was incidentally the biggest campaign with 99,999 tasks. The task was a simple visit to a link. The client promised future tasks for those who completed the task. More detail is provided on this task in section \[Sec:signup\].
The highest paid wage was \$3 for a task, but as can be seen from the above table, there were only 531 jobs offering between \$1.1 and \$3, while there were over a million jobs offering between 5 and 10 cents, making the higher paid tasks very rare indeed.
Campaign types and their analysis
=================================
Voting
------
Participation in voting (V) is a recurring activity on microworksers.com, with 68 campaigns posted featuring an aggregated 32,832 votes purchased. The top two voting campaigns seemed to be promoting a product and a sports team (2,130 and 2,000 tasks), the third was a giveaway voting for an expensive trip for couples, where the entrants were supposed to vote on the videos they uploaded about themselves. One entrant purchased 1,703 votes (the wording reveals that they bought the vote for themselves personally) for \$0.12 each:
``` {#votes caption="Buying" votes="" label="votes"}
Title: Video: vote
Payment: 0.12
Number of workers accepted: 1703
1. Go to <REDACTED URL>
Give the video a VOTE by clicking on the heart below the video
Required proof that task was finished?
1. Tell me how many votes I had after you've voted
```
Unfortunately while this can be seen as being clearly unethical, for a little more than \$ 200 it could have been economically viable if the entrant won the vote. And it is even plausible that a local vote could have been won with just 1,703 extra votes bought.
Other votes were for titles like best auto repair shop, best bakery or “tradie of the year in Australia”. There was also a census on how many New York City residents wanted to go on a date with a certain model. There were votes on temple photography, the best fintech firms, music mixes, the best female vocalists, several contests about the ranking of attractive persons, a vote on XXL Magazine, a vote on the best local charter flight provider, and so on. Some of these were clearly promoting a product or a performer or artist; others seemed to be clearly what we will call vanity-promotion.
The purchased votes ranged from dozens to about 1,000 at a time. It is hard to assess the overall efficiency of such campaigns, but it can certainly be said that for local contests, where the maximum number of people voting is expected to be measured in hundreds or thousands, it is very easy to rig contests this way, as it would cost only a few dollars.
Search and engage tasks {#Sec:search}
-----------------------
The common feature of these kinds of microtasks seems to be an attempt to manipulate the search results in search engines like Google, Yahoo, Bing or the search feature of Facebook. In most cases, a certain item is promoted, but in some rare cases, the goal actually seems to be to push unwanted result items back in the result list.
These campaigns appear to assume that search engines learn: if for given search term, a high number of users click on a particular result item, then that result item must be a good result for the search and therefore it will be listed early in the results listing. While the actual algorithms search engines use are proprietary and unpublished, it is known how they work in theory [@Buettcher2016]. It is thus plausible that they can be tricked in this way to a certain extent. We know that user behavior is taken into account in Google, for instance, as [@Clark2015] reported that Google’s novel AI solution, BrainRank, was the third most important factor (the technical term is “signal”) when ranking pages. We also know that it learns from user behavior, hence it is plausible these systems can be tricked through paid user behavior simulating genuine interest.
A search and engage campaign therefore hires a large number of microworkers for searching the given terms and clicking on the promoted item in the search results.
An example of this type of campaign is given below:
``` {#googlesearch caption="Google" search="" label="googlesearch"}
Title: Google search
Payment: 0.08
Number of workers accepted: 1600
Job description:
1. Open up Google.com (please use US version)
2. In Google, please search for this phrase: <REDACTED PERSON NAME>
3. Please click on any of the red boxed links you see in the attached file <the file is a Google result list screenshot, red rectangles designate what result items need to be promoted>
4. Stay on page for 1 minute
(...)
```
The top ten search and engage campaigns have the number of tasks offered as between 2,941 and 6,770. However, many of these campaigns seem to be part of the same project, making the biggest projects around 10,000–20,000 tasks.
An interesting tendency is that in many of these promotion campaigns, it seems that there is no marketed product involved, rather it is individuals concerned with their online persona who are the payers, in what is really another example of vanity-promotion.
The biggest search and engage project, with above 10,000 tasks, involved the promotion of a USA business executive’s Wikipedia entry, whose name happens to be the same as a famous USA American football player and also a former USA congressman. The project must have been a success as currently the promoted page comes out top in a Google search when searching for that name. Naturally, it is impossible to establish the causal relation between the campaign and the current ranking with any certainty, especially this long after the campaign.
Social media activity
---------------------
Paid social media activity involves task like creating Pinterest Pins, upvoting in Reddit, YouTube, or Google+, liking in Facebook, using Digg, Twitter, or Instagram, commenting on forums, upvoting on SoundCloud, Mixcloud or Datpiff, and so on. In terms of activity codes, this section covers B, C, F, H, L, N, T, W.
Campaign example \[facelikes\] is in this category. The campaigns are usually straightforward and easy to do, therefore the payment is usually very low. Clicking on like, upvote, etc. are the lowest paid tasks. For instance, the 77,104 Reddit upvotes purchased during the 365 days study period cost less than \$ 5,500 in total (see Table \[PlatBudg\]). The highest paying jobs in this category were those that require writing content that meets a set specification, e.g.:
``` {#youcomment caption="Youtube" comments="" label="youcomment"}
Title: YouTube: Comment 3x (1-3)
Payment: 0.30
Number of workers accepted: 90
Job description:
1. Go to www.youtube.com/channel/<REDACTED>.
2. Post a positive relevant comments on the videos found in first three links
Important: Comment must be at least 10-15 words long and cannot be generic and must include the following words <REDACTED Person name> and the word "Florida".
3. Stay on each video page for 1 minute
Required proof that task was finished?
1. YouTube display name
2. Copy of the comments you've posted
3. URL to YouTube videos where comments were posted
```
However, in other cases the freelancer is asked to copy-paste the comment content, and the job then pays less:
Title: YouTube: Comment 3x (<REDACTED>)
Payment: 0.12
Number of workers accepted: 440
Job description:
1. Go to the instruction page: <REDACTED URL>
2. Search Youtube.com for the key phrase
3. Copy-paste the supplied comments onto relevant video
4. Repeat steps 2 and 3 for 2 more videos (3 total)
Required proof that task was finished?
1. Your YouTube name
2. The search phrases you used
3. URLs of the 3 videos you commented on
For commenting, the most prominent platforms are YouTube (364 campaigns; 197,735 tasks, some paying for three comments), Instagram (101; 4,670), questions sites (81; 10,282), Facebook (30, 2,485), and a long tail of other forums (see some under Platforms; 118 campaigns).
Following a given account is done on Google+ (317 campaigns; 21,228 tasks) Instagram (54; 8,960), Twitter (20; 2325), Quora and Yahoo Answers (4; 466), and Google+ (1; 898). In must be noted that for Twitter, the clients often require the presence of some features from the freelancer, e.g.:
``` {#twittercheck caption="Twitter" account="" quality="" rules="" label="twittercheck"}
(...)To do this task, you need to have a Twitter account that meets the following requirements:- At least 100 followers - Your follower count needs to be at least double your following count (meaning - if you are following 100 users, you need to have at least 100*2 =200 followers) - The majority of the 20 most recent tweets are in English - At least 8 out of 20 most recent tweets have no links, are Not Retweets, and sound natural and interesting.(...)
```
This is obviously requested in an effort to imitate a real Twitter user and to not seem like a newly created one. These tasks pay bonuses too, meaning that the pay can reach as high as \$ 0.25.
Posting (P) and Tweeting (T) were grouped together for being very similar. P and T is most prevalent on Twitter (716 campaigns; 80,211 tasks) and Google+ (485; 28,819).
Connecting as a friend is mostly done on Facebook (13 campaigns; 9,190 tasks).
Liking/Upvoting (L) is an activity performed on Reddit (539 campaigns; 77,104 tasks), Facebook (485; 71,200), Mixcloud and SoundCloud and Datpiff (143; 11,773), YouTube (63; 23,993), Instagram and Google+ (both 13 campaigns, 5,366 and 5,401 tasks, respectively), and on some other platforms (49).
The 10 biggest like/upvote campaigns ranged between 1,830 and 7,417 offered tasks (here, Facebook, Reddit, Instagram, Google+ campaigns were all in the top 10). The content of these top campaigns unfortunately were redacted using the rotator technique (see later in the article), but some of the remaining involved promoting persons not noted on Wikipedia (vanity-promoting); and some niche products. Size seems to be a limitation again, just like with search and engage campaigns: for celebrities with hundreds of thousands of followers, even as much as 7 thousand new likes hardly seem to matter, and the promotion technique does not seem to scale up to higher numbers.
This limitation might be not there in the case of comments (C) campaigns. The biggest comments campaign was conducted on YouTube, very similar to example \[youcommentcc\], and it involved 21,140 tasks, three comments each, yielding more than 63,000 paid comments. The tenth-biggest involved 2,425 tasks, again three comments each. While a similar amount of likes would still represent just a fraction when it comes to comparing it to the likes received for the most popular YouTube videos, Facebook accounts, etc. For comments, the case is different, because only a small fraction of readers/visitors make comments. A cursory, non-representative investigation of the many YouTube videos reveals that it is very hard to find videos that have less than 20x viewers than comments. Thinking the other way around, 63,000 tendentious comments would suggest representing well over 1.2 million viewers, hence distorting the perception of actual viewers on what other’s opinions are in relation to the topic.
However, in local communities with a smaller overall size, it seems that even small (L) campaigns can make a difference among the competition. Consider this example related to warriorforum.com (its self-description is: “The world’s largest Internet Marketing Community and Marketplace.”)
``` {#warriorforum caption="Comments" on="" a="" small="" forum="" label="warriorforum"}
Title: Warriorforum Post: Comment
Payment: 0.10
Number of workers accepted: 30
Job description:
Must have a Warriorforum account at least 2 months old, or a very active account if you joined recently.
1. Go to: www.warriorforum.com/<REDACTED>
2. Post a comment/testimony relating to the post
Must be original/no copying others post
Tip on what you can post:
I learned a lot from the webinar,
I just join and I am excited, cannot wait to start working with <REDACTED>,
this is an awesome program, etc.
Required proof that task was finished?
1. Name used to leave comment
```
A cursory look at Warrior Forum reveals that the typical number of upvotes is around 10 and the number of comments (reply-s) is similar. The job offer above for 30 comments in example \[warriorforum\] would thus propel the entry among the top, while costing a mere \$ 3.3 for the client.
Smartphone apps {#Smartphone}
---------------
A distinct area of paid activity is installing smartphone applications, using/testing and rating them. There were 361 such campaigns with 16,934 tasks worth \$ 8,375.01. A typical campaign looks like this:
Title: Android App Testing (<REDACTED>): Download+Install+Honest Review
Payment: 0.50
Number of workers accepted: 370
Job description:
1. Install the app below: play.google.com/store/apps/<REDACTED>
2. Download the app
3. Open the app for 30 seconds and test
4. Rate the app
Optional: Leave 3, 4, or 5 stars
5. Write an honest review in the Microworkers proof box only
Required proof that task was finished?
1. Your Google username
2. Paste your review
We can interpret this campaign in several ways. The charitable interpretation would be that this is an honest test for the software. Even though the app is only required to run for 30 seconds, the freelancers would open it on dozens of different Android devices, with different capabilities, screens, API levels, etc. Thirty seconds is enough to run some self-assessment and report back to a server. This test could thus have some value from a Software Engineering perspective and may be a valid assignment.
However, this kind of test is surely already long overdue in the case of a published application. Issues around crashing apps and unwanted startup behavior should have been resolved way before then. The author speculates that these kinds of campaigns are instead promotional. The clients are usually careful not to explicitly order five stars or positive reviews (albeit there are counter-examples of such). Yet, in such a campaign, an initial, visible user base is created for the app. Again, we can assume that this is more useful in niches than in competition with mainstream applications, as the leading apps have millions of users already.
It must be mentioned that this kind of microtask carries security risks for the freelancer and for the general public. The fact that the freelancer installs apps for a fraction of a dollar presents an obvious opportunity for breaching their smartphones. Although the current number of tasks in these campaigns does not seem to be high enough, in theory it would be possible to create a zombie network for DOS attacks or similar purposes.
Signup {#Sec:signup}
------
A very common type of campaign is the signup (U) campaign. These campaigns involve creating an account meeting some client-specified requirements. In many cases, the microworkers will be required to hand over the account credentials. Example \[youtubeacc\] below was one of the biggest signup and account handover campaigns observed.
``` {#youtubeacc caption="Youtube" Account="" Creation="" label="youtubeacc"}
Title: YouTube: Create an Account
Payment: 0.10
Number of workers accepted: 2290
Job description:
YouTube: Create an Account
1. Go to www.youtube.com
2. Create a new account
3. Verify your email
4. Login to activate the account
Required proof that task was finished?
1. YouTube username or email
2. YouTube password
Your task will NOT be rated satisfied if your YouTube account requests phone verification.
```
The client in this case has acquired 2,290 YouTube accounts for a mere gross \$ 251.9. Example \[gmailacc\] from the introduction is a similar case, but for Gmail. The dangers posed by such campaigns are obvious. Besides promoting products, ideas and agendas, a cohort of 2,300 YouTube users can disrupt any smaller community on the platform and the use of fake accounts could facilitate the account owner to commit fraud or otherwise abuse the system.
What makes these mass account acquisitions very dangerous is that they are not easy to detect. Methods that are able to detect fake accounts typically only work if they are all created by the same person [@Xiao2015], but cannot be expected to work in this case as these accounts are created by real people. A landmark study by @Gurajala2015 involving the analysis of 62 million Twitter public user profiles relied on statistics about update frequency, reused profile pictures, and account creation days. Unfortunately, these factors can all be made to look genuine; for instance, the freelancers can be instructed to use profile pictures that are not reused; or possibly profile pictures themselves are acquired via microworkers (see example \[photos\]), and the creation times can be spread out with the help of “throttling”—a feature of the platform that allows only a certain number of tasks to be completed in a unit of time. Sometimes clients give instructions that enable the detection of such accounts, e.g., by requiring the freelancers to use the very same password. Also, it is probable that after handing over an account, the geolocation of the usage of that account is changed permanently, and so never again reflects the country of creation, which could be a factor in detection.
Not all signup campaigns seem to require account handover. For instance, the top four signup campaigns in terms of task numbers required a signup to two different website traffic providers and a polling site; involving 21,388, 19,952, 10,888, and 7,648 individual signups. Related to these campaigns was the biggest (99,999 workers) and cheapest (paying \$0.0) campaign observed, categorized as testing (Z), as technically it was a website spellcheck; its details are given in the following example:
``` {#recruit caption="A" recruitment="" campaign="" involving="" a="" small="" test="" to="" pass="" label="recruit"}
Title: Qualification Test: Find the Misspelled Word
Payment: 0.0
Number of workers accepted: 99 999
This is a qualification for a future website test which will pay $7.50 for about 11 minutes of work. This qualification test is to find workers with a great eye for detail.
1. Go to REDACTED URL
2. Find the misspelled word.
Hint: it is near the bottom
Required proof that task was finished?
1. The wrongly spelled word in its wrongly spelled form
```
All five sites (the aforementioned four traffic providers and the one in Example \[recruit\]) were categorized as “Other” on the platform, and were not commonly featured. The scale of these campaigns explain why the category Other is so prominent in the platform aggregation in Table \[PlatBudg\]. Also, all the sites are basically recruiting microworkers for their own platform. The nature of tasks to be done there seems to be are traffic generation (visiting sites), participation in paid market research by answering surveys, etc.
Among the next five in the top 10 signup campaigns (places 6-10 ranked by the number of tasks on offer) was example \[youtubeacc\], another account creation and handover involving 2,190 accounts to a site redacted with the rotator technique, plus three jobs requiring the signup of several thousands of users to various sites for unknown reasons.
Other interesting campaigns
---------------------------
This section covers several interesting campaigns that cannot easily be categorized in the other categories, many of them one-of-a-kind campaigns, and some of them seem rather strange and unexplained.
There was one campaign that requires the users to solve captchas. This is obviously to bypass a captcha-protected signup page. We can hypothesize that this is part of a human-in-the-loop automated account creator system.
There were several research campaigns observed. These are transparent and benign: the university or the research group is clearly present, there is usually a document attached as a brief for the research. The topic seems to be social psychology or web usability and ergonomy. The users are asked their gender and then made to do face expression recognition; evaluate risks; try out different webpage workflows, etc.
There is one observed snapchat promoter recruitment campaign (30 tasks x USD 0.5), see the following:
``` {#snapchat caption="Snapchat" recruitment="" label="snapchat"}
Objectives: I'm looking for cute girls who snap for marketing promotions (bonus possible).
Important: You must actually snap video and not just upload pictures from fake profiles.
1. Provide your Snapchat Username for proof (I will add you as a friend)
```
Some campaigns seem to be building stock photos, like example \[photos\] below (1000 x \$ 0.11 ). Another project required photos of windshields. Yet another project asked for a selfie of the freelancer, and the consent to use it, but only from those who had no beard.
``` {#photos caption="Acquiring" photos="" label="photos"}
Important: You must agree to allow us to use your photo for promotional purposes in order to complete this task.
1. Take a well lit, clear photo of an office. Important: Make sure no people are in the photo.
Notes:- I need a clear photo of the office showing computers, desks, etc.- Photo should not be a photo from the Internet, we search for all photos on the Internet before we approve
```
Finally, for the following campaign there is just no explanation:
``` {#anthem caption="An" unexplained="" campaign="" label="anthem"}
(30 x $1.75) Write 12 lines of lyrics for an Anthem based on your own individual traditions and struggles (Make it relevant to your life today) (...)
```
Techniques employed in campaigns
================================
As explained in the Limitations section, there are a number of invite-only campaigns on the site, called “hire groups”. These allow a client to select the freelancers, as contrasted to public campaigns that are open to anyone to participate. Also, this allows a per-employee task customization by providing a spreadsheet of input variable values. While also being feature rich, hire group campaigns are usually hidden from the public view.
Rotators are another way of per-employee customization and also allows hiding the content of the campaign from public display.
``` {#rotator caption="The" rotator="" technique="" label="rotator"}
Title: Forum: Sign up + Post + Screenshot
Payment: 0.14
Number of workers accepted: 96
1. Go to this link: bit.ly/<REDACTED URL ENDING1>
2. Search for blogs from this search link
3. Find blog, website or forum you can post a comment on
4. Go to this link: bit.ly/<REDACTED URL ENDING2> and then copy comment from this page and post this comment in the website blog or forum
Required proof that task was finished?
1. Your Forum Username
2. URL of the comment
3. Screenshot of posting
```
This technique allows the employer to customize the task per-employer without a hire group and to remove the instructions after the campaign is done without leaving a trace. Except for those freelancers who participated in the campaign, there is no way of knowing what sites, search keywords, or comments were involved in the job. The category “Redacted” among the platforms refer to this technique and not to data anonymization employed by the author in the examples in this article. Of course, there is no way of knowing if the client’s intention was just to rely on task customization, or to hide the campaign content, or both.
As explained at the Smartphone apps section, there might be ways for dressing up promotion campaigns as testing campaigns, by asking a couple of hundred users to install the app and then leave it there. Also, there might be search and engage campaigns masquerading as data collection and competition monitoring. In the case of some Amazon- and eBay-related campaigns, the freelancers are directed to search for different products, then to select from a given set of results, and then to collect prices, data, specifications, and to finally submit these as job proof. What makes these suspicious is that for an honest information campaign, it seems to be overly redundant to collect the same information many hundreds of times by many hundreds of microworkers. In reality, the point of these campaigns could be to make the microworker search and engage and then to spend time on the visited page while counting reviews and collecting information (the algorithm of a search engine might take the duration spent on a result page into account when adjusting itself), and then the accomplishment of the job can be conveniently verified by the client by looking at the collected data. Of course, these are just hypotheses for which there is no way to verify them.
Figure \[Fig1\] summarizes the promotional methods observed, together with the supporting techniques featured in various gray promotional campaigns:
![Elements of unethical online promotion campaigns.[]{data-label="Fig1"}](Fig1-Opmicro.png){width="1\hsize"}
Conclusion
==========
This article provides insights into the black market of likes, upvotes, comments, retweets, votes in contests, and search engine manipulation. The subject of the investigation was microworkers.com, which is not a black market itself per se, but it has light regulation of its campaigns and so can be used by clients to participate in black-market activities. Also, it clearly is only one of several venues for running such campaigns. @DeMicheli2013 have investigated several other players on the market (Fiverr, SeoClerks, InterTwitter, FanMeNow, LikedSocial, SocialPresence, SocializeUk, ViralMediaBoost) and they have found that the market size is probably several millions of dollars, making the share microworkers.com a tiny fraction. Other sites even recruit on microworkers.com for similar microtasks. However, thanks to the fact that on microworkers.com the client has to orchestrate the campaigns itself, we can get an insight how the other players in the market, that sell complete like and follower packages, might be operating.
The nature of the microworkers.com campaigns was explained in the sections above. About their efficiency, we concluded that it probably varies. The main limitation is that it seems to be hard to purchase more than some tens of thousands of items. As explained in the section on Social media activity covering likes/upvotes (**L**), these numbers do not make a big difference when it comes to widely discussed political topics or celebrities, as in this area millions of **L** items are not uncommon. However, in smaller communities, with normally dozens or hundreds of **L** items, they can make a huge difference. This is the context in which the effects of a total of 207,811 **L** tasks can be assessed. For instance, Reddit, on which 77,104 upvotes were purchased, is a platform where a couple of hundred or thousand purchased upvotes can go quite far, especially in thematic sub-Reddits. It might be noted that a similar number of downvotes would be much more significant as there are normally much less of these items—but no downvoting/dislike campaigns were observed.
For comments, we have to assume that the big observed campaigns, reaching 60,000 YouTube comments must be effective as these are quite high numbers when it comes to comments. It is of course unclear what the overall effect of tens of thousands of comments is on the thinking of the targeted audience. But it is enough to provide an apparent majority on almost any platform.
For online voting and contests, it seems that all kinds except the biggest contests can be rigged by microworker campaigns.
There are two areas where the efficiency is especially hard to assess: search and engage and app testing. For search and engage, over half a million tasks were observed and we must assume that the efficiency of these jobs really depends on the popularity of the topic in question. Also, the effect of these campaigns is really hard to track. In a similar way, a hypothesis was provided on how app developers on Android or iPhone might by trying to build an initial user base of a couple of hundred installs. The most popular applications have tens of millions of user and even their alternatives often have tens or hundreds of thousands (this also indicates just how hard the entry must be to that market). A better understanding of the app market places would be necessary to understand the significance of a couple of hundred individual users.
Finally, the knowledge that several thousand YouTube, Gmail, Snapchat and other accounts have been created and their usernames and passwords handed over during the period of observation is very troubling. Those accounts might be effectively used for large scale gray promotion campaigns and could also pose a security threat at the same time.
Future work could involve participatory research as a freelancer on this or other platforms to reveal the experience of a microworker as well as to discover more about the invite-only/hire-only campaigns. In cases of comment copy-pasting, the source and nature of the comments could also be learned. Another area could be investigating the logic and goals behind the traffic generator sites and unions that similarly to microworkers’ sites rely on the completion of menial tasks.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Guangning Tan, Nedialko S. Nedialkov\
McMaster University\
John D. Pryce\
Cardiff University\
title: 'Symbolic-Numeric Methods for Improving Structural Analysis of Differential-Algebraic Equation Systems'
---
abstract
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We prove a Second Main Theorem type inequality for any log-smooth projective pair $(X,D)$ such that $X\setminus D$ supports a complex polarized variation of Hodge structures. This can be viewed as a Nevanlinna theoretic analogue of the Arakelov inequalities for variations of Hodge structures due to Deligne, Peters and Jost-Zuo. As an application, we obtain in this context a criterion of hyperbolicity that we use to derive a vast generalization of a well-known hyperbolicity result of Nadel.
The first ingredient of our proof is a Second Main Theorem type inequality for any log-smooth projective pair $(X,D)$ such that $X\setminus D$ supports a metric whose holomorphic sectional curvature is bounded from above by a negative constant. The second ingredient of our proof is an explicit bound on the holomorphic sectional curvature of the Griffiths-Schmid metric constructed from a variation of Hodge structures.
As a byproduct of our approach, we also establish a Second Main Theorem type inequality for pairs $(X,D)$ such that $X\setminus D$ is hyperbolically embedded in $X$.
address:
- 'D. Brotbek: Institut Élie Cartan, Université de Lorraine, Vandœuvre les Nancy, France'
- 'Y. Brunebarbe: IMB-CNRS, Université de Bordeaux, Talence, France.'
author:
- Damian Brotbek
- Yohan Brunebarbe
bibliography:
- 'biblio.bib'
title: 'Arakelov-Nevanlinna inequalities for variations of Hodge structures and applications'
---
Introduction
============
Consider a smooth projective complex algebraic variety $X$ and a variation of Hodge structures $\mathbb{V} = (\cL, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ defined on the complementary of a divisor $D \subset X$, see Section \[Recollection in Hodge theory\] for a reminder in Hodge theory. Given a non-compact Riemann surface $B$ and a non-constant holomorphic map $f: B \rightarrow X$ such that $f(B) \not\subset D$, we would like to investigate how much the position of the image $f(B)$ in $X$ is constrained due to the existence of $\bV$. To address this problem we establish a close analog of what is traditionally called a *Second Main Theorem* in Nevanlinna theory. As a corollary, in case $B$ is algebraic, we obtain a criterion implying that the map $f$ is algebraic.\
In this paper, we will always assume that the Riemann surface $B$ is parabolic in the function-theoretic sense, meaning that any bounded subharmonic function defined on $B$ is constant. For example, any algebraic Riemann surface satisfies this property. It is known that a *non-compact* Riemann surface $B$ is parabolic if and only if it admits a parabolic exhaustion function, i.e. a continuous proper function $\sigma : B \rightarrow [0, + \infty)$ such that $ \log \sigma$ is harmonic outside a compact subset of $B$ (see for example [@Sto77 Theorem 10.12]). For instance $\C$ is parabolic, and in this case one can take $\sigma(z)=|z|$. More generally, for any affine algebraic Riemann surface $B$, any proper finite holomorphic map $\pi : B \rightarrow \bC$ yields the parabolic exhaustion function $z \mapsto |\pi(z)|$. Parabolic Riemann surfaces provide a suitable framework for value distribution theory (see for instance [@SNM66], [@Wu70], [@Griffiths-King], [@Sto77], [@Sto83] or [@P-S14]). We briefly recall the main notations here and refer to Section \[ssec:Nevanlinna\] for more details.\
Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$. One defines the *weighted Euler characteristic* of the pair $(B,\sigma)$ to be the function $$\mathfrak{X}_{B,\sigma}(r):=\int_1^r \chi(B(t)) \frac{dt}{t},$$ where $B(r):=\{p\in B\ ;\ \sigma(p)<r\} $ and $\chi(B(t))$ is the Euler characteristic of $B(t)$. If $\alpha$ is a current of type $(1,1)$ on $B$, one defines its *Nevanlinna characteristic function* by $$T_\alpha(r)=\int_1^r\left(\int_{B(t)}\alpha\right)\frac{dt}{t}.$$ If $ X$ is a smooth proper variety, $L$ a line bundle on $X$, $D$ a divisor on $ X$ and $f:B\to X$ a non-constant holomorphic map such that $f(B)\not\subset D$, then one defines the *characteristic function of $f$ with respect to $L$* and the *counting function of $f$ with respect to $D$* to be respectively the functions $$T_{f,L}(r):=T_{f^*C_1(L,h)}(r)\quad \text{and}\quad N^{[1]}_{f,D}(r):=T_{[f^{-1}(D)]}(r).$$ Here $[f^{-1}(D)]$ denotes the integration current with respect to the discrete subset $f^{-1}(D) \subset B$ and $C_1(L,h)$ is the first Chern form of $L$ with respect to a $\mathscr{C}^{\infty}$ metric $h$ on $L$ (the notation is somewhat abusive since the function $T_{f,L}(r)$ depends on the choice of the metric $h$, but only up to a bounded function, see Section \[ssec:Nevanlinna\]). The counting function describes the asymptotic behaviour of the (possibly infinite) intersection of $f(B)$ and $D$ without counting multiplicities.
An Arakelov-Nevanlinna inequality
---------------------------------
Let $X$ be a smooth projective complex algebraic variety and $\mathbb{V} = (\cL, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ be a variation of complex polarized Hodge structures of length $w$ defined on the complementary of a normal crossing divisor $D \subset X$. Assume that $\cL$ has unipotent monodromies around the irreducible components of $D$. We denote by $\bar \cF^p$ the canonical Deligne-Schmid extension of $\cF^p$ to $X$ for any integer $p$ and by $\bar L_{\mathbb{V}} = \otimes_p \det \bar \cF^p$ the canonical extension of the Griffiths line bundle of $\mathbb{V}$.
The following theorem is the main result in this paper (see also Theorem \[thm:Arakelov-Nevanlinna inequality\] for another Arakelov-Nevanlinna inequality).
\[thm:SMT for VHS\] Let $X, D, \bV$ and $\bar L_{\mathbb{V}}$ as above. Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$ and a non-constant holomorphic map $f: B \rightarrow X$ such that $f(B) \not\subset D$. Then, for any ample line bundle $A$ on $X$, there exists $C >0$ such that the inequality $$T_{f, \bar L_{\mathbb{V}} }(r) \leqslant \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot \left( - \mathfrak{X}_{B,\sigma}(r) + N^{[1]}_{f, D}(r) \right) + C \cdot \left( \log r + \log T_{f, A}(r) \right) \ $$ holds for all $r \in \bR_{\geq 1}$ outside a Borel subset of finite Lebesgue measure.
In the situation of Theorem \[thm:SMT for VHS\], assume moreover that $B$ is affine algebraic and that the holomorphic map $f: B \rightarrow X$ is algebraic, or equivalently that $B$ is the complementary of finitely many (but at least one) points in a compact Riemann surface $\bar B$ and that $f$ extends to a holomorphic map $\bar f: \bar B \to X$. Choose a finite algebraic map $\pi : B \rightarrow \bC$ and equip $B$ with the parabolic exhaustion function $\sigma : z \mapsto |\pi(z)| $. Then one easily verifies that the three functions $ T_{f, \,\bar L_{\mathbb{V}} }, \mathfrak{X}_{B,\sigma} $ and $N^{[1]}_{f, D}$ are equivalent to $\deg_{\bar B}(\bar f ^\ast \bar \cF^p) \cdot \log$, $ \chi(B) \cdot \log$ and $ \deg((f^\ast(D))_{red}) \cdot \log$ respectively. Therefore the preceding theorem is implied in this case by the following Arakelov inequality [@peters2000arakelovtype; @Jost-Zuo; @Bruni-level]: $$\deg_{\bar B}(\bar f ^\ast \bar L_{\mathbb{V}} ) \leqslant \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot \big( -\chi(B) + \deg((f^\ast(D))_{red}) \big) .$$
Arguing as in [@Bruni-level], one can prove that Theorem \[thm:SMT for VHS\] holds more generally when the monodromy of $\cL$ at infinity is quasi-unipotent, with $ \bar L_{\mathbb{V}}$ denoting the Griffiths parabolic line bundle of the variation $\bV$, cf. *loc. cit.*.
One of our motivations is the following conjecture of Griffiths (see [@N-W14 Conjecture 4.10.5] for an even more optimistic version): *If $X$ is a smooth projective variety, $D$ a simple normal crossing divisor on $X$ and $A$ an ample line bundle on $X$, then there exists a constant $\alpha >0$ such that for any Zariski dense entire curve $f:\C\to X$ and every $\ep>0$, the inequality* $$T_{f,K_{X}(D)}(r)\leqslant \alpha \cdot N_{f,D}(r)+\ep \cdot T_{f,A}(r).\label{eq:Griffiths}$$ *holds for all $r \in \bR_{\geq 1}$ outside a Borel subset of finite Lebesgue measure (the characteristic functions are defined with respect to the exhaustion function $z \mapsto |z|$).*
This conjecture has several far reaching consequences, for instance, it implies that if $K_{X}(D)$ is big (i.e. the pair $(X,D)$ is of log-general type) then every entire curve in $X\setminus D$ is algebraically degenerate. This is a version of the logarithmic Green-Griffiths-Lang conjecture.
It is therefore natural to try to establish the inequality or other similar inequalities, which are referred to as *Second Main Theorems* in Nevanlinna Theory, when one already knows that the pair $(X,D)$ satisfies the Green-Griffiths-Lang conjecture. This is for instance the case when $X\setminus D$ supports a complex polarized variation of Hodge structures whose period map is immersive at one point, cf. [@Brunebarbe-Cadorel].
That conjecture of Griffiths has the following algebraic analogue: *If $X$ is a projective variety and $D$ is a simple normal crossing divisor on $X$ such that the pair $(X,D)$ is of general type, then there exists constants $\alpha,\beta\in \R_+$ such that for any compact Riemann surface $C$ and every non-constant algebraic map $f:C\to X$ such that $f(C)\not\subset D$, one has* $$\deg_Cf^*K_{X}(D)\leqslant \alpha \cdot \left(\deg_C \left(f^*D \right)_{\rm red}+2g\left(C \right)-2 \right)+\beta.\label{eq:GriffithsAlg}$$ Both Griffiths’ conjecture and its algebraic counterpart are still widely open. For a more precise discussion on the analogies between value distribution theory for parabolic Riemann surfaces and its algebraic counterpart, we refer to [@Gas09].
From this point of view, Theorem \[thm:SMT for VHS\] can be viewed as a *Second Main Theorem* for varieties supporting a variation of Hodge structures, replacing the log-canonical bundle of the ambiant variety by the canonical bundle (in the sense of Griffiths) of the variation.
Examples
--------
Endowing $\bC^2$ with a non-degenerate hermitian sesquilinear form $h$ of signature $(1,1)$, and identifying the upper half plane $\bH$ with the space of positive lines in $\bC^2$, we can view $\bH$ as the parameter space of complex Hodge structures on $\bC^2$ with Hodge numbers $(1,1)$ and polarized by $h$, cf. Section \[Recollection in Hodge theory\]. This induces a variation of complex polarized Hodge structures on any quotient of $\bH$ by a torsion-free discrete subgroup of $SU(h) \simeq SU(1,1) \simeq SL(2, \bR)$. This applies in particular to any smooth complex algebraic curve $X$ with $\chi(X) <0$ (so that $X$ is the quotient of $\bH$ by a Fuchsian group) and Theorem \[thm:SMT for VHS\] specialized to that case recovers the Second Main Theorem from Nevanlinna Theory in one variable.
The preceding example generalizes as follows. Let $\cD$ be a bounded symmetric domain. Gross [@Gross94] (in the tube domain case) and Sheng-Zuo [@Sheng-Zuo10] (in general) have constructed an $\mathrm{Aut}(\cD)$-equivariant variation of complex polarized Hodge structures on $\cD$ of length $\operatorname{rank}(\cD)$ and whose smallest Hodge bundle can be identified with the automorphic line bundle of $\cD$. Moreover, both the canonical bundle of $\cD$ and the canonical bundle of $\bV$ are a positive tensor power of the automorphic line bundle of $\cD$. If $\Gamma$ is a net lattice in the group of holomorphic automorphisms of $\cD$, then the quotient $U := \Gamma \backslash \cD$ is (uniquely) a smooth complex algebraic variety and the variation $\bV$ descends to $U$. For any smooth compactification $X$ of $U$ such that $D = X - U$ is a normal crossing divisor, the canonical extension to $X$ of the Griffiths line bundle of $\mathbb{V}$ is proportional to $K_{X}(D)$. Therefore, if we specialize Theorem \[thm:SMT for VHS\] in that case, we recover the Second Main Theorem for compactified locally symmetric spaces due to Nadel [@Nad89] and to Aihara-Noguchi [@A-N91].
Theorem \[thm:SMT for VHS\] can be applied to any smooth complex algebraic variety $X$ which is a fine moduli space for a class of varieties that satisfy an infinitesimal Torelli theorem, e.g. smooth projective curves or smooth projective varieties with a trivial canonical bundle (with a suitable level structure).
Two Applications
----------------
The following applications refine some recent results of Deng [@deng2020big]. However, our approach differs from his: on the one hand, while Deng uses an ad hoc Finsler metric, we rely on the Griffiths-Schmid metric, and on the other hand, while Deng relies on the big Picard theorem from [@deng2019picard], we use our Theorem \[thm:SMT for VHS\].
### A criterion of hyperbolicity
Following [@Ariyan-Robert] we say that a complex algebraic variety $Y$ is Borel hyperbolic modulo a closed subvariety $Z \subset Y$ if, for every (reduced) complex algebraic variety $S$, any holomorphic map $f : S^{an} \rightarrow Y^{an}$ such that $f (S^{an}) \not \subset Z^{an}$ is algebraic. Observe that this implies that $Y$ is Brody hyperbolic modulo $Z$, i.e. that every non-constant holomorphic map $f : \bC \rightarrow Y^{an}$ satisfies $f(\bC) \subset Z^{an}$, since otherwise both $f$ and $f \circ \exp$ would be algebraic.
\[thm:criterion for Borel hyperbolicity\] Let $X$ be a smooth projective complex algebraic variety and let $\mathbb{V} = (\cL, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ be a variation of complex polarized Hodge structures of length $w$ defined on the complementary of a normal crossing divisor $D \subset X$. Assume that $\cL$ has unipotent monodromies around $D$ and let $\bar L_{\mathbb{V}}$ be the canonical extension of the Griffiths line bundle of $\mathbb{V}$. If one denotes by $\B_{+}$ the augmented base locus of the $\bQ$-line bundle $ \bar L_{\mathbb{V}} (- ( \frac{w^2}{4} \cdot \operatorname{rank}\cL) \cdot D)$, then $X$ is Borel hyperbolic modulo $D \cup \B_{+}$.
Recall that for any $\bQ$-line bundle $L$ on a projective complex variety $Y$, its augmented base locus $\B_{+}(L)$ is the intersection over all ample $\bQ$-divisor $A$ on $Y$ of the stable loci $\B(L(-A)) := \cap_{n \geq1} \mathrm{Bs}(L(-A)^{\otimes n})$.
### A generalization of a theorem of Nadel
Let $X$ be a (non necessarily smooth nor proper) complex algebraic variety equipped with a $\bZ$-local system $\cL$. For every prime number $p$ we have an induced $\bF_p$-local system $\cL \otimes_{\bZ} \bF_p$ and we denote by $X(p)$ the finite étale cover of $X$ that trivializes the local system of sets $\mathcal{I}som_{\bF_p}(\bF_p^{\operatorname{rank}\cL},\cL \otimes_{\bZ} \bF_p)$. If $X$ is connected, then after fixing a base point $x \in X$, the covering map $X(p) \rightarrow X$ corresponds to the action of $\pi_1(X,x)$ on the set of basis of the $\bF_p$-vector space $(\cL \otimes_{\bZ} \bF_p)_x$. Note that this action is not necessarily transitive, so that $X(p)$ might not be connected.
The following result is a vast generalization of the main result of [@Nad89].
\[thm:generalization of Nadel\] Assume that $\cL$ underlies a variation of complex polarized Hodge structures with a quasi-finite period map. Then, for all but finitely many prime numbers $p$, any proper algebraic variety $\overline{X(p)}$ that compactifies $X(p)$ is Borel hyperbolic modulo the boundary $\overline{X(p)} \backslash X(p)$.
A Second Main Theorem for pseudo-metrics with negative holomorphic sectional curvature
--------------------------------------------------------------------------------------
The first main input of the proof of Theorem \[thm:SMT for VHS\] is the following *Second Main Theorem* for pseudo-metrics with negative holomorphic sectional curvature (see also Theorem \[thm:SMTlocal\]).
\[thm:SMT for pseudo-metrics with negative holomorphic sectional curvature\] Let $X$ be a smooth projective variety and let $D$ be a reduced divisor on $X$. Suppose that $X \setminus D$ is endowed with a Finsler pseudo-metric $h$ of class $\mathscr{C}^{\infty}$ with holomorphic sectional curvature bounded above by $-\gamma<0$. Assume that the degeneracy set of $h$ is contained in a nowhere-dense closed analytic subset of $X \setminus D$. Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$ and a non-constant holomorphic map $f:B\to X$ whose image $f(B)$ is not contained in $D$ nor in the degeneracy set of $h$. If $\omega$ denotes the $(1,1)$-form associated to the induced pseudo-metric $f^\ast h$ on $B\setminus (f^{-1} (D))_{\rm red}$, then $\omega$ is locally integrable on $B$ and there exists a positive real number $C$ such that the following inequality $$T_{[\omega]}(r)\leqslant \frac{2\pi}{\gamma}\big(N^{[1]}_{f,D}(r) -\mathfrak{X}_{\sigma}(r)\big)+C \cdot (\log r+\log T_{[\omega]}(r))$$ holds for all $r \in \R_{\geq 1}$ outside a subset of finite Lebesgue measure.
Our proof of this result is very much inspired by the work of Nadel [@Nad89] and Aihara-Noguchi [@A-N91]. We generalize their arguments to parabolic Riemann surfaces by using a slight modification of the approach of Păun and Sibony [@P-S14] to Nevanlinna Theory.
The holomorphic sectional curvature of the Griffiths-Schmid metric
------------------------------------------------------------------
The second main input towards the proof of Theorem \[thm:SMT for VHS\] is the following result that describes a pseudo-metric with nice negative curvature properties on any manifold supporting a variation of Hodge structures. For the most part it is well-known to experts, but to our knowledge the upper bound on the holomorphic sectional curvature is new.
\[thm:holomorphic sectional curvature for VHS\] Let $(\mathcal{L}, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ be a variation of complex polarized Hodge structures of length $w$ on a complex manifold $S$. The first Chern form of its Griffiths line bundle $L_{\mathbb{V}} := \otimes_p \det \cF^p$ equipped with the hermitian metric induced by $h$ is a closed positive real $(1,1)$-form, which coincides after multiplication by $2 \pi$ with the pull-back of the Kähler form of the Griffiths-Schmid metric on the corresponding period domain. The corresponding Kähler pseudo-metric is non-degenerate on the Zariski-open subset of $S$ where the period map is immersive, has non-positive holomorphic bisectional curvature and its holomorphic sectional curvature $- \gamma$ satisfies $$\frac{1}{\gamma} \leqslant \frac{w^2}{4} \cdot \operatorname{rank}( \cL).$$
Note that when $(\mathcal{L}, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ is isomorphic to its dual (this holds for example if it is a variation of real polarized Hodge structures), then $$\frac{w^2}{4} \cdot \operatorname{rank}( \cL) = \frac{w}{2} \cdot \sum_{i= 1}^p \operatorname{rank}( \cF^p) .$$
Second Main Theorem for hyperbolically embedded complements
-----------------------------------------------------------
In a different direction, applying Theorem \[thm:SMT for pseudo-metrics with negative holomorphic sectional curvature\] to the Kobayashi metric yields a *Second Main Theorem* for pairs $(X,D)$ where $X\setminus D$ is hyperbolically embedded in $X$. While this result is independent of the other main applications of the present paper, it also goes in the direction of the Griffiths conjecture and we therefore felt that it was noteworthy to mention it here. The algebraic counterpart of this statement was obtained by Pacienza and Rousseau [@PR07].
Let $X$ be a smooth projective variety and let $H$ be a reduced divisor on $X$ such that $X\setminus H$ is hyperbolically embedded in $X$. Let $A$ be an ample line bundle on $X$. There exists a constant $\alpha >0$ such that for any non-compact parabolic Riemann surface $B$ equipped with a parabolic exhaustion function $\sigma$ and a non-constant holomorphic map $f:B\to X$ such that $f(B)\not\subset H$, there exists a positive real number $C$ such that the following inequality $$T_{f,A}(r)\leqslant \alpha\big(N^{(1)}_{f,H}(r)-\mathfrak{X}_\sigma(r)\big)+C \cdot \log r$$ holds for all $r \in \R_{\geq 1}$ outside a subset of finite Lebesgue measure.
Organization of the paper
-------------------------
In Section \[sec:Nevanlinna\] we give a proof of Theorem \[thm:SMT for pseudo-metrics with negative holomorphic sectional curvature\] that does not assume any prior knowledge in Nevanlinna Theory. In Section \[sec:Hodge\] we first recall the basic definitions from variational Hodge Theory and then prove Theorem \[thm:holomorphic sectional curvature for VHS\]. The proof of Theorem \[thm:SMT for VHS\] is given in Section \[sec:proof SMT for VHS\] and the applications (Theorem \[thm:criterion for Borel hyperbolicity\] and Theorem \[thm:generalization of Nadel\]) are given in Section \[sec:proof applications\]. Finally, we prove Theorem \[thm:SMTHyp\] in Section \[sec:SMTHyp\].
Second Main Theorem for pseudo-metrics with negative holomorphic sectional curvature {#sec:Nevanlinna}
====================================================================================
Metric and curvature
--------------------
Let us start by recalling some standard notations and definitions. Let $B$ be a Riemann surface. A pseudo-metric on $B$ is a map $h:T_B\times_B T_B\to \C$ of class $\mathscr{C}^{\infty}$, such that for any $x\in B$ the restricted map $h_x:T_{B,x}\times T_{B ,x}\to \C$ is a symmetric sesquilinear form. To such a pseudo-metric, we associate the pseudo-norm $\|\cdot\|_h:T_B\to \R^+$ defined by $\|\xi\|^2_h=h(\xi,\xi)$ for any $\xi \in T_B$. For any $x\in B$, the restriction of $\|\cdot\|_h$ to $T_{B,x}$ will be denoted by $\|\cdot\|_{h,x}:T_{B,x}\to \R^+$. The degeneracy set of the pseudo-metric $h$ will be denoted by $$\Sigma_h:=\left\{x\in B\ ; \ \|\cdot\|_{h,x}\equiv 0\right\}.$$ We will only consider pseudo-metrics whose degeneracy set is a discrete subset of $B$. Locally, given a holomorphic coordinate $z$ on some open subset $U\subset B$, letting $\lambda : U\to \R^+$, the function $$\lambda(z)=\left\|\frac{\partial}{\partial z}\right\|^2_{h,z},$$ one defines the associated $(1,1)$-form to the pseudo-metric $h$ to be locally defined by $$\omega=\frac{i}{2} \lambda dz\wedge d\bar{z}.$$ This defines a global $(1,1)$-form on $B$. The Ricci curvature of $h$ or equivalently, the Ricci curvature of $\omega$ is defined to be $$\operatorname{Ric}\omega=-2\pi dd^c\log \lambda=-i\partial\bar{\partial}\log\lambda = -i\frac{\partial^2\log\lambda}{\partial z\partial\bar{z}}dz\wedge d\bar{z}.$$ This is a $(1,1)$-form defined on $B\setminus \Sigma_h$. The Gaussian curvature of $h$ is the function $K:B\setminus \Sigma_h\to \R$ such that[^1] $$\operatorname{Ric}\omega=K \omega.$$
Let $\Sigma\subset B$ be a reduced subset (i.e. $\Sigma$ is a discrete subset of $B$). Given a point $p\in \Sigma$ and a local coordinate $z$ centered at $p$, writing $\omega=\frac{i}{2} \lambda dz\wedge d\bar{z}$, we say that:
- $\omega$ has a *logarithmic singularity* at $p$ if there exists $C\in \R^+$ such that $$\lambda\leqslant \frac{C}{|z|^2}.$$
- $\omega$ has a *Poincaré singularity* at $p$ if there exists $C\in \R^+$ such that $$\lambda\leqslant \frac{C}{|z|^2(\log|z|^2)^2}.$$
If $\log \lambda$ is locally integrable, then one can define the current $[\log \lambda]$. Therefore, if $\omega$ is such that in any local chart $\log \lambda$ is locally integrable (in which case, we say abusively that $\log \omega$ is locally integrable), one can define on $B$ the current $$\operatorname{Ric}[\omega]=-2\pi dd^c[\log\lambda].$$ As usual, this means that for any $\mathscr{C}^{\infty}$ function with compact support $\varphi$, one has $\operatorname{Ric}[\omega](\varphi):=-2\pi\int_B (\log \lambda) dd^c\varphi.$ One easily checks that this is a well-defined current on $B$.
A lemma for Ricci currents
--------------------------
Let $B$ be a Riemann surface and $\Sigma$ be a reduced divisor on $B$. Let $h$ be a pseudo-metric on $B\setminus \Sigma$ and let $\omega$ be the associated $(1,1)$-form. If $\log \omega$ is locally integrable, one can define as above $\operatorname{Ric}[\omega]$. On the other hand if the Ricci form $\operatorname{Ric}\omega$ (defined on $B\setminus (\Sigma\cup \Sigma_h)$) is locally integrable (on $B$), one can define the current $[\operatorname{Ric}\omega]$ on $B$. The first step is to compare $[\operatorname{Ric}\omega]$ and $\operatorname{Ric}[\omega]$. This is based on the following elementary lemma concerning subharmonic functions whose proof is left to the reader.
\[lem:SH\] Let $\psi$ be a subharmonic and $\mathscr{C}^\infty$ function on $\Delta^*$. Suppose that $\psi$ is bounded above. Then $\psi$ extends as a subharmonic function on $\Delta$. Moreover, the $(1,1)$-form $dd^c\psi$ is locally integrable on $\Delta$ and the following inequality holds in the sense of currents: $$[dd^c\psi]\leqslant dd^c[\psi].$$
Applying this lemma on the pair $(B,\Sigma)$, one obtains the following.
\[lem:RicciCurrents\] Let $B$ be a Riemann surface with a reduced divisor $\Sigma$. Let $h$ be a pseudo-metric on $B\setminus \Sigma$ and let $\omega$ be the associated $(1,1)$-form. Suppose that $\omega$ has at most logarithmic singularities around every point of $\Sigma$ and that $\log \omega$ is locally integrable over $B$. Suppose moreover that there exists a smooth $(1,1)$-form $\tilde\omega$ on $B$ such that over $B\setminus (\Sigma\cup \Sigma_h)$ one has $$\tilde{\omega}-\operatorname{Ric}\omega\geqslant 0.$$ Then:
1. Both currents $[\operatorname{Ric}\omega]$ and $\operatorname{Ric}[\omega]$ are well-defined,
2. And one has $$[-\operatorname{Ric}\omega]\leqslant 2\pi [\Sigma]+(-\operatorname{Ric}[\omega]).$$
The statement is local, so we restrict ourselves to a contractible neighborhood $V$ of a point $p\in B$. In such a neighborhood, let $z$ be a local coordinate that vanishes at $p$ and write $\omega=\frac{i}{2}\lambda dz\wedge d\bar{z}$. The function $\log\lambda$ is locally integrable by hypothesis. Moreover, the function $\varphi:z\mapsto \log(|z|^2 \lambda (z))$ is bounded above, since by hypothesis on the singularities of $\omega$ the function $z\mapsto |z|^2\lambda(z)$ is bounded above.
Applying the $dd^c$ lemma, we can suppose that $\tilde{\omega}=dd^c\alpha$ for some function $\alpha$ of class $\mathscr{C}^\infty$ . Applying Lemma \[lem:SH\] with $\psi=\varphi+\alpha$, we obtain that $dd^c\psi$ is locally integrable, therefore so is $dd^c\varphi=dd^c\log\lambda$ outside $p$. This shows that $[\operatorname{Ric}\omega]$ is well-defined. Moreover, we obtain $$[dd^c\alpha]+[dd^c\varphi]=[dd^c\psi]\leqslant dd^c[\psi] = dd^c[\alpha]+dd^c[\varphi].$$ Since $\alpha$ is of class $\mathscr{C}^\infty$, one has $[dd^c\alpha]=dd^c[\alpha]$, hence $$[-\operatorname{Ric}\omega]=2\pi [dd^c\varphi]\leqslant 2\pi dd^c[\varphi].$$ Moreover the Lelong-Poincaré lemma implies that on $V$, one has $$dd^c[\varphi]=dd^c[\log(|z|^2\lambda)]=dd^c[\log|z|^2]+dd^c[\log\lambda]=[\Sigma]-\frac{1}{2\pi}\operatorname{Ric}[\omega].$$ Altogether, we obtain the expected result.
Preliminaries on parabolic Riemann surfaces
-------------------------------------------
Let us first introduce some terminology concerning parabolic Riemann surfaces. A detailed presentation of parabolic Riemann surfaces can for instance be found in [@N-S70]. General references for Nevanlinna theory for parabolic manifolds are for instance the work of Stoll [@Sto77; @Sto83], and more recently the article [@P-S14] by Păun and Sibony. We have followed closely this last article in our proof of the *Second Main Theorem*, with some changes in the notations to make the analogy with the algebraic situation more transparent.
A non-compact Riemann surface $B$ is called parabolic if it admits a parabolic exhaustion function, i.e. a continuous proper function $\sigma : B \rightarrow [0, + \infty)$ such that $ \log \sigma$ is harmonic outside a compact subset of $B$.
If $(B,\sigma)$ is a non-compact Riemann surface equipped with a parabolic exhaustion function, then, for any $r > 0$, the open pseudo-ball of radius $r$ and the pseudo-sphere of radius $r$ will be denoted by $$B(r) :=\{p\in B\ ;\ \sigma(p)<r\},\ \text{and} \ S(r) :=\{p\in B \ ; \ \sigma(p)=r\}$$ respectively. When $r$ is a regular value of $\sigma$ (which is the case for almost all $r$), the pseudo-sphere $S(r)$ is smooth, and one considers on it the measure $$d\mu_r:=d^c\log\sigma|_{S(r)}.$$
1. The most standard example of parabolic Riemann surface is $\C$, endowed with the parabolic exhaustion function $\sigma(z)=|z|$. In this case $B(t)$ is the disc of radius $t$ centered at zero, and $d\mu_r=\frac{d\theta}{4\pi}$, where $\theta$ denotes the argument of $z$.
2. If there exists a proper finite morphism $\pi:B\to \C$, then $B$ is a parabolic Riemann surface and a parabolic exhaustion function is given by $\sigma(p)=|\pi(p)|$.
3. The previous example implies in particular that any affine Riemann surface is parabolic.
One has the following version of the Jensen’s formula:
\[Jensen formula\] Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$. Let $\varphi:B\to [-\infty,+\infty)$ be a function which locally near every point of $B$ can be written as the difference of two subharmonic functions. Then, for any $r>1$ large enough, one has $$\int_{1}^r\frac{dt}{t}\int_{B(t)}dd^c[\varphi]=\int_{S(r)}\varphi d\mu_r+O(1).$$
One defines the *weighted Euler characteristic* of $(B,\sigma)$ to be the function $$\mathfrak{X}_\sigma(r):=\int_1^r\chi(B(t))\frac{dt}{t},$$ where $\chi(B(t))$ is the Euler characteristic of $B(t)$. We will only need the following result.
\[weighted Euler characteristic\] Same notations as above. If $\xi\in \Gamma(B,T_B)$ is any never vanishing holomorphic vector field on $B$ (such a vector field exists since $B$ is non-compact), then one has $$\mathfrak{X}_\sigma(r)=- \int_{S(r)}\log|d\sigma(\xi)|^2d\mu_r+O(\log r).$$
The statement we give here slightly differs from the formula given in [@P-S14] because the formula given in *loc.cit.* contains some typo. For completeness and for the reader’s convenience, we provide here a more detailed version of the proof of this formula given in [@P-S14]. We emphasize that we claim no originality here, and that we simply fill in the technical details of the proof outlined in [@P-S14].
Since $\xi$ is of type $(1,0)$, one has $d_{\xi}\sigma=\partial_{\xi}\sigma$. Since $\xi$ is holomorphic and $\log \sigma$ is harmonic outside the compact subset $\overline{B}(r_0)$ for a well-choosen $r_0$, it follows that $\partial_\xi\log \sigma$ is holomorphic outside $\overline{B}(r_0)$. A direct computation shows that $\sigma \cdot \partial_{\xi}\log\sigma=\partial_{\xi}\sigma$. Therefore, $\partial_{\xi}\sigma$ vanishes only when $\partial_{\xi}\log\sigma$ does, and outside $\overline{B}(r_0)$ this happens only on a discrete subset (hence of measure zero). From now on, we take $r > r_0$ outside this subset.
Consider the vector field $v=\overline{\partial_{\xi}\sigma}\cdot \xi$, or more precisely the real vector field $v_{\mathbb{R}}$ associated to it via the isomorphisms $T^{1,0}\equiv T_{\R}$. This vector field has isolated singularities outside $\overline{B}(r_0)$, and, along the boundary $S(r)$, is pointing in the outward normal direction. To see this, we can work locally, with a holomorphic coordinate $z$ such that $\xi=\frac{\partial }{\partial z}$. In this coordinate, one has $\overline{\partial_{\xi}\sigma}\cdot \xi=\overline{\frac{\partial \sigma}{\partial z}}\frac{\partial }{\partial z}=\frac{1}{2}\left(\frac{\partial \sigma}{\partial x}+i\frac{\partial \sigma}{\partial y}\right)\frac{\partial }{\partial z}$, therefore under the isomorphism $T^{1,0}\equiv T_{\R}$, the corresponding real vector field $v$ is given by $\frac{\partial \sigma}{\partial x}\frac{\partial }{\partial x}+\frac{\partial \sigma}{\partial y}\frac{\partial }{\partial y}$. In particular $d_v\sigma=\left(\frac{\partial \sigma}{\partial x}\right)^2+\left(\frac{\partial \sigma}{\partial y}\right)^2\geqslant 0$, and is strictly positive on the boundary $S(r)$ with our choice of $r$. This shows that $v$ is pointing outwards $B(r)$, and moreover it is immediate that it is tangent to the normal direction.
Let $\tilde{v}$ be a $\mathscr{C}^\infty$ vector field on $B$ with isolated zeros which coincides with $v_{\mathbb{R}}$ outside $\overline{B}(r_0)$. Applying the Poincaré-Hopf index theorem to the surface with boundary $\overline{B}(r) $ and the vector field $\tilde{v}$, we get that $$\chi(B(r))=\sum_{p\in B(r)}{\rm index}_p(\tilde{v})=\sum_{p\in \overline{B}(r_0)}{\rm index}_p(\tilde{v})+\sum_{p\in B(r)\setminus \overline{B}(r_0)}{\rm index}_p(v_{\mathbb{R}})$$ We claim that for every $p\in B(r)\setminus\overline{B}(r_0)$, ${\rm index}_pv_{\mathbb{R}}=-{\rm ord}_p\partial_{\xi}\log\sigma$. To see this, recall that if $f$ is a holomorphic function, then ${\rm index}_0\left(f\frac{\partial }{\partial z}\right)_{\mathbb{R}}={\rm ord}_0f$, where $\left(f\frac{\partial }{\partial z}\right)_{\mathbb{R}}={\rm Re}(f)\frac{\partial}{\partial x}+{\rm Im}(f)\frac{\partial}{\partial y}$ is the real vector field associated to $f\frac{\partial}{\partial z}$. Therefore, in our situation, ${\rm index}_p(\partial_{\xi}\log\sigma\cdot \xi)_{\mathbb{R}}={\rm ord}_p\partial_{\xi}\log\sigma $. On the other hand, $${\rm index}_p(\partial_{\xi}\log\sigma\cdot \xi)_{\mathbb{R}}= {\rm index}_p\left(\frac{1}{\sigma}\partial_{\xi}\sigma\cdot \xi\right)_{\mathbb{R}}={\rm index}_p\left(\partial_{\xi}\sigma\cdot \xi\right)_{\mathbb{R}}=-{\rm index}_p\left(\overline{\partial_{\xi}\sigma}\cdot \xi\right)_{\mathbb{R}}=-{\rm index}_pv_{\mathbb{R}}.$$ Applying now the Lelong-Poincaré formula, we obtain $$-\sum_{p\in B(r)\setminus \overline{B}(r_0)}{\rm index}_p(v_{\mathbb{R}})=\sum_{p\in B(r)\setminus \overline{B}(r_0)}{\rm ord}_{p}\partial_{\xi}\log\sigma=\int_{B(r)\setminus \overline{B}(r_0)}dd^c[\log|\partial_{\xi}\log\sigma|^2],$$ and therefore $$\begin{aligned}
-\chi(B(r))&=&\int_{B(r)\setminus \overline{B}(r_0)}dd^c[\log|\partial_{\xi}\log\sigma|^2]+ C_1\label{eq:EulerNegative}\end{aligned}$$ where $$C_1=-\sum_{p\in \overline{B}(r_0)}{\rm index}_p(\tilde{v}).$$ Let $\varphi$ be a $\mathscr{C}^\infty$ function on $B$ that coincides with $\log|\partial_{\xi}\log \sigma|^2$ over $B\setminus \overline{B}(r_0)$, so that $$\int_{B(r)\setminus \overline{B}(r_0)}dd^c[\log|\partial_{\xi}\log\sigma|^2]=\int_{B(r)}dd^c[\varphi]-\int_{\overline{B}(r_0)}dd^c[\varphi].$$ Therefore, setting $C_2=-\int_{\overline{B}(r_0)}dd^c[\varphi]$ and $C=C_1+C_2$, Jensen’s formula implies $$\begin{aligned}
-\int_1^r\chi(B(t))\frac{dt}{t}=\int_1^r\left(\int_{B(t)}dd^c[\varphi]\right)\frac{dt}{t}+C\log r= \int_{S(r)}\varphi d\mu_r +C\log r+O(1). \end{aligned}$$ On the other hand, one has $$\begin{aligned}
\int_{S(r)}\varphi d\mu_r&=& \int_{S(r)}\log|\partial_{\xi}\log\sigma|^2d\mu_r=\int_{S(r)}\log\left|\frac{\partial_{\xi}\sigma}{\sigma}\right|^2d\mu_r\\
&=&\int_{S(r)}\log\left|\partial_{\xi}\sigma\right|^2d\mu_r-\int_{S(r)}\log\left|\sigma\right|^2d\mu_r\\
&=&\int_{S(r)}\log\left|\partial_{\xi}\sigma\right|^2d\mu_r+2\log r\int_{S(r)}d\mu_r =\int_{S(r)}\log\left|d_{\xi}\sigma\right|^2d\mu_r+C_3\log r.\end{aligned}$$ In the last line we used the fact $\int_{S(r)}d\mu_r $ is independant of $r$ (as an application of Stokes formula). It remains to put all this together to obtain the announced result.
Notations from Nevanlinna theory {#ssec:Nevanlinna}
--------------------------------
Parabolic Riemann surfaces provide a suitable framework for value distribution theory. In this section we provide a brief account on this theory and refer the reader to the works [@SNM66; @Wu70; @Sto77; @Sto83] or [@P-S14] for a more detailed presentation.\
Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$. For any current $\alpha$ of type $(1,1)$ on $B$, we define the *characteristic function of $\alpha$* to be the function $T_{\alpha}:[1,+\infty)\to \R$ defined by $$T_\alpha(r)=\int_1^r\left(\int_{B(t)}\alpha\right)\frac{dt}{t}.$$ If $\Sigma$ is a discrete set of points in $B$, we denote by $[\Sigma]$ the associated integration current on $B$, and set $$N_{\Sigma}:=T_{[\Sigma]}.$$ Suppose we are given a smooth projective variety $X$ and a non-constant holomorphic map $f:B\to X$. Let $D$ be a reduced effective divisor on $X$ such that $f(B)\not\subset D$. The *truncated counting function of $f$ with respect to $D$* is the function $$N^{[1]}_{f,D}(r)=N_{f^{-1}(D)}(r).$$ We emphasize that $f^{-1}(D)$ is the set theoretic inverse image, and therefore, this function doesn’t take into account the intersection multiplicities. Let $L$ be a line bundle on $X$, endow $L$ with a $\mathscr{C}^{\infty}$ hermitian metric $h$ and denote by $C_1(L,h)$ the first Chern form of the hermitian bundle $(L,h)$. Recall that $C_1(L,h)$ is the $(1,1)$-form locally defined by $-dd^c\log\|s\|_h^2$, where $s$ is any nowhere-vanishing local holomorphic section of $L$. The *characteristic function of $f$ with respect to $L$* is the function $$T_{f,L}(r):=T_{f^*C_1(L,h)}(r).$$ This function depends on the choice of the metric $h$ only up to a bounded function. Indeed, if $h_1, h_2$ are two $\mathscr{C}^{\infty}$ metrics on $L$, then there exists a $\mathscr{C}^{\infty}$ function $\varphi:X\to \R_+^*$ such that $$C_1(L,h_1)-C_1(L,h_2)=dd^c\varphi.$$ By compactness of $X$, the function $\varphi$ is bounded and Jensen’s formula implies that the function $T_{f^*dd^c\varphi}$ is also bounded. Similar results can be obtained if one allows metrics with mild singularities. If $(L,h_1)$ is a hermitian line bundle, a *singular metric $h_2$* is of the form $h_2=e^{-\varphi} h_1$ where $\varphi:X\to \R\cup\{\pm\infty\}$ is a locally integrable function. In this situation, we define the *curvature current of the line bundle $L$ endowed with the singular metric $h_2$* to be $$C_1(L,h_2):=C_1(L,h_1)+dd^c[\varphi].$$
Let $X$ be a smooth projective variety. Let $L$ be a line bundle on $X$ and let $h_1$ be a smooth hermitian metric on $L$. Let $h_2$ be a singular metric on $L$. If the curvature current $C_1(L,h_2)$ is positive in the sense of current, then for any non-compact parabolic Riemann surface equipped with a parabolic exhaustion function and any non-constant holomorphic map $f:B\to X$, one has $$T_{f^*C_1(L,h_2)}(r)\leqslant T_{f^*C_1(L,h_1)}(r)+ O(1).$$
Since the curvature current $C_1(L,h_2)$ is positive in the sense of current, it follows that the function $\varphi$ is quasi-plurisubharmonic. Recall that this means that locally, $\varphi$ is the sum of a plurisubharmonic function and a smooth function. From this and the compactness of $X$, it follows that $\varphi$ is bounded above. One can now apply Jensen’s formula to $\varphi$ and we obtain $$T_{f^*C_1(L,h_2)}(r)-T_{f^*C_1(L,h_1)}(r)=\int_1^r\left(\int_{B(t)}dd^c\varphi\right)\frac{dt}{t}=\int_{S(r)}\varphi d\mu_r+O(1)\leqslant O(1).$$
\[lem:singular metric with log growth\] Let $X$ be a smooth projective variety. Let $L$ be a line bundle on $X$ and let $h_1$ be a smooth hermitian metric on $L$. Let $h_2=e^{-\varphi}h_1$ be a singular metric on $L$. Let $D$ be an effective divisor on $X$ and let $s_D\in H^0(X,\mathscr{O}_X(D))$ be a global section such that $D = \{s_D=0 \}$. Let $\|\cdot\|_D$ be the norm associated to a hermitian metric on the line bundle $\mathscr{O}_X(D)$. If locally on $X$ there exists $\beta,C\in \mathbb{\R_+}$ such that $$e^{-\varphi}\leqslant C \log(\|s_D\|_D^{-2\beta})$$ then for any ample line bundle $A$ on $X$, any non-compact parabolic Riemann surface $B$ equipped with a parabolic exhaustion function and any non-constant holomorphic map $f:B\to X$ such that $f(B)\not\subset D$, one has $$T_{f^*C_1(L,h_1)}(r)\leqslant T_{f^*C_1(L,h_2)}(r)+O(\log( T_{f,A}(r))).$$
Let $(U_i)_i$ be a finite set of open subsets of $X$ on which the inequality of the hypothesis holds, let $C_i,\beta_i$ be the associated constants, and let $C=\max\{C_i\}$ and $\beta=\max\{\beta_i\}$. Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function and let $f:B\to X$ be a non-constant holomorphic map such that $f(B)\not\subset D$. The hypothesis implies that $-\varphi\circ f\leqslant \log C+ \log(\beta)+\log (\log(\|s_D\circ f\|^{-2}_D))$. Applying Jensen’s formula and the hypothesis one obtains $$\begin{aligned}
\int_1^r\left(\int_{B(t)}-f^*dd^c[\varphi]\right)\frac{dt}{t}&=&\int_{S(r)}\varphi\circ f d \mu_r+O(1)\leqslant \int_{S(r)}\log (\log(\|s_D\circ f\|^{-2}_D)) d\mu_r+O(1)\\
&\leqslant& \log\int_{S(r)}\log(\|s_D\circ f\|^{-2}_D)d\mu_r+O(1)=\log \left( T_{f^*dd^c[\log(\|s_D\|^{-2}_D)]}(r) \right)\\
&=&\log \big(-T_{[f^*D]}(r)+T_{f^*C_1(\OO_X(D),\|\cdot \|_D)}\big)\leqslant \log (T_{f,\OO_X(D)}(r))=O(\log T_{f,A}(r)).\end{aligned}$$ From this relation, the announced inequality follows immediately from the definition of the curvature of the singular metric $h_2$.
Let us conclude this section by recalling some classical notations from Nevanlinna theory that shall be needed to state our results. Given two functions $g,h : [1,+\infty) \to \R$ we write $$g(r)\leqslant_{\rm exc} h(r)$$ if there exists a Borel subset $E\subset [1,+\infty)$ of finite Lebesgue measure such that $g(r)\leqslant h(r)$ for all $r\in [1,+\infty)\setminus E$. Moreover, given $s:[1,+\infty)\to \R$, we shall write $g(t)\leqslant_{\rm exc}h(t)+O(s(t))$ if there exists $C\in \R_+$ such that $g(t)\leqslant_{\rm exc} h(t)+Cs(t)$.
This notation is motivated from the use of the Borel lemma (see Lemma 1.2.1 in [@N-W14]) :
Let $\varphi:[1,+\infty)\to \R$ a monotone increasing function. Then, for any $\ep>0$, $$\varphi'(r)\leqslant_{\rm exc} \varphi(r)^{1+\ep}.$$
One of the main goal of Nevanlinna theory is to compare the characteristic functions and the counting functions for different line bundles and divisors. For instance, as a consequence of the *First Main Theorem* in Nevanlinna theory, one obtains *Nevanlinna’s inequality*, which guarantees that, if $X$ is a smooth projective variety and $D$ is an effective divisor on $X$, then, for any non-compact parabolic Riemann surface $B$ equipped with a parabolic exhaustion function $\sigma$ and for any non-constant holomorphic map $f:B\to X$ such that $f(B)\not\subset D$, one has $$N_{f,D}(r)\leqslant T_{f,\OO_X(D)}(r)+O(1).$$
In an opposite direction, we say that a *Second Main Theorem* holds for a pair $(X,D)$ and a line bundle $L$ on $X$ if there exists a constant $\alpha$ and an ample line bundle $A$ on $X$ such that for any non-compact parabolic Riemann surface $B$ equipped with a parabolic exhaustion function $\sigma$ and any non-constant holomorphic map $f:B\to X$ such that $f(B)\not\subset D$, one has $$T_{f,L}(r)\leqslant_{\rm exc}\alpha (N^{[1]}_{f,D}(r) - \mathfrak{X}_{\sigma}(r))+O(\log r+\log T_{f,A}(r)).\label{eq:SMTGeneralVersion}$$ One easily checks that if this inequality holds for some ample line bundle $A$, then it holds for any ample line bundle on $X$. The term $O(\log r+\log T_{f,A}(r))$ should be thought of as an error term.
A version of the logarithmic derivative lemma
---------------------------------------------
\[lem:RicciNegligeable\] Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$, and let $\xi\in \Gamma(B,T_B)$ be a global trivialization for $T_B$. Let $\Sigma$ be a reduced divisor on $B$. Let $h$ be a pseudo-metric on $B\setminus \Sigma$ such that the function $\log \|\xi\|_h$ can be locally written as the difference of two subharmonic functions. Let $\omega$ denotes the associated $(1,1)$-form and suppose that $\omega$ is locally integrable. Then $\operatorname{Ric}[\omega]$ is well-defined and one has $$\begin{aligned}
T_{-\operatorname{Ric}[\omega]}(r)
& \leqslant_{\rm exc}& -2\pi \mathfrak{X}_{\sigma}(r)+O\left(\log r+ \log T_{[\omega]}(r)\right).\end{aligned}$$
Consider the function $$\varphi=\log(\|\xi\|^2_h)$$ on $B\setminus \Sigma$, which by hypothesis is locally the difference of two subharmonic functions on $B$.
By definition, one has $$\operatorname{Ric}[\omega]=-2\pi dd^c[\varphi].$$ By Jensen’s formula (Proposition \[Jensen formula\]), one has, for any large enough regular value $r$ of $\sigma$, $$\frac{1}{2\pi}T_{-\operatorname{Ric}[\omega]}(r)=\int_1^{r}\left(\int_{B(t)}dd^c[\varphi]\right)\frac{dt}{t}=\int_{S(r)}\varphi d\mu_r+O(1)$$ Therefore we are reduced to prove that $$\int_{S(r)}\varphi d\mu_r= - \mathfrak{X}_\sigma(r)+O\left(\log r+ \log T_{[\omega]}(r)\right).$$ One the one hand, one has $$\begin{aligned}
\int_{S(r)}\varphi d\mu_r&=&\int_{S(r)}\log(\|\xi\|^2_h) d\mu_r=\int_{S(r)}\log\left(\frac{\|\xi\|^2_h|d_\xi(\sigma)|^2}{|d_\xi(\sigma)|^2}\right) d\mu_r\\
&=& \int_{S(r)}\log\left(\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2}\right) d\mu_r+ \int_{S(r)}\log\left(|d_\xi(\sigma)|^2\right) d\mu_r\end{aligned}$$ On the other hand, it follows from Proposition \[weighted Euler characteristic\] that $$\int_{S(r)}\log\left(|d_\xi(\sigma)|^2\right) d\mu_r= - \mathfrak{X}_\sigma(r)+ O(\log(r)).$$
It therefore remains to bound the other term. First observe that it follows from Fubini’s theorem that for any smooth $1$-form $\varphi$ on $B$, one has $$\int_{B(r)}d\sigma\wedge \varphi=\int_{1}^r\left(\int_{S(t)}\varphi\right)dt.$$ Letting $\varphi = \frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2} d^c\sigma$ and differentiating with respect to $r$, we obtain that $$\int_{S(r)}\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2} d^c\sigma = \frac{d}{dr}\int_{ B(r)}\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2} d\sigma\wedge d^c\sigma.$$
Therefore, by the concavity of the logarithm, $$\begin{aligned}
\int_{S(r)}\log\left(\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2}\right) d\mu_r&\leqslant & \log \int_{S(r)}\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2} d\mu_r= \log \frac{1}{r}\int_{S(r)}\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2} d^c\sigma\\
& =& \log \frac{1}{r}\frac{d}{dr}\int_{ B(r)}\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2} d\sigma\wedge d^c\sigma.\end{aligned}$$
On the other hand, one has $$\frac{\|\xi\|^2}{|d_\xi(\sigma)|^2}d\sigma\wedge d^c\sigma= \frac{1}{\pi}\omega.$$ Indeed, if one takes a local coordinate $z$ such that $\xi=\frac{\partial}{\partial z}$ and we write $\omega=\frac{i}{2}\lambda dz\wedge d\bar{z}$. One has $\lambda=\|\xi\|^2_{h}$ and $d_{\xi}\sigma=\frac{\partial \sigma}{\partial z}$. Moreover, since $\sigma$ is real, $$d\sigma\wedge d^c\sigma=\frac{i}{4\pi}\left(\frac{\partial \sigma}{\partial z}dz+\frac{\partial \sigma}{\partial \bar{z}}d\bar{z}\right)\left(\frac{\partial \sigma}{\partial \bar{z}}d\bar{z}-\frac{\partial \sigma}{\partial z}dz\right)=\frac{i}{2\pi}\frac{\partial \sigma}{\partial z}\frac{\partial \sigma}{\partial \bar{z}}dz\wedge d\bar{z}=\frac{i}{2\pi}\left|\frac{\partial \sigma}{\partial z}\right|^2dz\wedge d\bar{z}.$$ Hence $$\frac{\|\xi\|^2}{|d_\xi(\sigma)|^2}d\sigma\wedge d^c\sigma= \frac{\lambda}{\left|\frac{\partial \sigma}{\partial z}\right|^2}\frac{i}{2\pi}\left|\frac{\partial \sigma}{\partial z}\right|^2dz\wedge d\bar{z}=\frac{i}{2\pi}\lambda dz\wedge d\bar{z}=\frac{1}{\pi}\omega.$$ By applying Borel’s lemma twice, we obtain that for any $\ep>0$, one has $$\begin{aligned}
\frac{1}{\pi r}\frac{d}{dr}\int_{ B(r)}\omega&\leqslant_{{\rm exc} }& \frac{1}{\pi r}\left(\int_{ B(r)}\omega\right)^{1+\ep}= \frac{r^\ep}{\pi}\left(\frac{1}{r}\int_{ B(r)}\omega\right)^{1+\ep}=\frac{r^\ep}{\pi}\left(\frac{d}{dr}T_{\omega}(r)\right)^{(1+\ep)}\\
&\leqslant_{\rm exc}& \frac{r^\ep}{\pi}T_{\omega}(r)^{(1+\ep)^2}\end{aligned}$$ Altogether this implies that $$\int_{S(r)}\log\left(\frac{\|\xi\|^2_h}{|d_\xi(\sigma)|^2}\right) d\mu_r \leqslant_{\rm exc} \log(\frac{r^\ep}{\pi}T_{\omega}(r)^{(1+\ep)^2})=\ep \log(r)+(1+\ep)^2\log T_{\omega}(r)-\log\pi,$$ which implies the result.
Combining Lemma \[lem:RicciCurrents\] and Lemma \[lem:RicciNegligeable\] one obtains the following
\[cor:Tautological\] Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$, and let $\xi\in \Gamma(B,T_B)$ be a global trivialization for $T_B$. Let $\Sigma$ be a reduced divisor on $B$. Let $h$ be a pseudo-metric on $B\setminus \Sigma$ such that the function $\log \|\xi\|_h$ can be locally written as the difference of two subharmonic functions and let $\omega$ denotes the associated $(1,1)$-form. Suppose moreover that $\omega$ has at most logarithmic singularities around points of $\Sigma$. Then, $$\begin{aligned}
T_{[-\operatorname{Ric}\omega]}(r)&\leqslant 2\pi N_\Sigma(r)+T_{-\operatorname{Ric}[\omega]}(r)+O(1)\\
&\leqslant_{{\rm exc}} 2\pi \big(N_\Sigma(r)- \mathfrak{X}_{\sigma}(r)\big)+O( \log r+\log T_{[\omega]}(r))\end{aligned}$$
This result can be thought of as an intrinsic version of *McQuillan’s tautological inequality*. Although it is not completely immediate, it is possible, with a suitable choice of metric, to derive from the previous corollary the usual version of the tautological inequality. Since this approach still requires some technical estimates and will not be further used in the present article, we don’t give the details here. We refer to [@Gas09] for a detailed presentation of the tautological inequality.
Second Main Theorem for metrics with negative holomorphic sectional curvature
-----------------------------------------------------------------------------
In this section, we observe that the hypothesis of Corollary \[cor:Tautological\] are verified under an assumption on the Gaussian curvature of the pseudo-metric $h$.
\[lem:HSC\] Let $B$ be a non-compact Riemann surface endowed with a reduced divisor $\Sigma$. Let $h$ be a pseudo-metric on $B\setminus \Sigma$ such that the holomorphic sectional curvature of $h$ is bounded above by a negative constant $-\gamma<0$ over $B\setminus(\Sigma\cup\Sigma_h)$. Let $\omega$ denote the $(1,1)$-form associated to $h$. Then:
1. The form $\omega$ has at most a Poincaré singularity at the points of $\Sigma$.
2. If $\xi\in \Gamma(B,T_B)$ is a never-vanishing holomorphic vector field, then $\log\|\xi\|_h$ can locally be written as the difference of two subharmonic functions.
The first claim is an immediate consequence of the Schwarz lemma in view of our hypothesis on the Gaussian curvature. To prove the second assertion, one can restrict ourself to small enough neighborhoods of points of $\Sigma$. So fix $p\in \Sigma$ and take a small enough neighborhood $W$ of $p$ with a local holomorphic coordinate $z$ centered at $p$ such that $\xi=\frac{\partial }{\partial z}$ and such that $\Sigma\cap W=\{p\}$. Set $\lambda(z)=\|\xi(z)\|^2$ for all $z\in W$. Since $\omega$ has a Poincaré singularity at $p$, it follows a fortiori that $z \mapsto |z|^2\lambda(z)$ is bounded above around $p$. On the other hand, by our curvature assumption, one has $-\operatorname{Ric}\omega\geqslant \gamma \omega$ on $W\setminus\{p\}$. Therefore the function $z \mapsto \log(|z|^2\lambda(z))$ is subharmonic on $W\setminus \{p\}$, and thus extends as a subharmonic function $\varphi$ on $W$. It follows that $z \mapsto \log\lambda(z)=\varphi(z) -\log|z|^2$ is indeed the difference of two subharmonic functions on $W$.
\[thm:SMTlocal\] Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function and a reduced divisor $\Sigma \subset B$. Let $h$ be a pseudo-metric on $B\setminus \Sigma$ such that the Gaussian curvature of $h$ is bounded above by a negative constant $-\gamma<0$ over $B\setminus(\Sigma\cup\Sigma_h)$. If $\omega$ denotes the $(1,1)$-form associated to $h$, then $\omega$ is locally integrable on $B$ and one has $$T_{[\omega]}(r)\leqslant_{{\rm exc}} \frac{2\pi}{\gamma}\big(N_\Sigma(r) -\mathfrak{X}_{\sigma}(r)\big)+O(\log r+\log T_{[\omega]}(r)).$$
Observe first that both currents $[\omega]$ and $[\operatorname{Ric}\omega]$ are well-defined in view of Lemma \[lem:RicciCurrents\] and Lemma \[lem:HSC\]. By hypothesis, the Gaussian curvature $K_h$ of $h$ satisfies $K_h\leqslant -\gamma$. Therefore, outside $\Sigma_h$, one has $$\operatorname{Ric}\omega=K_h\omega\leqslant -\gamma \omega$$ and thus, $\gamma \omega\leqslant -\operatorname{Ric}\omega$. At the level of currents, it follows that $$[\omega]\leqslant \frac{1}{\gamma}[-\operatorname{Ric}\omega].$$ From this, one obtains $$T_{[\omega]}(r)\leqslant \frac{1}{\gamma}T_{[-\operatorname{Ric}\omega]}(r).$$ It suffices then to apply Corollary \[cor:Tautological\].
From this result one immediately infers a *Second Main Theorem* for varieties having a metric with negative holomorphic sectional curvature. Let us first recall a definition. Let $U$ be a complex manifold equipped with a Finsler pseudo-metric $h$ of class $\mathscr{C}^{\infty}$. Let $x \in U$ in the complementary of the degeneracy set of $h$. The holomorphic sectional curvature of $h$ at $x$ in a direction $\xi\in T_{U,x}\setminus\{0\}$ is defined to be $${\rm HSC}_{x,\xi}(h):=\sup K_{f^*h}(0)$$ where the supremum is taken over all holomorphic maps $f:\Delta\to U$ such that $f(0)=x$ and such that $\xi\in \C\cdot f'(0)$. If ${\rm HSC}(\omega)\leqslant -\gamma$ for some $\gamma>0$, we say that the holomorphic sectional curvature is bounded above by $-\gamma$. In that case, if $B$ is a Riemann surface and $f:B\to U$ is a holomorphic map such that $f^*h$ is a pseudo-metric on $B$, then its Gaussian curvature is bounded above by $-\gamma$. Therefore, we obtain the following
\[thm:SMTglobal\] Let $X$ be a smooth projective variety and let $D$ be a reduced divisor on $X$. Suppose that $X \setminus D$ is endowed with a Finsler pseudo-metric $h$ of class $\mathscr{C}^{\infty}$ with holomorphic sectional curvature bounded above by $-\gamma<0$. Assume that the degeneracy set of $h$ is contained in a nowhere-dense closed analytic subset of $X \setminus D$. Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$ and a non-constant holomorphic map $f:B\to X$ whose image $f(B)$ is not contained in $D$ nor in the degeneracy set of $h$. If $\omega$ denotes the $(1,1)$-form associated to the induced pseudo-metric $f^\ast h$ on $B\setminus (f^{-1} (D))_{\rm red}$, then $\omega$ is locally integrable on $B$ and one has $$T_{[\omega]}(r)\leqslant_{{\rm exc}} \frac{2\pi}{\gamma}\big(N^{[1]}_{f,D}(r) -\mathfrak{X}_{\sigma}(r)\big)+O(\log r+\log T_{[\omega]}(r)).\label{eq:SMTglobal}$$
In the following formulation, the error term does not depend on $\omega$.
Same setting as in Theorem \[thm:SMTglobal\]. Then, for any $\ep >0$, one has $$(1-\ep)T_{[\omega]}(r)\leqslant_{{\rm exc}} \frac{2\pi}{\gamma}\big(N_{f,D}^{[1]}(r)-\mathfrak{X}_{\sigma}(r)\big)+O(\log r).$$
Just observe that the function $T_{[\omega]}$ is unbounded and increasing, hence for any $C, \ep>0$ we have $$C \log T_{[\omega]}(r)\leqslant_{\rm exc} \ep T_{[\omega]}(r).$$
Recollection in Hodge theory {#Recollection in Hodge theory}
============================
\[sec:Hodge\]
Complex polarized Hodge structure
---------------------------------
A complex polarized Hodge structure (of weight zero) on a finite-dimensional complex vector space $V$ is the data of a non-degenerate hermitian form $h$ on $V$ (the polarization) and of a decomposition $V = \bigoplus_{p \in \bZ} {V}^p$ (the Hodge decomposition) which is orthogonal for $h$ and such that the restriction of $h$ to $ {V}^p$ is positive definite for $p$ even and negative definite for $p$ odd. The $r_p := \dim V^p$ are called the Hodge numbers. The Hodge metric on $V$ is the positive-definite hermitian metric $h_H$ obtained from $h$ by imposing that the Hodge decomposition $V = \bigoplus_{p \in \bZ} {V}^p$ is $h_H$-orthogonal and setting $h_H := (-1)^p \cdot h$ on $V^p$.\
The Hodge filtration is the decreasing filtration $\{F^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}\}$ on $V$ defined by $F^p := \bigoplus_{q \geq p} V^q$. The Hodge decomposition is determined by the Hodge filtration thanks to the formula $V^p = F^p \cap (F^{p+1})^\perp$. Here $(F^{p+1})^\perp$ denotes the orthogonal with respect to the polarization $h$, which is clearly equal to the orthogonal of $F^{p+1}$ with respect to the Hodge metric $h_H$.\
If $[a, b]$ is the smallest interval such that $V^p = 0$ for $p \notin [a, b]$, then the length of $V$ is by definition the integer $w= b - a$. Note that by shifting the numbering of the Hodge decomposition, one can always assume that $a = 0$ and $b = w$.\
The category of complex polarized Hodge structures is Abelian, semisimple and admits tensor products and internal Hom. In particular, $\End(V)$ inherits a complex polarized Hodge structure with decomposition $\End(V) = \bigoplus_{p \in \bZ} {\End(V)}^p$ where $\End(V)^p = \bigoplus_{s - r = p } \Hom(V^r, V^s)$.\
We often denote abusively by the same symbols $h$ or $h_H$ the metric that they induce on any object obtained from $V$ by a tensorial construction.
Variation of Hodge structures
-----------------------------
A variation of complex polarized Hodge structures $\bV$ on a (reduced) complex analytic space $S$ consists in a complex local system $\cL$ on $S$ equipped with a non-degenerate hermitian form $h : \cL \otimes_{\bC} \bar \cL \rightarrow \underline{\bC}_S$ and a locally split finite filtration $\cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}$ of $\cV := \cL \otimes_\bC \cO_S$ by analytic coherent subsheaves such that
- for every $s \in S$, the triple $(\cL_s, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}_s, h_s)$ defines a complex polarized Hodge structure;
- (Griffiths’ transversality) letting $\nabla := id \otimes d$, we have $\nabla( \cF^p) \subset \cF^{p-1} \otimes_{\cO_S} \Omega_S^1$ for all $p$.
Note that when $S$ is smooth, the data of the complex local system $\cL$ is equivalent to the data of the holomorphic vector bundle $\cV := \cL \otimes_\bC \cO_S$ equipped with the integrable connection $\nabla := id \otimes d$.
The length of the complex polarized Hodge structure on the complex vector space $\mathcal{L}_s$ is independent of the point $s \in S$ and is called the length of $\bV$.
Example
-------
If $f : X \to S$ is a smooth projective holomorphic map, then the choice of a relatively ample line bundle on $X$ endows canonically the complex local system $R^k f_\ast \underline{\mathbb{C}}_X $ with a structure of variation of complex polarized Hodge structures for every $k \geq 0$.
Associated system of Hodge bundles
----------------------------------
Starting with a variation of complex polarized Hodge structures $(\mathcal{V}, \nabla, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ on a complex manifold $S$, we define $\cE^p := \cF^p / \cF^{p+1}$ and $\cE := \bigoplus_p \cE^p$. It follows from Griffiths transversality that $\nabla$ induces an $\cO_S$-linear morphism $ \cE^p \rightarrow \Omega^1_S \otimes_{\cO_S} \cE^{p-1}$ for every $p$. We denote by $\phi_p$ the corresponding element of $ \Omega^1_S \otimes_{\cO_S} \mathcal{H}om(\cE^p , \cE^{p-1})$ and set $\phi := \oplus_p \phi_p \in \Omega^1_S (\mathcal{E}nd(\cE))$. The pair $(\cE, \phi)$ together with the decompositions $\cE := \bigoplus_p \cE^p$ and $\phi := \oplus_p \phi_p $ is called the system of Hodge bundles associated to the variation of complex polarized Hodge structures $(\mathcal{V}, \nabla, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$. In what follows, we will often consider the so-called Higgs field $\phi \in \Omega^1_S (\mathcal{E}nd(\cE))$ as an $\cO_S$-linear morphism $ \mathcal{T}_S \rightarrow \mathcal{E}nd(\cE)$ that we will abusively denote by the same symbol, and similarly for the $\phi_p$’s.\
The holomorphic vector bundle $\cE$ comes equipped with the positive-definite hermitian metric $h_H$ that we call the Hodge metric. The decomposition $\cE := \bigoplus_p \cE^p$ is orthogonal for $h_H$. The curvature of $(\cE, h_H)$ has been computed by Griffiths (the real structure plays no role in the computations):
\[curvature computation\] The curvature form $\Theta$ of $(\cE, h_H)$ satisfies $$h_{H}(\Theta_{(\cE, h_H)}(X, \bar Y) e, f) = h_{H}( \phi(X)e, \phi(Y) f) - h_{H}(\phi^\ast(\bar Y) e, \phi^\ast (\bar X) f)$$ for any vector fields $X$ and $Y$ of type $(1,0)$ and any sections $e$ and $f$ of $\cE$.
In the formula above, $\phi^\ast$ denotes the adjoint of $\phi$, meaning that $\phi^\ast(X)$ and $\phi(\bar X)$ are adjoint with respect to the induced Hodge metric on $\cE nd (\cE)$, for every tangent vector $X$ of type $(1,0)$. The decomposition $\phi = \oplus_p \phi_p $ corresponds to the decomposition $\phi^\ast = \oplus_p \phi^\ast_p $ where $\phi^\ast_p$ is a $(0,1)$-form with values in $\cH om(\cE^{p-1}, \cE^p)$.
Induced system of Hodge bundles on the endomorphism bundle
----------------------------------------------------------
Let $\mathbb{V}$ be a variation of complex polarized Hodge structures on a complex manifold $S$, and let $(\cE = \bigoplus_p \cE^p, \phi := \oplus_p \phi_p )$ be its associated system of Hodge bundles. By functoriality, the holomorphic vector bundle $\cE nd(\cE)$ has a structure of system of Hodge bundles $\mathcal{E}nd(\cE) = \bigoplus_p \mathcal{E}nd(\cE)^p$ with Higgs field $\Phi$, which is associated to the variation of complex polarized Hodge structures $\End(\mathbb{V})$. One has $\mathcal{E}nd(\cE)^p = \bigoplus_k \mathcal{H}om(\cE^k , \cE^{k+p})$, and for any local holomorphic section $\Psi$ of $\mathcal{E}nd(\cE)$ and any local holomorphic section $X$ of $ \mathcal{T}_S$, we have $$\Phi(X)(\Psi)= \phi(X) \circ \Psi - \Psi \circ \phi(X)= [\phi(X) , \Psi] .$$ Similarly, the adjoint $\Phi^\ast$ of $\Phi$ is given by $$\Phi^\ast(X)(\Psi)= \phi^\ast(X) \circ \Psi - \Psi \circ \phi^\ast(X)= [\phi^\ast(X) , \Psi] .$$ In particular, we observe the following immediate application of Proposition \[curvature computation\].
\[bisectional curvature\] For any any local holomorphic section $X$ of $ \mathcal{T}_S$ and any local holomorphic section $\Psi$ of $\mathcal{E}nd(\cE)$ such that $ \Phi(X)(\Psi)= 0$, we have $$h_{H}(\Theta_{(\cE nd(\cE), h_H)}(X, \bar X) \Psi, \Psi) = - | [ \phi^\ast(\bar X), \Psi ] |_{h_H}^2 .$$
Note that this applies in particular to $\Psi = \phi(Y)$ for any holomorphic vector field $Y$ since $[\phi(X) , \phi(Y)] = 0$.
The Griffiths-Schmid pseudo-metric and its curvature
----------------------------------------------------
So far the computations hold more generally for any Higgs bundle $(\cE, \phi)$ equipped with a harmonic metric $h_H$. We will now take advantage of the decomposition $\cE = \bigoplus_p \cE^p$.
For any integer $p$, we denote by $h_p$ the pseudo-metric on $ \mathcal{T}_S$ obtained as the pull-back by $\phi_p$ of the Hodge metric on $\mathcal{H}om(\cE^p , \cE^{p-1})$. The sum of these pseudo-metrics, which coincides with the pseudo-metric on $ \mathcal{T}_S$ obtained as the pull-back by $\phi$ of the Hodge metric on $\mathcal{E}nd(\cE)$, is called the Griffiths-Schmid pseudo-metric on $S$ induced by $\bV$.
Note that by its very definition, the locus where the Griffiths-Schmid pseudo-metric is non-degenerate coincide with the Zariski-open (but maybe empty) subset of $S$ where the $\cO_S$-linear morphism $\phi : \cT_S \rightarrow \cE nd(\cE)$ is injective, or in other words with the locus where the period map associated to $\bV$ is immersive.\
Recall that one gets a correspondence by associating to any real $(1,1)$-form $\omega$ on $S$ the hermitian sesquilinear form $h$ on $\cT_S$ which satisfies $h(X,X) = - i \cdot \omega(X, \bar X)$ for any tangent vector of type $(1,0)$ (as usual we identify the holomorphic bundle with $T_{\bC}^{1,0}$). The following result is an easy consequence of Griffiths computation of the curvature of the Hodge metrics.
Through the correspondence between real $(1,1)$-form and hermitian sesquilinear forms on $\cT_S$, the curvature form of the Chern connection of $(\det \cF^p , h_H)$ corresponds to $h_p$. A fortiori, the curvature form of the Chern connection of the Griffiths line bundle $L_{\mathbb{V}} := \otimes_p \det \cF^p$ equipped with the Hodge metric corresponds to the Griffiths-Schmid pseudo-metric.
In other words, the Griffiths-Schmid pseudo-metric on $S$ is Kähler and its Kähler form is equal to the Chern curvature form of $(L_{\mathbb{V}} , h_H)$.
Over the locus where it is non-degenerate (which is Zariski-open but possibly empty), the Griffiths-Schmid metric has non-positive holomorphic bisectional curvature and negative holomorphic sectional curvature $- \gamma$ with $$\frac{1}{\gamma} \leqslant \frac{w^2}{4} \cdot \rank( \cV).$$
We restrict ourselves to the locus in $S$ where the period map is immersive. It follows readily from Proposition \[bisectional curvature\] that for any tangent vectors $X$ and $Y$ of type $(1,0)$, one has $$h_{H}(\Theta_{(\cE nd(\cE), h_H)}(X, \bar X) \phi(Y), \phi(Y)) = - | [ \phi^\ast(\bar X), \phi(Y) ] |_{h_H}^2 .$$ Since the Griffiths curvature decreases in subbundles, it follows that the Griffiths-Schmid metric has non-positive holomorphic bisectional curvature. Moreover, it follows that its holomorphic sectional curvature in the direction $X$ is less or equal to: $$\frac{1}{|X|_{h_H}^4} \cdot h_{H}(\Theta_{(\cE nd(\cE), h_H)}(X, \bar X) \phi(X), \phi(X)) =- \frac{1}{|X|_{h_H}^4} \cdot | [ \phi^\ast(\bar X), \phi(X) ] |_{h_H}^2 .$$ Therefore, we can rely on the next result to conclude the proof.
\[computation sectional curvature\] If $(V, F^p ,h)$ is a complex polarized Hodge structure of length $w$, then for any $u \in \End(V)^{-1} $: $$| [u, u^\ast] |^2_{h_H} \geq \frac{4}{w^2 \cdot \dim V} \cdot | u|^4_{h_H} .$$ Here $u ^\ast$ denotes the adjoint of $u$ with respect to the Hodge metric $h_H$.
Note that when $(V, F^p ,h)$ is autodual (this holds for example if it is a real polarized Hodge structure), then $ \frac{4}{w^2 \cdot \dim V} = \frac{2}{w \cdot \sum_{i= 1}^{w} \rank( F^p) }$.\
Before giving the proof of the lemma, we recall two well-known elementary applications of the Cauchy-Schwarz lemma.
\[Cauchy-Schwarz\] If $V$ is complex vector space equipped with a positive-definite hermitian sesquilinear form $h$ and $u = u^\ast \in \End(V)$, then the following inequality holds $$\dim(V) \cdot \tr(u^2) \geq \left( \tr(u) \right)^2,$$ with equality exactly when $u$ is an homothety.
This is a direct application of the Cauchy-Schwarz inequality $$\left( \tr(u v^\ast) \right) ^2 \leq \tr(u u ^\ast) \cdot \tr(v v^\ast)$$ with $v = id$.
\[inequality\] For any real numbers $a_1, \ldots, a_n$ and any positive real numbers $\lambda_1, \ldots, \lambda_n$, the following inequality holds: $$\sum_i \frac{a_i^2}{\lambda_i} \geq \frac{1}{\sum_i \lambda_i} \cdot \left( \sum_i a_i \right)^2,$$ with equality if and only if the $\frac{a_1}{\lambda_1} = \frac{a_2}{\lambda_2}= \cdots = \frac{a_n}{\lambda_n}$.
In view of the expression of the Hodge metric on $\End(V)$, we need to prove the following inequality:
$$\tr \left([u, u^\ast]^2 \right) \geq \frac{4}{w^2 \cdot \dim V} \cdot \left( \tr \left(u \circ u^\ast \right) \right) ^2 .$$
Let $r_i := \dim(V^i)$ for every integer $i$. Up to renumbering, one can assume that $r_i = 0$ if $i < 0$ or $i > w$. If one writes $ u = \oplus_{i} u_i$ according to the decomposition $$\End (V)^{-1} = \bigoplus_{i} \Hom \left(V^i, V^{i-1} \right),$$ so that $u_i = 0$ if $i >w$ or $i< 1$, we have that $ [u, u^\ast] = \oplus_{i} \left(u_{i+1} \circ u_{i+1}^\ast - u_{i}^\ast \circ u_{i} \right)$. Since the decomposition above is orthogonal, it follows that $$\tr \left( [u, u^\ast]^2 \right) = \sum \limits_{i=1}^w \tr \left( \left(u_{i+1} \circ u_{i+1}^\ast - u_{i}^\ast \circ u_{i} \right)^2 \right) \geq \sum_{i=1}^w \frac{1}{r_i} \cdot \left( \tr \left(u_{i+1} \circ u_{i+1}^\ast - u_i^\ast \circ u_i \right) \right)^2 ,$$ where the last inequality is obtained by applying Lemma \[Cauchy-Schwarz\]. Letting $a_i := \tr \left( u_i \circ u_i^\ast \right) = \tr \left( u_{i}^\ast \circ u_{i} \right)$, we get the inequality $$\tr \left( [u, u^\ast]^2 \right) \geq \sum_{i=1}^w \frac{1}{r_i} \cdot \left( a_{i+1} - a_i \right)^2 .$$
On the other hand, with the help of Lemma \[inequality\], we have for any integer $p$: $$\sum_{i \geq p} \frac{1}{r_i} \cdot \left(a_{i+1} - a_i \right)^2 \geq \frac{1}{ \sum_{i \geq p} r_i} \cdot \left( \sum_{i \geq p} (a_{i+1} - a_i) \right)^2 = \frac{a_p^2}{f_p}$$ and $$\sum_{i < p} \frac{1}{r_i} \cdot \left(a_{i+1} - a_i \right)^2 \geq \frac{1}{ \sum_{i < p} r_i} \cdot \left( \sum_{i < p} \left( a_{i+1} - a_i \right) \right)^2 = \frac{a_p^2}{\dim V - f_p} ,$$ hence by summing the two we get: $$\sum_{i=1}^w \frac{1}{r_i} \cdot \left( a_{i+1} - a_i \right)^2 \geq \left( \frac{1}{f_p} + \frac{1}{\dim V - f_p} \right) \cdot a_p^2 .$$ Another application of Lemma \[inequality\] shows that $$\frac{1}{f_p} + \frac{1}{\dim V - f_p} \geq \frac{4}{\dim V} ,$$ and by letting $p$ varying between $1$ and $w$, we get that $$w \cdot \left( \sum_{i=1}^w \frac{1}{r_i} \cdot (a_{i+1} - a_i)^2 \right) \geq \frac{4}{\dim V} \cdot \sum_{p = 1}^w a_p^2 .$$ A final application of Lemma \[inequality\] gives that $$\sum_{p = 1}^w a_p^2 \geq \frac{1}{w} \cdot \left( \sum_{p = 1}^w a_p\right) ^2 ,$$ so we finally get that $$\sum_{i=1}^w \frac{1}{r_i} \cdot \left( a_{i+1} - a_i \right)^2 \geq \frac{4}{w^2 \cdot \dim V} \cdot \left( \sum_{p = 1}^w a_p\right) ^2$$ and this finishes the proof.
Proof of the main Theorem {#sec:proof SMT for VHS}
=========================
In this section we prove Theorem \[thm:SMT for VHS\]. As a first step, by putting together Theorem \[thm:SMT for pseudo-metrics with negative holomorphic sectional curvature\] and Theorem \[thm:holomorphic sectional curvature for VHS\], we immediately deduce the following Arakelov-Nevanlinna inequality.
\[thm:Arakelov-Nevanlinna inequality\] Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$, $\Sigma \subset B$ a discrete subset of points and $\mathbb{V} = (\cL, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ be a variation of complex polarized Hodge structures of length $w$ on $ B - \Sigma$ with a non-constant period map. Then the first Chern form of the holomorphic line bundle $ L_{\mathbb{V}} = \otimes_p \det \cF^p$ equipped with the hermitian metric induced by $h$ extends as a current $ [L_{\mathbb{V}} ] $ on $B$, and there exists a positive real number $C$ such that the following inequality $$T_{[L_{\mathbb{V}} ]}(r) \leqslant \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot ( - \mathfrak{X}_{\sigma}(r) + N_{\Sigma}(r))+ C \cdot (\log T_{[L_{\mathbb{V}}]}(r)+ \log r) \ $$ holds for all $r \in \bR_{\geq 1}$ outside a subset of finite Lebesgue measure.
Note that in this statement we make no assumption on the monodromy of $\cL$.\
In view of Theorem \[thm:Arakelov-Nevanlinna inequality\], the proof of Theorem \[thm:SMT for VHS\] is now a consequence of the following result.
Let $X$ be a smooth projective complex algebraic variety and $\mathbb{V} = (\cL, \cF^{{\,\begin{picture}(-1,1)(-1,-3)\circle*{3}\end{picture}\ }}, h)$ be a variation of complex polarized Hodge structures defined on the complementary of a normal crossing divisor $D \subset X$. Assume that $\cL$ has unipotent monodromies around the irreducible components of $D$. We denote by $\bar \cF^p$ the canonical Deligne-Schmid extension of $\cF^p$ to $X$ for any integer $p$ and by $\bar L_{\mathbb{V}} = \otimes_p \det \bar \cF^p$ the canonical extension of the Griffiths line bundle of $\mathbb{V}$. Let $A$ be an ample line bundle on $ X$, $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function $\sigma$ and $f:B\to X$ be a non-constant holomorphic map such that $f(B)\not\subset D$. Then the first Chern form of the line bundle $ L_{\mathbb{V}}$ equipped with the hermitian metric induced by $h$ extends as a current $ [L_{\mathbb{V}} ] $ on $B$, and for any choice of a smooth hermitian metric on $\bar L_{\mathbb{V}} $ used to compute its Nevanlinna characteristic function, one has $$T_{f, \bar L_{\mathbb{V}} }(r) \leqslant T_{[L_{\mathbb{V}} ]}(r) +O(\log( T_{f,A}(r))).$$
This follows immediately from Lemma \[lem:singular metric with log growth\], since the assumption that $\cL$ has unipotent monodromies around the irreducible components of $D$ ensures that the (singular) Hodge metric on $\bar L_{\mathbb{V}} $ satisfies the growth assumption of the lemma.
Borel hyperbolicity {#sec:proof applications}
===================
In this section we prove Theorem \[thm:criterion for Borel hyperbolicity\] and Theorem \[thm:generalization of Nadel\]. We keep the notations of the statements.
Thanks to the main result of [@Ariyan-Robert], it is sufficient to prove that for any smooth complex algebraic curve $C$, any non-constant holomorphic map $f : C^{an} \rightarrow { X}^{an}$ such that $f (C^{an}) \not \subset (D \cup \B_{+})^{an}$ is algebraic.
Since $f (C^{an}) \not \subset (\B_{+})^{an}$, there exists an ample $\bQ$-divisor $A$ on $X$ and a section of some power of $ \bar L_{\mathbb{V}} (- ( \frac{w^2}{4} \cdot \operatorname{rank}\cL) \cdot D -A)$ that does not vanish on $f (C^{an})$. With the help of the *First Main Theorem*, it follows that for every $r \geq 1$: $$T_{f, A}(r) \leqslant T_{f, \bar L_{\mathbb{V}} (- ( \frac{w^2}{4} \cdot \operatorname{rank}\cL) \cdot D)}(r) \leqslant T_{f, \bar L_{\mathbb{V}}}(r) - \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot N^{[1]}_{f, D}(r) + O(1) .$$ On the other hand, by applying Theorem \[thm:SMT for VHS\] to $C^{an}$ equipped with the parabolic exhaustion function $\sigma$ constructed from a proper algebraic map $C \rightarrow \mathbb{A}^1$, we obtain that $$T_{f, \bar L_{\mathbb{V}}}(r) - \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot N^{[1]}_{f, D}(r) \leqslant_{{\rm exc}} \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot \left(- \mathfrak{X}_{B,\sigma}(r) \right) + O \left( \log r + \log T_{f, A}(r) \right) . \ $$ Therefore, we get from the two preceding inequalities that $$T_{f, A}(r) \leqslant_{{\rm exc}} \frac{w^2 \cdot \operatorname{rank}\cL}{4} \cdot \left(- \mathfrak{X}_{B,\sigma}(r) \right) + O \left( \log r + \log T_{f, A}(r) \right).$$ Since $-\mathfrak{X}_{\sigma}(r)=O(\log r)$, it follows from this inequality that there exists an ample line bundle $A^\prime$ on $X$ such that the inequality $ T_{f, A^\prime}(r) \leqslant \log r $ holds for all $r \in \bR_{\geq 1}$ outside a Borel subset of finite Lebesgue measure. Since $ T_{f, A^\prime}(r)$ is a convex increasing function in $\log r$, the function $r \mapsto T_{f, A^\prime}(r) / \log r$ is monotone increasing, so that the inequality $ T_{f, A^\prime}(r) \leqslant \log r $ holds in fact for all $r \in \bR_{\geq 1}$ sufficiently big. We conclude by the criterion [@Griffiths-King Proposition 5.9] that $f$ is algebraic.
First note the two following immediate results.
Let $X$ be a complex algebraic variety and $\bar X, \bar X^\prime$ two compactifications of $X$. Then $\bar X$ is Borel hyperbolic modulo $\bar X - X$ if and only if $\bar X^\prime$ is Borel hyperbolic modulo $\bar X^\prime - X$.
Let $\bar X$ be a complex algebraic variety and $\bar Y, \bar Z \subset \bar X$ two closed subvarieties. Then $\bar X$ is Borel hyperbolic modulo $\bar Z$ if and only if $\bar X$ is Borel hyperbolic modulo $\bar Z \cup \bar Y$ and $\bar Y$ is Borel hyperbolic modulo $\bar Y \cap \bar Z$.
By applying the second lemma with $\bar X$ a compactification of $X$, $\bar Z := \bar X - X$ and $\bar Y $ a closed subvariety that contains the singular locus of $\bar X$ and such that $\bar X \setminus \bar Y$ is affine, one sees that it is sufficient to consider the case where $X$ is smooth and affine. Fix a smooth projective compactification $\bar X$ of $X$ such that $D := \bar X - X$ is a normal crossing divisor. For every prime number $p$, let $\bar X(p)$ denote the normalization of $\bar X$ in the fraction field of $X(p)$ and $D(p) := \bar X(p) - X(p)$. Thanks to [@Bruni-level Proposition 2.4], the augmented base locus of the canonical extension of the Griffiths parabolic line bundle $ \bar L_{\mathbb{V}}$ is contained in $D$. A fortiori, if one takes a sufficiently small $\epsilon >0$, the augmented base locus of $ \bar L_{\mathbb{V}}(- \epsilon \cdot D)$ is also contained in $D$. On the other hand, by [@Bruni-level Theorem 5.1], the map $\pi_p : \bar X(p) \rightarrow \bar X$ ramifies over every irreducible component of $D$ with an order divisible by $p$ for almost all prime numbers $p$. Therefore, for almost all prime numbers $p$, the augmented base locus of the line bundle $(\pi_p^\ast \bar L_{\mathbb{V}})(- \epsilon \cdot p \cdot D(p)) = \bar L_{\pi_p ^\ast \mathbb{V}}(- \epsilon \cdot p \cdot D(p))$ is contained in $D(p)$, and we conclude using Theorem \[thm:criterion for Borel hyperbolicity\].
Second Main Theorem for hyperbolically embedded complements {#sec:SMTHyp}
===========================================================
In this section, we give another application of the results of Section \[sec:Nevanlinna\] by establishing a *Second Main Theorem* for hyperbolically embedded complements. In Section \[ssec:Kobayashi\], we prove a *Second Main Theorem* for the Kobayashi metric of the complement of a reduced divisor in a parabolic Riemann surface in case this complement is hyperbolic. Then, in Section \[ssec:SMTHyp\], we use this to establish a *Second Main Theorem* for pairs $(X,D)$ such that $X\setminus D$ is hyperbolically embedded in $X$.
Applications to the Kobayashi metric {#ssec:Kobayashi}
------------------------------------
Let $B$ be a Riemann surface and let $\Sigma$ be a reduced divisor on $B$. If $U:=B\setminus \Sigma$ is hyperbolic, then the universal cover of $U$ is given by the unit disc $\Delta$, and the $(1,1)$-form $\omega_{\Delta}=\frac{i}{2}\frac{1}{(1-|z|^2)^2}dz\wedge d\bar{z}$ associated to the Poincaré metric on $\Delta$ descends to the $(1,1)$-form $\omega_U$ associated to the Kobayashi metric on $U$. The Gaussian curvature $K$ of the Kobayashi metric on $U$ is constant and verifies $K=-4$. We therefore have $$\omega_U=-\frac{1}{4}\operatorname{Ric}\omega_U,$$ and the Schwarz lemma implies that $\omega_U$ has Poincaré singularities around every point of $\Sigma$. Moreover, one has $$=\frac{1}{4}[-\operatorname{Ric}\omega_U]= \frac{\pi}{2}[\Sigma]-\frac{1}{4}\operatorname{Ric}[\omega_U].\label{eq:RicciKoba}$$ The fact that the left hand side is smaller than the right hand side is a consequence of the Schwarz lemma. To see that one has in fact an equality , one can use the precise behavior of the Kobayashi metric around the points of $\Sigma$ (see for instance Section IV in [@FK92]).
In case $B$ is compact, integrating , one obtains the following (see for instance Section IV in [@FK92]) $$\int_B[\omega_U]= \frac{\pi}{2}(\deg \Sigma+2g(B)-2).\label{eq:GaussBonnet}$$ Lemma \[lem:RicciNegligeable\] allows us to give the following Nevanlinna theoretic analogue of this result.
\[cor:SMTKoba\] Let $B$ be a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function and let $\Sigma$ be a reduced divisor on $B$ such that $U:=B\setminus \Sigma$ is hyperbolic. Let $\omega_U$ be the $(1,1)$-form associated to the Kobayashi metric on $U$. Then, $\omega_U$ is locally integrable and one has $$T_{[\omega_U]}(r)\leqslant_{{\rm exc}} \frac{\pi}{2}\big(N_{\Sigma}(r) -\mathfrak{X}_{\sigma}(r)\big) +O\left(\log r+ \log T_{[\omega_U]}(r)\right).$$ In particular, for any $\ep >0$ one has $$(1-\ep)T_{[\omega_U]}\leqslant_{{\rm exc}} \frac{\pi}{2}(N_{\Sigma}(r)-\mathfrak{X}_{\sigma}(r)) + O\left(\log r \right).$$
For the first statement, one integrates and applies Lemma \[lem:RicciNegligeable\] to obtain $$\begin{aligned}
T_{[\omega_U]}(r)&=&\frac{\pi}{2}N_{\Sigma}(r)-\frac{1}{4}\operatorname{Ric}[\omega_U]\leqslant_{\rm exc}\frac{\pi}{2}\big(N_{\Sigma}(r)- \mathfrak{X}_{\sigma}(r)\big)+O\left(\log r+ \log T_{[\omega_U]}(r)\right) \end{aligned}$$ The second statement follows since for every $\ep>0$, $$\log T_{[\omega_U]}(r)\leqslant_{{\rm exc}}\ep T_{[\omega_U]}(r).$$
Second Main Theorem for hyperbolically embedded varieties {#ssec:SMTHyp}
---------------------------------------------------------
Let $X$ be a smooth complex projective variety. Recall that an open subset $U\subset X$ is hyperbolically embedded in $X$ if for every hermitian $h$ metric on $X$, there exists $\eta>0$ such that $\eta h\leqslant F_U$, where $F_U$ denotes the Kobayashi-Royden infinitesimal pseudo-metric on $U$. We refer to Chapter 3 of [@Kob98] for the definition of the Kobayashi-Royden pseudo-metric and to Theorem 3.3.3 in [@Kob98] or Proposition 16 in [@PR07] for the proof of the equivalent characterization we use here.
The algebraic counterpart of the *Second Main Theorem* we shall prove below was established by Pacienza and Rousseau in [@PR07 Theorem 5]. In order to emphasize the analogies between the analytic and the algebraic sides both in statements and proofs, we state and reprove this result of Pacienza and Rousseau in the next theorem.
\[thm:SMTHyp\] Let $X$ be a smooth projective variety and let $H$ be a reduced divisor on $X$ such that $X\setminus H$ is hyperbolically embedded in $X$. Let $A$ be an ample line bundle on $X$. Then
1. (Pacienza-Rousseau). There exists a constant $\alpha_{\rm alg} >0$ such that for any projective curve $B$ and any non-constant algebraic map $f:B\to X$ such that $f(B)\not\subset H$, one has $$\deg f^*A\leqslant \alpha_{\rm alg}\big(\deg (f^*H)_{\rm red} - \chi(B)\big).$$
2. There exists a constant $\alpha_{\rm an} >0$ such that for any non-compact parabolic Riemann surface $B$ equipped with a parabolic exhaustion function and any non-constant holomorphic map $f:B\to X$ such that $f(B)\not\subset H$, one has $$T_{f,A}(r)\leqslant_{\rm exc} \alpha_{\rm an}\big(N^{(1)}_{f,H}(r) - \mathfrak{X}_\sigma(r)\big)+O(\log r).$$
Let $B$ be either a smooth projective curve or a parabolic Riemann surface with a non-constant holomorphic map $f:B\to X$ such that $f(B)\not\subset H$. Let $ \Sigma :=(f^*H)_{\rm red}$ be the set theoretical inverse image of the divisor $H$. Let us set $U=B\setminus \Sigma$ and $V=X\setminus H$. By restriction, $f$ induces a non-constant holomorphic map ${f}|_U:U\to V$. Since $V$ is hyperbolic, so is $U$. Let us denote by $\omega_{U}$ the $(1,1)$-form associated to the Kobayashi metric on $U$ and let us denote by $\|\cdot\|_U$ the associated norm. Let us denote by $F_V$ the Kobayashi Royden infinitesimal pseudo-norm on $V$. The distance decreasing property of the Kobayashi-Royden pseudo-metric implies that $$f^*F_{V}\leqslant \|\cdot\|_{h_{U}}.$$
Let $A$ be an ample line bundle on $X$. Since $A$ is ample, it admits a hermitian metric with positive curvature $\omega_A$. Since $\omega_A$ is a positive $(1,1)$-form, it induces a hermitian metric $h_A$ on $X$. Since $V$ is hyperbolically embedded in $X$, there exists a positive real number $\eta>0$ such that $$\eta \|\cdot\|_{h_A}|_{V}\leqslant F_{V}.$$ Therefore, one has $$\eta f^*\|\cdot\|_{h_A}|_{U}\leqslant \|\cdot\|_{{U}}.$$ At the level of forms, this yields $$\eta f^*\omega_A|_{U}\leqslant \omega_{U}.\label{eq:FormulePreuve}$$ If $B$ is projective, it suffices to integrate and apply $$\eta \deg f^*A=\eta \int_{B}f^*\omega_A=\eta \int_{U}f^*\omega_A\leqslant \int_{U}\omega_{U}=\int_{B}[\omega_{U}]\leqslant \frac{\pi}{2}(\deg \Sigma+2g(B)-2).$$ Taking $\alpha_{\rm alg}=\frac{\pi}{2 \eta}$ yields the first assertion.
If $B$ is a non-compact parabolic Riemann surface equipped with a parabolic exhaustion function, it suffices to apply $\int_{1}^r\int_{B(t)}\cdot \frac{dt}{t}$ to and apply Corollary to obtain that for any $\ep>0$ $$\eta T_{f,A}(r) \leqslant T_{[\omega_{U}]}(r)\leqslant_{\rm exc} \frac{\pi}{2(1-\ep)} \big(N^{(1)}_{f,H}(r)- \mathfrak{X}_{\sigma}(r)\big) + O(\log r).$$ It suffices to take $\alpha_{\rm an}:=\frac{\pi}{2\eta(1-\ep)}$ to conclude the proof of the second assertion.
[^1]: In this article, we use the $d^c$ operator defined by $d^c=\frac{i}{4\pi}(\bar{\partial}-\partial)$, so that $dd^c=\frac{i}{2\pi}\partial\bar{\partial}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article, we show that a hypersurface of the nearly K[ä]{}hler ${\mathbb{C}\mathrm{P}}^3$ or ${\mathbb{F}_{1,2}}$ cannot have its shape operator and induced almost contact structure commute together. This settles the question for six-dimensional homogeneous nearly K[ä]{}hler manifolds, as the cases of ${\mathbb{S}}^6$ and ${\mathbb{S}}^3 \times {\mathbb{S}}^3$ were previously solved, and provides a counterpart to the more classical question for the complex space forms ${\mathbb{C}\mathrm{P}}^n$ and ${\mathbb{C}\mathrm{H}}^n$. The proof relies heavily on the construction of ${\mathbb{C}\mathrm{P}}^3$ and ${\mathbb{F}_{1,2}}$ as twistor spaces of ${\mathbb{S}^{4}}$ and ${\mathbb{C}\mathrm{P}}^2$'
address: 'Univ. Brest, CNRS UMR 6205, LMBA, F-29238 Brest, France'
author:
- 'G. Deschamps'
- 'E. Loubeau'
title: 'Hypersurfaces of nearly K[ä]{}hler twistor spaces'
---
Introduction
============
The classical study of real hypersurfaces in complex space forms has led to extensive lists by Takagi [@Takagi1; @Takagi2] (for ${\mathbb{C}\mathrm{P}}^n$) and Montiel [@Montiel] (for ${\mathbb{C}\mathrm{H}}^n$).
Driven by the number of principal curvatures and the importance of Hopf hypersurfaces, i.e. when the ambient complex structure maps the normal vector field to a principal direction, hypersurfaces of ${\mathbb{C}\mathrm{P}}^n$ or ${\mathbb{C}\mathrm{H}}^n$ where the shape operator $A$ and the induced almost contact structure $\varphi$ commute constitute a remarkable class, amenable to classification.
Indeed, by [@Cecil2015 Th. 6.19] their principal curvatures must be constant and in twos or threes. Moreover, they must belong to type A of the Takagi-Montiel lists (cf. [@Cecil2015 Th. 8.37] as well)
An almost Hermitian manifold $(Z,\mathbb I,g)$ is called nearly K[ä]{}hler [@Gray1970] if $\nabla \mathbb I$ is anti-symmetric. The best-known (non-K[ä]{}hler) example is the round sphere ${\mathbb{S}}^6$ with its canonical metric and the structure that comes from octonion multiplication.
In view of the classical theory for complex space forms, it is natural to ask which hypersurfaces of nearly K[ä]{}hler manifolds satisfy $A\varphi =\varphi A$.
Nearly K[ä]{}hler manifolds enjoy many topological and geometric properties akin to K[ä]{}hler geometry (cf. [@Gray1976]) and have known a recent revival of interest with the structure theorem of Nagy [@Nagy2002] in 2002, which shows that six-dimensional nearly K[ä]{}hler manifolds act as building blocks, and Butruille’s 2005 classification of homogeneous nearly K[ä]{}hler six-manifolds [@Butruille2005], namely ${\mathbb{S}}^6$, ${\mathbb{S}}^3 \times {\mathbb{S}}^3$, ${\mathbb{C}\mathrm{P}}^3$ and ${\mathbb{F}_{1,2}}$.
While the explicit construction of the nearly K[ä]{}hler structure (and metric) on ${\mathbb{S}}^3 \times {\mathbb{S}}^3$ is rather involved and ad-hoc, the ${\mathbb{C}\mathrm{P}}^3$ and ${\mathbb{F}_{1,2}}$ examples both have their origin in twistor theory, as twistor spaces of ${\mathbb{S}}^4$ and ${\mathbb{C}\mathrm{P}}^2$.
In the case of the four-dimensional sphere, as its unitary frame bundle is ${{\rm SO}}(5)$, its twistor space is the associated bundle $${{\rm SO}}(5) \times_{{{\rm SO}}(4)} {{\rm SO}}(4)/{{\rm U}}(2) \simeq {{\rm SO}}(5)/{{\rm U}}(2)$$ which is ${\mathbb{C}\mathrm{P}}^3$ and the twistor projection ${\mathbb{C}\mathrm{P}}^3 \to {\mathbb{S}}^4$ is ${\mathrm{span}}_{{\mathbb{C}}} v \mapsto {\mathrm{span}}_{{\mathbb{H}}} v$, where ${\mathbb{S}}^4 \simeq {\mathbb{H}}\mathrm{P}^1$ by the Hopf map. When the spaces are equipped with their canonical metrics, this projection is a Riemannian submersion.
For the two-dimensional complex projective space, one considers $y\in {\mathbb{C}\mathrm{P}}^2$ and $(x,y,z)$ mutually orthogonal complex lines in ${\mathbb{C}}^3$. Identifying a complex structure in $T_{y} {\mathbb{C}\mathrm{P}}^2$ with a choice of holomorphic and anti-holomorphic bundles, one shows that $$\label{hom1}
T^{1,0}_{y} {\mathbb{C}\mathrm{P}}^2 = {\mathrm{Hom}}(y,x) \oplus {\mathrm{Hom}}(y,z)$$ and $$\label{hom2}
T^{0,1}_{y} {\mathbb{C}\mathrm{P}}^2 = {\mathrm{Hom}}(x,y) \oplus {\mathrm{Hom}}(z,y) .$$ There is a one-one correspondence between triples $(x,y,z)$ and couples $(l,p)$, where $l$ is a complex line and $p$ a complex plane in ${\mathbb{C}}^3$ with $l \subset p$, i.e. the flag manifold ${\mathbb{F}_{1,2}}$.
Since ${\mathbb{C}\mathrm{P}}^2$ is self-dual [@Atiyah], the integrable almost Hermitian structure is defined by taking the standard Hermitian structure on $T^{1,0}_{y} {\mathbb{C}\mathrm{P}}^2$ and its opposite on $T^{0,1}_{y} {\mathbb{C}\mathrm{P}}^2 $. Because of this orientation reversal, this identifies ${\mathbb{F}_{1,2}}$ with $Z(\overline{{\mathbb{C}\mathrm{P}}^2})$, and the twistor projection $(l\subset p) \in {\mathbb{F}_{1,2}}\mapsto l^\perp \cap p \in {\mathbb{C}\mathrm{P}}^2$ is also a Riemannian submersion.
There is a general procedure, due to [@Eells-Salamon] and [@Nagy2002], to produce nearly K[ä]{}hler manifolds: If $(Z,{\mathbb I_1}, g_1)$ is a K[ä]{}hler manifold with a Riemannian foliation $\mathcal{F}$, which induces an (${\mathbb I_1}$-invariant) integrable distribution ${\mathcal{V}}$ and its orthogonal complement ${\mathcal{H}}$, then the Riemannian metric $$g_2 (X,Y) = \frac12 g_1 (X,Y) \quad \forall X,Y \in {\mathcal{V}}$$ and $$g_2 (X,Y) = g_1 (X,Y) \quad \forall X,Y \in {\mathcal{H}}$$ together with the almost complex structure $${\mathbb I_2}X = - {\mathbb I_1}X \quad \forall X \in {\mathcal{V}}\qquad \text{and} \qquad
{\mathbb I_2}X = {\mathbb I_1}X \quad \forall X\in {\mathcal{H}}$$ make $(Z,{\mathbb I_2},g_2)$ into a nearly K[ä]{}hler manifold.
According to Hitchin [@Hitchin], ${\mathbb{C}\mathrm{P}}^3$ and ${\mathbb{F}_{1,2}}$ are the only compact twistor spaces $(Z,{\mathbb I_1}, g_1)$ to be K[ä]{}hler and, therefore, the only ones to admit a nearly K[ä]{}hler structure.
Let $(Z,{\mathbb I_2},g)$ be a nearly K[ä]{}hler manifold and $H \hookrightarrow Z$ a hypersurface. Call $N$ the unit normal to $H$ and then define an almost contact (metric) structure $\varphi$ on $H$ by: $$\varphi X = {\mathbb I_2}X - g({\mathbb I_2}X,N)N, \quad \forall X\in TH.$$ One easily verifies that $$g(\varphi X, \varphi Y) = g(X,Y)\quad \forall X,Y\in TH\cap ({\mathbb I_2}N)^\perp,$$ or more generally $$g(\varphi A, B) = g(A, \varphi B)\quad \forall A,B\in TH,$$ as well as $$\varphi ({\mathbb I_2}N)=0.$$ The other fundamental tensor is the shape operator $A$ of $H$: $$AX = - \nabla_{X}^{Z} N ,$$ so that $$\nabla_{X}^{Z} Y = \nabla_{X}^{H} Y + g(AX,Y)N .$$
An immediate remark on hypersurfaces of nearly K[ä]{}hler manifolds which satisfy $A\varphi =\varphi A$ is that the Hopf vector field ${\mathbb I_2}N$ has to be an eigenvector of $A$ (of eigenvalue $\mu$) and the eigenspaces of $A|_{({\mathbb I_2}N)^{\perp}}$ must be ${\mathbb I_2}$-stable.
In dimension $6$, since we have a full classification of homogeneous nearly K[ä]{}hler [@Butruille2005], the first two cases of the list, ${\mathbb{S}}^6$ and ${\mathbb{S}}^3 \times {\mathbb{S}}^3$ have already been investigated.
Combining results of [@Berndt1995] and [@Martins2001], shows that the only hypersurfaces of ${\mathbb{S}}^6$ with $A\varphi =\varphi A$ are (open parts of) geodesic spheres.
For the nearly K[ä]{}hler ${\mathbb{S}}^3 \times {\mathbb{S}}^3$, that is equipped with the right metric and almost complex structure, its hypersurfaces with $A\varphi =\varphi A$ must be locally given by the canonical immersion of ${\mathbb{S}}^2 \times {\mathbb{S}}^3$ in ${\mathbb{S}}^3 \times {\mathbb{S}}^3$ ([@Hu2017]). Note that this classification contains three immersions but, by [@Moruz2018], they turn out to be all isometric one to the other.
There exists an almost contact counterpart to the nearly K[ä]{}hler condition, coined nearly cosymplectic (and defined by $\nabla\varphi$ being antisymmetric). By [@Blair1971], they must satisfy $A\varphi =\varphi A$ but while ${\mathbb{S}}^5 \hookrightarrow {\mathbb{S}}^6$ is well-known to be nearly cosymplectic, the hypersurface ${\mathbb{S}}^2 \times {\mathbb{S}}^3 \hookrightarrow{\mathbb{S}}^3 \times {\mathbb{S}}^3$ is not, as a quick inspection of the eigenvalues of its shape operator reveals.
The objective of this article is to extend these results to the remaining two homogeneous nearly K[ä]{}hler six-manifolds and to prove the following theorem.
Let $Z(M)$ be the nearly K[ä]{}hler manifold ${\mathbb{C}\mathrm{P}}^3$ or ${\mathbb{F}_{1,2}}$. Then there exists no hypersurface $H \hookrightarrow Z(M)$ such that its shape operator $A$ and the induced almost contact structure $\varphi$ commute: $$A\varphi =\varphi A .$$
A direct consequence is that this construction produces only one example of nearly cosymplectic almost contact hypersurface.
The only nearly cosymplectic hypersurface of a homogeneous $6$-dimensional nearly K[ä]{}hler manifold is ${\mathbb{S}}^5 \hookrightarrow {\mathbb{S}}^6$.
As a byproduct of the Theorem, we obtain information on the eigenvalues of the shape operator $A$.
There is no totally geodesic or totally umbilical hypersurface of the nearly K[ä]{}hler manifolds ${\mathbb{C}\mathrm{P}}^3$ or ${\mathbb{F}_{1,2}}$.
Curvature properties of nearly K[ä]{}hler ${\mathbb{C}\mathrm{P}}^3$ and ${\mathbb{F}_{1,2}}$
=============================================================================================
Throughout the rest of this article, we specialize to the cases $M={\mathbb{S}^{4}}$ and $M=\overline{{\mathbb{C}\mathrm{P}}^2}$. Let $Z(M)$ be the twistor space of $M$, equiped with the Riemannian metric [@Apostolov98] $$g_t = \pi^* g_M + t g_{{\mathbb{C}\mathrm{P}}^1}, \quad (t>0) .$$ Two almost complex structures can be defined on $Z(M)$: First the Atiyah-Hitchin-Singer structure ${\mathbb I_1}$ on $T_{(x_0,I)} Z(M)$, with $x_0 \in M$ and $I$ a complex structure on $T_{x_0} M$, defined by $$\left\{\begin{array}{ccc}
{\mathbb I_1}X &=& IX \quad \text{if } X\in {\mathcal{H}}\\
{\mathbb I_1}Y&=& \mathbb{J}_{{\mathbb{C}}\mathrm{P}^1} Y,\quad \text{if } Y\in {\mathcal{V}}\end{array}\right.$$ where we identify vectors tangent to $M$ with their horizontal lifts in ${\mathcal{H}}\subset T_{(x_0,I)} Z(M)$; Second the Eells-Salamon structure [@Eells-Salamon]: $$\left\{\begin{array}{ccc}
{\mathbb I_2}&=& {\mathbb I_1}\quad \text{on } {\mathcal{H}}\\
{\mathbb I_2}&=& - {\mathbb I_1}\quad \text{on } {\mathcal{V}}.
\end{array}\right.$$
Then, as the cases we consider are anti-self dual, [@Atiyah] shows that $(Z(M), g_t, {\mathbb I_1})$ is a K[ä]{}hler manifold for $t= \frac{12}{s}$, ($s = {\mathrm{scal}}_{(M,g_M)}$), while [@Friedrich; @Nagy2002] prove that $(Z(M),g_t, {\mathbb I_2})$ is nearly K[ä]{}hler for $t= \frac{6}{s}$.
The next proposition relates the curvature tensors of the twistor space and the base manifold, in terms of the nearly K[ä]{}hler structure. This will lead to crucial curvature properties in Lemma \[lemma2\].
\[prop1\] Let $Z(M)$ be the twistor space of ${\mathbb{S}^{4}}$ or ${\mathbb{C}\mathrm{P}}^2$. Write $g= g_{\tfrac{6}{s}}$ so that $(Z(M),{\mathbb I_2}, g)$ is nearly K[ä]{}hler and denote by $R$ and $R^M$ the respective curvature tensors of $(Z(M),g)$ and $(M,g_M)$.
Let $X,Y,Z,T \in T_p Z(M)$ ($p\in Z(M)$) then
$$\begin{aligned}
&R(X,Y,Z,T)=R^M\Big(d\pi(X),d\pi(Y),d\pi(Z),d\pi(T)\Big)\\
&+2(b+2a)g^h(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)
+(b+2a)g^h(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)\\
&-(b+2a)g^h(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)
+(c-5b)\Big(g^h(X,Z)g^h(Y,T)-g^h(X,T)g^h(Y,Z)\Big)\\
&-2ag^h(\mathbb I_2X,Y)g(\mathbb I_2 Z,T)
-2a g(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)
-ag^h(\mathbb I_2X,Z)g(\mathbb I_2Y,T)\\
&-ag(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)
+ag^h(\mathbb I_2X,T)g(\mathbb I_2Y,Z)+ag(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)\\
&+(b-c)\Big(g^h(X,Z)g(Y,T)+g(X,Z)g^h(Y,T) -g^h(X,T)g(Y,Z)-g(X,T)g^h(Y,Z)\Big)\\
&+c\Big(g(X,Z)g(Y,T)-g(X,T)g(Y,Z)\Big),\end{aligned}$$
where $$\begin{aligned}
a&= \frac{s}{24}-t\Big(\frac{s}{24}\Big)^2 ; b= t\Big(\frac{s}{24}\Big)^2 ;c= \frac1t ,\end{aligned}$$ with $t = \frac{6}{s}$
We rely on the formula of [@Apostolov98]: $$\begin{aligned}
&R(X,Y,Z,T)=R^M\Big(d\pi(X),d\pi(Y),d\pi(Z),d\pi(T)\Big)\\
&+2ag^h(\mathbb I_1X,Y)g^v(\mathbb I_1Z,T)+2ag^v(\mathbb I_1X,Y)g^h(\mathbb I_1Z,T)
+ag^h(\mathbb I_1X,Z)g^v(\mathbb I_1Y,T)\\
&+ag^v(\mathbb I_1X,Z)g^h(\mathbb I_1Y,T)
-ag^h(\mathbb I_1X,T)g^v(\mathbb I_1Y,Z)-ag^v(\mathbb I_1X,T)g^h(\mathbb I_1Y,Z)\\
&+2bg^h(\mathbb I_1X,Y)g^h(\mathbb I_1Z,T)+bg^h(\mathbb I_1X,Z)g^h(\mathbb I_1Y,T)
-bg^h(\mathbb I_1X,T)g^h(\mathbb I_1Y,Z)\\
&+bg^h(X,Z)g^v(Y,T)+bg^v(X,Z)g^h(Y,T)
-bg^h(X,T)g^v(Y,Z)-bg^v(X,T)g^h(Y,Z)\\
&-3bg^h(X,Z)g^h(Y,T)+cg^v(X,Z)g^v(Y,T)
+3bg^h(X,T)g^h(Y,Z)-cg^v(X,T)g^v(Y,Z),\end{aligned}$$ where $\mathbb I_1$ is the K[ä]{}hler structure on $Z(M)$.
Since ${\mathbb I_1}$ and ${\mathbb I_2}$ agree on the horizontal distribution and are opposite on ${\mathcal{V}}$, we have
$$\begin{aligned}
&R(X,Y,Z,T)=R^M\Big(d\pi(X),d\pi(Y),d\pi(Z),d\pi(T)\Big)\\
&-2ag^h(\mathbb I_2X,Y)g^v(\mathbb I_2Z,T)-2ag^v(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)
-ag^h(\mathbb I_2X,Z)g^v(\mathbb I_2Y,T)\\
&-ag^v(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)
+ag^h(\mathbb I_2X,T)g^v(\mathbb I_2Y,Z)+ag^v(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)\\
&+2bg^h(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)
+bg^h(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)
-bg^h(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)\\
&+bg^h(X,Z)g^v(Y,T)+bg^v(X,Z)g^h(Y,T)
-bg^h(X,T)g^v(Y,Z)-bg^v(X,T)g^h(Y,Z)\\
&-3bg^h(X,Z)g^h(Y,T)+cg^v(X,Z)g^v(Y,T)
+3bg^h(X,T)g^h(Y,Z)-cg^v(X,T)g^v(Y,Z)\\\end{aligned}$$
We use the shorthand $g^h(X,Y) = g(X^h , Y^h)$ and $g^v(X,Y) = g(X^v , Y^v)$ for $X= X^h + X^v$ its decomposition in the horizontal and vertical distributions. Since $g(X,Y) = g^h(X,Y) + g^v(X,Y)$, we have $$\begin{aligned}
&R(X,Y,Z,T)=R^M\Big(d\pi(X),d\pi(Y),d\pi(Z),d\pi(T)\Big)\\
&+4ag^h(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)+2ag^h(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)
-2ag^h(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)\\
&-2ag^h(\mathbb I_2X,Y)g(\mathbb I_2 Z,T)-2a g(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)
-ag^h(\mathbb I_2X,Z)g(\mathbb I_2Y,T)\\
&-ag(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)
+ag^h(\mathbb I_2X,T)g(\mathbb I_2Y,Z)+ag(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)\\
&+2bg^h(\mathbb I_2X,Y)g^h(\mathbb I_2Z,T)
+bg^h(\mathbb I_2X,Z)g^h(\mathbb I_2Y,T)
-bg^h(\mathbb I_2X,T)g^h(\mathbb I_2Y,Z)\\
&-2bg^h(X,Z)g^h(Y,T)+2bg^h(X,T)g^h(Y,Z)
+bg^h(X,Z)g(Y,T)+bg(X,Z)g^h(Y,T)\\
&-bg^h(X,T)g(Y,Z)-bg(X,T)g^h(Y,Z)+(c-3b)g^h(X,Z)g^h(Y,T)\\
&-(c-3b)g^h(X,T)g^h(Y,Z)\\
&+c(g(X,Z)g(Y,T)-g(X,Z)g^h(Y,T)-g^h(X,Z)g(Y,T))\\
&-c(g(X,T)g(Y,Z)-g(X,T)g^h(Y,Z)-g^h(X,T)g(Y,Z)), \end{aligned}$$ and reorganising terms yields the proposition.
Let $H \hookrightarrow Z(M)$ be a hypersurface of $(Z(M),{\mathbb I_2},g)$ satisfying $$\label{eq-comm}
A\varphi = \varphi A.$$ We call $N$ the normal to $H$ and Equation implies that ${\mathbb I_2}N$ is an eigenvector of $A$ (of eigenvalue $\mu$). We denote by $\lambda$ an eigenvalue of $A$ and observe that the eigenspace $E_\lambda\cap (\mathbb I_2 N)^\perp$ is ${\mathbb I_2}$-invariant.
\[lemma2\] Let $X\in E_\lambda \cap (\mathbb I_2 N)^\perp$ then $$R({\mathbb I_2}N,X,{\mathbb I_2}X,N) = - R({\mathbb I_2}N,{\mathbb I_2}X,X,N),$$ and $R({\mathbb I_2}N,X,{\mathbb I_2}X,N) = \lambda (\lambda - \mu) \Vert X \Vert^2$.
Since both $X$ and ${\mathbb I_2}X$ belong to $E_\lambda$, the Codazzi Equation gives $$\begin{aligned}
R({\mathbb I_2}N,X,{\mathbb I_2}X,N) &= g((\nabla_{{\mathbb I_2}N}A)(X)-(\nabla_{X} A)({\mathbb I_2}N),{\mathbb I_2}X)\\
&= -g((\nabla_{X} A)({\mathbb I_2}N),{\mathbb I_2}X)\end{aligned}$$ since $A|_{E_\lambda} = \lambda {\mathrm{id}}_{E_\lambda}$.
Therefore $$R({\mathbb I_2}N,X,{\mathbb I_2}X,N) = (\lambda - \mu) g(\nabla_X {\mathbb I_2}N, {\mathbb I_2}X),$$ which is ${\mathbb I_2}$-invariant since $$\begin{aligned}
\lambda \Vert X\Vert^2 = g(AX,X) &= - g( \nabla_X N , X) = g ( (\nabla_{X} {\mathbb I_2})({\mathbb I_2}N) + {\mathbb I_2}\nabla_{X} {\mathbb I_2}N , X ) \\
&= - g ( (\nabla_{{\mathbb I_2}N} {\mathbb I_2})(X), X) - g(\nabla_X {\mathbb I_2}N , {\mathbb I_2}X) \\
&= - g(\nabla_X {\mathbb I_2}N , {\mathbb I_2}X).\end{aligned}$$
Motivated by the results of Lemma \[lemma2\], we use Proposition \[prop1\] to obtain the following curvature expression.
\[cor1\] If $X$ is a vector field in $(N,\mathbb I_2 N)^\perp$, we have $$\begin{aligned}
&R(\mathbb I_2 N,\mathbb I_2X,X,N)=R^M\Big(d\pi(\mathbb I_2 N),d\pi(\mathbb I_2 X),d\pi(X),d\pi(N)\Big)\\
&+(b+2a)g^h(N,X)^2+(-4a+3b-c)g^h(N,\mathbb I_2 X)^2
+a(\Vert N^h \Vert^2 \Vert X\Vert^2+\Vert X^h\Vert^2)\\
&-(2a+b)\Vert N^h \Vert^2\Vert X^h\Vert^2.\end{aligned}$$
From this symmetry of the curvature tensor, we can eliminate vertical normal vector fields.
\[proppage6\] Let $H$ be a hypersurface of $Z(M)$ such that $A\varphi = \varphi A$. Then the normal vector $N$ cannot be vertical.
If $N$ is vertical then so is ${\mathbb I_2}N$ and all eigenvectors orthogonal to it must be horizontal. But for such an eigenvector $X$ associated to the eigenvalue $\lambda$ and orthogonal to ${\mathbb I_2}N$ $$\begin{aligned}
\lambda \Vert X\Vert^2& =g(AX,X)=g(-\nabla_{X}N,X)=g(N,\nabla_{X}X)\\
&=g(N,\frac12 [X,X])=0,\end{aligned}$$ by O’Neill [@Oneill] since $X$ is horizontal. However, this implies, by Lemma \[lemma2\], that $R(\mathbb I_2N,\mathbb I_2X,X,N)$ is zero, that is, by $N^h=0$ and Corollary \[cor1\] $$a \Vert X^h\Vert^2 =0 ,$$ which contradicts the fact that $X$ is horizontal, since $a\neq 0$.
The case $(M,Z(M))= ({\mathbb{S}^{4}}, {\mathbb{C}\mathrm{P}}^3)$
=================================================================
As ${\mathrm{scal}}_{{\mathbb{S}^{4}}} = 12$, the constants in Proposition \[prop1\] take on the values $ a= \frac{3}{8}, b=\frac{1}{8}$ and $c= 2$. Since $$R^{{\mathbb{S}^{4}}}(U,V,W,T)=g(V,W)g(U,T)-g(U,W)g(V,T) \quad \forall U,V,W,T \in T_p {\mathbb{S}^{4}},$$ we have $$\begin{aligned}
\pi^* R^{{\mathbb{S}^{4}}}(\mathbb I_2 N,\mathbb I_2 X, X,N)&=g^h(\mathbb I_2 X,N)^2 .\end{aligned}$$
From Corollary \[cor1\], we obtain that for any $X\in (N,\mathbb I_2 N)^\perp$ $$\begin{aligned}
R(\mathbb I_2 N,\mathbb I_2 X, X,N) &=
\frac{7}8g^h(X,N)^2-\frac{17}8 g^h(\mathbb I_2 X,N)^2
+\frac38\Big(\Vert X^h\Vert^2+\Vert N^h\Vert^2\Vert X\Vert^2\Big)\\
&-\frac78\Vert X^h\Vert^2\Vert N^h\Vert^2\end{aligned}$$ and $$\begin{aligned}
- R(\mathbb I_2 N,X,\mathbb I_2 X,N) &=
-\frac{17}8g^h(X,N)^2+\frac{7}8 g^h(\mathbb I_2 X,N)^2
+\frac38\Big(\Vert X^h\Vert^2+\Vert N^h\Vert^2\Vert X\Vert^2\Big)\\
&-\frac78\Vert X^h\Vert^2\Vert N^h\Vert^2 .\end{aligned}$$ By Lemma \[lemma2\], when $X\in E_\lambda \cap (\mathbb I_2 N)^\perp$ $$g^h(X,N)^2=g^h(\mathbb I_2X, N)^2.$$ As this must remain true for the eigenvector $X+\mathbb I_2 X$, we infer that $$\label{Eq}
g^h(X,N)=g^h(\mathbb I_2X, N)=0.$$
We then easily prove that the vertical component of the normal vector field must be zero.
\[lemmaA\] Let $H$ be a hypersurface of ${\mathbb{C}\mathrm{P}}^3$ such that $A\varphi = \varphi A$. Then the normal vector field must be horizontal.
If $N^v\neq 0$, then $(N^v ,{\mathbb I_2}N^v)$ is a basis of the vertical distribution. But Equation forces $$g^v(X,N)=g^v(\mathbb I_2X, N)=0$$ as $X$ is an eigenvector orthogonal to $N$ and ${\mathbb I_2}N$, so $X$ must be horizontal. Since this applies to all eigenvectors of $A$ in $(\mathbb I_2 N)^\perp$, they must be horizontal and orthogonal to $N$, hence $N^h$ must vanish, and we conclude with Proposition \[proppage6\].
The complementary contingency is resolved using tools from twistor theory.
\[lemmaB\] Let $H$ be a hypersurface of ${\mathbb{C}\mathrm{P}}^3$ such that $A\varphi = \varphi A$. Then $N$ cannot be horizontal.
If $N^v =0$, then for any horizontal $X$, we have by O’Neill $$(AX)^v = (-\nabla_X N )^v = -\frac12 ([X,N])^v.$$ Let $p= (x_0 ,I) \in H \subset {\mathbb{C}\mathrm{P}}^3 = Z({\mathbb{S}^{4}})$, $x_0\in {\mathbb{S}^{4}}$ and $I$ a complex structure on $T_{x_0}{\mathbb{S}^{4}}$. Take a positive orthonormal frame $(e_1,e_2,e_3,e_4)$ of $T_{x_0}{\mathbb{S}^{4}}$ such that at $p$ : $$e_1= d\pi(N), e_2 = I e_1, e_3\in (e_1,e_2)^\perp , e_4=I e_3.$$ Let ${\mathcal{V}}_p$ be the vertical space at $p\in {\mathbb{C}\mathrm{P}}^3$, i.e. the tangent space to the fibre. We identify $\bigwedge^2 T_{x_0}{\mathbb{S}^{4}}$ with ${\mathfrak{so}}(T_{x_0}{\mathbb{S}^{4}})$, then there exists a surjection [@deBartolomeis] $$\begin{aligned}
{\mathfrak{so}}(T_{x_0}{\mathbb{S}^{4}}) &\to {\mathcal{V}}_{(x_0,I)}\\
A &\mapsto \widehat A:=[I,A]=IA-AI.\end{aligned}$$ Denote by $(I^+,J^+,K^+,I^-,J^-,K^-)$ the basis of $\bigwedge^2 T_{x_0}{\mathbb{S}^{4}}$ with
\[system\] $
\left\{
\begin{array}{ccc}
I^+&=&e_1\wedge e_2+e_3\wedge e_4\\
J^+&=&e_1\wedge e_3-e_2\wedge e_4\\
K^+&=&e_1\wedge e_4+e_2\wedge e_3
\end{array}
\right.
$ and $
\left\{
\begin{array}{ccc}
I^-&=&e_1\wedge e_2-e_3\wedge e_4\\
J^-&=&e_1\wedge e_3+e_2\wedge e_4\\
K^-&=&-e_1\wedge e_4+e_2\wedge e_3
\end{array}
\right.
$
so that $I^+ = I$. From [@deBartolomeis], we know that $$(AX)^v =\frac12 [X,N]^v = \frac12 {\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{S}^{4}}}(X\wedge N)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{S}^{4}}}(X\wedge N)}{\tmpbox}}.$$ One can easily check that $$R^{{\mathbb{S}^{4}}}(e_3 \wedge N) = R^{{\mathbb{S}^{4}}}(e_3 \wedge e_1) = e_1 \wedge e_3 = \frac12 (J^+ + J^-),$$ so ${\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{S}^{4}}}(e_3 \wedge N)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{S}^{4}}}(e_3 \wedge N)}{\tmpbox}} = K^+$.
Identifying $e_2$, $e_3$ and $e_4$ with their horizontal lifts, we have $(Ae_3)^v = - \frac12 K^+$. Similarly $(Ae_4)^v = \frac12 J^+ .$
Therefore the block matrix of $A$ in the basis $\Big(\{e_2\},\{ e_3,e_4\},\{J^+,K^+\}\Big)$ is $$A=\left(\begin{array}{ccc}\mu &0&0\\0&E&F\\0&F&G\end{array}\right)
\textrm{ with } F=-\frac12 \left(\begin{array}{ccc}0&-1\\1&0\end{array}\right),$$ while $\varphi$ is $$\left(\begin{array}{ccc}0&0&0\\0&I&0\\0&0&-I\end{array}\right) \textrm{ with } I=\left(\begin{array}{ccc}0&-1\\1&0\end{array}\right) .$$ As, by hypothesis $A\varphi = \varphi A$, a straightforward computation shows this to be impossible.
Combining Propositions \[lemmaA\] and \[lemmaB\] shows the ${\mathbb{C}\mathrm{P}}^3$ case of the Theorem.
The case $(M,Z(M))= (\overline{{\mathbb{C}\mathrm{P}}^2} , {\mathbb{F}_{1,2}})$
===============================================================================
The curvature tensor of $({\mathbb{C}\mathrm{P}}^2, g_{{\mathbb{C}\mathrm{P}}^2}, {\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}})$ is $$\begin{aligned}
R^{{\mathbb{C}\mathrm{P}}^2}(U,V,W,S)&=g(U,S)g(V,W)-g(U,W)g(V,S)-g(U,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}W)g(V,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}S)\\
&+g(U,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}S)g(V,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}W)-2g(U,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}V)g(W,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}S).\end{aligned}$$ for $U,V,W$ and $S$ in $T_{x_0}{\mathbb{C}\mathrm{P}}^2$ ($x_0 \in {\mathbb{C}\mathrm{P}}^2$). We still denote by $\mathbb J_{{\mathbb{C}\mathrm{P}}^2}$ the almost complex structure induced on the horizontal distribution ${\mathcal{H}}$, hence, $$\begin{aligned}
\pi^*R^{{\mathbb{C}\mathrm{P}}^2}(\mathbb I_2N,\mathbb I_2X,X,N)&= g^h(N,\mathbb I_2X)^2 +
2g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2}N,X)^2-g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2}N,\mathbb I_2X)^2\\
&+g^h(\mathbb I_2N,\mathbb J_{{\mathbb{C}\mathrm{P}}^2}N)g^h(\mathbb I_2X,\mathbb J_{{\mathbb{C}\mathrm{P}}^2}X),\end{aligned}$$ and, as ${\mathrm{scal}}_{{\mathbb{C}\mathrm{P}}^2} = 24$ and $t=\frac14$, $a=\tfrac34$, $b=\tfrac14$ and $c=4$. From Proposition \[prop1\], we obtain\[page8\] $$\begin{aligned}
&R(\mathbb I_2 N,\mathbb I_2X,X,N)=\tfrac74 g^h(N,X)^2 - \tfrac{21}{4} g^h(N,\mathbb I_2 X)^2
+2g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2} N,X)^2
-g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2}N,\mathbb I_2X)^2 \\
&+ \tfrac34 (\Vert N^h \Vert^2+\Vert X^h\Vert^2)-\tfrac74 \Vert N^h \Vert^2\Vert X^h\Vert^2
+g^h(\mathbb I_2N,\mathbb J_{{\mathbb{C}\mathrm{P}}^2}N)g^h(\mathbb I_2X,\mathbb J_{{\mathbb{C}\mathrm{P}}^2}X).\end{aligned}$$ We deduce, by Lemma \[lemma2\]:
\[lemma\*\] If $X\in E_{\lambda}$ then $$\label{equation*}
7\Big(g^h(N,X)^2-g^h(N,\mathbb I_2 X)^2\Big)+3\Big(g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2} N,X)^2-g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2} N,\mathbb I_2 X)^2\Big)=0.$$
The next result is key to our argument since it reduces the type of the vector field normal to $H$ to just two possibilities.
\[proposition2\] Let $H$ be a hypersurface of ${\mathbb{F}_{1,2}}$ such that $A\varphi = \varphi A$. Then the normal vector $N$ must be either vertical or horizontal.
The proof of Proposition \[proposition2\] consists of a series of lemmas.
Assume that $N$ is neither vertical nor horizontal. We consider a basis of the $T_p {\mathbb{S}^{4}}$ given by $$\begin{aligned}
e_1&= \frac{d\pi(N^h)}{|N^h|} ; e_2= I e_1 ;\\
e_3 &= \begin{cases}
\textrm{unitary part of ${\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}e_1$ that is normal to $(e_1,e_2)$, if non-zero,}\\
\textrm{any unit vector in $(e_1,e_2)^\perp$, otherwise}
\end{cases}.\\
e_4&= I e_3 .
\end{aligned}$$ Recall that since ${\mathbb{C}\mathrm{P}}^2$ is self-dual, ${\mathbb{F}_{1,2}}= Z(\overline{{\mathbb{C}\mathrm{P}}^2})$ and ${\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}\in \bigwedge^{2}_{-}(\overline{{\mathbb{C}\mathrm{P}}^2})$, so (using the same notation as on page ) we can consider $\tilde{c},\tilde{s} \in {\mathbb{R}}$, with $\tilde{c}^2 + \tilde{s}^2 = 1$, such that ${\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}= \tilde{c} I^- + \tilde{s} J^-$, which, in the basis $(e_1,e_2,e_3,e_4)$, translates as $$I =I^+=\left(\begin{array}{rrrr}
0&-1&0&0\\
1&0&0&0\\
0&0&0&-1\\
0&0&1&0
\end{array}
\right),
{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}=\left(\begin{array}{rrrr}
0&-\tilde{c}&-\tilde{s}&0\\
\tilde{c}&0&0&-\tilde{s}\\
\tilde{s}&0&0&\tilde{c}\\
0&\tilde{s}&-\tilde{c}&0
\end{array}
\right).$$
We first describe the solutions to Equation in Lemma \[lemma\*\] in the basis we just constructed.
\[lemma13\] If $\tilde{s}\neq 0$, $d\pi(E_\lambda\cap (\mathbb I_2 N)^\perp)$ is included in $${\mathrm{Vect}}\left(\left(\begin{array}{c}1\\0\\0 \\\delta_-\end{array}\right),\left(\begin{array}{c}0\\1\\-\delta_-\\0\end{array}\right)\right)
\bigcup
{\mathrm{Vect}}\left(\left(\begin{array}{c}1\\0\\0 \\\delta_+\end{array}\right),\left(\begin{array}{c}0\\1\\-\delta_+\\0\end{array}\right)\right).$$ with $\delta_\pm=\frac{6\tilde{c}\tilde{s}\pm\sqrt{84}\tilde{s}}{6\tilde{s}^2}$.\
For $\tilde{s}=0$, $d\pi(E_\lambda\cap (\mathbb I_2 N)^\perp)$ is included in ${\mathrm{Vect}}(e_3, e_4)$.
Assume $X\in E_{\lambda}\cap (\mathbb I_2 N)^\perp$, with $X^h = (x,y,z,t)$ its coordinates in the basis $(e_1,e_2,e_3,e_4)$ (identifying vectors tangent to the base manifold with their horizontal lifts). Then by Lemma \[lemma\*\], we have $$\begin{aligned}
0&=7\Big(g^h(X,N)^2-g^h(X,\mathbb I_2N)^2\Big)+3\Big(g(X,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N)^2-g(X,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}\mathbb I_2 N)^2\Big)\\
&=7(x^2-y^2)+3\Big((\tilde{c}y+\tilde{s}z)^2-(-\tilde{c}x+\tilde{s}t)^2\Big)\\
&=(7-3\tilde{c}^2)(x^2-y^2)+3\tilde{s}^2(z^2-t^2)+6\tilde{c}\tilde{s}(xt+yz).\end{aligned}$$ Re-writing this system with the eigenvector $X+\mathbb I_2 X$, yields $$\begin{aligned}
0&=7g^h(X,N)g^h(X,\mathbb I_2N)+3g^h(X,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N)g^h(X,{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}\mathbb I_2N)\\
&=7xy+3(\tilde{c}y+\tilde{s}z)(-\tilde{c}x+\tilde{s}t)\\
&=(7-3\tilde{c}^2)xy+3\tilde{s}^2zt-3\tilde{c}\tilde{s}(xz-yt),\end{aligned}$$ so $X^h = (x,y,z,t)$ must satisfy the system $$\label{dagger}
\begin{cases}
(7-3\tilde{c}^2)(x^2-y^2)+3\tilde{s}^2(z^2-t^2)+6\tilde{c}\tilde{s}(xt+yz)=0 ,\\
(7-3\tilde{c}^2)xy+3\tilde{s}^2zt-3\tilde{c}\tilde{s}(xz-yt)=0 .
\end{cases}$$ We work with complex numbers $z_1=x+iy$ and $z_2=z+it$ to re-write as a polynomial in $z_2$: $$3\tilde{s}^2z_2^2-6i\tilde{c}\tilde{s}z_1z_2+(7-3\tilde{c}^2)z_1^2=0.$$ If $\tilde{s}\neq 0$, its roots are $z_2=\frac{6i\tilde{c}\tilde{s} z_1\pm \sqrt{84}i\tilde{s}z_1}{6\tilde{s}^2}=i\delta_\pm z_1$.\
Note that $\delta_-\delta_+=-\frac{7-3\tilde{c}^2}{3\tilde{s}^2}$, so neither $\delta_-$ nor $\delta_+$ can vanish.
If $\tilde{s}=0$ then the set of solutions is $\{ z_1 =0\}$.
This description forces the number of eigenvalues of $A|_{({\mathbb I_2}N)^{\perp}}$.
The shape operator $A$ of the hypersurface $H$, restricted to $({\mathbb I_2}N)^{\perp}$, admits two distinct eigenvalues $\lambda_1$ and $\lambda_2$.
Lemma \[lemma13\] implies that the dimension of $d\pi(E_\lambda\cap (\mathbb I_2 N)^\perp)$ must be at most two, and since it is $\mathbb I_2$-invariant and $\mathbb I_2 N$ cannot be neither vertical nor horizontal, the dimension of $E_\lambda\cap (\mathbb I_2 N)^\perp$ is exactly two.
Next we prove that the horizontal parts of the eigenspaces are in direct sum.
\[lemma12\] If $N^v \neq 0$, then $$d\pi:T_{(x_0 , I)} {\mathbb{F}_{1,2}}\cap (N,\mathbb I_2N)^\perp\to T_{x_0}{\mathbb{C}\mathrm{P}}^2$$ is an isomorphism. In particular, as $T_{(x_0 , I)} {\mathbb{F}_{1,2}}\cap (N,\mathbb I_2N)^\perp$ decomposes into a direct sum of eigenspaces of $A$, we have $$d\pi(E_{\lambda_1})\oplus d\pi(E_{\lambda_2})=T_{x_0}{\mathbb{C}\mathrm{P}}^2.$$
As $T {\mathbb{F}_{1,2}}= {\mathcal{H}}\oplus {\mathcal{V}}$, at a point $z = (x_0 , I) \in H \subset {\mathbb{F}_{1,2}}$, write $N=(N^h,N^v)$ and ${\mathbb I_2}N = (IN^h, -IN^v)$ in their horizontal and vertical components. If $(X^h,X^v)\in T_{(x_0 , I)} {\mathbb{F}_{1,2}}\cap (N,\mathbb I_2N)^\perp$, we have $$\left\{\begin{array}{ccc}
(N^h,X^h)+(N^v,X^v)&=&0\\
(I N^h,X^h)-(I N^v,X^v)&=&0 ,
\end{array}\right.$$ and clearly if $N^v\neq 0$ then $d\pi$ is injective.
Observe that, when $N^v\neq 0$, for reasons of dimensions, the case $\tilde{s}=0$ is excluded by Lemma \[lemma12\].
We fully describe $E_{\lambda}$ by obtaining its vertical part.
Assume that neither $N^h$ nor $N^v$ vanishes, then in the basis $$\Big((N^h,I^+N^h,J^+N^h,K^+N^h),(N^v,IN^v)\Big)$$ of $T_p {\mathbb{F}_{1,2}}$, then one of the eigenspaces of $A|_{({\mathbb I_2}N)^{\perp}}$ is $$E_{\lambda}={\mathrm{Vect}}\left(\begin{array}{c}\left(\begin{array}{c}\Vert N^v\Vert^2\\0\\0\\ \delta_+\Vert N^v\Vert^2\end{array}\right)\\
\\\left(\begin{array}{c}-\Vert N^h\Vert^2\\0\end{array}\right)\end{array}
,\begin{array}{c}\left(\begin{array}{c}0\\\Vert N^v\Vert^2\\-\delta_+\Vert N^v\Vert^2\\0\end{array}\right)\\
\\\left(\begin{array}{c}0\\-\Vert N^h\Vert^2\end{array}\right)\end{array}\right),$$ while the other corresponds to $\delta_-$.
From Lemmas \[lemma12\] and \[lemma13\], without loss of generality, we know that $$d\pi(E_{\lambda})=
{\mathrm{Vect}}\left(\left(\begin{array}{c}1\\0\\0 \\\delta_+\end{array}\right),\left(\begin{array}{c}0\\1\\-\delta_+\\0\end{array}\right)\right).$$ so in the basis $\Big((N^h,I^+N^h,J^+N^h,K^+N^h),(N^v,IN^v)\Big)$, as $E_\lambda \perp {\mathrm{Vect}}(N, {\mathbb I_2}N)$, we necessarily have the above description of $E_{\lambda}$.
To conclude the proof of Proposition \[proposition2\], first recall that a nearly K[ä]{}hler manifold satisfies [@Gray1969] $$\label{star}
\Vert (\nabla_X\mathbb I_2)N\Vert^2=R(X,N,X,N)+R(\mathbb I_2N,\mathbb I_2X,X,N),$$ and, moreover, in dimension six, we have [@Gray1970] $$\label{starstar}
\Vert (\nabla_X\mathbb I_2)N\Vert^2=\alpha\Vert X\Vert^2 ,$$ where $\alpha =1$ for ${\mathbb{F}_{1,2}}= Z(\overline{{\mathbb{C}\mathrm{P}}^2})$, since ${\mathrm{scal}}_{{\mathbb{F}_{1,2}}} = 24$ (cf. [@Davidov1991]).
If $X$ is a vector field in $E_\lambda$, we have by Proposition \[prop1\] $$\begin{aligned}
&R(X,N,X,N) =R^{{\mathbb{C}\mathrm{P}}^2} \Big(d\pi(X),d\pi(N),d\pi(X),d\pi(N)\Big)\\
&-\tfrac{11}{4}g^h(N,X)^2+\tfrac{21}{4}g^h(N,\mathbb I_2 X)^2
-\tfrac{15}{4}(\Vert N^h \Vert^2 \Vert X\Vert^2 +\Vert X^h\Vert^2)\\
&+\tfrac{11}{4}\Vert N^h \Vert^2\Vert X^h\Vert^2 +4 \Vert X\Vert^2 ,\end{aligned}$$ with $$\pi^*R^{{\mathbb{C}\mathrm{P}}^2}(X,N,X,N)=-\Vert X^h\Vert^2\Vert N^h\Vert^2+g^h(N,X)^2-3g^h({\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,X)^2$$ hence $$\begin{aligned}
&R(X,N,X,N)=-\tfrac74 g^h(N,X)^2+ \tfrac{21}{4}g^h(N,\mathbb I_2 X)^2\\
&-3g^h(\mathbb J_{{\mathbb{C}\mathrm{P}}^2} N,X)^2 - \tfrac{15}{4}(\Vert N^h \Vert^2\Vert X\Vert^2+\Vert X^h\Vert^2)+\tfrac74 \Vert N^h \Vert^2\Vert X^h\Vert^2
+4\Vert X\Vert^2\end{aligned}$$ Second, observe that from page , we have that $$\begin{aligned}
&R(X,N,X,N)+R(\mathbb I_2N,\mathbb I_2X,X,N)= -g^h({\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,X)^2-g^h(\mathbb I_2{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,X)^2\\
&-3(\Vert N^h\Vert^2 \Vert X\Vert^2+\Vert X^h\Vert^2)
+g^h(\mathbb I_2{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,N)g^h(\mathbb I_2{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}X,X)+4\Vert X\Vert^2.\end{aligned}$$ For $X\in E_\lambda$ of the form $$X =\left(\begin{array}{c}\left(\begin{array}{c}\Vert N^v\Vert^2\\0\\0\\ \delta_+\Vert N^v\Vert^2\end{array}\right)\\
\\\left(\begin{array}{c}-\Vert N^h\Vert^2\\0\end{array}\right)\end{array}\right)$$ we compute that: $$\begin{aligned}
&g^h({\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,X)^2=0 ; g^h(\mathbb I_2{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,X)^2=(-\tilde{c}+\tilde{s}\delta_+)^2\Vert N^h\Vert^4\Vert N^v\Vert^4 ; \\
&\Vert X\Vert^2=(1+\delta_+^2)\Vert N^h\Vert^2\Vert N^v\Vert^4+\Vert N^h\Vert^4\Vert N^v\Vert^2=\delta_+^2\Vert N^h\Vert^2\Vert N^v\Vert^4+\Vert N^h\Vert^2\Vert N^v\Vert^2 ;\\
&\Vert X^h\Vert^2=(1+\delta_+^2)\Vert N^h\Vert^2\Vert N^v\Vert^4 ;
g^h(\mathbb I_2{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}N,N)=-\tilde{c}\Vert N^h\Vert^2 ;\\
&g^h(\mathbb I_2{\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}X,X)=\Big((-\tilde{c}+\tilde{s}\delta_+)+\delta_+(\tilde{s}+\tilde{c}\delta_+)\Big)\Vert N^h\Vert^2\Vert N^v\Vert^4 .\end{aligned}$$ Then Equation yields $$-4\delta_+^2\Vert N^h\Vert^4\Vert N^v\Vert^4=0,$$ and this is impossible by the observation at the end of the proof of Lemma \[lemma13\].\
This concludes the proof of Proposition \[proposition2\].
We can now exclude the remaining case, since $N$ vertical has already been ruled out by Proposition \[proppage6\].
\[prop15\] Let $H$ be a hypersurface of ${\mathbb{F}_{1,2}}$ such that $A\varphi = \varphi A$. Then the normal vector $N$ cannot be horizontal.
If $N$ were horizontal, then by O’Neill’s formula [@Oneill], for any horizontal vector $X$ $$(AX)^v=(-\nabla_X N)^v=-\frac12 [X,N]^v ,$$ and by [@deBartolomeis], $[X,N]^v = {\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{C}\mathrm{P}}^2}(X\wedge N)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{C}\mathrm{P}}^2}(X\wedge N)}{\tmpbox}}$. Let $p\in H \subset{\mathbb{F}_{1,2}}$, $p= (x_0, I), x_0 \in {\mathbb{C}\mathrm{P}}^2$. We identify vectors tangent to ${\mathbb{C}\mathrm{P}}^2$ at $x_0$ with their horizontal lifts in $T_p {\mathbb{F}_{1,2}}$. Let $(e_1,e_2,e_3,e_4)$ be an orthonormal basis of $T_{x_0}{\mathbb{C}\mathrm{P}}^2$ adapted to our problem, i.e. $$e_1=d\pi(N),\quad I=I^+\textrm{ et } {\mathbb{J}_{{\mathbb{C}}\mathrm{P}^2}}=\tilde{c}I^-+\tilde{s}J^-.$$ where
$
\left\{
\begin{array}{ccc}
I^+&=&e_1\wedge e_2+e_3\wedge e_4\\
J^+&=&e_1\wedge e_3-e_2\wedge e_4\\
K^+&=&e_1\wedge e_4+e_2\wedge e_3
\end{array}
\right.
$ and $
\left\{
\begin{array}{ccc}
I^-&=&e_1\wedge e_2-e_3\wedge e_4\\
J^-&=&e_1\wedge e_3+e_2\wedge e_4\\
K^-&=&-e_1\wedge e_4+e_2\wedge e_3
\end{array}
\right.
$
As in the case of ${\mathbb{S}^{4}}$, there exists a surjection $$\begin{aligned}
{\mathfrak{so}}(T_{x_0}{\mathbb{C}\mathrm{P}}^2) &\to {\mathcal{V}}_{(x_0,I)}\\
A &\mapsto \widehat A:=[I,A]=IA-AI.\end{aligned}$$ and $$[X,N]^v ={\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{C}\mathrm{P}}^2}(X\wedge N)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{C}\mathrm{P}}^2}(X\wedge N)}{\tmpbox}}$$ One easily obtains $$\begin{aligned}
R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)e_1&=-(1+3\tilde{s}^2)e_3-3\tilde{c}\tilde{s}e_2 ;\\
R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)e_2&=(1-3\tilde{s}^2)e_4 +3\tilde{c}\tilde{s}e_1 ;\\
R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)e_3&=3\tilde{c}\tilde{s} e_4 +(3\tilde{s}^2+1)e_1 ;\\
R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)e_4&=-3\tilde{c}\tilde{s} e_3 - (1-3\tilde{s}^2)e_2 ,\end{aligned}$$ so that $$R^{{\mathbb{C}\mathrm{P}}^2}(e_1\wedge e_3)=-3\tilde{c}\tilde{s} I^--(1+3\tilde{s}^2)e_1\wedge e_3+(1-3\tilde{s}^2)e_2\wedge e_4\\$$ and substituting $e_1\wedge e_3=\frac12 (J^++ J^-)$ and $e_2\wedge e_4=\frac12(J^- - J^+)$, we obtain $${\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)}{\tmpbox}}=\Big[I^+,R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)\Big]=-2 K^+.$$ Similarly $$\begin{aligned}
R^{{\mathbb{C}\mathrm{P}}^2}(e_1 \wedge e_4)&=-K^+ .\end{aligned}$$ As elements of $\bigwedge^+$ and $\bigwedge^-$ commute with each other $$\begin{aligned}
\mathcal V(A e_3)&=-\frac12{\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_3)}{\tmpbox}}= K^+\\
\mathcal V(A e_4)&=-\frac12{\savestack{\tmpbox}{\stretchto{ \scaleto{ \scalerel*[\widthof{\ensuremath{R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_4)}}]{\kern-.6pt\bigwedge\kern-.6pt} {\rule[-\textheight/2]{1ex}{\textheight}} }{\textheight}}{0.5ex}}\stackon[1pt]{R^{{\mathbb{C}\mathrm{P}}^2}(e_1,e_4)}{\tmpbox}}=- J^+\end{aligned}$$ In the basis $\Big(\{e_2\},\{e_3,e_4\},\{J^+,K^+\}\Big)$ the endomorphisms $A$ and $\varphi$ have the following block matrices $$A=\left(\begin{array}{ccc}\mu &0&0\\
&&\\
0&E&F\\
&&\\
0&F&G\end{array}\right)
\quad\textrm{ with
$F=\left(\begin{array}{cc}
0&-1\\
1&0
\end{array}
\right)
$. }$$ and $$\varphi=\left(\begin{array}{ccc}0&0&0\\0&I&0\\0&0&-I\end{array}\right) \textrm{ with } I=\left(\begin{array}{ccc}0&-1\\1&0\end{array}\right)$$ By hypothesis, $A$ and $\varphi$ commute, which contradicts the form of block $F$.
[10]{}
V. Apostolov, G. Grantcharov, and S. Ivanov. Hermitian structures on twistor spaces. , 16(3):291–308, 1998.
M. F. Atiyah, N. J. Hitchin, and I. M. Singer. Self-duality in four-dimensional [R]{}iemannian geometry. , 362(1711):425–461, 1978.
J. Berndt, J. Bolton, and L. M. Woodward. Almost complex curves and [H]{}opf hypersurfaces in the nearly [K]{}ähler 6-sphere. , 56(3):237–247, 1995.
D. E. Blair. Almost contact manifolds with [K]{}illing structure tensors. , 39:285–292, 1971.
J.-B. Butruille. Classification des vari[é]{}t[é]{}s approximativement kähl[é]{}riennes homog[è]{}nes. , 27(3):201–225, 2005.
Th. E. Cecil and P. J. Ryan. . Springer New York, 2015.
J. Davidov and O. Muskarov. On the [R]{}iemannian curvature of a twistor space. , 58(3-4):319–332, 1991.
P. de Bartolomeis and A. Nannicini. Introduction to differential geometry of twistor spaces. In [*Geometric theory of singular phenomena in partial differential equations ([C]{}ortona, 1995)*]{}, Sympos. Math., XXXVIII, pages 91–160. Cambridge Univ. Press, Cambridge, 1998.
J. Eells and S. Salamon. Twistorial construction of harmonic maps of surfaces into four-manifolds. , 12(4):589–640, 1985.
Th. Friedrich and H. Kurke. Compact four-dimensional self-dual [E]{}instein manifolds with positive scalar curvature. , 106(1):271–299, 1982.
A. Gray. Almost complex submanifolds of the six sphere. , 20(1):277–277, 1969.
A. Gray. Nearly [K]{}ähler manifolds. , 4(3):283–309, 1970.
A. Gray. The structure of nearly [K]{}ähler manifolds. , 223(3):233–248, 1976.
N. J. Hitchin. Kählerian twistor spaces. , 43(1):133–150, 1981.
Z. Hu, Z. Yao, and Y. Zhang. On some hypersurfaces of the homogeneous nearly [K]{}ähler $\mathbb{S}^3 \times \mathbb{S}^3$. , 291(2-3):343–373, 2017.
J. Kenedy Martins. Congruence of hypersurfaces in ${\mathbb{S}}^6$ and in $\mathbb{C}\mathrm{P}^n$. , 32(1):83–105, 2001.
S. Montiel. Real hypersurfaces of a complex hyperbolic space. , 37(3):515–535, 1985.
M. Moruz and L. Vrancken. Properties of the nearly [K]{}ähler $\mathbb{S}^3 \times
\mathbb{S}^3$. , 103(117):147–158, 2018.
P.-A. Nagy. Nearly [K]{}ähler geometry and [R]{}iemannian foliations. , 6(3):481–504, 2002.
B. O’Neill. The fundamental equations of a submersion. , 13:459–469, 1966.
R. Takagi. Real hypersurfaces in a complex projective space with constant principal curvatures. , 27:43–53, 1975.
R. Takagi. Real hypersurfaces in a complex projective space with constant principal curvatures. [II]{}. , 27(4):507–516, 1975.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We re-analyze hadronic decays of $B$ mesons to a pseudoscalar ($P$) and a tensor meson ($T$), or a vector meson ($V$) and a tensor meson, through a $b \to c$ transition. We discuss possible large uncertainties to branching ratios (BR’s) of the relevant modes, mainly arising from uncertainties to the hadronic form factors for the $B \to T$ transition. The BR’s and CP asymmetries for $B \to PT$ and $VT$ decays are then calculated by using the form factors given in the ISGW2 model (the improved version of the original Isgur-Scora-Grinstein-Wise (ISGW) model). We find that the estimated BR’s of many modes are increased by an order of magnitude, compared to the previous results calculated within the ISGW model.'
address: 'Department of Physics and IPAP, Yonsei University, Seoul, 120-749, Korea'
author:
- 'C. S. Kim[^1], Jong-Phil Lee[^2], and Sechul Oh[^3]'
title: |
Hadronic decays of $B$ involving a tensor meson\
through a $b \to c$ transition
---
epsf.tex (\#1 width \#2)[=\#2 ]{}
\#1\#2\#3[[*ibid.*]{} [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Eur. Phys. J. C [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Int. J. Mod. Phys. A [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Mod. Phys. Lett. A [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Nucl. Phys. [**B\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Phys. Lett. B [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Phys. Rev. D [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Phys. Rev. Lett. [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Phys. Rep. [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[Z. Phys. [**\#1**]{}, \#3 (\#2)]{} \#1\#2\#3[[*ibid*]{}. [**\#1**]{}, \#3 (\#2)]{}
Introduction
============
In the next few years plenty of new experimental data on rare decays of $B$ mesons will be available from $B$ factory experiments such as Belle, BaBar, BTeV, LHC-B and so on. Experimentally several tensor mesons have been observed [@pdg], such as the isovector $a_2$(1320), the isoscalars $f_2$(1270), $f_2^{\prime}$(1525), $f_2$(2010), $f_2$(2300), $f_2$(2340), $\chi_{c2}(1P)$, $\chi_{b2}(1P)$ and $\chi_{c2}(2P)$, and the isospinors $K_2^*$(1430) and $D_2^*$(2460). Experimental data on the branching ratios (BR’s) for $B$ decays involving a tensor meson ($T$) in the final state provide only upper bounds[@pdg]: for instance, for a $b \to c$ transition, $$\begin{aligned}
{\cal B} (B^+ \rightarrow \pi^+ D_2^*(2460)^0) &<&
1.3 \times 10^{-3}, \nonumber \\
{\cal B} (B^0 \rightarrow \pi^+ D_2^*(2460)^-) &<&
2.2 \times 10^{-3}, \nonumber \\
{\cal B} (B^{+} \rightarrow \rho^+ D_2^*(2460)^{0}) &<& 4.7 \times
10^{-3}, \nonumber
\\ {\cal B} (B^{0} \rightarrow \rho^+ D_2^*(2460)^{-}) &<&
4.9 \times 10^{-3}~.
\label{expdata}\end{aligned}$$ Recently the process $B \to K_2^* \gamma$ has been observed for the first time by the CLEO Collaboration with a branching ratio of $(1.66^{+0.59}_{-0.53} \pm 0.13) \times 10^{-5}$ [@cleo], and by the Belle Collaboration with ${\cal B}(B\to K_2^*\gamma)=(1.50^{+0.58 +0.11}_{-0.53 -0.13}) \times 10^{-5}$ [@belle].
Two-body hadronic $B$ decays involving a tensor meson $T$ ($J^P = 2^+$) in the final state have long been studied [@kv; @cm; @mrc; @btopt; @btovt] using the non-relativistic quark model of Isgur, Scora, Grinstein and Wise (ISGW) [@isgw] with the factorization ansatz. Some of those works [@kv; @cm; @mrc] studied $B$ decays involving a $b \rightarrow c$ transition, which include the Cabibbo-Kobayashi-Maskawa (CKM)-favored $B$ decays and the CKM-suppressed $B$ decays. The estimated branching ratios of those decay modes strongly depend on the properties of hadronic form factors. A characteristic feature of the form factors given in the original ISGW model [@isgw] is that values of the form factors decrease exponentially as a function of $(t_m - t)$, where $t \equiv (p_B -p_T)^2$ is the momentum transfer and $t_m \equiv (m_B -m_T)^2$ is the maximum possible momentum transfer in the $B$ meson rest frame for a $B \to T$ transition. The authors in Ref. [@kv] used the form factors calculated at the maximum momentum transfer $t_m$ for allowed transitions, assuming that in the relevant transitions the momentum transfer $(t)$ is close to the maximum momentum transfer $(t_m)$. In contrast, other authors [@mrc; @btopt; @btovt] used the form factors with their exponentially decreasing behavior as a function of $(t_m -t)$. In particular, in our previous works [@btopt; @btovt], the exponentially decreasing behavior of the form factors was assumed to predict the BR’s of charmless decays $B \to PT$ and $B \to VT$ ($P$ and $V$ denote a pseudoscalar and a vector meson, respectively)[^4].
Because the exponentially decreasing behavior of the form factors in the ISGW model is less justified, and the assumption of $t_m
\approx t$ seems to be too naive, it is very important to carefully study all the relevant processes using more *reliable* and *consistent* values of the form factors, if available. In fact, the ISGW model has been improved to the ISGW2 model [@isgw2], whose feature includes a more accurate parametrization of the form factors which have a more realistic behavior at large $(t_m -t)$ by making the replacement of the exponentially decreasing term to a certain polynomial term. The improved ISGW2 model also incorporates more reliable features, say the constraints of heavy quark symmetry, relativistic corrections, hyperfine distortions of wave functions, and so forth [@isgw2].
In this work we re-analyze $B \to PT$ and $B \to VT$ decays through a $b \to c$ transition[^5], using the hadronic form factors calculated in the ISGW2 model. We first discuss possible large uncertainties to the BR’s of the relevant modes, mainly arising from uncertainties to the hadronic form factors which are heavily model-dependent. Then, using the form factors obtained in the ISGW2 model, we calculate the BR’s, ratios of ${\cal B}(B\to VT)/{\cal B}(B\to PT)$ and CP asymmetries for $B \to PT$ and $B \to VT$. We make comments on the difference between our results and the previous results obtained using the relevant form factors calculated in the original ISGW model.
This work is organized as follows. In Sec. II we discuss uncertainties relevant to the hadronic form factors. Our framework is introduced in Sec. III. We present our analysis of $B \to PT$ and $B \to VT$ decays in Sec. IV. Finally, our results are summarized in Sec. V.
Uncertainties relevant to hadronic form factors
===============================================
The decay rate $(\Gamma)$ of $B \to PT$ or $B \to VT$ strongly depends on the relevant hadronic form factors for $B \to T$ transitions. For instance, in $B \to PT$ decays, the decay rate is $$\begin{aligned}
\Gamma (B \to PT) \propto (F^{B \to T})^2 ~,\end{aligned}$$ where $$\begin{aligned}
F^{B \to T} = k +(m_B^2 -m_T^2) b_+ +m_P^2 b_-~.\end{aligned}$$ (See the next section for definitions of the form factors $k$, $b_+$, and $b_-$.) Table I shows the values of the form factors $F^{B \to T}$ calculated in three cases: (i) at $q^2 = m_D^2$ $(q^{\mu} \equiv p_B^{\mu} -p_T^{\mu})$, (ii) at the maximum momentum transfer $t_m \equiv (m_B -m_T)^2$ in the ISGW model, and (iii) at $q^2 = m_D^2$ in the ISGW2 model. We note that $|F^{B \to T}| \approx 0.2$ at $t_m$, while $|F^{B \to T}| \approx 0.05$ at $q^2 = m_D^2$ in the ISGW model, where $T = a_2,~ f_2,~ f^{\prime}_2$. The value of $|F^{B \to T}|$ calculated at $t_m$ is about 4 times larger than that calculated at $q^2 = m_D^2$. Thus, the decay rate of a relevant process (e.g., $B \to D a_2$, $D f_2$, $D f_2^{\prime}$, etc) evaluated by using the former value of the form factor (evaluated at $t_m$) would be roughly 16 times larger than that obtained using the latter value of the form factor (at $q^2 = m_D^2$). For $B \to VT$ decays, the similar argument holds. It is obvious that the uncertainty relevant to the hadronic form factors can seriously spoil theoretical estimates of the BR’s of $B \to PT$ and $B \to VT$ decays. More reliable values of the form factors are definitely needed.
As previously mentioned, a crucial improvement of the ISGW2 model is that the form factors in this model have a more realistic and reasonable behavior at large $(t_m -t)$. Thus, one no longer needs to *naively assume* $t \approx t_m$ in $B \to PT$ and $VT$ processes. The value of $|F^{B \to T}|$ obtained at $q^2 = m_D^2$ in the ISGW2 model is in between that obtained at $q^2 = m_D^2$ and that calculated at $t_m$ in the ISGW model (except $|F^{B \to K^*}|$). In fact, from Table I, we see that for $B \to a_2$ and $B \to f_2$ transitions, $|F^{B \to T}|$ obtained at $t_m$ is about 2 times larger than that obtained at $q^2 = m_D^2$ in the ISGW2 model, which would lead to overestimation of the relevant decay rates. Compared to $|F^{B \to T}|$ obtained at $q^2 = m_D^2$ in the ISGW model, the values obtained in the ISGW2 model are about $2 - 6$ times larger, which would result in roughly $4 - 36$ times larger decay rates.
Framework
=========
The relevant $\Delta B =1$ effective Hamiltonian for hadronic $B$ decays can be written as $$\begin{aligned}
H_{eff}^{q} &=& {G_F \over \sqrt{2}} \left[ V_{ub}V^{*}_{uq} (c_1 O^q_{1u}
+c_2 O^q_{2u}) + V_{cb}V^{*}_{cq} (c_1 O^q_{1c} +c_2 O^q_{2c}) \right.
\nonumber \\
&-& \left. \sum_{i=3}^{10} \left( V_{ub} V^*_{uq} c_i^u +V_{cb} V^*_{cq} c_i^c
+V_{tb} V^*_{tq} c_i^t \right) O_i^q \right] \nonumber \\
&+& H.C.~ ,\end{aligned}$$ where $O^q_i$’s are defined as $$\begin{aligned}
O^q_{1f} &=& \bar q \gamma_{\mu} L f \bar f \gamma^{\mu} L b, \ \
O^q_{2f} = \bar q_{\alpha} \gamma_{\mu} L f_{\beta} \bar f_{\beta} \gamma^{\mu}
L b_{\alpha}~, \nonumber \\
O^q_{3(5)} &=& \bar q \gamma_{\mu} L b \sum_{q^{\prime}} \bar q^{\prime}
\gamma^{\mu} L(R) q^{\prime}, \ \
O^q_{4(6)} = \bar q_{\alpha} \gamma_{\mu} L b_{\beta} \sum_{q^{\prime}}
\bar q^{\prime}_{\beta} \gamma^{\mu}
L(R) q^{\prime}_{\alpha}~, \nonumber \\
O^q_{7(9)} &=& {3 \over 2} \bar q \gamma_{\mu} L b \sum_{q^{\prime}}
e_{q^{\prime}} \bar q^{\prime} \gamma^{\mu}
R(L) q^{\prime} , \ \
O^q_{8(10)} ={3 \over 2} \bar q_{\alpha} \gamma_{\mu} L b_{\beta}
\sum_{q^{\prime}} e_{q^{\prime}} \bar
q^{\prime}_{\beta} \gamma^{\mu} R(L) q^{\prime}_{\alpha}~ ,\end{aligned}$$ where $L(R) = (1 \mp \gamma_5)$, $f$ can be $u$ or $c$ quark, $q$ can be $d$ or $s$ quark, and $q^{\prime}$ is summed over $u$, $d$, $s$, and $c$ quarks. $\alpha$ and $\beta$ are the color indices. $T^a$ is the SU(3) generator with the normalization ${\rm Tr}(T^a T^b) = \delta^{ab}/2$. $G^{\mu \nu}_a$ and $F^{\mu
\nu}$ are the gluon and photon field strength, and $c_i$’s are the Wilson coefficients (WC’s). We use the improved effective WC’s given in Ref.[@improvedwc], where the renormalization scheme- and scale-dependence of the WC’s are discussed and resolved. The regularization scale is taken to be $\mu=m_b$ [@ddo]. The operators $O_1$, $O_2$ are the tree level and QCD corrected operators, $O_{3-6}$ are the gluon induced strong penguin operators, and finally $O_{7-10}$ are the electroweak penguin operators due to $\gamma$ and $Z$ exchange, and the box diagrams at loop level.
We use the improved ISGW2 quark model to analyze two-body nonleptonic decay processes $B \to PT$ and $VT$ in the framework of generalized factorization. We describe the parameterizations of the hadronic matrix elements in $B \to PT$ and $VT$ decays: [@isgw; @isgw2] $$\begin{aligned}
\langle 0 | A^{\mu} | P \rangle &=& i f_P p_P^{\mu} ~, \\
\langle 0 | V^{\mu} | V \rangle &=& f_{_V} m_{_V} \epsilon^{\mu} ~, \\
\langle T | j^{\mu} | B \rangle &=& i h(m_P^2) \epsilon^{\mu \nu
\rho \sigma} \epsilon^*_{\nu \alpha} p_B^{\alpha} (p_B
+p_T)_{\rho} (p_B -p_T)_{\sigma} + k(m_P^2) \epsilon^{* \mu \nu}
(p_B)_{\nu} \nonumber \\
&\mbox{}& + \epsilon^*_{\alpha \beta} p_B^{\alpha} p_B^{\beta} [
b_+(m_P^2) (p_B +p_T)^{\mu} +b_-(m_P^2) (p_B -p_T)^{\mu} ]~,
\label{formfactor}\end{aligned}$$ where $j^{\mu} = V^{\mu} -A^{\mu}$. $V^{\mu}$ and $A^{\mu}$ denote a vector and an axial-vector current, respectively. $f_P~(f_V)$ denotes the decay constant of the relevant pseudoscalar (vector) meson. $h(m_{P(V)}^2)$, $k(m_{P(V)}^2)$, $b_+(m_{P(V)}^2)$, and $b_-(m_{P(V)}^2)$ express the form factors for the $B \rightarrow T$ transition, which have been calculated at $q^2 =m_{P(V)}^2$ $(q^{\mu} \equiv p_B^{\mu} -p_T^{\mu})$ in the ISGW2 quark model. $p_B$ and $p_T$ denote the momentum of the $B$ meson and the tensor meson, respectively. Using the above parameterizations, the decay amplitudes for $B \to PT$ and $B \to
VT$ are [@btopt; @btovt; @Oh] $${\cal A}(B \to PT) \sim F^{B \to T}(m_P^2)~,~~~~~
{\cal A}(B\to VT) \sim \epsilon^{*\alpha\beta} F^{B \to T}_{\alpha\beta}(m_V^2)~,$$ where $$\begin{aligned}
F^{B \to T}(m_P^2) &=& k(m_P^2)+(m_B^2-m_T^2)b_+(m_P^2)+m_P^2b_-(m_P^2)~,
\label{FBT} \\
F^{B \to T}_{\alpha\beta}(m_V^2) &=& \epsilon^*_\mu(p_B+p_T)_\rho
\Big[ih(m_V^2)\cdot \epsilon^{\mu\nu\rho\sigma}
g_{\alpha\nu}(p_V)_\beta(p_V)_\sigma
+k(m_V^2)\cdot\delta^\mu_\alpha\delta^\rho_\beta\nonumber\\
&& +b_+(m_V^2) \cdot(p_V)_\alpha(p_V)_\beta g^{\mu\rho}\Big]~.
\label{FBTv}\end{aligned}$$ For our numerical analysis, we use the following values of the decay constants (in MeV) [@sv]: $f_{\pi} = 132$, $f_{K} =
162$, $f_{D} = 252$, $f_{D_s} = 280$, $f_{\eta_c} = 393$, $f_{\rho} = 216$, $f_{K^*} = 222$, $f_{D^*} = 249$, $f_{D_s^*} =
270$, $f_{J/\psi} = 405$. The running quark masses (in MeV) at $m_b$ scale are used as follows [@do]: $m_u =3.6$, $m_d=6.6$, and $m_s=100$.
An important feature of the ISGW2 model is that a more accurate parametrization of the form factors $h$, $k$, $b_+$, and $b_-$ is adopted by making the replacement, for $B \to T$ transition, $$\begin{aligned}
{\rm exp}[- ({\rm constant}) \cdot (t_m -t)]~ \Rightarrow ~[1 + ({\rm constant})
\cdot (t_m -t)]^{-3}~,\end{aligned}$$ where $t \equiv (p_B -p_T)^2$ is the momentum transfer and $t_m \equiv (m_B -m_T)^2$ is the maximum possible momentum transfer in the $B$ meson rest frame. As a result, the form factors have a more realistic behavior at large $(t_m -t)$.
We note that the matrix element $\langle 0 | j^{\mu} | T \rangle$ vanishes: $$\langle 0 | j^{\mu} | T \rangle = p_\nu \epsilon^{\mu \nu}
(p_{_T}, \lambda) + p_{_T}^\mu \epsilon^\nu_{~\nu} (p_{_T},
\lambda) =0~,
\label{fT}$$ because the trace of the polarization tensor $\epsilon^{\mu \nu}$ of the tensor meson $T$ vanishes and the auxiliary condition holds, $p_T^{\mu} \epsilon_{\mu \nu} =0$ [@epsilon]. Thus, in the generalized factorization scheme, any decay amplitude for $B \rightarrow PT$ (or $VT$) is simply proportional to the decay constant $f_P$ (or $f_V$) and a certain linear combination of the form factors $F^{B \rightarrow T}$ (or $F^{B \to T}_{\alpha \beta}$), [*i.e.*]{}, there is no such amplitude proportional to $f_T \times F^{B \rightarrow P}$ (or $F^{B \to T}_{\alpha \beta}$) (see Appendix).
Analyses and Results
====================
We calculate the BR’s of $B \to PT$ and $B \to VT$ decays, whose quark level processes are the $b \to c$ transition. Among the relevant decay modes, many processes involve a tree diagram only; their decay amplitudes are proportional to the CKM elements $V_{cb}^* V_{ud}$ or $V_{cb}^* V_{us}$ (see Tables II$-$V). Other processes involve both tree and (strong and electroweak) penguin diagrams; their tree amplitudes are proportional to the CKM elements $V_{cb}^* V_{cd}$ or $V_{cb}^* V_{cs}$. But, the penguin diagram contribution is much smaller than the tree contribution. Expressions for all the amplitudes having both the tree and penguin terms are presented in Appendix, calculated in the generalized factorization scheme. (For expressions of the other amplitudes, see Refs. [@kv; @cm; @mrc].) Among the relevant modes, some processes, such as $B^{+ (0)} \to \pi^+ \bar D_2^{*0(-)}$, $B^{+ (0)} \to D_s^+ \bar D_2^{*0(-)}$, etc, are the CKM-favored decays whose amplitudes are proportional to the CKM elements $V_{cb}^* V_{ud}$ (or $V_{cb}^* V_{cs}$ at tree level). Other processes, such as $B^{+ (0)} \to K^+ \bar D_2^{*0(-)}$, $B^{+ (0)} \to D^+ \bar D_2^{*0(-)}$, etc., are the CKM-suppressed decays whose amplitudes are proportional to $V_{cb}^* V_{us}$ (or $V_{cb}^* V_{cd}$ at tree level). As commented in the footnote of Sec. I, we also calculate the BR’s of some CKM-suppressed processes involving the $b \to u$ transition, such as $B^{+(0)} \to D_s^+ a_2^{0(-)}$, $B^{+(0)} \to D^0 K_2^{*+(0)}$, and so on; these processes involve a tree diagram only and their amplitudes are proportional to $V_{ub}^* V_{cs}$.
Tables II and III show the BR’s of $B \to PT$ processes for $\Delta S =0$ and $|\Delta S| =1$, respectively ($S$ denotes the strangeness quantum number). Similarly, Tables IV and V show the BR’s of $B \to VT$ for $\Delta S =0$ and $|\Delta S| =1$, respectively. In the tables, the results are shown for three different values of the parameter $\xi \equiv 1/ N_c$ ($N_c$ denotes the effective number of color)[^6]: $\xi =0.1,~ 0.3,~
0.5$ . For comparison, the BR’s are also calculated using $a_1 =
1.15$ and $a_2 =0.26$ whose values are obtained from a fit to $B
\to PP$ and $B \to PV$ data [@a1a2], where the QCD coefficients are $a_1 \equiv c_1 + \xi c_2$ and $a_2 \equiv c_2
+\xi c_1$ ($c_1$ and $c_2$ are the effective WC’s).
The decay amplitudes of all the modes shown in Tables II$-$V are (dominantly) proportional to either $a_1$ (color-favored) or $a_2$ (color-suppressed) only. The value of $a_1 \equiv c_1 + \xi c_2$ does not vary much as $\xi$ varies: $a_1= 1.132$ for $\xi =0.1$, $a_1= 1.059$ for $\xi =0.3$, and $a_1= 0.986$ for $\xi =0.5$. In contrast, the value of $a_2 \equiv c_2 +\xi c_1$ varies as follows: $a_2= -0.248$ for $\xi =0.1$, $a_2= -0.015$ for $\xi
=0.3$, and $a_2= 0.219$ for $\xi =0.5$. We note that the value of $a_2$ for $\xi =0.3$ is about an order of magnitude smaller than that for $\xi =0.1$ or $\xi =0.5$. It would lead to the estimation that the BR’s of the decay modes, whose amplitudes are proportional to $a_2$, are very small (i.e., about two orders of magnitude smaller) for $\xi =0.3$. However, compared with the values of $a_1$ and $a_2$ obtained from $B \to PP$ and $B \to PV$ data, the value of $a_2$ for $\xi= 0.3$ seems to be too small, while the values of $a_1$ and $a_2$ for $\xi= 0.5$ are quite consistent with those values. (For $\xi= 0.1$, the value of $a_1$ fits well to that obtained from the data, but $a_2$ has the opposite sign to that deduced from the data. However, the sign of $a_2$ has no (or negligible) effect on our results since each decay amplitude is (dominantly) proportional to only one QCD coefficient (i.e., either $a_1$ or $a_2$)).
As expected, the BR’s of both the CKM-favored and color-favored processes are generally large. In $B \to PT$ decays, the BR of $B^{+(0)} \to \pi^+ \bar D_2^{*0(-)}$ is about $3 \times 10^{-4}$, and the BR of $B^{+(0)} \to D^+_s \bar D_2^{*0(-)}$ is about $4 \times 10^{-4}$. In $B \to VT$ decays, the BR of $B^{+(0)} \to \rho^+ \bar D_2^{*0(-)}$ is $(7-9) \times 10^{-4}$, and the BR of $B^{+(0)} \to D^{*+}_s \bar D_2^{*0(-)}$ is about $1 \times 10^{-3}$. The BR’s of the CKM-favored and color-suppressed modes are $O(10^{-4}) - O(10^{-5})$, except ${\cal B}(B^0 \to \bar D^0 f_2^{\prime}) \sim O(10^{-7})$ and ${\cal B}(B^0 \to \bar D^{*0} f_2^{\prime}) \sim O(10^{-6})$. The BR’s of the CKM-suppressed modes are relatively smaller, $O(10^{-5}) - O(10^{-8})$. From Tables II$-$V, we see that the BR’s of the decay modes such as $B^{+(0)} \to \bar D^{(*)0} a_2^{+(0)}$, $B^0 \to \bar D^{(*)0} f_2^{(\prime)}$, $B^{+(0)} \to \eta_c (J/\psi) a_2^{+(0)}$, $B^0 \to \eta_c (J/\psi) f_2^{(\prime)}$, etc, for $\xi =0.3$ are about two orders of magnitudes smaller than those for $\xi =0.1$ or $\xi =0.5$. This occurs because the decay amplitudes of all those modes are (dominantly) proportional to $a_2$ (see Appendix), as explained above.
We note that for many processes our predictions are larger than the BR’s given in Ref. [@mrc]. In particular, for the processes whose amplitudes are proportional to $V^*_{ub} V_{cs}$, our results are about an order of magnitude larger than the BR’s given in [@mrc]; for instance, for $B \to PT$ such as $B^{+(0)} \to D_s^+ a_2^{0(-)}$, $B^+ \to D_s^+ f_2^{(\prime)}$, $B^{+(0)} \to D^0 K_2^{*+(0)}$, and for $B \to VT$ such as $B^{+(0)} \to D_s^{*+} a_2^{0(-)}$, $B^+ \to D_s^{*+} f_2^{(\prime)}$, $B^{+(0)} \to D^{*0} K_2^{*+(0)}$.
In Table VI, we show the ratios of ${\cal B}(B \to VT) / {\cal B}(B \to PT)$. The ratios are roughly 3 for the processes which involve a tree diagram only and whose amplitudes are proportional to $a_1$ (via the external $W$ emission); for instance, ${\cal B}(B^{+(0)}\to \rho^{+} \Dbar_2^{*0(-)})
/{\cal B}(B^{+(0)}\to \pi^{+} \Dbar_2^{*0(-)}) \approx 3$. This is naively expected from the fact that massive vector particles have three polarization states. But, for the processes which involve both tree and penguin diagrams, the ratios deviate from 3; e.g., ${\cal B}(B^{+(0)}\to D^{*+} \Dbar_2^{*0(-)})
/{\cal B}(B^{+(0)}\to D^{+} \Dbar_2^{*0(-)}) \approx 2.3$. For the processes whose amplitudes are proportional to $a_2$ (via the internal $W$ emission), the ratios are $\sim 1.6$, except the processes involving $J/\psi$ or $\eta_c$ in the final state; e.g., ${\cal B}(B^{+(0)}\to \Dbar^{*0} a_2^{+(0)})
/{\cal B}(B^{+(0)}\to \Dbar^{0} a_2^{+(0)}) \approx 1.6$. We note that for the processes involving $J/\psi$ or $\eta_c$ in the final state, the ratios substantially vary as $\xi$ varies from 0.1 to 0.3 to 0.5. This is because the penguin contribution to the decay amplitudes involving $\eta_c$ differs from that to the amplitudes involving $J/\psi$; the penguin effect to the former amplitudes is proportional to the combination of the QCD coefficients $(a_3 -a_5 +a_7 -a_9)$, while the penguin effect to the latter amplitudes is proportional to $(a_3 +a_5 +a_7 +a_9)$. We also compute CP asymmetries ${\mathcal A}_{CP}$ for $B \to PT$ and $B \to VT$ decays, defined by $$\begin{aligned}
{\mathcal A}_{CP} = {{\mathcal B}(B \rightarrow f) -{\mathcal
B}(\bar B \rightarrow \bar f) \over {\mathcal B}(B \rightarrow f)
+{\mathcal B}(\bar B \rightarrow \bar f)}~,\end{aligned}$$ where $f$ and $\bar f$ denote a generic final state and its CP-conjugate state. Since in the relevant modes the tree contribution is very much dominant compared to the penguin contribution, the asymmetries are relatively small[^7]. We note that for a non-vanishing ${\mathcal A}_{CP}$ for a process and its CP-conjugate process, there should exist both the weak phase and the strong phase differences between their tree and penguin amplitudes. Thus, ${\mathcal A}_{CP}$’s vanish for the processes involving $V^*_{cb} V_{cs}$ and $V^*_{tb} V_{ts}$, since there is no weak phase in their amplitudes; e.g., ${\mathcal A}_{CP}
(B^{+(0)} \to D_s^+ \bar D_2^{*0(-)}) = 0$. We present our result in Table VII. The CP asymmetries are shown for different values of $\xi$. For all the relevant modes, the CP asymmetries are expected to be a few percent.
Conclusions
===========
We have studied the decay modes $B \to PT$ and $B \to VT$ whose quark level processes are the $b \to c$ transition. Due to large uncertainties to the relevant hadronic form factors which are model-dependent, the previously estimated BR’s could be spoiled by large uncertainties. Using more *reliable* and *consistent* values of the form factors given in the improved version (ISGW2) of ISGW model, we re-calculate the BR’s of all the relevant modes and find that for many modes our results are much larger than those given in the previous work using the ISGW model.
Our results show that the BR’s of some processes are quite large: in $B \to VT$, the BR’s of $B^{+(0)} \to D^{*+}_s \bar D_2^{*0(-)}$ and $B^{+(0)} \to \rho^+ \bar D_2^{*0(-)}$ are $\sim 10^{-3}$, and in $B \to PT$, the BR’s of $B^{+(0)} \to D^+_s \bar D_2^{*0(-)}$ and $B^{+(0)} \to \pi^+ \bar D_2^{*0(-)}$ are $(3-4) \times 10^{-4}$. (These results are roughly consistent with those obtained under the naive assumption of $t \approx t_m$ in the ISGW model.) The estimated BR’s of $B^{+(0)} \to \pi^+ \bar D_2^{*0(-)}$ and $B^{+(0)} \to \rho^+ \bar D_2^{*0(-)}$ are about a factor of $(4 - 5)$ smaller than the present experimental upper bounds shown in Eq. (\[expdata\]), and so far there is no known experimental data on the modes $B \to D^{(*)}_s \bar D_2^*$. Observations of these processes in $B$ experiments such as Belle, BaBar, BTeV and LHC-B will be crucial in testing the ISGW2 model as well as validity of the factorization scheme.\
**ACKNOWLEDGEMENTS**
We thank G. Cvetic for careful reading of the manuscript and his valuable comments. The work of C.S.K. was supported in part by CHEP-SRC Program, Grant No. 20015-111-02-2 and Grant No. R02-2002-000-00168-0 from BRP of the KOSEF, and in part by Grant No. 2001-042-D00022 of the KRF. The work of J.-P.L. was supported by the BK21 Program. The work of S.O. was supported by the KRF Grants, Project No. 2001-042-D00022.
[\[001\]]{}
Particle Data Group, D. E. Groom [*et al*]{}, Eur. Phys. J. C [**15**]{}, 1 (2000). CLEO Collaboration, T. E. Coan, [*et al.*]{}, Phys. Rev. Lett. [**84**]{}, 5283 (2000). Belle Collaboration, A. Ishikawa, talk at XXXVIIth Rencontres de Moriond on Electroweak Interactions and Unified Theories, Les Arcs, France, March 9$-$16, 2002.
A. C. Katoch and R. C. Verma, Phys. Rev. D [**52**]{}, 1717 (1995); Erratum - *ibid.* [**55**]{}, 7316 (1997). G. López Castro and J. H. Muñoz, Phys. Rev. D [**55**]{}, 5581 (1997). J. H. Muñoz, A. A. Rojas, and G. López Castro, Phys. Rev. D [**59**]{}, 077504 (1999). C. S. Kim, B. H. Lim, and Sechul Oh, Eur. Phys. J. C [**22**]{}, 683 (2002). C. S. Kim, B. H. Lim, and Sechul Oh, Eur. Phys. J. C [**22**]{}, 695 (2002). N. Isgur, D. Scora, B. Grinstein, and M. B. Wise, Phys. Rev. D [**39**]{}, 799 (1989). Belle Collaboration, A. Garmash, to appear in the Proc. of the 4th Internatioal Conference on $B$ Physics & CP Violation, Ise-Shima, Japan, February 19$-$23, 2001 \[hep-ex/0104018\]; Belle Collaboration, K. Abe [*et al.*]{}, contributed to Lepton Photon 01, Rome, Italy, July 23$-$28, 2001 \[hep-ex/0107051\]; Belle Collaboration, A. Garmash [*et al.*]{}, \[hep-ex/0201007\]. C. S. Kim, Jong-Phil Lee, and Sechul Oh, \[hep-ph/0205263\]. D. Scora, N. Isgur, Phys. Rev. D [**52**]{}, 2783 (1995). A. Ali and C. Greub, Phys. Rev. D [**57**]{}, 2996 (1998); Y.-H. Chen, H.-Y. Cheng, B. Tseng, and K.-C. Yang, Phys. Rev. D [**60**]{}, 094014 (1999). N. G. Deshpande, B. Dutta, and Sechul Oh, Phys. Rev. D [**57**]{}, 5723 (1998); N. G. Deshpande, B. Dutta, and Sechul Oh, Phys. Lett. B [**473**]{}, 141 (2000). Sechul Oh, ; M. Gronau and J.L. Rosner, . K. K. Sharma and R. C. Verma, hep-ph/9801202 (1998). B. Dutta and Sechul Oh, Phys. Rev. D [**63**]{}, 054016 (2001). D. Spehler and S. F. Novaes, Phys. Rev. D [**44**]{}, 3990 (1991). M. S. Alam [*et al.*]{}, Phys. Rev. D [**50**]{}, 43 (1994), and references therein.
[**APPENDIX**]{}
In this Appendix, we present expressions for the decay amplitudes of $B \rightarrow PT$ and $VT$ modes which have both the tree and penguin contributions (shown in Tables II$-$V). Below we use $F^{B \to T}$, $F^{B \to T}_{\alpha \beta}$ and $X_{q q^{\prime}}$, defined by Eqs. (\[FBT\]) and (\[FBTv\]), and $$\begin{aligned}
X_{q q^{\prime}} = {m_P^2 \over (m_b +m_{q^{\prime}}) (m_q
+m_{q^{\prime}})}~.\end{aligned}$$
\(1) $B \rightarrow PT$ ($\Delta S = 0$) decays. $$\begin{aligned}
A(B^+ \rightarrow D^+ \bar D_2^{*0}) &=& i {G_F \over \sqrt{2}} f_{D^+}
\epsilon^*_{\mu \nu} p^{\mu}_B p^{\nu}_B F^{B \rightarrow \bar D_2^{*0}}
(m^2_{D^+}) \left\{ V^*_{cb} V_{cd} a_1 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} [a_4
+a_{10} -2 (a_6 +a_8) X_{dc} ] \right\},
\\
A(B^0 \rightarrow D^+ D_2^{*-}) &=& i {G_F \over \sqrt{2}} f_{D^+}
\epsilon^*_{\mu \nu} p^{\mu}_B p^{\nu}_B F^{B \rightarrow D_2^{*-}}
(m^2_{D^+}) \left\{ V^*_{cb} V_{cd} a_1 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} [a_4
+a_{10} -2 (a_6 +a_8) X_{dc} ] \right\},
\\
A(B^+ \rightarrow \eta_c a^+_2) &=& i {G_F \over \sqrt{2}}
f_{\eta_c} \epsilon^*_{\mu \nu}
p^{\mu}_B p^{\nu}_B F^{B \rightarrow a^+_2} (m^2_{\eta_c}) \left\{
V^*_{cb} V_{cd} a_2 - V^*_{tb} V_{td} (a_3 -a_5 +a_7 -a_9 ) \right\},
\\
A(B^0 \rightarrow \eta_c a^0_2) &=& i {G_F \over 2}
f_{\eta_c} \epsilon^*_{\mu \nu}
p^{\mu}_B p^{\nu}_B F^{B \rightarrow a^0_2} (m^2_{\eta_c}) \left\{
V^*_{cb} V_{cd} a_2 - V^*_{tb} V_{td} (a_3 -a_5 +a_7 -a_9 ) \right\},
\\
A(B^0 \rightarrow \eta_c f_2) &=& i {G_F \over 2} \cos \phi_T
f_{\eta_c} \epsilon^*_{\mu \nu}
p^{\mu}_B p^{\nu}_B F^{B \rightarrow f_2} (m^2_{\eta_c}) \left\{
V^*_{cb} V_{cd} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} (a_3 -a_5 +a_7 -a_9 ) \right\},
\\
A(B^0 \rightarrow \eta_c f_2^{\prime}) &=& i {G_F \over 2} \sin \phi_T
f_{\eta_c} \epsilon^*_{\mu \nu}
p^{\mu}_B p^{\nu}_B F^{B \rightarrow f_2^{\prime}} (m^2_{\eta_c}) \left\{
V^*_{cb} V_{cd} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} (a_3 -a_5 +a_7 -a_9 ) \right\}.\end{aligned}$$
\(2) $B \rightarrow PT$ ($|\Delta S| = 1$) decays. $$\begin{aligned}
A(B^+ \rightarrow D^+_s \bar D_2^{*0}) &=& i {G_F \over \sqrt{2}} f_{D^+_s}
\epsilon^*_{\mu \nu} p^{\mu}_B p^{\nu}_B F^{B \rightarrow \bar D_2^{*0}}
(m^2_{D^+_s}) \left\{ V^*_{cb} V_{cs} a_1 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{ts} [a_4
+a_{10} -2 (a_6 +a_8) X_{sc} ] \right\},
\\
A(B^0 \rightarrow D^+_s D_2^{*-}) &=& i {G_F \over \sqrt{2}} f_{D^+_s}
\epsilon^*_{\mu \nu} p^{\mu}_B p^{\nu}_B F^{B \rightarrow D_2^{*-}}
(m^2_{D^+_s}) \left\{ V^*_{cb} V_{cs} a_1 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{ts} [a_4
+a_{10} -2 (a_6 +a_8) X_{sc} ] \right\},
\\
A(B^+ \rightarrow \eta_c K_2^{*+}) &=& i {G_F \over \sqrt{2}}
f_{\eta_c} \epsilon^*_{\mu \nu}
p^{\mu}_B p^{\nu}_B F^{B \rightarrow K_2^{*+}} (m^2_{\eta_c}) \left\{
V^*_{cb} V_{cs} a_2 - V^*_{tb} V_{ts} (a_3 -a_5 +a_7 -a_9 ) \right\},
\\
A(B^0 \rightarrow \eta_c K_2^{*0}) &=& i {G_F \over \sqrt{2}}
f_{\eta_c} \epsilon^*_{\mu \nu}
p^{\mu}_B p^{\nu}_B F^{B \rightarrow K_2^{*0}} (m^2_{\eta_c}) \left\{
V^*_{cb} V_{cs} a_2 - V^*_{tb} V_{ts} (a_3 -a_5 +a_7 -a_9 ) \right\}.\end{aligned}$$
\(3) $B \rightarrow VT$ ($\Delta S = 0$) decays. $$\begin{aligned}
A(B^+ \rightarrow D^{*+} \bar D_2^{*0}) &=& i {G_F \over \sqrt{2}}
( m_{D^{*+}} f_{D^{*+}} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow \bar D_2^{*0}} (m^2_{D^{*+}})) \left\{ V^*_{cb} V_{cd} a_1
- V^*_{tb} V_{td} (a_4 + a_{10}) \right\},
\\
A(B^0 \rightarrow D^{*+} \bar D_2^{*-}) &=& i {G_F \over \sqrt{2}}
( m_{D^{*+}} f_{D^{*+}} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow \bar D_2^{*-}} (m^2_{D^{*+}})) \left\{ V^*_{cb} V_{cd} a_1
- V^*_{tb} V_{td} (a_4 + a_{10}) \right\},
\\
A(B^+ \rightarrow J/\psi a^+_2) &=& i {G_F \over \sqrt{2}}
( m_{J/\psi} f_{J/\psi} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow a^+_2} (m^2_{J/\psi})) \left\{
V^*_{cb} V_{cd} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} (a_3 +a_5 +a_7 +a_9 ) \right\},
\\
A(B^0 \rightarrow J/\psi a^0_2) &=& i {G_F \over 2}
( m_{J/\psi} f_{JJ/\psi} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow a^0_2} (m^2_{J/\psi})) \left\{
V^*_{cb} V_{cd} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} (a_3 +a_5 +a_7 +a_9 ) \right\},
\\
A(B^0 \rightarrow J/\psi f_2) &=& i {G_F \over 2} \cos \phi_T
( m_{J/\psi} f_{J/\psi} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow f_2} (m^2_{J/\psi})) \left\{
V^*_{cb} V_{cd} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} (a_3 +a_5 +a_7 +a_9 ) \right\},
\\
A(B^0 \rightarrow J/\psi f_2^{\prime}) &=& i {G_F \over 2} \sin \phi_T
( m_{J/\psi} f_{J/\psi} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow f_2^{\prime}} (m^2_{J/\psi})) \left\{
V^*_{cb} V_{cd} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{td} (a_3 +a_5 +a_7 +a_9 ) \right\}.\end{aligned}$$
\(4) $B \rightarrow VT$ ($|\Delta S| = 1$) decays. $$\begin{aligned}
A(B^+ \rightarrow D^{*+}_s \bar D_2^{*0}) &=& i {G_F \over \sqrt{2}}
( m_{D^{*+}_s} f_{D^{*+}_s} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow \bar D_2^{*0}} (m^2_{D^{*+}_s})) \left\{ V^*_{cb} V_{cs} a_1
- V^*_{tb} V_{ts} (a_4 + a_{10}) \right\},
\\
A(B^0 \rightarrow D^{*+}_s D_2^{*-}) &=& i {G_F \over \sqrt{2}}
( m_{D^{*+}_s} f_{D^{*+}_s} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow \bar D_2^{*-}} (m^2_{D^{*+}_s})) \left\{ V^*_{cb} V_{cs} a_1
- V^*_{tb} V_{ts} (a_4 + a_{10}) \right\},
\\
A(B^+ \rightarrow J/\psi K_2^{*+}) &=& i {G_F \over \sqrt{2}}
( m_{J/\psi} f_{J/\psi} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow K_2^{*+}} (m^2_{J/\psi})) \left\{
V^*_{cb} V_{cs} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{ts} (a_3 +a_5 +a_7 +a_9 ) \right\},
\\
A(B^0 \rightarrow J/\psi K_2^{*0}) &=& i {G_F \over \sqrt{2}}
( m_{J/\psi} f_{J/\psi} \epsilon^{* \alpha \beta} F_{\alpha \beta}^{B
\rightarrow K_2^{*0}} (m^2_{J/\psi})) \left\{
V^*_{cb} V_{cs} a_2 \right.
\nonumber \\
&\mbox{}& \left. - V^*_{tb} V_{ts} (a_3 +a_5 +a_7 +a_9 ) \right\}.\end{aligned}$$
Form factor for $B \to T$ ISGW($m_D^2$) ISGW($t_m$) ISGW2
--------------------------- --------------- ------------- -------
$F^{B \to a_2}$ $-0.046$ $-0.203$ 0.101
$F^{B \to f_2}$ $-0.045$ $-0.205$ 0.099
$F^{B \to f_2^{\prime}}$ $-0.052$ $-0.191$ 0.134
$F^{B \to K_2^*}$ $-0.049$ $-0.111$ 0.131
$F^{B \to D_2^*}$ $-0.060$ 0.378 0.367
: Form factors for $B \to T$ transitions calculated at $q^2
= m_D^2$ $(q^{\mu} \equiv p_B^{\mu} -p_T^{\mu})$, at the maximum momentum transfer $t_m \equiv (m_B -m_T)^2$ in the ISGW model, and at $q^2 = m_D^2$ in the ISGW2 model, respectively.
Decay mode $\xi=0.1$ $\xi=0.3$ $\xi=0.5$ $a_1=1.15$, $a_2=0.26$
----------------------------- ------------------ -------------- -------------- ------------------------
$\propto V_{cb}^*V_{ud}$
$B^+\to\pi^+ \Dbar_2^{*0}$ 339.63 297.22 257.64 350.83
$B^+\to \Dbar^0 a_2^{+}$ 92.82 0.32 72.27 101.86
$B^0\to \pi^{+} D_2^{*-}$ 318.96 279.13 241.96 329.48
$B^0\to \Dbar^0 a_2^0$ 43.55 0.15 33.91 47.79
$B^0\to \Dbar^0 f_2$ 48.56 0.17 37.81 53.29
$B^0\to \Dbar^0 f_2'$ 0.57 0.002 0.44 0.62
$\propto V_{cb}^*V_{cd}$
$B^+\to D^{+} \Dbar_2^{*0}$ 22.23 19.45 16.86 22.68
$B^+\to \eta_c a_2^+$ 4.17 0.004 3.73 4.89
$B^0\to D^{+} D_2^{*-}$ 20.87 18.27 15.83 21.30
$B^0\to \eta_c a_2^0$ 1.96 0.002 1.75 2.30
$B^0\to \eta_c f_2$ 2.27 0.002 2.03 2.67
$B^0\to \eta_c f_2'$ 0.019 0.00002 0.017 0.02
: Branching ratios of $B\to PT$ with $\Delta S=0$ in units of $10^{-6}$, calculated in the ISGW2 model.
Decay mode $\xi=0.1$ $\xi=0.3$ $\xi=0.5$ $a_1=1.15$, $a_2=0.26$
------------------------------ ------------------ -------------- -------------- ------------------------
$\propto V_{cb}^*V_{cs}$
$B^+\to D_s^{+}\Dbar_2^{*0}$ 493.04 431.56 373.61 493.04
$B^+\to \eta_c K_2^{*+}$ 81.19 0.042 88.71 105.39
$B^0\to D_s^{+} D_2^{*-}$ 462.95 405.22 350.81 462.95
$B^0\to \eta_c K_2^{*0}$ 74.46 0.038 81.35 96.64
$\propto V_{cb}^*V_{us}$
$B^+\to K^{+} \Dbar_2^{*0}$ 24.64 21.56 18.69 25.45
$B^+\to \Dbar^{0} K_2^{*+}$ 6.69 0.023 5.21 7.34
$B^0\to K^{+} D_2^{*-}$ 23.14 20.25 17.56 23.91
$B^0\to \Dbar^{0} K_2^{*0}$ 6.19 0.021 4.82 6.80
$\propto V_{ub}^*V_{cs}$
$B^+\to D_s^{+} a_2^0$ 9.14 8.00 6.93 9.44
$B^+\to D_s^{+} f_2$ 10.20 8.96 7.74 10.54
$B^+\to D_s^{+} f_2'$ 0.12 0.10 0.09 0.12
$B^+\to D^0 K_2^{*+}$ 1.07 0.004 0.83 1.17
$B^0\to D_s^+ a_2^-$ 17.15 15.01 13.01 17.71
$B^0\to D^{0} K_2^{*0}$ 0.99 0.003 0.77 1.08
: Branching ratios of $B\to PT$ with $|\Delta S|=1$ in units of $10^{-6}$, calculated in the ISGW2 model.
Decay mode $\xi=0.1$ $\xi=0.3$ $\xi=0.5$ $a_1=1.15$, $a_2=0.26$
-------------------------------- ----------------- -------------- -------------- ------------------------
$\propto V_{cb}^*V_{ud}$
$B^+\to \rho^{+} \Dbar_2^{*0}$ 950.15 831.51 720.77 981.48
$B^+\to \Dbar^{*0} a_2^+$ 151.77 0.53 118.16 166.54
$B^0\to \rho^{+} D_2^{*-}$ 892.23 780.82 676.83 921.64
$B^0\to \Dbar^{*0} a_2^0$ 71.20 0.25 55.44 78.14
$B^0\to \Dbar^{*0} f_2$ 76.82 0.27 59.81 84.30
$B^0\to \Dbar^{*0} f_2'$ 0.95 0.003 0.74 1.05
$\propto V_{cb}^*V_{cd}$
$B^+\to D^{*+} \Dbar_2^{*0}$ 50.05 43.78 37.94 53.25
$B^+\to J/\psi a_2^+$ 14.21 0.059 10.78 16.41
$B^0\to D^{*+} D_2^{*-}$ 46.98 41.10 35.61 49.99
$B^0\to J/\psi a_2^0$ 6.67 0.028 5.60 7.70
$B^0\to J/\psi f_2$ 7.28 0.03 5.53 8.41
$B^0\to J/\psi f_2'$ 0.074 0.0003 0.056 0.09
: Branching ratios of $B\to VT$ with $\Delta S=0$ in units of $10^{-6}$, calculated in the ISGW2 model.
Decay mode $\xi=0.1$ $\xi=0.3$ $\xi=0.5$ $a_1=1.15$, $a_2=0.26$
-------------------------------- ----------------- -------------- -------------- ------------------------
$\propto V_{cb}^*V_{cs}$
$B^+\to D_s^{*+} \Dbar_2^{*0}$ 1080.17 944.61 818.12 1200.8
$B^+\to J/\psi K_2^{*+}$ 307.66 1.66 224.02 383.62
$B^0\to D_s^{*+} D_2^{*-}$ 1013.89 886.64 767.92 1127.12
$B^0\to J/\psi K_2^{*0}$ 284.10 1.53 206.87 354.25
$\propto V_{cb}^*V_{us}$
$B^+\to K^{*+} \Dbar_2^{*0}$ 50.64 44.31 38.41 52.31
$B^+\to \Dbar^{*0} K_2^{*+}$ 11.04 0.038 8.59 12.11
$B^0\to K^{*+} D_2^{*-}$ 47.55 41.61 36.07 49.12
$B^0\to \Dbar^{*0} K_2^{*0}$ 10.25 0.035 7.98 11.24
$\propto V_{ub}^*V_{cs}$
$B^+\to D_s^{*+} a_2^0$ 15.00 13.13 11.38 15.49
$B^+\to D_s^{*+} f_2$ 16.17 14.15 12.26 16.70
$B^+\to D_s^{*+} f_2'$ 0.20 0.17 0.15 0.21
$B^+\to D^{*0} K_2^{*+}$ 1.76 0.006 1.37 1.93
$B^0\to D_s^{*+} a_2^-$ 28.15 24.63 21.35 29.08
$B^0\to D^{*0} K_2^{*0}$ 1.63 0.006 1.27 1.79
: Branching ratios of $B\to VT$ with $|\Delta S|=1$ in units of $10^{-6}$, calculated in the ISGW2 model.
-----------------------------------------------------------------------------------------------------------
Ratio $\xi =0.1$ $\xi =0.3$ $\xi =0.5$
-------------------------------------------------------------------- ------------ ------------ ------------
${\cal B}(B^{+(0)}\to \rho^{+} \Dbar_2^{*0(-)}) 2.80 2.80 2.80
/{\cal B}(B^{+(0)}\to \pi^{+} \Dbar_2^{*0(-)})$
${\cal B}(B^{+(0)}\to \Dbar^{*0} a_2^{+(0)}) 1.64 1.64 1.63
/{\cal B}(B^{+(0)}\to \Dbar^{0} a_2^{+(0)})$
${\cal B}(B^0\to \Dbar^{*0} f_2)/{\cal B}(B^0\to \Dbar^{0} f_2)$ 1.58 1.58 1.58
${\cal B}(B^0\to \Dbar^{*0} f_2')/{\cal B}(B^0\to \Dbar^{0} f_2')$ 1.68 1.68 1.68
${\cal B}(B^{+(0)}\to D^{*+} \Dbar_2^{*0(-)}) 2.25 2.25 2.25
/{\cal B}(B^{+(0)}\to D^{+} \Dbar_2^{*0(-)})$
${\cal B}(B^{+(0)}\to J/\psi a_2^{+(0)}) 3.41 14.34 2.89
/{\cal B}(B^{+(0)}\to \eta_c a_2^{+(0)})$
${\cal B}(B^0\to J/\psi f_2)/{\cal B}(B^0\to \eta_c f_2)$ 3.21 13.50 2.72
${\cal B}(B^0\to J/\psi f_2')/{\cal B}(B^0\to \eta_c f_2')$ 3.83 16.11 3.24
${\cal B}(B^{+(0)}\to D_s^{*+} \Dbar_2^{*0(-)}) 2.19 2.19 2.19
/{\cal B}(B^{+(0)}\to D_s^{+} \Dbar_2^{*0(-)})$
${\cal B}(B^{+(0)}\to J/\psi K_2^{*+(0)}) /{\cal B}(B^{+(0)}\to 3.79 39.82 2.53
\eta_c K_2^{*+(0)})$
${\cal B}(B^{+(0)}\to K^{*+} \Dbar_2^{*0(-)}) 2.06 2.06 2.06
/{\cal B}(B^{+(0)}\to K^{+} \Dbar_2^{*0(-)})$
${\cal B}(B^{+(0)}\to \Dbar^{*0} K_2^{*+(0)}) /{\cal 1.65 1.65 1.65
B}(B^{+(0)}\to \Dbar^{0} K_2^{*+(0)})$
${\cal B}(B^{+(0)}\to D_s^{*+} a_2^{0(-)}) 1.64 1.64 1.64
/{\cal B}(B^{+(0)}\to D_s^+ a_2^{0(-)})$
${\cal B}(B^+\to D_s^{*+} f_2)/{\cal B}(B^+\to D_s^{+} f_2)$ 1.58 1.58 1.58
${\cal B}(B^+\to D_s^{*+} f_2')/{\cal B}(B^+\to D_s^{+} f_2')$ 1.69 1.69 1.69
${\cal B}(B^{+(0)}\to D^{*0} K_2^{*+(0)}) /{\cal B}(B^{+(0)}\to 1.65 1.65 1.65
D^{0} K_2^{*+(0)})$
-----------------------------------------------------------------------------------------------------------
: Ratios of ${\cal B}(B\to VT)/{\cal B}(B\to PT)$
Decay mode $\xi =0.1$ $\xi =0.3$ $\xi =0.5$
--------------------------------------- ------------ ------------ ------------ -- -- --
$B^{+(0)} \to D^{+} \Dbar_2^{*0 (-)}$ $0.001$ $0.001$ $0.001$
$B^{+(0)} \to D^{*+} \Dbar_2^{*0(-)}$ $-0.004$ $-0.004$ $-0.004$
$B^{+(0)} \to J/\psi a_2^{+(0)}$ $-0.0082$ $-0.0045$ $-0.0087$
$B^0 \to J/\psi f_2^{(\prime)}$ $-0.0082$ $-0.0045$ $-0.0087$
: CP asymmetries for $B \to PT$ and $B \to VT$
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
[^4]: Recently the Belle Collaboration measured the BR of $B^+ \to K^+ \pi^+ \pi^-$, where two known candidate states for a $\pi^+ \pi^-$ invariant mass around 1300 MeV are $f_2 (1270)$ and $f_0 (1370)$ [@bellePT]. Because our previous result using the ISGW model predicts a rather small BR for $B^+ \to f_2(1270) K^+$ [@btopt], they concluded that the measurements would provide evidence for a significant nonfactorizable effect, if the peak were due to $f_2 (1270)$. However, our recent result using the improved version of the model (ISGW2) shows that the BR of $B^+ \to f_2(1270) K^+$ is enhanced by an order of magnitude [@newbtoptvt].
[^5]: In addition, we also study a few $B \to PT$ and $VT$ modes involving a $b \to u$ transition, such as $B \to D_s^{(*)} a_2$, $B \to D_s^{(*)} f_2^{(\prime)}$, and $B \to D^{(*)} K_2^*$.
[^6]: In the frameworks of the QCD factorization and the perturbative QCD approaches, nonfactorizable effects vary for different four-quark operators: e.g., $\xi$ is different for tree- and penguin-dominated processes. But within our generalized factorization framework, the $\xi$ is assumed universal.
[^7]: In addition to the strong phases, there can be other possible sources for the strong phases: for example, in the QCD factorization a large strong phase for the WC, $a_2$, can be induced by hard gluon exchange between final meson states, and in the perturbative QCD approach large absorptive parts can be generated from the weak annihilation diagrams. But, because in the relevant modes the effect from the tree is much larger than that from the penguin, as just mentioned in the text, the resultant asymmetries would remain relatively small.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Yosuke Enokida
- Atsushi Suzuki
- Kenji Yamanishi
title: Geodesic Update on Hyperbolic Space and its Application to Embeddings
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Our aim is to investigate the tidal interaction in High-mass X-ray Binaries and Symbiotic stars in order to determine in which objects the rotation of the mass donors is synchronized or pseudosynchronized with the orbital motion of the compact companion. We find that the Be/X-ray binaries are not synchronized and the orbital periods of the systems are greater than the rotational periods of the mass donors. The giant and supergiant High-mass X-ray binaries and symbiotic stars are close to synchronization. We compare the rotation of mass donors in symbiotics with the projected rotational velocities of field giants and find that the M giants in S-type symbiotics rotate on average 1.5 times faster than the field M giants. We find that the projected rotational velocity of the red giant in symbiotic star MWC 560 is $= 8.2 \pm 1.5$ , and estimate its rotational period to be $P_{\rm rot}$ = 144 - 306 days. Using the theoretical predictions of tidal interaction and pseudosynchronization, we estimate the orbital eccentricity $e=0.68-0.82$.'
author:
- 'Kiril Stoyanov & Radoslav Zamanov'
title: 'Rotation of the Mass Donors in High-mass X-ray Binaries and Symbiotic Stars'
---
[2]{}
Introduction
============
A High-mass X-ray binary system consists of a compact object (a neutron star or a black hole) accreting material from an O or B companion star. They are divided into Be/X-ray binaries (main-sequence star as a companion) and Giant/Supergiant X-ray binaries (giant or supergiant star as a companion). Accretion of matter is different for both types of X-ray binaries. In the Be/X-ray binaries, the compact object crosses the circumstellar disc and accretes matter from that disk. In the giant/supergiant X-ray binaries, the mass donor ejects a slow and dense wind radially outflowing from the equator and the compact object directly accretes the stellar wind through Bondi-Hoyle-Lyttleton accretion.
Symbiotic stars are interacting binaries, consisting of an evolved giant (either a normal red giant in S-types symbiotics or a Mira-type variable in D-types symbiotics) transferring mass to a hot and luminous white dwarf or neutron star. The symbiotic stars are surrounded by a rich and luminous nebula resulting from the presence of both an evolved giant with a heavy mass-loss and of a hot companion abundant in ionizing photons and often emanating its own wind.
Synchronization and Pseudosynchronization
=========================================
In a binary with a circular orbit the rotational period of the primary, P$_{rot}$, reaches an equilibrium value at the orbital period, $P_{orb}=P_{rot}$. In other words the synchronous rotation (synchronization) means that the rotational period is equal to the orbital period. In a binary with an eccentric orbit, the corresponding equilibrium is reached at a value of $P_{rot}$ which is less than $P_{orb}$, the amount less being a function solely of the orbital eccentricity $e$. In practice, in a binary with an eccentric orbit the tidal force acts to synchronize the rotation of the mass donor with the motion of the compact object at the periastron - pseudosynchronous rotation. To calculate the period of pseudosynchronization, P$_{ps}$, we use (Hut 1981): $$P_{ps} = \frac{(1+3e^2+\frac{3}{8}e^4)(1-e^2)^\frac{3}{2}}{1+\frac{15}{2}e^2+
\frac{45}{8}e^4+\frac{5}{16}e^6} P_{orb}.
\label{Eq-ps}$$
Stars with radiative envelopes
------------------------------
Following Hurley, Tout & Pols (2002) the circularization timescale for stars with radiative envelopes is: $$\frac{1}{{\tau}_{circ}}=
\frac{21}{2} \left( \frac{G M_1}{R_1^3} \right)^\frac{1}{2}
q_2 \left(1 + q_2 \right)^\frac{11}{6} E_2 \left(\frac{R_1}{a} \right)^\frac{21}{2},
\label{circ_re}$$ where M$_1$ and R$_1$ are the mass and the radius of the primary respectively, q$_2$ is the mass ratio M$_2/$M$_1$, and $a$ is the semi-major axis. The second-order tidal coefficient E$_2$ = 1.592 $\times 10^{-9} M_1^{2.84}$.
The synchronization time scale is given as, $$\tau_{sync} = K \tau_{circ},
\label{sync_re}$$ where K is: $$K \approx \frac{0.015}{r_g} \, \frac{1 + q_2}{q_2} \left( \frac{R_1}{a} \right)^2.$$ For the gyration radius of the primary $r_g$ we adopt $r_g\approx0.16$ for giants, and $r_g\approx0.25$ for main sequence stars (Claret & Gimenez, 1989).
Stars with convective envelopes
-------------------------------
Following Hurley, Tout & Pols (2002) the synchronization timescale for stars with convective envelopes is: $$\tau_{\rm syn} \approx 800 \left( \frac{ M_{\rm 1} R_{\rm 1}}{ L_{\rm 1}}\right)^{1/3}
\frac{M_{\rm 1}^2 (\frac{M_{\rm 1}}{M_{\rm 2}} + 1)^2}{R_{\rm 1}^6} P_{\rm orb}^4\ \; {\rm yr},
\label{sync}$$ where $L_{\rm 1}$ is the luminosity of the giant.
The circularization time scale is: $$\frac{1}{{\tau}_{\rm circ}} = \frac{21}{2} \left(\frac{k}{T}\right) q{_2} (1+q{_2})
\left(\frac{R_{\rm 1}}{a}\right)^8 .
\label{circ}$$ In Eq. \[circ\], $(k/T)$ is: $$\left(\frac{k}{T}\right) = \frac{2}{21} \frac{f_{\rm conv}}{{\tau}_{\rm conv}} \frac{M_{\rm env}}{M_{\rm 1}}
\; {\rm yr}^{-1} ,
\label{KT}$$ where $R_{\rm env}$ is the depth of the convective envelope, $M_{\rm env}$ is the envelope’s mass, and $${\tau}_{\rm conv} = 0.4311 \left(\frac{M_{\rm env}R_{\rm env}(R_{\rm 1}-\frac{1}{2}R_{\rm env})}{3 L_{\rm 1}}\right) ^{\frac{1}{3}} {\rm yr}
\label{tau}$$ is the eddy turnover time scale (the time scale on which the largest convective cells turnover). The numerical factor $f_{\rm conv}$ is $${f}_{\rm conv} = {\rm min} \left[1, \left( \frac{P_{\rm tid}}{2{\tau}_{\rm conv}} \right) ^2 \right],
\label{fconv}$$ where $P_{\rm tid}$ is the tidal pumping time scale given by $$\frac{1}{P_{\rm tid}} = \left|\frac{1}{P_{\rm orb}} - \frac{1}{P_{\rm rot}}\right|.
\label{ptid}$$
The pseudosynchronization timescale is $\tau_{\rm ps}$ = (7/3($\alpha$ - 3)) $\tau_{\rm circ}$, where $\alpha$ is a dimensionless quantity, representing the ratio of the orbital and rotational angular momentum:
$$\alpha = \frac {q_2}{1+q_2} \frac {1}{r_{\rm g}^{2}} \left( \frac {a}{R_{\rm 1}} \right) ^2 .$$
For a red giant we adopt $r_{\rm g} \approx 0.3$ (Claret, 2007).
In all equations, the masses, the radii and the lumunosities are in solar units.
High-mass X-ray Binaries
========================
The orbital and stellar parameters of 13 High-mass X-ray binaries are given in Table 1 and Table 2 in Stoyanov & Zamanov (2009). We add 2 more objects - 4U 2206+54 and MWC 148. The orbital and stellar parameters are taken from Rib[ó]{} et. al. (2006) and Casares et al. (2012) respectively.
Using Eq.\[circ\_re\] and Eq.\[sync\_re\] we estimate the circularization and synchronization timescales. The results are given in Table \[t-times\]. The lifetime of a star on the main sequence can be estimated as $\tau_{MS} = 10^{10} ({M_\odot}/{M})^{2.5} \; \; {\rm years}$ (Hansen & Kawaler, 1994). Comparing these lifetimes with $\tau_{sync}$ from Table \[t-times\], we see that among the Be/X-ray binaries only for LSI+61$^0$303 is $\tau_{sync} \sim \tau_{MS}$. This is the only Be/X-ray binary for which we can expect considerable changes of the rotation of the primary during the lifetime of the Be star.
The lifetime of the giant is comparable or longer then $\tau_{circ}$ and $\tau_{sync}$ for the giant/supergiant systems with short orbital periods. The exceptions are V725 Tau and BP Cru, for which $\tau_{sync}$ and $\tau_{circ}$ are longer than the lifetime of the giant/supergiant stage.
On Fig.1 in Stoyanov & Zamanov (2009) is plotted P$_{rot}$ versus P$_{ps}$. The giant/supergiant systems are located close to the line P$_{ps} = P_{rot}$, while those with mass donors from spectral class V are far away from the equilibrium.
In the Be/X-ray systems BQ Cam, V635 Cas, V725 Tau and 4U 2206+54, the tidal force spinning down the donor star. For the system LSI+61$^0$303, the rotation of the mass donor is close to pseudosynchronization. This is the only Be/X-ray binary in which $\tau_{sync}$ is comparable with the life-time of the binary. In the binaries X Per and MWC 148, the neutron star is far away from the Be star and the tidal force is weak.
Giant and supergiant systems are close to (pseudo)synchronization. In these binaries the rotation of the mass donors is influenced by the presence of the compact object. In LMC X-4 and Cen X-3, the mass donors are synchronized and the orbits are circularized. With respect to the rotation of the mass donor, V725 Tau is similar to the Be/X-ray binaries. Cyg X-1 is synchronized and almost circularized. V830 Cen is pseudosynchronized but not circularized yet. The systems LSI$+65^0010$ and Vela X-1 are close to pseudosynchronization and the tidal force accelerates the rotation of the mass donors. In the case of SMC X-1, the tidal force acts as a decelerator of the rotation of the mass donor. In BP Cru, a gas stream from the mass donor exists, probably resulting from the strong tidal force and spin-up of the mass donor (Leahy & Kostka, 2008).
[lllc]{} object & $\tau_{sync}$ \[yr\] & $\tau_{circ}$ \[yr\] & lifetime \[yr\]\
\
LSI+61$^0$303\*& 2.810$^7$ & 2.410$^8$ &5.610$^7$\
X Per & 6.210$^{17}$ & 1.810$^{21}$ &1.110$^7$\
BQ Cam & 3.510$^{11}$ & 7.610$^{13}$ &3.910$^6$\
V635 Cas & 1.410$^{11}$ & 9.510$^{12}$ &7.310$^6$\
4U 2206+54 & 4.910$^9$ & 3.710$^{11}$ &7.310$^7$\
MWC 148 & 1.210$^{17}$ & 5.110$^{20}$ &9.810$^7$\
\
V725 Tau & 2.810$^{12}$ & 8.010$^{14}$ &4$10^5$\
LMC X-4 & 4.510$^2$ & 7.710$^2$ &1$10^6$\
Cen X-3 & 2.310$^3$ & 4.210$^3$ &5$10^5$\
\
V830 Cen & 7.510$^3$ & 1.410$^4$ & 110$^6$\
LSI+65$^0$010 & 1.310$^4$ & 3.910$^4$ & 110$^6$\
Vela X-1 & 1.010$^4$ & 2.810$^4$ & 3.910$^5$\
SMC X-1 & 3.310$^4$ & 8.210$^4$ & 8.810$^5$\
BP Cru & 1.810$^6$ & 8.810$^6$ & 810$^4$\
Cyg X-1 & $<1$ & $<1$ & 110$^5$\
\*assuming neutron star as a secondary component.
S-type symbiotic stars
======================
43 symbiotic stars have been observed with FEROS spectrograph at the 2.2m ESO telescope of the La Silla Observatory (Zamanov et al. 2007). The data for the rotation of 55 field red giants are taken from the literature. M giants in S-type symbiotics rotate faster than the field M giants. Histograms of the available data for the red giants are plotted in Fig.2 in Zamanov & Stoyanov (2012). For the field M0III-M6III giants we calculate a mean $=$5.0 , median $=$4.3 , and standard deviation of the mean $\sigma=4.0$ . For the M0III-M6III giants in symbiotics, we get a mean $=7.8$ , median $=$8.0 , and standard deviation of the mean $\sigma=$2.1 .
There are 5 objects in our sample that deviate from the synchronization. These objects are RS Oph, MWC 560, CH Cyg, CD-43$^\circ$14304 and Z And. In three of them collimated jets are detected: Z And (Skopal et al. 2009); CH Cyg (Crocker et al. 2002), MWC 560 (Tomov et al. 1990). Additionally to the jets, ejection of blobs are detected from RS Oph and CH Cyg (Iijima et al. 1994). This confirms the suggestions that in the jet-ejecting symbiotics the mass donors rotate faster than the orbital periods. Probably there is a link between the jets and the mass donor rotation.
On Fig.\[stoyanov-fig1\] are plotted together the High-mass X-ray binaries and the S-type symbiotic stars. It shows that none of the objects in our sample is above the line of synchronization.
Orbital eccentricity of MWC 560
===============================
MWC 560 is a symbiotic star, which consists of a red giant and a white dwarf (Tomov et al. 1990). The most spectacular features of this object are the collimated ejections of matter with velocities of up to $\sim 6000$ (Tomov et al. 1992) and the resemblance of its emission line spectrum to that of the low-redshift quasars (Zamanov & Marziani, 2002). The jet ejections are along the line of sight and the system is seen almost pole-on ($i < 16^\circ$). This makes it difficult to obtain the orbital eccentricity of the system in a conventional way.
For the system we adopt $R_{\rm g} = 140 \pm 7~R_\odot$, $L_{\rm g} \sim 2400~L_\odot$, $M_{\rm g} = 1.7~M_\odot$, $M_{\rm wd}=0.65~M_\odot$, $P_{\rm orb}=1931 \pm 162$ day (Gromadzki et al. 2007), $R_{\rm env}=0.9~R_{\rm g}$ and $M_{\rm env}= 1.0~M_\odot$ (Herwig 2005). With the above values of the parameters assumed, we derive the semi-major axis of the orbit to be $a \approx 860~R_\odot$. Using these parameters, we calculate from Eq. \[sync\] and Eq. \[circ\] the synchronization and circularization time scales: $\tau_{\rm sync} = 2.6 \times 10^4$ yr and $\tau_{\rm circ} = 3.1 \times 10^6$ yr. The typical lifetime of a symbiotic star is $\tau_{\rm ss} \sim 10^5$ yr (Yungelson et al., 1995). From the rate of accretion on the white dwarf, $\dot M_{\rm acc} \approx 5 \times 10^{-7}~M_\odot$ (Schmid et al. 2001), we can estimate, that it will take $10^6$ yr to accrete $\sim 0.5~M_\odot$ from the envelope of the red giant companion. Because the giant also losses mass via stellar wind, we find that the lifetime of the symbiotic phase of MWC 560 should be $\tau_{\rm ss} \leq 10^6$ yr. For MWC 560 we have therefore the situation in which $\tau_{\rm ps} < \tau_{\rm syn} < \tau_{\rm ss} < \tau_{\rm circ}$. This means that the symbiotic phase is long enough that the tidal forces can (pseudo)synchronize the rotation of the red giant. On the other hand, the value of $\tau_{\rm circ}$ demonstrates that the symbiotic lifetime of MWC 560 is shorter than the circularization time, and therefore the orbit can be eccentric. This is in agreement with the observational evidences found by Fekel et al. (2007) that the symbiotic stars with P$_{\rm orb} > 800$ days tend to have eccentric orbits. The above implies that in MWC 560, the red giant is probably synchronized, but the orbit is not circularized. To determine the orbital eccentricity of MWC 560, we need to calculate $P_{\rm rot}$ for the mass donor. We analyzed 21 high resolution spectra of MWC 560 and obtained value for $= 8.2 \pm 1.5$ . Using $R_{\rm g} = 140 \pm 7~R_\odot$ and $\; i=12^\circ-16^\circ$, we calculate $P_{\rm rot}=144 - 306$ days. This value is less than the orbital period. MWC 560 should be close to synchronization or pseudosynchronization, and $P_{\rm rot} =P_{\rm ps}$. Using Eq. \[Eq-ps\] we can estimate the orbital eccentricity to be $e=0.68-0.82$.
Conclusions
===========
Using rotational velocity measurements and the theory of synchronization/pseudosynchronization we:
\(1) find that the Be/X-ray binaries are far away from (pseudo)synchronization. The tidal force in the Be/X-ray binaries acts as a decelerator of the rotation of the mass donors. The only Be/X-ray binary which is close to pseudosynchronization is the LSI$+61^0303$. The objects containing mass donors of spectral class I and III typically have $P_{rot} \sim P_{ps}$ and are close to (pseudo)synchronization;
\(2) demonstrate that the M giants in symbiotic stars rotate faster than the field giants. Most symbiotics with orbital period less than 1000 d are synchronized;
\(3) show that the High-mass X-ray binaries and the S-type symbiotic stars are either on the line of synchronization or they are under the line. None of the objects in our sample is above the line of synchronization.
\(4) calculate that the orbit of the symbiotic star MWC 560 should be highly eccentric, with $e\sim$ 0.7.
We thank the anonymous referee for constructive comments. This work was supported by the OP “HRD“, ESF and Bulgarian Ministry of Education, Youth and Science under the contract BG051PO001-3.3.06-0047.
[99]{}
Casares, J., Rib[ó]{}, M., Ribas, I., Paredes, J. M., Vilardell, F., & Negueruela, I.: 2012, MNRAS, 421, 1103 Claret, A.: 2007, A&A 467, 1389 Claret, A., & Gimenez, A.: 1989, A&AS, 81, 37 Crocker, M. M., Davis, R. J., Spencer, R. E., et al.: 2002, MNRAS, 335, 1100 Fekel, F. C., Hinkle, K. H., Joyce, R. R., Wood, P. R., Lebzelter, T.: 2007, AJ 133, 17 Gromadzki, M., Miko[ł]{}ajewska, J., Whitelock, P. A., Marang, F.: 2007, A&A 463, 703 Hansen, C. J., & Kawaler, S. D.: 1994, [*Stellar Interiors. Physical Principles, Structure, and Evolution*]{}, Springer-Verlag Herwig, F.: 2005, ARA&A 43, 435 Hurley, J. R., Tout, C. A., & Pols, O. R.: 2002, MNRAS, 329, 897 Hut, P.: 1981, A&A, 99, 126 Iijima, T., Strafella, F., Sabbadin, F., & Bianchini, A.: 1994, A&A, 283, 919 Leahy, D. A., & Kostka, M.: 2008, MNRAS, 384, 747 Rib[ó]{}, M., Negueruela, I., Blay, P., Torrej[ó]{}n, J. M., & Reig, P.: 2006, A&A, 449, 687 Schmid, H. M., Kaufer, A., Camenzind, M., Rivinius, T., Stahl, O., Szeifert, T., Tubbesing, S., Wolf, B.: 2001, A&A 377, 206 Skopal, A., Pribulla, T., Budaj, J., et al.: 2009, ApJ, 690, 1222 Stoyanov, K. A., & Zamanov, R. K.: 2009, Astronomische Nachrichten, 330, 727 Tomov, T., Zamanov, R., Kolev, D., Georgiev, L., Antov, A., Mikolajewski, M., Esipov, V.: 1992, MNRAS 258, 23 Tomov, T., Kolev, D., Zamanov, R., Georgiev, L., Antov, A.: 1990, Nature 346, 637 Yungelson, L., Livio, M., Tutukov, A., & Kenyon, S. J.: 1995, ApJ 447, 656 Zamanov, R. K., Bode, M. F., Melo, C. H. F., et al.: 2007, MNRAS, 380, 1053 Zamanov, R., Marziani, P.: 2002, ApJL 571, L77 Zamanov, R. K., & Stoyanov, K. A.: 2012, Bulgarian Astronomical Journal, 18, 41
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a generalization of the random matrix ensembles, including all classical ensembles. Then we derive the joint density function of the generalized ensemble by one simple formula, which give a direct and unified way to compute the density functions for all classical ensembles and various kinds of new ensembles. An integration formula associated with the generalized ensemble is also given. We also give a classification scheme of the generalized ensembles, which will include all classical ensembles and some new ensembles which were not considered before.'
address:
- 'School of mathematical science, Peking University, Beijing, 100871, P. R. China'
- 'School of mathematical science, Peking University, Beijing, 100871, P. R. China '
- 'School of mathematics and physics, Zhejiang Normal University, Zhejiang Jinhua, 321004, P. R. China '
author:
- Jinpeng An
- Zhengdong Wang
- Kuihua Yan
title: 'A Generalization of random matrix ensemble I: general theory$^1$'
---
0.5cm
Introduction
============
0.5cm One of the most fundamental problems in the theory of random matrices is to derive the joint density functions for the eigenvalues (or equivalently, the measures associated with the eigenvalue distributions) of various types of matrix ensembles. In his monograph [@Me], Mehta summarized the classical analysis methods by which the density functions for various types of ensembles were derived case by case. But a systematical method to compute the density functions was desired.
0.3cm The first achievement in this direction was made by Dyson [@Dy], who introduced an idea of expressing various kinds of circular ensemble in terms of symmetric spaces with invariant probability measures. From then on, guided by Dyson’s idea, many authors observed new random matrix ensembles in terms of Cartan’s classification of Riemannian symmetric spaces, and obtained the joint density functions for such ensembles using the integration formula on symmetric space (see, for example, [@AZ; @Ca1; @Ca2; @Du; @Iv; @TBFM; @Zi]). Here we mention the recent work of Dueñez [@Du] briefly. Dueñez explored the random matrix ensembles which correspond to infinite families of compact irreducible Riemannian symmetric spaces of type I, including circular orthogonal and symplectic ensembles and various kinds of Jacobi ensembles. Using an integration formula associated with the $KAK$ decomposition of compact groups, he obtained the induced measure on the space of eigenvalues associated with the underlying symmetric space, and then derived the eigenvalue distribution of the corresponding random matrix ensemble. These methods of deriving the eigenvalue distributions of random matrix ensembles by means of Riemannian symmetric spaces were summarized by the excellent review article of Caselle and Magnea [@CM].
0.3cm In this paper we provide a generalization of the random matrix ensembles, including all classical ensembles, and then give an unified way to derive the joint density function for the eigenvalue distribution by one simple formula. The proof of this formula make no use of integration formula. In fact, the corresponding integration formula can be derived from this formula as corollary. We also give a classification scheme of the generalized random matrix ensembles, which will include all classical ensembles and some new ensembles which were not considered before.
0.3cm More precisely, Let $\sigma:G\times X\rightarrow X$ be a smooth action of a Lie group $G$ on a Riemannian manifold $X$, preserving the induced Riemannian measure $dx$. Let $p(x)$ be a $G$-invariant smooth function on $X$, and consider the measure $p(x)dx$ on $X$, which is not necessary a finite measure. We choose a closed submanifold $Y$ of $X$ consisting of representation points for almost all $G$-orbits in $X$. The Riemannian structure on $X$ induces a Riemannian measure $dy$ on $Y$. Let $K$ be the closed subgroup of $G$ which fixes all points in $Y$, then the map $\sigma$ reduces to a map $\varphi:G/K\times
Y\rightarrow X$. Suppose there is a $G$-invariant measure $d\mu$ on $G/K$, and suppose $\dim(G/K\times Y)=\dim X$, then it can be proved that the pull back measure $\varphi^*(p(x)dx)$ of the measure $p(x)dx$ is of the form $\varphi^*(p(x)dx)=d\mu d\nu$ for some measure $d\nu$ on $Y$, which is just the measure associated with the eigenvalue distribution. The measure $d\nu$ can be expressed as the form $d\nu(y)=\mathcal{P}(y)dy$ for some function $\mathcal{P}(y)$ on $Y$, which is just the joint density function. We write $\mathcal{P}(y)$ as the form $\mathcal{P}(y)=p(y)J(y)$, then under some orthogonality condition (that is $T_yY\perp
T_yO_y$ for almost all $y\in Y$), we can compute the factor $J(y)$ by the following formula $$\label{E:main}
J(y)=C|\det\Psi_y|,$$ where $C$ is a constant, which can also be computed explicitly. This formula is the main result of this paper, the density function $\mathcal{P}(y)$ and the eigenvalue distribution $d\nu$ are determined by it. Here the map $\Psi_y:{\mathfrak{l}}\rightarrow T_yO_y$ is defined by $$\Psi_y(\xi)=\frac{d}{dt}\Big|_{t=0}\hskip 0.1cm \sigma_{\exp
t\xi}(y),$$ where ${\mathfrak{l}}$ is a linear subspace of the Lie algebra ${\mathfrak{g}}$ of $G$ such that ${\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{l}}$, ${\mathfrak{k}}$ is the Lie algebra of $K$. We call the system $(G,\sigma,X,p(x)dx,Y,dy)$ a *generalized random matrix ensemble*. The measure $d\nu$ and the function $\mathcal{P}(y)$ on $Y$ are called *generalized eigenvalue distribution* and *generalized joint density function*, respectively. Using Formula , one can derive the joint density function for Gaussian ensemble, chiral ensemble, new transfer matrix ensembles, circular ensemble, Jacobi ensemble, and some other new generalized ensembles. The precise deriving process will be the content of a sequel paper [@AWY]. Here we should point out that the proof of Formula is not difficult, but this formula is very effective and available. The derivation of all concrete examples in [@AWY], including all classical random matrix ensembles, will be based on it.
0.3cm Once the eigenvalue distribution $d\nu$ is derived by Formula , under a covering condition, we can get the associated integration formula. The Weyl integration formula for compact Lie groups, the Harish-Chandra’s integration formula for complex semisimple Lie groups and real reductive groups, the integration formulae on Riemannian symmetric spaces of noncompact and compact types which were appeared in [@He], as well as their Lie algebra versions are all of particular cases of it (see [@AWY]).
0.3cm Now let us give a sketch of each section of this paper. In §2 we will develop some geometrical preliminaries on the geometry of $G$-space which will be required to establish the generalized ensemble. After presenting four conditions, that is, the invariance condition, the transversality condition, the dimension condition, and the orthogonality condition, on which the definition of generalized ensemble will be based, we will prove in Theorem \[T:det\] a primary form of Formula .
0.3cm §3 will be devoted to the integration over $G$-spaces, which will be needed when we derive the integration formula associated with the generalized random matrix ensemble. Based on the four conditions presented in §2 and a covering condition, we will prove an integration formula in Theorem \[T:Weylbianhuan\], which converts the integration over a $G$-space to the integration by first integrating over each $G$-orbit, and then integrating over the orbits space. Two criterions on when the covering condition holds will also be given.
0.3cm Prepared by the preliminaries of §2 and §3, In §4 we will give the precise definition of the generalized random matrix ensemble, as well as the associated generalized eigenvalue distribution and generalized joint density function. In Theorem \[T:distribution\] the Formula will be presented, from which the associated eigenvalue distribution measure and density function will be derived for various concrete examples of the generalized ensemble in an unified way in [@AWY].
0.3cm In §5 we will give a classification scheme of generalized ensembles, that is, the linear ensemble, the nonlinear noncompact ensemble, the compact ensemble, the group and algebra ensembles, as well as the pseudo-group and pseudo-algebra ensembles. According to this classification scheme, Gaussian ensemble and chiral ensemble are included in linear ensemble, new transfer matrix ensembles is included in nonlinear noncompact ensemble, circular and Jacobi ensembles are included in compact ensemble. Some new ensembles which were not considered before will also be included.
0.5cm
Geometry of $G$-spaces
======================
0.5cm In this section we develop some geometrical preliminaries which will be needed to establish our theory of the generalized random matrix ensembles.
0.3cm First we make some preparation about measures on manifolds. Let $M$ be an $n$-dimensional smooth manifold. A measure $dx$ on $M$ is called *smooth* (or *quasi-smooth*) if on any local coordinate chart $(U;
x_1,\cdots,x_n)$ of $M$, $dx$ has the form $dx=f(x)dx_1\cdots
dx_n$, where $f$ is a smooth function on $U$ and $f>0$ (or $f\geq
0$), $dx_1\cdots dx_n$ is the Lebesgue measure on $\mathbb{R}^n$. Note that the smooth measures on $M$ are unique up to multiplying a positive smooth function on $M$, so the concept of set of measure zero makes sense, which is independent of the choice of smooth measure.
0.3cm Let $M, N$ be two $n$-dimensional smooth manifolds, and let $\varphi:M\rightarrow N$ be a smooth map. If $dy$ is a smooth (or quasi-smooth) measure on $N$ which can be expressed locally as $dy=f(y)dy_1\cdots dy_n$, we can define the *pull bake* $\varphi^*(dy)$ of $dy$ locally as $$\label{E:pullback}
\varphi^* (dy)=f(\varphi(x))\left|\det\left(\frac{\partial
y}{\partial x}\right)\right|dx_1\cdots dx_n.$$ It is easily to check that the definition is compatible when we choose different coordinate charts, and $\varphi^*(dy)$ is a quasi-smooth measure on $M$. We can not expect $\varphi^*(dy)$ is smooth in general, even if $dy$ is smooth, since $\varphi$ may have critical points. But if $\varphi$ is a local diffeomorphism and $dy$ is smooth, then $\varphi^*(dy)$ is smooth.
0.3cm If $M, N$ are Riemannian manifolds and $dx, dy$ are the associated Riemannian measures, then we can express the pull back measure $\varphi^*(dy)$ globally. To do this, first we need some comments on the “determinant" of a linear map between two different inner product vector spaces of the same dimension. Suppose $V,W$ are two $n$-dimensional vector spaces with inner products. For $n$ vectors $v_1,\cdots,v_n\in V$, let $a_{ij}=\langle v_i,v_j\rangle$ for $1\leq i,j\leq n$, and define $Vol(v_1,\cdots,v_n)=\sqrt{\det(a_{ij})}$. Note that if $v_1,\cdots,v_n$ is an orthogonal basis, then $Vol(v_1,\cdots,v_n)=|v_1|\cdots|v_n|$. For vectors in $W$ we define the same things. Suppose $A:V\rightarrow W$ is a linear map, define $$\label{E:det}
|\det A|=\frac{Vol(Av_1,\cdots,Av_n)}{Vol(v_1,\cdots,v_n)},$$ where $v_1,\cdots,v_n$ is a basis of $V$. It is easily to check that the definition is independent of the choice of the basis $v_1,\cdots,v_n$. In the special case that $v_1,\cdots,v_n$ is an orthogonal basis of $V$ and $Av_1,\cdots,Av_n$ are mutually orthogonal, then $$\label{E:det2}
|\det A|=\frac{|Av_1|\cdots|Av_n|}{|v_1|\cdots|v_n|}.$$ Note that we can only expect the norm of the determinant $|\det A|$ is well defined, since the sign “$\pm$" depends on the choice of orientations of $V$ and $W$.
\[P:pullback\] Suppose $M, N$ are two $n$-dimensional Riemannian manifolds with the associated Riemannian measures $dx, dy$, repsectively. If $\varphi:M\rightarrow N$ is a smooth map, then $$\label{E:Riemmannpullback}
\varphi^* (dy)=|\det(d\varphi)_x|dx.$$
Suppose that in local coordinate charts the Riemannian metrics on $M$ and $N$ are $ds^2=\Sigma_{ij}g_{ij}(x)dx_idx_j$ and $d\widetilde{s}^2=\Sigma_{ij}\widetilde{g}_{ij}(y)dy_idy_j$, respectively, where $g_{ij}(x)=\langle\frac{\partial}{\partial
x_i},\frac{\partial}{\partial x_j}\rangle$ and $\widetilde{g}_{ij}(y)=\langle\frac{\partial}{\partial
y_i},\frac{\partial}{\partial y_j}\rangle$. Then by definition, the Riemannian measures $dx,dy$ are $dx=\sqrt{\det\big(g_{ij}(x)\big)}dx_1\cdots dx_n$ and $dy=\sqrt{\det\big(\widetilde{g}_{ij}(y)\big)}dy_1\cdots dy_n$. We have $$\begin{aligned}
|\det(d\varphi)_x|^2
=&\frac{Vol\left((d\varphi)_x(\frac{\partial}{\partial
x_1}),\cdots,(d\varphi)_x(\frac{\partial}{\partial
x_n})\right)^2}{Vol\left(\frac{\partial}{\partial
x_1},\cdots,\frac{\partial}{\partial x_n}\right)^2}\\
=&\frac{\det\left(\left\langle\sum_k\frac{\partial y_k}{\partial
x_i}(\frac{\partial}{\partial
y_k})_{\varphi(x)},\sum_l\frac{\partial y_l}{\partial
x_j}(\frac{\partial}{\partial
y_l})_{\varphi(x)}\right\rangle\right)}{\det\left(\left\langle\frac{\partial}{\partial
x_i},\frac{\partial}{\partial x_j}\right\rangle\right)}\\
=&\frac{\det\left(\sum_{kl}\frac{\partial y_k}{\partial
x_i}\frac{\partial y_l}{\partial
x_j}\widetilde{g}_{kl}(\varphi(x))\right)}{\det\big(g_{ij}(x)\big)}\\
=&\frac{\det\left(\left(\frac{\partial y_k}{\partial x_i}\right)^t
\Big(\widetilde{g}_{kl}(\varphi(x))\Big)\left(\frac{\partial
y_l}{\partial
x_j}\right)\right)}{\det\big(g_{ij}(x)\big)}\\
=&\frac{\left(\det\left(\frac{\partial y}{\partial
x}\right)\right)^2\det\big(\widetilde{g}_{ij}(\varphi(x))\big)}{\det\big(g_{ij}(x)\big)}.\end{aligned}$$ Hence $$\begin{aligned}
\varphi^*(dy)=&\sqrt{\det\big(\widetilde{g}_{ij}(\varphi(x))\big)}\left|\det\left(\frac{\partial
y}{\partial x}\right)\right|dx_1\cdots dx_n\\
=&|\det(d\varphi)_x|\sqrt{\det\big(g_{ij}(x)\big)}dx_1\cdots dx_n\\
=&|\det(d\varphi)_x|dx.\end{aligned}$$
Now we come to the main geometric problems which will be concerned in the following sections. Let $G$ be a Lie group, which acts on an $n$-dimensional smooth manifold $X$. The action is denoted by $\sigma: G\times X\rightarrow X$, and we denote $\sigma_g(x)=\sigma(g,x)$. Our first goal is, roughly speaking, to choose a representation point in each $G$-orbit $O_x = \{
\sigma_g(x): g\in G \}$, and the representation points should depend smoothly on the orbits. But in general this aim can only be achieved partially. So suppose we have a closed submanifold $Y$ of $X$, which consists of the representation points of the orbits in one’s mind, such that $Y$ intersects “almost all" orbits transversally. More precisely, we suppose there are closed zero measure subsets $X_{\mathrm{z}}\subset X$, $Y_{\mathrm{z}}\subset Y$. Let $X' = X
\setminus X_{\mathrm{z}}$, $Y' = Y \setminus Y_{\mathrm{z}}$, and suppose that
0.3cm (*invariance condition*) $X' = {\displaystyle}\bigcup_{y\in Y'} O_y$.
0.3cm (*transversality condition*) $T_y X=T_y O_y\oplus T_y Y$, $\forall y\in
Y'$.
0.3cm [It]{} is clear that (a) implies $Y'=Y\cap
X'$, and then $Y_{\mathrm{z}}=Y\cap X_{\mathrm{z}}$. Notice that $X'$ and $Y'$ are open and dense submanifolds of $X$ and $Y$, respectively. So $\forall y\in Y'$, $T_y X' = T_y X$, $T_y Y' = T_y Y$.
0.3cm Let $K=\{g\in G:\sigma_g(y)=y, \forall y\in Y\}$, then $K$ is a closed subgroup of $G$. For $g\in G$, we denote $[g]=gK$ in $G/K$. The map $\sigma: G\times X\rightarrow X$ reduces to a map $\varphi: G/K\times Y\rightarrow X$ by $\varphi([g],y)=\sigma_g(y)$, and then induces a map $G/K\times
Y'\rightarrow X'$ by restriction, which we also denote by $\varphi$. By the assumption above, $\varphi:G/K\times
Y'\rightarrow X'$ is surjective. For $x\in X$, Let $G_x=\{g\in
G:\sigma_g(x)=x\}$ be the isotropic subgroup associated with $x$. Then $K\subset G_y, \forall y\in Y$. Let $dx, dy$ be smooth measures on $X$ and $Y$, respectively. We suppose $dx$ is $G$-invariant. In the following we suppose that
0.3cm (*dimension condition*) $\mathrm{dim}G_y=\mathrm{dim}K, \quad \forall
y\in Y'$.
0.3cm [This]{} means that $\forall y\in Y'$, $G_y$ and $K$ have the same Lie algebras, and the only difference between $G_y$ and $K$ is that $G_y$ may have more components than $K$. Then for some $y\in Y'$, we have $$\begin{aligned}
\mathrm{dim}X
=&\mathrm{dim}T_y X\\
=&\mathrm{dim}T_y Y+\mathrm{dim}T_y
O_y\\
=&\mathrm{dim}Y+\mathrm{dim}G-\mathrm{dim}G_y\\
=&\mathrm{dim}Y+ \mathrm{dim}G-\mathrm{dim}K .\end{aligned}$$ So $\varphi$ is a map between manifolds of the same dimension, and the pull back $\varphi^*(dx)$ of $dx$ makes sense. Suppose also that there is a $G$-invariant smooth measure $d\mu$ on $G/K$, then the product measure $d\mu dy$ on $G/K\times Y$ is smooth, so $$\label{F:J(g,y)}
\varphi^*(dx)=J([g],y)d\mu dy$$ for some $J\in C^\infty(G/K\times Y)$ with $J\geq 0$.
The $G$-invariant smooth measure $d\mu$ on $G/K$ exists if and only if $\Delta_G|_K=\Delta_K$, where $\Delta_G$ and $\Delta_K$ are the modular functions on $G$ and $K$, respectively, see, e.g., Knapp [@Kn], Section 8.3. For concrete examples in the following sections, this condition always hold.
\[P:J\] The smooth function $J\in C^\infty(G/K\times Y)$ is independent of the first variable $[g]\in G/K$. So we can rewrite formula as $$\label{F:J}
\varphi^*(dx)=J(y)d\mu dy$$ where $J\in C^\infty(Y)$ with $J\geq 0$.
We denote the natural action of $h\in G$ on $G/K$ also by $l_h$, then one can easily verify that $\sigma_h\circ\varphi=\varphi\circ(l_h\times id)$. By the $G$-invariance of $dx$ and $d\mu$, we have $$\begin{aligned}
&J([g],y)d\mu dy\\
=&\varphi^*(dx)\\
=&\varphi^*\circ\sigma_h^*(dx)\\
=&(l_h\times id)^*\circ\varphi^*(dx)\\
=&(l_h\times id)^*(J([g],y)d\mu dy)\\
=&J(h[g],y)(l_h^*(d\mu)\times id^*(dy))\\
=&J([hg],y)d\mu dy.\end{aligned}$$ So $J([g],y)=J([hg],y)$ for all $g,h\in G$, which means $J$ is independent of the first variable.
\[C:J\] There exists a quasi-smooth measure $d\nu$ on $Y$ such that $$\label{F:dmudnu}
\varphi^*(dx)=d\mu d\nu.$$ The measure $d\nu$ is given by $$\label{F:dnu}
d\nu(y)=J(y)dy.$$
The factor $J(y)$ can also be given by more general smooth measures $u(x)dx$ and $v(y)dy$ on $X$ and $Y$. A direct calculation yields the following
\[P:bian\] Suppose conditions (a), (b), and (c) hold. If the measures vary by $dx'=u(x)dx$, $dy'=v(y)dy$, and $d\mu'=\lambda d\mu$, where $u, v$ are positive smooth functions on $X, Y$, respectively, $u$ is $G$-invariant, and $\lambda$ is a positive constant, then $J(y)$ varies by $$J'(y)=\frac{u(y)}{\lambda v(y)}J(y).$$
Now we suppose that there is a Riemannian structure on $X$ such that $dx$ and $dy$ are the induced Riemannian measures on $X$ and $Y$, respectively. We suppose the following orthogonality condition holds
0.3cm (*orthogonality condition*) $T_yY\perp T_yO_y$, $\forall y\in Y'$.
0.3cm [Then]{} we can compute the factor $J(y)$ in a simple way by the following theorem.
0.3cm Let ${\mathfrak{l}}$ be a linear subspace of the Lie algebra ${\mathfrak{g}}$ of $G$ such that ${\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{l}}$, where ${\mathfrak{k}}$ is the Lie algebra of $K$. Let $\pi:G\rightarrow G/K$ be the natural projection, then $(d\pi)_e|_{\mathfrak{l}}:{\mathfrak{l}}\rightarrow T_{[e]}(G/K)$ is an isomorphism. We endow a Riemannian structure on $G/K$ such that the associated Riemannian measure is $d\mu$, then it also induces an inner product on $T_{[e]}(G/K)$. For $y\in Y$, we define a linear map $\Psi_y:{\mathfrak{l}}\rightarrow T_yO_y$ by $$\label{E:Psi}
\Psi_y(\xi)=\frac{d}{dt}\Big |_{t=0}\sigma_{\exp t\xi}(y) , \quad
\forall \xi\in {\mathfrak{l}}.$$ If $y\in Y'$, then $\dim{\mathfrak{l}}=\dim T_y O_y$. We choose an inner product on ${\mathfrak{l}}$, and endow the inner product on $T_y O_y$ induced from the Riamannian structure on $X$. Then the “determinants" $|\det\Psi_y|$ and $|\det((d\pi)_e|_{\mathfrak{l}})|$ make sense.
\[T:det\] Under the above assumptions, we have $$\label{F:J=det}
J(y)=C|\det\Psi_y|,$$ for $y\in Y'$, where $C=|\det((d\pi)_e|_{\mathfrak{l}})|^{-1}$ is a constant.
By the transversality condition (b), the tangent map $$(d\varphi)_{([e],y)}:T_{([e],y)}(G/K\times Y)\rightarrow T_yX$$ of $\varphi$ at the point $([e], y)$ ($y\in Y'$) can be regarded as $$(d\varphi)_{([e],y)}:T_{[e]}(G/K) \oplus T_yY \rightarrow T_yO_y
\oplus T_yY .$$ Denote $\widetilde{\Psi}_y=(d\varphi)_{([e],y)}|_{T_{[e]}(G/K)}:T_{[e]}(G/K)\rightarrow
T_yO_y$, then it is obvious that $\Psi_y=\widetilde{\Psi}_y\circ(d\pi)_e|_{\mathfrak{l}}$, and one can easily show that in the matrix form, $$(d\varphi)_{([e],y)}=\left(\begin{array}{cc}\widetilde{\Psi}_y&0\\0&id
\end{array}\right).$$
0.3cm Since $d\mu$ is the associated Riemannian measure on $G/K$, the product measure $d\mu dy$ is the the associated Riemannian measure on the product Riemannian manifold $G/K\times
Y'$. By Proposition \[P:pullback\] and the orthogonality condition (d), $$\begin{aligned}
J(y)=&\left|\det(d\varphi)_{([e],y)}\right|\\
=&\left|\det\left(\begin{array}{cc}\widetilde{\Psi}_y&0\\0&id
\end{array}\right)\right|\\
=&|\det\widetilde{\Psi}_y|\\
=&|\det(\Psi_y\circ((d\pi)_e|_{\mathfrak{l}})^{-1})|\\
=&C|\det\Psi_y|,\end{aligned}$$ where $C=|\det((d\pi)_e|_{\mathfrak{l}})|^{-1}$. This proves the theorem.
Although formula only hold on $Y'$, since $Y'$ is dense in $Y$ and $J\in C^\infty(Y)$, we can get $J(y)$ for all $y\in Y$ by smooth continuation.
0.5cm
Integrations over $G$-spaces
============================
0.5cm Occasionally we will be interested in some kinds of integration formulae. In this section we give some preliminaries on integrations. The reader who has more interest in the eigenvalue distributions of the generalized random matrix ensembles may skip this section and go to §4 directly.
0.3cm The following proposition generalizes the change of variables formula for multiple integration.
\[P:jifenbianhuan\] Let $\varphi: M\rightarrow N$ be a smooth map between two $n$-dimensional smooth manifolds $M$ and $N$, $dy$ a smooth measure on $N$. If $\varphi$ is a local diffeomorphism and is a $d$-sheeted covering map, then for any $f\in C^\infty(N)$ with $f\geq0$ or with $f\in L^1(N,dy)$, we have $$\label{F:jifenbianhuan}
\int_N f(y) dy=\frac{1}{d}\int_M f(\varphi(x)) \varphi^*(dy).$$
It is a standard argument using partition of unity, the details is omitted here.
Formula seems like a formula which relates degree of a map and integration of volume forms on manifold. When $M, N$ are compact and oriented, then under the conditions of Proposition \[P:jifenbianhuan\], up to a “$\pm$" sign, formula says nothing but of this. But, in general, the integration of differential forms is not suitable for us. What we will need is a change of variables formula which should ignore the negative sign.
As in the previous section, Let $X$ be a $G$-space, where $X$ is an $n$-dimensional smooth manifold, $G$ is a Lie group. Then we have the reduced map $\varphi: G/K\times Y\rightarrow X$. Suppose $dx, dy$, and $d\mu$ are smooth measures on $X,Y$, and $G/K$, respectively, with $dx$ and $d\mu$ to be $G$-invariant. Our goal is to convert the integration over $X$ to the integration over $Y$. Suppose the conditions (a), (b), and (c) hold. We hope the map $\varphi: G/K\times Y'\rightarrow X'$ satisfies the conditions as given in Proposition \[P:jifenbianhuan\].
0.3cm
\[P:localdiffeo\] Suppose conditions (a), (b) and (c) hold. Then $\varphi: G/K\times
Y'\rightarrow X'$ is a local diffeomorphism.
0.3cm
0.2cm Let $e$ be the unit element in $G$. For $([e],y)\in G/K\times Y'$, $d\varphi_{([e],y)}(0,v)=v, \forall v\in T_y Y'$, so $T_y
Y'\subset Im(d\varphi_{([e],y)})$. Furthermore, $\varphi|_{G/K\times\{y\}}:G/K\times\{y\}\rightarrow O_y\cong
G/G_y$ is a local diffeomorphism, so $T_y O_y\subset
Im(d\varphi_{([e],y)})$. Thus $d\varphi_{([e],y)}$ is surjective. But $\mathrm{dim}(G/K\times Y')=\mathrm{dim}X'$, so $d\varphi_{([e],y)}$ is in fact an isomorphism. For general $([g],y)\in G/K\times Y'$, notice that $\varphi\circ
l_g=\sigma_g\circ\varphi$, where $l_g([h],y)=([gh],y)$, so $d\varphi_{([g],y)}\circ
(dl_g)_{([e],y)}=(d\sigma_g)_{([e],y)}\circ d\varphi_{([e],y)}$, and $d\varphi_{([e],y)}$ is isomorphic implies $d\varphi_{([g],y)}$ is isomorphic. Thus $\varphi$ is everywhere regular, and hence is a local diffeomorphism.
To make Proposition \[P:jifenbianhuan\] available, we endow the following covering condition.
0.3cm (*covering condition*) The map $\varphi: G/K\times Y'\rightarrow X'$ is a $d$-sheeted covering map, with $d<+\infty$.
0.3cm
\[T:Weylbianhuan\] Suppose conditions (a), (b), (c), and (e) hold. Then we have $$\label{F:Weylbianhuan}
\int_X f(x)
dx=\frac{1}{d}\int_Y\left(\int_{G/K}f(\sigma_g(y))d\mu([g])\right)
J(y)dy$$ for all $f\in C^\infty(X)$ with $f\geq0$ or with $f\in L^1(X,dx)$, where $J\in C^\infty(Y)$ is determined by Formula (\[F:J\]).
0.3cm
By Proposition \[P:localdiffeo\], $\varphi: G/K\times
Y'\rightarrow X'$ is a local diffeomorphism. By the covering condition (e), $\varphi$ is a $d$-sheeted covering map. So by Proposition \[P:jifenbianhuan\], for $f\in C^\infty(X)$ with $f\geq0$ or $f\in L^1(X,dx)$, we have $$\begin{aligned}
\int_X f(x)dx
=&\int_{X'} f(x)dx\\
=&\frac{1}{d}\int_{G/K\times Y'}f(\varphi([g],y))\varphi^*(dx)\\
=&\frac{1}{d}\int_{G/K\times Y'}f(\sigma_g(y))J(y)d\mu([g])dy\\
=&\frac{1}{d}\int_{Y'}\left(\int_{G/K}f(\sigma_g(y))d\mu([g])\right)J(y)dy\\
=&\frac{1}{d}\int_Y\left(\int_{G/K}f(\sigma_g(y))d\mu([g])\right)J(y)dy\end{aligned}$$
0.3cm
\[C:leihanshubianhuan\] Under the same conditions as in the above Theorem, if furthermore $f(\sigma_g(x))=f(x), \forall g\in G, x\in X$, then $$\label{F:leihanshubianhuan}
\int_X f(x) dx=\frac{\mu(G/K)}{d}\int_Y f(y)J(y)dy.$$
0.3cm To make the above conclusion more available, we give some criterions on when the map $\varphi: G/K\times Y'\rightarrow
X'$ is a covering map.
0.3cm
\[P:covering\] Let $M,N$ be smooth $n$-dimensional manifolds. Then an everywhere regular smooth map $\varphi : M\rightarrow N$ is a $d$-sheeted covering map if and only if for each $y\in N$, $\varphi^{-1}(y)$ has $d$ points.
0.3cm
0.2cm The “$\Rightarrow$" part is obvious. We prove the “$\Leftarrow$" part.
0.3cm For $y\in N$, let $\varphi^{-1}(y)=\{x_1,\cdots,x_d\}$. Since $\varphi$ is everywhere regular, there exists open neighborhood $U_i$ of $x_i$, $i=1,\cdots,d$, such that $U_i\cap U_j=\emptyset$ for $i\neq j$, and $\varphi_i=\varphi|_{U_i}:U_i\rightarrow\varphi(U_i)$ is a diffeomorphism. Let $V=\bigcap_{i=1}^d\varphi(U_i)$, and let $V_i=\varphi_i^{-1}(V)$, then $\varphi|_{V_i}$ is also a diffeomorphism onto $V$. We conclude that $\varphi^{-1}(V)=\bigcup_{i=1}^d V_i$. In fact, $\forall
z\in\varphi^{-1}(V)$, let $z_i=\varphi_i^{-1}(\varphi(z))$, then $z_i\in\varphi^{-1}(\varphi(z))$ and $z_i\neq z_j$ for $i\neq j$. But $z\in\varphi^{-1}(\varphi(z))$ and $|\varphi^{-1}(\varphi(z))|=d$, this force $z=z_{i_0}$ for some $i_0$. Hence $z\in\bigcup_{i=1}^d V_i$. Therefore $\varphi^{-1}(V)=\bigcup_{i=1}^d V_i$. The Lemma is proved.
0.3cm
\[C:covering\] Suppose conditions (a), (b) and (c) hold. If furthermore $\exists
d\in \mathbb{N}$ such that $\forall y\in Y'$,
0.3cm (1) the isotropic subgroup 0.1cm $G_y=K$, 0.6cm (2) $|O_y\cap Y'|=d$,
0.3cm then $\varphi: G/K\times Y'\rightarrow X'$ is a $d$-sheeted covering map.
0.3cm
By Proposition \[P:localdiffeo\], $\varphi$ is a local diffeomorphism. So by the above Proposition, we need only to show that for each $x\in X'$, $\varphi^{-1}(x)$ has $d$ points.
0.3cm For $x\in Y'$, suppose $O_x\cap
Y'=\{y_1,\cdots,y_d\}$. Then there exists $g_i\in G$ such that $\sigma_{g_i}(y_i)=x$ for each $i\in\{1,\cdots,d\}$. Then $([g_i],y_i)\in\varphi^{-1}(x)$. On the other hand, if $([g],y)\in\varphi^{-1}(x)$, then $y=y_{i_0}$ for some $i_0\in\{1,\cdots,d\}$. Now we have $\sigma_{gg_{i_0}^{-1}}(x)=\sigma_{g}(y_{i_0})=x$, that is $gg_{i_0}^{-1}\in G_x=K$, so $[g]=[g_{i_0}]$ and $([g],y)=([g_{i_0}],y_{i_0})$. Thus $\varphi^{-1}(x)=\{([g_1],y_1),\cdots,([g_d],y_d)\}$.
0.3cm In general for $x\in X'$, suppose $\sigma_h(x)\in Y'$ for some $h\in G$, then the relation $\varphi^{-1}(\sigma_{h}(x))=l_h(\varphi^{-1}(x))$ reduces the general case to the above one.
0.3cm Both Proposition \[P:covering\] and Corollary \[C:covering\] will be used in the following sections when we consider concrete examples.
0.3cm
The converse of Corollary \[C:covering\] is not true. That is, the isotropic subgroups $G_y$ associated with $y\in Y'$ may vary “suddenly", even if $Y'$ is connected. For example, The group $SO(n)$ acts on $\mathbb{R}P^n$ smoothly if we regard $\mathbb{R}P^n$ as the quotient space by gluing the opposite points on the boundary of the closed unit ball $B^n$. Let $X_{\mathrm{z}}$ be the image of $\{0\}$, $Y$ be the image of the segment $\{(x,0,\cdots,0):|x|\leq 1\}$, then the conditions (a), (b), (c), and (e) hold. The isotropic subgroup associated with the image of a point in $Y'$ which is an interior point of $B^n$ is ${\mathrm{diag}}(1,SO(n-1))$, but for the image of the point $(1,0,\cdots,0)$, its isotropic subgroup is ${\mathrm{diag}}(\pm1,O^\pm(n-1))$ (here $O^\pm(n-1)=\{g\in O(n-1):\det
g=\pm1\}$ ). Other examples with the similar phenomena will appeared in [@AWY] when we consider the group ensemble associated with complex semisimple Lie groups. When the phenomena of sudden variation of the isotropic subgroups happens, whether we can in general make them to be of the same by enlarging the set $X_{\mathrm{z}}$ is an open problem.
0.5cm
Generalized random matrix ensembles
===================================
0.5cm Now we are prepared to establish the generalized random matrix ensembles.
0.3cm Let $G$ be a Lie group which acts on an $n$-dimensional smooth manifold $X$ by $\sigma: G\times
X\rightarrow X$. For the convenience, we suppose $X$ is a Riemannian manifold. Suppose the induced Riemannian measure $dx$ is $G$-invariant (note that we do not require the Riemannian structure on $X$ to be $G$-invarinant). Let $Y$ be a closed submanifold of $X$ which is endowed the induced Riemannian measure $dy$, and let $K=\{g\in G:\sigma_g(y)=y, \forall y\in Y\}$. As in §2, we form the map $\varphi: G/K\times Y\rightarrow X$ by $\varphi([g],y)=\sigma_g(y)$. Let $X_{\mathrm{z}}\subset X$, $Y_{\mathrm{z}}\subset
Y$ be closed zero measure subsets of $X$ and $Y$, respectively. Denote $X' = X \setminus X_{\mathrm{z}}$, $Y' = Y \setminus Y_{\mathrm{z}}$. We suppose the conditions (a), (b), (c), and (d) of §2 hold. For the reader’s convenience, we list them below.
0.3cm (*invariance condition*) $X' = {\displaystyle}\bigcup_{y\in Y'} O_y$.
0.3cm (*transversality condition*) $T_y X=T_y O_y\oplus T_y Y$, $\forall y\in
Y'$.
0.3cm (*dimension condition*) $\mathrm{dim}G_y=\mathrm{dim}K, \quad \forall
y\in Y'$.
0.3cm (*orthogonality condition*) $T_yY\perp T_yO_y$, $\forall y\in Y'$.
0.3cm [Suppose]{} $d\mu$ is a $G$-invariant smooth measure on $G/K$, and suppose $p(x)$ is a $G$-invariant smooth function on $X$. Then by Corollary \[C:J\], there is a quasi-smooth measure $d\nu$ on $Y$ such that $$\label{E:distribution}
\varphi^*(p(x)dx)=d\mu d\nu.$$
\[D:generalized\] Let the conditions and notations be as above. Then the system $(G,\sigma,X,p(x)dx,Y,dy)$ is called a *generalized random matrix ensemble*. The manifolds $X$ and $Y$ are called the *integration manifold* and the *eigenvalue manifold*, respectively. The measure $d\nu$ on $Y$ determined by is called the *generalized eigenvalue distribution*.
Recall that in §2 we have defined the map $\Psi_y : {\mathfrak{l}}\rightarrow T_yO_y$ by $$\Psi_y (\xi ) = \frac{d}{dt}\Big |_{t=0}\hskip 0.1cm \sigma_{\exp
t\xi}(y) , \quad \forall \xi\in {\mathfrak{l}},$$ where ${\mathfrak{l}}$ is a linear subspace of ${\mathfrak{g}}$ such that ${\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{l}}$. Thanks to the preliminaries in §2, we can compute the generalized eigenvalue distribution directly according to the following theorem.
\[T:distribution\] Let $(G,\sigma,X,p(x)dx,Y,dy)$ be a generalized random matrix ensemble. Then the generalized eigenvalue distribution $d\nu$ is given by $$\label{E:distribution=}
d\nu(y)=\mathcal{P}(y)dy=p(y)J(y)dy,$$ where $$\label{E:factor=det}
J(y)= C |\det\Psi_y|,$$ here $C=|\det((d\pi)_e|_{\mathfrak{l}})|^{-1}$.
This follows directly from Proposition \[P:J\], Corollary \[C:J\], Proposition \[P:bian\] and Theorem \[T:det\].
The function $\mathcal{P}(y)=p(y)J(y)$ determined by formula is called the *generalized joint density function*.
0.3cm One of the most fundamental problems in the random matrix theory is to compute the eigenvalue distribution $d\nu$. In our generalized scheme, it is given by formulae and . Note that the power of is reflected by the fact that it provides a direct and unified method to compute the eigenvalue distributions of various kinds of random matrix ensembles. In the sequel paper [@AWY], we will see that all the classical ensembles are included in the generalized scheme, and the corresponding eigenvalue distributions can be derived from and . We will also present various kinds of generalized ensemble which were not considered before, and compute their eigenvalue distributions explicitly.
0.3cm Now we consider the integration formula associated with the generalized random matrix ensemble. As in §3, we assume the following covering condition holds.
0.3cm (*covering condition*) The map $\varphi: G/K\times Y'\rightarrow X'$ is a $d$-sheeted covering map, with $d<+\infty$.
\[T:random-integration\] Let $(G,\sigma,X,p(x)dx,Y,dy)$ be a generalized random matrix ensemble. Suppose the covering condition (e) holds. Then we have the following integration formula $$\label{E:integration}
\int_X f(x)p(x)
dx=\frac{1}{d}\int_Y\left(\int_{G/K}f(\sigma_g(y))d\mu([g])\right)
d\nu(y)$$ for all $f\in C^\infty(X)$ with $f\geq0$ or with $f\in
L^1(X,p(x)dx)$. If moreover $f(\sigma_g(x))=f(x), \forall g\in G,
x\in X$, then $$\label{E:classintegration}
\int_X f(x)p(x) dx=\frac{\mu(G/K)}{d}\int_Y f(y)d\nu(y).$$
It is obvious by Theorem \[T:Weylbianhuan\] and Corollary \[C:leihanshubianhuan\].
In formula , if the measure $p(x)dx$ is a probability measure, and we let $f=1$, we get $\frac{\mu(G/K)}{d}\int_Y d\nu(y)=1$. So if $G/K$ is compact, we can normalized the measure $d\mu$ such that $\mu(G/K)=d$, then the generalized eigenvalue distribution $d\nu$ is a probability measure.
The condition $f\in C^\infty(X)$ in Theorem \[T:random-integration\] is superfluous. In fact, it is sufficient to assume $f$ is measurable. The same is true for Proposition \[P:jifenbianhuan\] and Theorem \[T:Weylbianhuan\].
0.5cm
A classification scheme of generalized ensembles
================================================
0.5cm In this section we give a classification scheme of the generalized random matrix ensembles, that is, 0.2cm [(1)]{} Linear ensemble,\
(2) Nonlinear noncompact ensemble,\
(3) Compact ensemble,\
(4) Group ensemble,\
(5) Algebra ensembles,\
(6) Pseudo-group ensemble,\
(7) Pseudo-algebra ensemble.
0.3cm First we define the linear ensemble and the nonlinear noncompact ensemble. Let $G$ be a real reductive Lie group with Lie algebra ${\mathfrak{g}}$ in the sense of Knapp [@Kn], Section 7.2. Then $G$ admits a global Cartan involution $\Theta$, which induces a Cartan involution $\theta$ of ${\mathfrak{g}}$. Let the corresponding Cartan decomposition of ${\mathfrak{g}}$ is ${\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{p}}$. Let $K=\{g\in
G:\Theta(g)=g\}$, $P=\exp({\mathfrak{p}})$, then $K$ is a maximal compact subgroup of $G$ with Lie algebra ${\mathfrak{k}}$, $P$ is a closed submanifold of $G$ satisfies $T_eP={\mathfrak{p}}$. The spaces ${\mathfrak{p}}$ and $P$ are invariant under the adjoint action ${\mathrm{A}}={\mathrm{Ad}}|_K$ and the conjugate action $\sigma$ of $K$, respectively. Let ${\mathfrak{a}}$ be a maximal abelian subspace of ${\mathfrak{p}}$, and let $A$ be the connected subgroup of $G$ with Lie algebra ${\mathfrak{a}}$. Let $M=\{k\in
K:{\mathrm{A}}_k(\eta)=\eta, \forall \eta\in{\mathfrak{a}}\}=\{k\in K:\sigma_k(a)=a,
\forall a\in A\}$. It can be shown that there are Riemannian structures on ${\mathfrak{p}}$ and $P$ inducing $K$-invariant Riemannian measures $dX$ on ${\mathfrak{p}}$ and $dx$ on $P$. They also induce Riemannian measures $dY$ on ${\mathfrak{a}}$ and $da$ on $A$. There is also a $K$-invariant smooth measure $d\mu$ on $K/M$. Let $p_1(\xi)$ and $p_2(x)$ be $K$-invariant positive smooth functions on ${\mathfrak{p}}$ and $P$, then it can be proved that the systems $(K,{\mathrm{A}},{\mathfrak{p}},p_1(\xi)dX(\xi),{\mathfrak{a}},dY)$ and $(K,\sigma,P,p_2(x)dx,A,da)$ are generalized random matrix ensembles, which we called *linear ensemble* and *nonlinear noncompact ensemble*, respectively. It can be shown that the Gaussian ensemble and the chiral ensemble are particular examples of linear ensemble, and the new transfer matrix ensembles are particular examples of nonlinear noncompact ensemble.
0.3cm Next we define the compact ensemble. Let $G$ be a connected compact Lie group $G$ with Lie algebra ${\mathfrak{g}}$. Suppose $\Theta$ is a global involutive of $G$ with the induced involution $\theta=d\Theta$ of ${\mathfrak{g}}$. Let $K=\{g\in G:\Theta(g)=g\}$, and let ${\mathfrak{p}}$ be the eigenspace of $\theta$ corresponding the eigenvalue $-1$. Let $P=\exp({\mathfrak{p}})$, then $P$ is invariant under the conjugate action $\sigma$ of $K$. It was proved in [@AW] that $P$ is a closed submanifold of $G$ satisfies $T_eP={\mathfrak{p}}$, which is just the identity component of the set $\{g\in G:\Theta(g)=g^{-1}\}$. Let ${\mathfrak{a}}$ be a maximal abelian subspace of ${\mathfrak{p}}$, and let $A$ be the torus with Lie algebra ${\mathfrak{a}}$. There is a Riemannian structure on $P$ induces a $K$-invariant Riemannian measure $dx$ on $P$ and a Riemannian measure $da$ on $A$. Let $M=\{k\in K:\sigma_k(a)=a,
\forall a\in A\}$, then there is a $K$-invariant smooth measure $d\mu$ on $K/M$. Let $p(x)$ be a $K$-invariant positive smooth function on $P$, then it can be proved that the system $(K,\sigma,P,p(x)dx,A,da)$ is a generalized random matrix ensemble, which we call it *compact ensemble*. It can be shown that the circular ensemble and the Jacobi ensembles are particular examples of compact ensemble.
0.3cm Let $G$ be an unimodular Lie group $G$ with Lie algebra ${\mathfrak{g}}$. Then there are Riemannian structures on $G$ and ${\mathfrak{g}}$ inducing a $\sigma$-invariant Riemannian measure $dg$ on $G$ and an ${\mathrm{Ad}}$-invariant Riemannian measure $dX$ on ${\mathfrak{g}}$, where $\sigma$ denotes the conjugate action of $G$ on itself. Let $p_1(g)$ and $p_2(\xi)$ be two function on $G$ and ${\mathfrak{g}}$, respectively, which are invariant under the corresponding actions of $G$. If there exists a closed submanifold $Y$ of $G$ such that $(G,\sigma,G,p(g)dg,Y,dy)$ is a generalized random matrix ensemble, where $dy$ is the induced Riemannian measure on $Y$, then we call it a *group ensemble*. And if there exists a closed submanifold ${\mathfrak{y}}$ of ${\mathfrak{g}}$ such that $(G,{\mathrm{Ad}},{\mathfrak{g}},p(\xi)dX(\xi),{\mathfrak{y}},dY)$ is a generalized random matrix ensemble, where $dY$ is the induced Riemannian measure on ${\mathfrak{y}}$, then we call it an *algebra ensemble*. Among all the unimodular Lie groups, the connected compact Lie group and the connected complex semisimple Lie group are of particular interest. For a connected compact Lie group $G$, we can let the submanifold $Y$ of $G$ be a maximal torus $T$ of $G$, and let the submanifold ${\mathfrak{y}}$ of ${\mathfrak{g}}$ be the Lie algebra of $T$. For a connected complex semisimple Lie group $G$, we can let the submanifold ${\mathfrak{y}}$ of ${\mathfrak{g}}$ be a Cartan subalgebra of ${\mathfrak{g}}$, and let the submanifold $Y$ of $G$ be the corresponding Cartan subgroup of $G$. For these cases, it can be proved that the systems $(G,\sigma,G,p(g)dg,Y,dy)$ and $(G,{\mathrm{Ad}},{\mathfrak{g}},p(\xi)dX(\xi),{\mathfrak{y}},dY)$ are generalized random matrix ensembles.
0.3cm Now we define the pseudo-group ensemble and the pseudo-algebra ensembles, which are related to real reductive groups. Let $G$ be a real reductive group with lie algebra ${\mathfrak{g}}$. Let $\theta$ be a Cartan involution of ${\mathfrak{g}}$, and let ${\mathfrak{h}}_1,\cdots,{\mathfrak{h}}_m$ be a maximal set of mutually nonconjugate $\theta$ stable Cartan subalgebras of ${\mathfrak{g}}$ with the corresponding Cartan subgroups $H_1,\cdots,H_m$ of $G$. Denote the sets of all regular elements in $G$ and ${\mathfrak{g}}$ by $G_r$ and ${\mathfrak{g}}_r$. Let $H'_j=H_j\cap G_r$, ${\mathfrak{h}}'_j={\mathfrak{h}}_j\cap{\mathfrak{g}}_r$. Then it is known that $G_r=\bigsqcup_{j=1}^m\bigcup_{g\in G}\sigma_g(H'_j)$ (see [@Kn], Theorem 7.108), ${\mathfrak{g}}_r=\bigsqcup_{j=1}^m\bigcup_{g\in
G}{\mathrm{Ad}}_g({\mathfrak{h}}'_j)$ (see [@Wa], Proposition 1.3.4.1), here the symbol “$\bigsqcup$" means disjoint union. Each $\bigcup_{g\in
G}\sigma_g(H'_j)$ is an open set in $G$, and each $\bigcup_{g\in
G}{\mathrm{Ad}}_g({\mathfrak{h}}'_j)$ is an open set in ${\mathfrak{g}}$. Let $G_j=\overline{\bigcup_{g\in G}\sigma_g(H'_j)}$, ${\mathfrak{g}}_j=\overline{\bigcup_{g\in G}{\mathrm{Ad}}_g({\mathfrak{h}}'_j)}$. It can be shown that some suitable Riemannian structures on $G$ and ${\mathfrak{g}}$ induce a $\sigma$-invariant measure $dg_j$ on $G_j$ and an ${\mathrm{Ad}}$-invariant measure $dX_j$ on ${\mathfrak{g}}_j$ for each $j$, and they also induce a Riemannian measure $dh_j$ on $H_j$ and a Riemannian measure $dY_j$ on ${\mathfrak{h}}_j$. It is known that $Z(H_j)=\{g\in
G:\sigma_g(h)=h,\forall h\in H_j\}$, $H_j=\{g\in
G:{\mathrm{Ad}}_g(\xi)=\xi,\forall \xi\in{\mathfrak{h}}_j\}$. Let $d\mu'_j, d\mu_j$ be $G$-invariant measures on $G/Z(H_j)$ and $G/H_j$, respectively. In general, the spaces $G_j$ and ${\mathfrak{g}}_j$ may have singularities. But this doesn’t matter, since they are closures of open submanifolds in $G$ and ${\mathfrak{g}}$, whose boundaries have measure zero. If we ignore this ambiguity, then it can be proved that $(G,\sigma,G_j,dg_j,H_j,dh_j)$ and $(G,{\mathrm{Ad}},{\mathfrak{g}}_j,dX_j,{\mathfrak{h}}_j,dY_j)$ are generalized random matrix ensembles, which we called *pseudo-group ensemble* and *pseudo-algebra ensemble*, respectively.
0.3cm Due to the generality of the definition, our classification could not exhaust all kinds of generalized ensemble. But it would include all kinds of classical random matrix ensembles and some new examples of generalized ensembles, which will be analyzed explicitly in the sequel paper [@AWY].
1.0cm
[99]{} 0.5cm
An, J., Wang, Z., *On the realization of Riemannian symmetric spaces in Lie groups*, preprint, math.GT/0502028.
An, J., Wang, Z., Yan, K., *A Generalization of random matrix ensemble II: concrete examples and integration formulae*, preprint.
Altland, A., Zirnbauer, M. R., *Novel symmetry classes in mesoscopic normalconducting-superconducting hybrid structures*, Phys. Rev. B, 55, 1142-1161, 1997.
Caselle, M., *On the distribution of transmission eigenvalues in disordered wires*, arXiv:cond-mat/9410097, 1994.
Caselle, M., *A new classification scheme for random matrix theories*, arXiv:cond-mat/9610017, 1996.
Caselle, M., Magnea, U., *Random matrix theory and symmetric spaces*, Phys. Rep. 394, 41-156, 2004.
Dueñez, E., *Random matrix ensembles associated to compact symmetric spaces*, Comm. Math. Phys. 244, 29–61, 2004.
Dyson, F. J., *Correlations between the eigenvalues of a random matrix*, Comm. Math. Phys. 19, 235-250, 1970.
Helgason, S., *Groups and geometric analysis: integral geometry, invariant differential operators, and spherical functions*, American Mathematical Society, 2000.
Ivanov, D. A., *Random-matrix ensembles in p-wave vortices*, arXiv:cond-mat/0103089, 2001.
Knapp, A. W., *Lie groups beyond an introduction*, 2nd edition, Birkhäuser, Boston, 2002.
Mehta, M. L., *Random matrices*, Academic Press, San Diego, 1991.
Titov, M., Brouwer, P. W., Furusaki, A., Mudry, C., *Fokker-Planck equations and density of states in disordered quantum wires*, Phys. Rev. B, 63, 235318, 2001
Warner, G., *Harmonic analysis on semi-simple Lie groups, I* , Springer-Verlag, New York-Heidelberg, 1972.
Zirnbauer, M. R., *Riemannian symmetric superspaces and their origin in random-matrix theory*, J. Math. Phys., 37(10), 4986–5018, 1996.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The median absolute deviation (MAD) is a robust measure of scale that is simple to implement and easy to interpret. Motivated by this, we introduce interval estimators of the MAD to make reliable inferences for dispersion for a single population and ratios and differences of MADs for comparing two populations. Our simulation results show that the coverage probabilities of the intervals are very close to the nominal coverage for a variety of distributions. We have used partial influence functions to investigate the robustness properties of the difference and ratios of independent MADs.'
author:
- |
Chandima N. P. G. Arachchige\
Department of Mathematics and Statistics, La Trobe University\
<[email protected]>\
\
Luke A. Prendergast\
Department of Mathematics and Statistics, La Trobe University\
<[email protected]>
bibliography:
- 'ref.bib'
title: Confidence intervals for median absolute deviations
---
[**Keywords:**]{} asymptotic variance, partial influence functions, robust
Introduction {#sec:Intro}
============
The median absolute deviation is a robust measure of dispersion [MAD, see e.g. @hamp-1974; @hampel1986robust]. Defined as the median of the absolute residuals from the median, the MAD is a suitable scale measure to accompany the median. [@hamp-1974] referred to the $\mad$ as the “median deviation" and it had first received attention even as early as [@gauss1816demonstratio], and later rediscovered by [@hampel1968contribution]. The MAD is the *most robust* estimator of scale as measured by robustness measures such as the break-down point and gross error sensitivity [@hamp-1974]. The breakdown point of an estimator is the proportion of contamination that the estimator can handle before providing unreliable results and for the MAD this is equal to 1/2 (the maximum). The $\mad$ estimator has what is known as a bounded influence function so that the amount of influence any observational type can exert on the estimator is limited. More will be said on the influence function later.
[@arachchige2019interval] showed that excellent coverages for interval estimators of ratios of interquantile ranges can be achieved. This makes these intervals more suitable than those for ratio of variances when normality cannot be assumed. Then, [@arachchige2019CV] considered interval estimators for robust versions of the coefficient of variation, one of which uses the MAD in place of the standard deviation (and the median to replace the mean). Motivated by these good coverage properties, we consider interval estimators for the MAD and for ratios and differences of independent MADs as robust alternatives to intervals based on sample variances. To the best of our knowledge, and not to confuse the MAD with the *mean* absolute deviation for which interval estimators with good coverage have been introduced by [@bonett2003confidence], no one has introduced these interval estimators for the MAD. The very good coverage properties, that we will highlight later, ensure inferences about dispersion based on the MAD are possible.
In Section 2 we provide some necessary notations before considering influence functions for ratios of MADs. In Section 3 we consider confidence intervals for MADs, differences of MADs and ratios of MADs with coverage properties explored via simulations in Section 4. Examples are also considered in Section 4 and we conclude in Section 5.
Notations and influence functions {#sec:MAD}
=================================
Let $X$ denote a random variable and $F$ its distribution function. Then [@hamp-1974] defined the median absolute deviation (MAD) as $$\mad(X) =\text{med}\mid{X-M}\mid~,
\label{eq:MAD}$$ where ‘med’ denotes the median and $M=\text{med}(X)=F^{-1}(0.5)$ is the population median. Let $X_1,\ldots,X_n$ denote a random sample of $n$ observations. Then the MAD estimate is simply the median of the absolute residuals from the sample median. That is, for $m$ denoting the sample median, $\widehat{\text{MAD}}$ is the sample median of the $|X_1-m|,\ldots,|X_n-m|$. While inference, for a single MAD may be of interest, it is often that case that comparison of dispersion measures, such as the MAD, is needed to compare two populations.
Consider two independent random variables $X\sim F_1$ and $Y\sim F_2$ and let us consider $\mad(X)$ and $\mad(Y)$. Then, the population squared ratio of MADs, which we denote as $R_M$, and associated estimator can be define as $$R_M = \left[\frac{\mad(X)}{\mad(Y)}\right]^2\;\;\text{and}\;\; \widehat{R}_M = \left[\frac{\widehat{\mad(X)}}{\widehat{\mad(Y)}}\right]^2.
\label{eq:R_m}$$
Here we have suggested the squared ratio of MADs since it is the analogue to the ratio of variances and, in fact, equal to ratio of variances for some distributions (e.g. normal). However, the ratio of MADs may also be used. Another possibility is the difference of MADs, $D_M$, where $$D_M = \mad(X) - \mad(Y) \;\;\text{and}\;\; \widehat{D}_M = \widehat{\mad(X)} - \widehat{\mad(Y)}
\label{eq:D_m}~.$$
Influence function and partial influence functions {#sec:IFandPIF}
--------------------------------------------------
Define the contamination distribution to be $F_{\epsilon}=(1-\epsilon)F+\epsilon \Delta_{x}$, where $\epsilon \in [0,1]$ is the proportion of contamination and $\Delta_{x}$ has all of its mass at the contaminant $x$. Consider an estimator functional $\Tf$ such that $\Tf(F)=\theta$ and $\Tf(F_n)=\widehat{\theta}$ where $F_n$ denotes the empirical distribution function for sample of $n$ observations. The relative influence on $\Tf(F)$ of $\epsilon$ proportion of contaminated observations at $x$ is given by, $[\Tf(F_\epsilon)-\Tf(F)]/\epsilon$, where $\Tf(F_{\epsilon})=(1-\epsilon)\Tf(F)+\epsilon \Delta_{x}$. Then, the influence function [IF @hamp-1974] is defined as, $$\IF (x;\Tf,F)=\lim_{\epsilon\downarrow 0} \frac{\Tf(F_\epsilon) - \Tf(F)}{\epsilon}\equiv \frac {\partial }{\partial \epsilon }\Tf(F_\epsilon)\Big|_{\epsilon = 0}.$$
When more than one population exists, the IF is determined by contaminating one population while the other population remains uncontaminated. [@pires2002partial] defines this notion as “partial IFs" (PIFs) and in our context with two populations we have two PIFs. The first PIF of the estimator with functional $\Tf$ at ($F_1,\ F_2$) is $$\begin{aligned}
{\text{PIF}}_1 (x;\Tf,F_1,F_2) = \lim_{\epsilon \to 0}\left[\frac{\Tf[(1-\epsilon)F_1+\epsilon\Delta_{x_0},F_2]-\Tf(F_1,F_2)}{\epsilon}\right]
\label{eq:PIF}\end{aligned}$$ and with ${\text{PIF}}_2 (x;\Tf,F_1,F_2)$ defined similarly.
Now, consider the functional for the standardized MAD denoted by $\MADf$ so that $\MADf(F)=\mad_X$. [@hamp-1974] gives the influence function for the $\mad$ when $F$ is the normal distribution and further details can be found on page 107 of [@hampel1986robust]. Let $f=F'$ denote the density function then, assuming $f(M)$ and $2[f(M+\text{MAD}_X)+f(M-\text{MAD}_X)]$ are nonzero, a general form of the IF for the MAD exists; e.g. see page 137 of [@huber1981robust] or page 16 of [@andersen2008modern]. This is given as $$\IF (x;\,\MADf ,F) = \displaystyle\frac{\left[\text{sign}(x-M)-\text{MAD}_X\right]-\displaystyle\frac{f(M+\text{MAD}_X)-f(M-\text{MAD}_X)}{f(M)}\text{sign}(x-M)}{2[f(M+\text{MAD}_X)+f(M-\text{MAD}_X)]}. \label{eq:IFmad3}$$
### Partial influence functions of the difference and squared ratio of MADs {#sec:PIF_Rm}
Let $\D_M$ be the functional for the difference of $\mad$s so that, $$\D_M(F_1,F_2)= \MADf(F_1)-\MADf(F_2) ~$$ then the PIFs are ${\text{PIF}}_1 (x;\D_M,F_1,F_2) = \IF (x;\MADf,F_1) $ and ${\text{PIF}}_2 (x;\D_M,F_1,F_2) = -\IF (x;\MADf,F_2)$. These are trivial and previous studies on robustness of the MAD may be considered for this context. We therefore do not explore the difference PIFs further.
Let $\R_M$ be the functional for the squared ratio of $\mad$s so that, $$\R_M(F_1,F_2)=\left[\frac{\MADf(F_1)}{\MADf(F_2)}\right]^2 ~.$$ Then the PIFs for the squared ratio of MADs are given below.
For ${\text{PIF}}(x;T,F_1,F_2)$ as defined in , the PIFs of $\R_M$ are $$\begin{aligned}
{\text{PIF}}_1 (x;\R_M,F_1,F_2) &= \frac{2\,\R_M(F_1,F_2)}{\MADf(F_1)}\IF (x;\MADf,F_1) ,\nonumber\\
{\text{PIF}}_2 (x;\R_M,F_1,F_2) &= -\frac{2\,\R_M(F_1,F_2)}{\MADf(F_2)}\IF (x;\MADf,F_2).
\end{aligned}$$ \[th:PIF\_Rm\]
The proof of Theorem \[th:PIF\_Rm\] is in Appendix \[app:proof\_of\_PIF\_Rm\] and we consider some examples of the first PIF next.
### Partial influence functions comparison {#sec:PIF_Comp}
![PIF$_1$ comparisons for (A) two exponential populations both with rates 0.5, 1 and 1.5 and (B) two log-normal populations both with $\mu$=0 and $\sigma$=0.5,1,1.5.[]{data-label="fig:PIF_Com"}](PIF2.pdf){width="\linewidth"}
Figure \[fig:PIF\_Com\] depicts the ${\text{PIF}}$s of the first population for the squared ratio of $\mad$s and the ratio of variances [see @arachchige2019interval for these]. In Plot A we consider the ratio of variances and squared ratio of MADs for two exponential distributions, both with rates equal to 0.5, or 1 or 1.5. Similarly, in Plot B we do this for two log normal distributions both with $\mu$=0 and $\sigma=0.5$ or 1 or 1.5. Since the numerator and denominator distributions are the same, both are estimators of one and therefore the PIFs are comparable. As expected, the ${\text{PIF}}$s of the ratio of variances is unbounded indicating that outliers can exert large influence on the estimator. The ${\text{PIF}}$s of the squared ratio of MAD is bounded and the influence of any large outliers is limited, and far less than for the ratio of variances. For the exponential distribution, the ${\text{PIF}}$s of ratio of variances do not depend on the rate parameter. However, for the log-normal distribution the PIF for the ratio of variances increases quickly with increasing $\sigma$.
Asymptotic confidence intervals {#sec:CIs}
===============================
In their discussion of intervals for the mean absolute deviations, [@bonett2003confidence] provide suggestions for median absolution deviations from a fixed point, $h$. They suggest using intervals for the median and where the data used is the transformed $|X_i-h|$s. When $h$ is the population median, i.e. $h=M$, and this median is known, simulations (not shown) result in good coverage that is close to nominal. However, when $M$ is not known and needs to be estimated, this approach typically results in coverage that is too low (e.g. less than 0.8 for a nominal 0.95). In this section we therefore provide confidence intervals that have good coverage properties, as shown by our simulations that follow.
Asymptotic normality and associated variance of the $\mad$ can be found in [@falk1997asymptotic] who provide the asymptotic joint normality between the median and $\mad$ estimators. We again let $\text{MAD}_X=\MADf(F)$ and also let $\MADf(F_{n}) = \widehat{\text{MAD}}_X$. Then, if $F$ is continuous near, and differentiable at, the median $M$, $M-\mad_X$ and $M + \mad_X$ with $f(M)>0$ and $ B_1 =f(M - \mad_X) + f(M + \mad_X) > 0,$ we have $$\sqrt{n}\left(\widehat{\text{MAD}}_X-\mad_X\right) {\stackrel{\text{\tiny approx.}}{\sim}}N(0, \asv)~,
\label{eq:ASyNormal_mMAD}$$ where ‘${\stackrel{\text{\tiny approx.}}{\sim}}$’ denotes ‘approximately distributed’. The asymptotic variance of the MAD estimator is $$\begin{aligned}
\asv = \asv(\MADf; F)=\frac{1}{4B_1^2}\left[1+\frac{B_2}{\left[f(M)\right]^2}\right] ~,
\label{eq:ASV_MAD}\end{aligned}$$ where $B_1$ is given above and $B_2=B_3^2 + 4B_3f(M) \left[1-F(M+\mad_X)-F(M-\mad_X)\right]$ with $B_3 =f(M-\mad_X)-f(M+\mad_X)$.
We used the $\asv$ in and the Delta method [see e.g., chapter 3 of @Das-2008] to derive the asymptotic variance of the ratios of $\mad$s. The asymptotic variance of $\sqrt{n_1 + n_2}\R_M(F_{n_1}, F_{n_2})$ is $$\asv (\R_M;n_1,n_2)= 4\R^2_M(F_1,F_2)\Bigg[\frac{\asv(\MADf,F_1)}{w_1\,\MADf^2 (F_1)}+\frac{\asv(\MADf,F_2)}{w_2\,\MADf^2 (F_2)} \Bigg]~
\label{eq:ASV_Rm}$$ where $w_i=n_i/(n_1+n_2)$ for $i=1,2$.
Since the two populations are independent, deriving the asymptotic variance of the difference of MAD is straightforward. $$\asv (\D_M;n_1, n_2)= \asv(\MADf,F_1) + \asv(\MADf,F_2).
\label{eq:ASV_Dm}$$
Throughout, let $\widehat{\text{ASD}}(\cdot)=\sqrt{\widehat{\text{ASV}}(\cdot)}$ denote the estimated asymptotic standard deviation estimate. Note that the ASV depends on both $f$ and $F$, the density and distribution functions. There are several options to estimate these, but we choose to use the very flexible Generalized Lambda Distribution (GLD) which, for the FKML parameterization [@freimer1988study], is defined in terms of its quantile function, $Q(p)$, $$Q(p)=\lambda_1+\lambda_2^{-1}\left\{\lambda_3^{-1}(p^{\lambda_3} - 1)-\lambda_4^{-1}[(1-p)^{\lambda_4} - 1]\right\} ~,$$ where $\lambda_1$, $\lambda_2$, $\lambda_3$ and $\lambda_4$ are the location, inverse scale and two shape parameters respectively. To estimate the GLD parameters we use a recent approach introduced by [@dedduwakumara2019efficient] which is computationally efficient making it useful for our simulations that follow. However, other estimators can also be used. We then use these parameter estimates with the density and distribution functions for the GLD in R `gld` package [@gld-king].
Based on asymptotic normality of the MAD [e.g. @falk1997asymptotic], an asymptotic $(1-\alpha)\%$ confidence interval for $\mad$ is given as $$[L, U]_{\mad} = \Bigg[\widehat{\mad}_X\pm z_{1-\alpha/2}\;\frac{\wasd{(\MADf,F_{n})}}{\sqrt{n}\,}\Bigg]~,
\label{eq:CI_MAD}$$ where the $z_{1-\alpha/2}$ is the $(1-\alpha/2)\times $100 percentile of the standard normal distribution.
When constructing the interval estimator for the squared ratio of $\mad$s, we first derive the confidence interval for the log transformed ratio and then exponentiate to return to the ratio scale. Let $\mathcal{W}(F_1,F_2)=\ln[\R_M(F_1,F_2)]$ then, using the Delta method, it is straightforward to show that $\asv (\mathcal{W},F_1,F_2) \doteq \asv(\R_M,F_1,F_2)/[\R_M(F_1,F_2)]^2$ . Then a $(1-\alpha)\%$ confidence interval estimator for $R_M$ is given as $$[L, U]_{R_M} = \exp\Bigg[\ln(\widehat{R}_M)\pm z_{1-\alpha/2}\;\frac{\wasd{(\R_M,F_{n_1},F_{n_2})}}{\widehat{R}_M\sqrt{n_1+n_2}\,}\Bigg]~,
\label{eq:CI_Rm}$$ where $\widehat{R}_M$ is the squared ratio of MADs estimator and the $\asv$ is in .
Finally, a $(1-\alpha)\%$ confidence interval for the difference in MADs is simply $$[L, U]_{D_M} = \widehat{D}_M\pm z_{1-\alpha/2}\;\frac{\wasd{(\D_M,F_{n_1},F_{n_2})}}{\sqrt{n_1+n_2}\,}~,
\label{eq:CI_Dm}$$ where $\widehat{D}_M$ is the difference of MADs estimator and the $\asv$ can be found in .
Simulations and Examples {#sec:SimandEx}
========================
We begin by conducting simulations to assess the coverage properties of the interval estimations for data generated from several distributions. As pointed out earlier, we have used a new estimator of the GLD parameters provided by [@dedduwakumara2019simple] since it exhibits very good performance and is very efficient making it useful for our simulations. In Appendix A.2, we provide R code for the interval estimators using readily available estimators for the GLD from the `gld` package [@gld-king]. In that code we have opted for Titterington’s method [@titterington1985comment] since it to has good performance, albeit is more time consuming.
Simulations {#sec:Sim}
-----------
To investigate the performance of the MAD, squared ratio of $\mad$s and difference of $\mad$s intervals we consider simulated coverage probability and the average confidence interval width as performance measures. We have selected the log normal (LN), exponential (EXP), chi-square ($\chi^2_5$) and Pareto (PAR) distributions with different sample sizes of $n=50, 100, 200, 500, 1000$. Each simulation consists of 10,000 trials.
Sample size $X\sim $ EXP(1) $X\sim \chi^2_5$ $X\sim$ PAR(1,7)
-------------- -------------- ----------------- ------------------ ------------------
True $\mad=$ 0.599 0.481 1.895 0.075
50 0.938 (1.43) 0.936 (1.93) 0.927 (1.25\*) 0.939 (0.34)
100 0.940 (0.37) 0.939 (0.29) 0.938 (0.91) 0.939 (0.05)
200 0.938 (0.26) 0.947 (0.20) 0.942 (0.65) 0.944 (0.03)
500 0.945 (0.16) 0.948 (0.12) 0.947 (0.41) 0.949 (0.02)
1000 0.946 (0.12) 0.951 (0.09) 0.944 (0.29) 0.947 (0.01)
: Simulated coverage probabilities (and widths in parentheses) for the 95% confidence interval for the $\mad$ (\* denotes median width reported due to excessively large widths for a small number of intervals that skew the mean).
\[tab:CP\_MAD\]
Simulated coverages and widths for the interval estimator of $\mad$s, from , are provided in Table \[tab:CP\_MAD\] for several distributions. The coverage probabilities are all close to the nominal level of 0.95, even for $n=50$ where coverages were approximately in the vicinity of 0.93-0.94. Coverages become closer to the nominal level as the sample size increases and, as expected the interval widths decrease with increasing sample size.
----------- ------------- ------------------ ----------------- ------------------ ------------------
$X\sim $ EXP(1) $X\sim \chi^2_5$ $X\sim$ PAR(1,7)
Measure $Y\sim $ LN(0,1) $Y\sim$ EXP(1) $Y\sim \chi^2_2$ $Y\sim$ PAR(1,3)
True $R_M=$ 1 1 3.876 0.148
True $D_M=$ 0 0 0.932 -0.119
50,50 $R_M$ 0.958 (3.71\*) 0.971 (4.03\*) 0.955 (12.14\*) 0.978 (0.91\*)
$D_M$ 0.967 (2.55) 0.972 (3.49) 0.956 (1.54\*) 0.967 (1.17)
100,100 $R_M$ 0.949 (2.23) 0.958 (1.87\*) 0.954 (6.48\*) 0.960 (0.33\*)
$D_M$ 0.954 (0.52) 0.958 (0.42) 0.952 (1.08) 0.951 (0.16)
200,200 $R_M$ 0.953 (1.37) 0.946 (1.28) 0.950 (4.51) 0.952 (0.22)
$D_M$ 0.945 (0.37) 0.950 (0.28) 0.950 (0.76) 0.947 (0.10)
200,500 $R_M$ 0.946 (1.09) 0.951 (1.02) 0.950 (3.47) 0.952 (0.17)
$D_M$ 0.945 (0.31) 0.951 (0.23) 0.946 (0.69) 0.956 (0.07)
500,500 $R_M$ 0.946 (0.81) 0.952 (0.75) 0.949 (2.69) 0.950 (0.12)
$D_M$ 0.948 (0.23) 0.953 (0.17) 0.950 (0.48) 0.947 (0.06)
500,1000 $R_M$ 0.947 (0.69) 0.952 (0.64) 0.948 (2.23) 0.951 (0.10)
$D_M$ 0.947 (0.20) 0.949 (0.15) 0.949 (0.45) 0.948 (0.04)
1000,1000 $R_M$ 0.947 (0.56) 0.949 (0.52) 0.949 (1.87) 0.950 (0.09)
$D_M$ 0.944 (0.16) 0.950 (0.12) 0.952 (0.34) 0.948 (0.04)
----------- ------------- ------------------ ----------------- ------------------ ------------------
: Simulated coverage probabilities (and widths in parentheses) for the 95% confidence interval for the squared ratio of $\mad$s ($R_M$) and difference of $\mad$s ($D_M$) (\* Median width reported due to excessively large widths for a small number, between 1% and 2%, of intervals).
\[tab:CP\_RmDm\]
Simulated coverages for interval estimators of squared ratio of $\mad$s and difference of $\mad$s are provided in Table \[tab:CP\_RmDm\] for several distributions. Results show excellent coverages compared to the coverages of F-test [the coverage probabilities for interval estimator of the $F$-test can be found in Table 3 of @arachchige2019interval] which are poor due to the violation of underlying normality assumptions). Coverages are very close to the nominal 0.95 for both the squared $\mad$ ratio and difference of $\mad$ for all the selected distributions, including smaller sample sizes. There are some slightly conservative coverages only for $n=50$ and for other sample sizes the coverages become very close. For smaller sample sizes a very small number of the intervals were very wide (between 1% and 2%) so we report the median width instead.
Prostate data example {#sec:Ex}
---------------------
The prostate data set, which is available in the `depthTools` package [@dt], is a normalized subset of the [@singh2002gene] prostate data set. The data consists of gene expressions for the 100 most variable genes for 25 normal and 25 tumoral samples.
We selected three genes that are interesting when comparing intervals for ratios of variances and those based on the MADs. These three genes are three of the six that were considered by [@arachchige2019interval]. The genes and their abbreviations we consider are Glucose-6-phosphate dehydrogenase (G6pd), HDKFZp564A072 and calcium-binding protein A4 (S100cbpA4). Box plots of the genes are provided in Figure \[fig:gene\] where we note that, ignoring extreme outliers, the spread for the bulk of the data looks similar for G6pd and very different for HDKFZp564A072 and S100cbpA4.
--------------- ------- ----------------- ------- ----------------- ------- ------------------
Gene Est. CI Est. CI Est. CI
G6pd 6.496 (2.863, 14.742) 1.000 (0.268, 3.734) 0.000 (-0.185, 0.185)
HDKFZp564A072 1.930 (0.850, 4.379) 5.013 (1.211, 20.761) 0.213 ( 0.035, 0.391)
S100cbpA4 1.748 (0.770, 3.968) 8.725 (1.440, 52.856) 0.301 ( -0.013, 0.615)
--------------- ------- ----------------- ------- ----------------- ------- ------------------
: 95% asymptotic confidence intervals (CI) for the ratios of variances resulting from the $F$-test ($F$), the ratio of $\mad$s ($R_M$) and difference of $\mad$s ($D_M$) for the three selected genes.
\[tab:prost\_CI\]
In Table \[tab:prost\_CI\] we provide the point estimate and asymptotic 95% confidence intervals for the ratio of variances (from the $F$-test assuming underlying normality), the squared ratios of $\mad$s and difference of $\mad$s for the three selected genes. When ignoring two outliers for G6pd the spread looks similar, however the interval for the ratio of variances suggests a large difference in variance between the two. This is not the case for the MAD intervals where the point and intervals estimates suggest little difference. For HDKFZp564A072 and S100cbpA4 the intervals tell a different story. The ratio of variance intervals do not find a significant difference, while the MAD intervals do, or in the case of the difference very close to. We favor the findings from the MAD due to the obvious difference in spread for the bulk of the data as depicted in the box plots. This difference in findings is likely due to the group with smaller spread for most data, have extreme outliers that increases the sample variance so that it is similar to the sample variance for the other group. The MADs are not affected by these outliers. [@arachchige2019interval] provide similar contrasting results when comparing an asymptotic interval for the ratio of variances and intervals based on the interquantile range.
Summary and discussion {#sec:Summary}
======================
The MAD is a robust estimator of scale exhibiting good robustness properties. We have considered interval estimators for the MAD, ratios of $\mad$s and differences of $\mad$s. Simulation results for the interval estimators showed excellent coverages even for small sample sizes such as $n=50$ for all distributions we considered. Our example reveals that different conclusions can be made by using ratios of $\mad$s and differences of $\mad$s compared to intervals for the ratio of variances which is influenced by outliers. Future extensions to this work would be to consider intervals for alternatives to the MAD [e.g. see @rousseeuw1993alternatives].
Appendix {#sec:Appendix}
========
Proof of Theorem \[th:PIF\_Rm\] {#app:proof_of_PIF_Rm}
-------------------------------
A power series expansion of $\MADf(F_\epsilon)$ can be written as $$\MADf(F_\epsilon)=\MADf(F) + \epsilon \IF(x;\MADf,F)+O(\epsilon^2)~.$$ Let $F_\epsilon=(1-\epsilon)F_1+\epsilon\Delta_{x}$, then we have $$\left[\MADf(F_\epsilon)\right]^2=\MADf^2 (F_1) + 2\epsilon \MADf(F_1)\IF(x;\MADf,F_1)+ O(\epsilon^2)~.
\label{eq:MAD2}$$
Therefore, the first PIF is
$$\begin{aligned}
{\text{PIF}}_1 (x; \R_M,F_1,F_2)=&\lim_{\epsilon \downarrow 0}\Bigg\{\frac{\MADf^2 (F_1) + 2\epsilon \MADf(F_1)\IF(x;\MADf,F_1)+ O(\epsilon^2) -\MADf^2 (F_1)}{\epsilon \MADf^2 (F_2)}\Bigg\}\\\end{aligned}$$
For the second PIF set $F_\epsilon=(1-\epsilon)F_2+\epsilon\Delta_{x}$. Then $$\begin{aligned}
{\text{PIF}}_2 (x; \R_M,F_1,F_2)=&\lim_{\epsilon \downarrow 0}\Bigg\{\frac{\MADf^2 (F_1)\left[\MADf^2 (F_\epsilon)\right]^{-1} -\MADf^2 (F_1)/\MADf^2 (F_2)}{\epsilon}\Bigg\}\\
=&\lim_{\epsilon \downarrow 0}\Bigg\{\frac{\MADf^2 (F_1)\MADf^2 (F_2) -\MADf^2 (F_1)\MADf^2 (F_\epsilon)}{\epsilon\MADf^2 (F_2)\MADf^2 (F_\epsilon)}\Bigg\}\\
=&\lim_{\epsilon \downarrow 0}\Bigg\{\frac{-2\epsilon\MADf^2 (F_1)\MADf(F_2)\IF(x;\MADf,F_2) + O(\epsilon^2)}{\epsilon\MADf^2 (F_2)\MADf^2 (F_\epsilon)}\Bigg\}\end{aligned}$$ Recall the $\IF(x;\MADf,F)$ in and evaluated at $F_1$ and $F_2$. Finally, the ${\text{PIF}}_1$ and ${\text{PIF}}_2$ can be obtained by taking the limit by noting that $\lim_{\epsilon\downarrow 0}[O(\epsilon^2)/\epsilon]=0$.
R code for interval estimators {#app:proof_of_PIF_Rm}
------------------------------
# This codes uses the gld R package for estimation of the GLD since it is
# readily available in R.
library(gld)
mad <- function(x) median(abs(x - median(x)))
asv.mad <- function(x, method = "TM"){
lambda <- fit.fkml(x, method = method)$lambda
m <- median(x)
mad.x <- mad(x)
fFinv <- dgl(c(m - mad.x, m + mad.x, m), lambda1 = lambda)
FFinv <- pgl(c(m - mad.x, m + mad.x), lambda1 = lambda)
A <- fFinv[1] + fFinv[2]
C <- fFinv[1] - fFinv[2]
B <- C^2 + 4*C*fFinv[3]*(1 - FFinv[2] - FFinv[1])
(1/(4 * A^2))*(1 + B/fFinv[3]^2)
}
ci.mad <- function(x, y = NULL, gld.est = "TM",
two.samp.diff = TRUE, conf.level = 0.95){
alpha <- 1 - conf.level
z <- qnorm(1 - alpha/2)
x <- x[!is.na(x)]
est <- mad.x <- mad(x)
n.x <- length(x)
asv.x <- asv.mad(x, method = gld.est)
if(is.null(y)){
ci <- mad.x + c(-z, z)*sqrt(asv.x/n.x)
} else{
y <- y[!is.na(y)]
mad.y <- mad(y)
n.y <- length(y)
asv.y <- asv.mad(y, method = gld.est)
if(two.samp.diff){
est <- mad.x - mad.y
ci <- est + c(-z, z)*sqrt(asv.x/n.x + asv.y/n.y)
} else{
est <- (mad.x/mad.y)^2
log.est <- log(est)
var.est <- 4 * est * ((1/mad.y^2)*asv.x/n.x + (est/mad.y^2)*asv.y/n.y)
Var.log.est <- (1 / est^2) * var.est
ci <- exp(log.est + c(-z, z) * sqrt(Var.log.est))
}
}
list(Estimate = est, conf.int = ci)
}
x <- rlnorm(100)
y <- rlnorm(200, meanlog = 1.2)
ci.mad(x) # single sample
ci.mad(x, y) # two sample difference
ci.mad(x, y, two.samp.diff = FALSE) # two sample squared ratio
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, a data hiding scheme ready for Internet applications is proposed. An existing scheme based on chaotic iterations is improved, to respond to some major Internet security concerns, such as digital rights management, communication over hidden channels, and social search engines. By using Reed Solomon error correcting codes and wavelets domain, we show that this data hiding scheme can be improved to solve issues and requirements raised by these Internet fields.'
author:
-
bibliography:
- 'internet2010papier2.bib'
title: An improved watermarking scheme for Internet applications
---
Information hiding; Internet security; Wavelets; Error correcting codes; Social search engines; Digital rights management.
Introduction
============
Information hiding has recently become a major digital technology, especially with the increasing importance and widespread distribution of digital media through the Internet. It encompasses steganography[@Ker06] and digital watermarking. The aim of watermarking is to slightly alter some digital documents, like pictures or movies, for a large panel of reasons, such as: copyright protection, control utilization, data description, integrity checking, or content authentication. Various qualities are required for a watermarking method, depending on the aims to reach: discretion, robustness against attacks [@AdelsbachKS06], *etc.* Many watermarking schemes have been proposed in recent years, which can be classified into two categories: spatial domain [@Wu2007] and frequency domain watermarking [@Cong2006], [@Dawei2004]. In spatial domain watermarking, a great number of bits can be embedded without inducing too clearly visible artifacts, while frequency domain watermarking has been shown to be quite robust against JPEG compression, filtering, noise pollution and so on. More recently, chaotic methods have been proposed to encrypt the watermark, or embed it into the carrier image, to improve security.
Information hiding is now an integral part of Internet technologies. In the field of social search engines, for example, contents like pictures or movies are tagged with descriptive labels by contributors, and search results are determined by these descriptions. These collaborative taggings, used for example in Flickr [@Frick] and Delicious [@Delicious] websites, contribute to the development of a Semantic Web, in which every Web page contains machine-readable metadata that describe its content. Information hiding technologies can be used for embedding these metadata. The advantage of its use is the possibility to realize social search without websites and databases: descriptions are directly embedded into media, whatever their formats. Robustness is required in this situation, as descriptions should resist to modifications like resizing, compression, and format conversion.
The Internet security field is also concerned by watermarking technologies. Steganography and cryptography are supposed to be used by terrorists to communicate through the Internet. Furthermore, in the areas of defense or in industrial espionage, many information leaks using steganographic techniques have been discovered. Lastly, watermarking is often cited as a possible solution to digital rights managements issues, to counteract piracy of digital work in an Internet based entertainment world [@Nakashima2003].
In this paper, the desire is to improve the robustness of the watermarking scheme proposed in [@guyeux10ter], to respond to Internet security concerns recalled above. The robustness of the watermarking scheme through geometric attacks is improved by using Reed Solomon correcting codes, whereas the capacity to withstand JPEG compression and noise pollution attacks is enlarged by embedding the watermark into the wavelets domain. Due to its improved robustness, this scheme is suitable for tagging multimedia contents in a social web search context. Additionally, the proposed scheme possesses various properties of chaos and is secure (see [@guyeux10ter]), so it is suitable when desiring to establish a hidden communication channel through the Internet, or for digital rights management. Lastly, watermark encryption and authentication are possible, which enlarge the variety of use in Internet security applications.
The rest of this paper is organized as follows. Firstly, some basic definitions concerning chaotic iterations and topological chaos are given in Section \[Section:Basic recalls\]. The data hiding scheme used in this paper is recalled in the same section. In Section \[Geometric\], the way to use Reed Solomon error correcting codes to improve robustness against geometric attacks is given. Then it is explained in Section \[frequency\] how to improve robustness against frequency domain attacks by using wavelets coefficients into our scheme. The paper ends with a conclusion section where the contribution is summed up and the planned future work is discussed.
Basic recalls {#Section:Basic recalls}
=============
This section is devoted to the recall of the data hiding scheme, which will be improved in Sections \[Geometric\] and \[frequency\]. To do so, basic notations and terminologies in the fields of chaotic iterations and topological chaos are introduced.
Chaotic iterations and Devaney’s chaos
--------------------------------------
### Chaotic iterations
In the sequel $S^{n}$ denotes the $n^{th}$ term of a sequence $S$, $V_{i}$ denotes the $i^{th}$ component of a vector $V$, and $f^{k}=f\circ ...\circ f$ is for the $k^{th}$ composition of a function $f$. Finally, the following notation is used: $\llbracket1;N\rrbracket=\{1,2,\hdots,N\}$.
Let us consider a *system* of a finite number $\mathsf{M}$ of *cells*, so that each cell has a boolean *state*. Then a sequence of length $\mathsf{M}$ of boolean states of the cells corresponds to a particular *state of the system*. A sequence which elements belong in $\llbracket 1;\mathsf{M} \rrbracket $ is called a *strategy*. The set of all strategies is denoted by $\mathbb{S}.$
Let $S\in \mathbb{S}$. The *shift* function is defined by $\sigma :(S^{n})_{n\in \mathds{N}}\in \mathbb{S}\longrightarrow (S^{n+1})_{n\in \mathds{N}}\in \mathbb{S}$ and the *initial function* $i$ is the map which associates to a sequence, its first term: $i:(S^{n})_{n\in \mathds{N}}\in \mathbb{S}\longrightarrow S^{0}\in \llbracket1;\mathsf{M}\rrbracket$.
The set $\mathds{B}$ denoting $\{0,1\}$, let $f:\mathds{B}^{\mathsf{M}%
}\longrightarrow \mathds{B}^{\mathsf{M}}$ be a function and $S\in \mathbb{S}
$ be a strategy. Then, the so-called *chaotic iterations* are defined [@Robert1986] by $x^0\in \mathds{B}^{\mathsf{M}}$ and $\forall n\in \mathds{N}^{\ast }$, $\forall i\in \llbracket1;\mathsf{M}\rrbracket$,
$$\begin{array}{l}
x_i^n=\left\{
\begin{array}{ll}
x_i^{n-1} & \text{ if }S^n\neq i, \\
\left(f(x^{n-1})\right)_{S^n} & \text{ if }S^n=i.%
\end{array}%
\right.%
\end{array}%$$
In other words, at the $n^{th}$ iteration, only the $S^{n}-$th cell is iterated.
### Devaney’s chaotic dynamical systems {#Devaney's chaotic dynamical systems}
Consider a metric space $(\mathcal{X},d)$ and a continuous function $f:%
\mathcal{X}\longrightarrow \mathcal{X}$. $f$ is said to be *topologically transitive* if, for any pair of open sets $U,V\subset \mathcal{%
X}$, there exists $k>0$ such that $f^{k}(U)\cap V\neq \varnothing $. $(%
\mathcal{X},f)$ is said to be *regular* if the set of periodic points is dense in $\mathcal{X}$. $f$ has *sensitive dependence on initial conditions* if there exists $\delta >0$ such that, for any $x\in \mathcal{X}$ and any neighborhood $V$ of $x$, there exists $y\in V$ and $n\geqslant 0$ such that $|f^{n}(x)-f^{n}(y)|>\delta $. $\delta $ is called the *constant of sensitivity* of $f$.
Quoting Devaney in [@Devaney], a function $f:\mathcal{X}\longrightarrow \mathcal{X}$ is said to be “chaotic” on $\mathcal{X}$ if $(\mathcal{X} ,f)$ is regular, topologically transitive, and has sensitive dependence on initial conditions. When $f$ is chaotic, then the system $(\mathcal{X}, f)$ is highly unpredictable because of regularity and sensitive dependence on initial conditions. Moreover, it cannot be simplified (broken down or decomposed into two subsystems which do not interact) because of topological transitivity. These chaotic dynamical systems then present behaviors very similar to physical noise sources.
In [@guyeux10], a rigorous theoretical framework has been introduced for the study of chaotic iterations. It has been proven that chaotic iterations (CIs) presented above satisfy topological chaos properties, which leads to improve the security of data hiding schemes based on CIs.
Definition of a chaos-based data hiding scheme {#sec:Algo}
----------------------------------------------
### Most and least significant coefficients
Let us define the notions of most and least significant coefficients of an image.
\[definitionMSC\] For a given image, most significant coefficients (in short MSCs), are coefficients that allow the description of the relevant part of the image, *i.e.*, its richest part (in terms of embedding information), through a sequence of bits.
For example, in a spatial description of a grayscale image, a definition of MSCs can be the sequence constituted by the first four bits of each pixel (see Figure \[fig:MSCLC\]). In a discrete cosine frequency domain description, each $8\times 8$ block of the carrier image is mapped onto a list of 64 coefficients. The energy of the image is mostly contained in a determined part of themselves, which can constitute a possible sequence of MSCs.
\[definitionLSC\] By least significant coefficients (LSCs), we mean a translation of some insignificant parts of a medium in a sequence of bits (insignificant can be understand as: “which can be altered without sensitive damages”).
These LSCs can be, for example, the last three bits of the gray level of each pixel (see Figure \[fig:MSCLC\]). Discrete cosine, Fourier, and wavelet transforms can be used also to generate LSCs and MSCs. Moreover, these definitions can be extended to other types of media.
![Example of most and least significant coefficients of Lena.[]{data-label="fig:MSCLC"}](lena512){width="3.5cm"}
\(a) Lena.
![Example of most and least significant coefficients of Lena.[]{data-label="fig:MSCLC"}](lena_msb_678){width="3.5cm"}
\(b) MSCs of Lena.
![Example of most and least significant coefficients of Lena.[]{data-label="fig:MSCLC"}](lena_lsb_1234_facteur17){width="3.5cm"}
\(c) LSCs of Lena ($\times 17$).
LSCs are used during the embedding stage. Indeed, some of the least significant coefficients of the carrier image will be chaotically chosen and switched, or replaced by the bits of the watermark. The MSCs are only useful in case of authentication; mixture and embedding stages depend on them. Hence, a coefficient should not be defined at the same time as a MSC and a LSC: the last can be altered while the first is needed to extract the watermark.
### Stages of the scheme
Our data hiding scheme consists of two stages: (1) mixture of the watermark and (2) its embedding.
#### Watermark mixture
Firstly, for safety reasons, the watermark can be mixed before its embedding into the image. A common way to achieve this stage is to use the bitwise exclusive or (XOR), for example between the watermark and a pseudo-random binary sequence provided by the generator defined in [@guyeux09bis]. In this paper, we introduce a new mixture scheme based on chaotic iterations. Its chaotic strategy will be highly sensitive to the MSCs, in the case of an authenticated watermarking.
#### Watermark embedding
Some LSCs will be switched, or substituted by the bits of the possibly mixed watermark. To choose the sequence of LSCs to be altered, a number of integers, less than or equal to the number $\mathsf{M}$ of LSCs corresponding to a chaotic sequence $U$, is generated from the chaotic strategy used in the mixture stage. Thus, the $U^{k}$-th least significant coefficient of the carrier image is either switched, or substituted by the $k^{th}$ bit of the possibly mixed watermark. In case of authentication, such a procedure leads to a choice of the LSCs which are highly dependent on the MSCs [@guyeux10].
On the one hand, when the switch is chosen, the watermarked image is obtained from the original image whose LSBs $L = \mathds{B}^{\mathsf{M}}$ are replaced by the result of some chaotic iterations. Here, the iterate function is the vectorial boolean negation, $$f_0:(x_1,...,x_\mathsf{M}) \in \mathds{B}^\mathsf{M} \longmapsto (\overline{x_1},...,\overline{x_\mathsf{M}}) \in \mathds{B}^\mathsf{M},$$ the initial state is $L$, and the strategy is equal to $U$. In this case, the whole embedding stage satisfies topological chaos properties (see [@guyeux10]), but the original medium is needed to extract the watermark. On the other hand, when the selected LSCs are substituted by the watermark, its extraction can be done without the original cover (blind watermarking). In this case, the selection of LSBs still remains chaotic because of the use of a chaotic map, but the whole process does not satisfy topological chaos [@guyeux10]. The use of chaotic iterations is reduced to the mixture of the watermark. See the following sections for more detail.
#### Extraction
The chaotic strategy can be regenerated even in the case of an authenticated watermarking, because the MSCs have not been changed during the embedding stage. Thus, the few altered LSCs can be found, the mixed watermark can be rebuilt, and the original watermark can be obtained. In case of a switch, the result of the previous chaotic iterations on the watermarked image should be the original cover. The probability of being watermarked decreases when the number of differences increase.
If the watermarked image is attacked, then the MSCs will change. Consequently, in case of authentication and due to the high sensitivity of the embedding sequence, the LSCs designed to receive the watermark will be completely different. Hence, the result of the recovery will have no similarity with the original watermark.
The chaos-based data hiding scheme is summed up in Figure \[fig:organigramme\].
![The chaos-based data hiding decision tree.[]{data-label="fig:organigramme"}](organigramme22){width="8cm"}
Improving robustness against geometric attacks {#Geometric}
==============================================
In this section, we are interested in improving our scheme to make its use relevant in a social web search context. The idea is to embed the tag of a given image into its pixel values. As neither the cover image nor the tag should be required during a search, the LSBs will be replaced (not switched), by the tags. Authentication is not required, as man-in-the-middle attacks are not supposed to occur. However tags are vulnerable to involuntary attacks such as rotation or resizing, so we need to improve robustness against geometric attacks. To do so, the embedding domain will be the spatial domain. This choice leads to a large given payload, so a lot of tags can be embedded into the cover image. Additionally, we will use Reed-Solomon error correcting codes, to reinforce the capacity to extract the watermark from a tagged image, even though it has been altered. As an illustrative example, we will show how to embed the description of the well-known Lena into its own image. Let us remark that the same procedure can be applied to create a hidden channel for communicating through a given web page, for example, by inserting messages in the background image of this website. In this situation, it is recommended to add the encryption stage to improve the security of the hidden channel.
In this illustrative example, the following text extracted from Wikipedia’s description of Lena will be inserted into its own image:
`Lena (Soderberg), a standard test image`\
`originally cropped from the November `\
`1972 issue of Playboy magazine.`
The cover image will be the Figure \[fig:MSCLC\](a), which is a $256 \times 256$ grayscale image. The text to embed is converted into 756 binary digits by using the ASCII table: each of the 109 characters are coded with 7 digits, thus obtaining the following bits flow (called a system):
100110011001011101110110000101000000101000101001111011 111100100110010111100101100010110010111100101100111...
20000 binary digits are computed from a logistic map, with parameters $\mu = 3.999999$, $x^0 = 0.65$, and those binary digits are grouped ten by ten ($10 = \lceil \log_2(756)\rceil$) to obtain an integer sequence $S$ lesser than or equal to 756. So, chaotic iterations are applied to the above system, with chaotic strategy $S$ and the vectorial boolean negation, to obtain the following encrypted message:
001000111110001110001101110111111000011011010011000101 001011110000110110011010010001110101101100010110101...
In this example, there is no authentication step, but Reed-Solomon error correction codes are used to increase the robustness. Here, two layers of Reed–Solomon coding, respectively (32,24)-RS and (24,16)-RS codes, are separated by a 3-way convolutional interleaver operation, to obtain a scheme similar to the Cross-Interleaved Reed Solomon Coding (CIRC) of the compact disc. The message to embed is the result of this coding operation: a 2112 binary stream, starting by:
010110100101100000100001000111000010011100111111010001 110111100000010110001101010111011000010011001001110...
These 2112 bits will be embedded into Lena, an image constituted by $256 \times 256 \times 8 = 524288$ bits (8 bits per pixel). To do so, we will consider the two least significant bits of each pixel as LSCs: a few of them will be replaced by the bits of the watermark. To select these bits to replace, the strategy $S$ of the encryption stage is used again, to generate a sequence of triplets $(x^n, y^n, z^n)_{n\in \mathds{N}}$ in such a way that $x^n, y^n \in \llbracket 0 ; 255 \rrbracket^\mathds{N}$, and $z^n \in \{1 ; 2 \}^\mathds{N}$. This generation is realized as follows:
$$\left\{
\begin{array}{lll}
x^0 & = & 11, \\
y^0 & = & 23, \\
z^0 & = & 1, \\
\end{array}
\right.$$
and
$$\left\{
\begin{array}{llll}
x^{n+1} & = & 2 x^n +S^{3n}+n & (\text{mod } 255), \\
y^{n+1} & = & 2 y^n +S^{3n+1}+n & (\text{mod } 255), \\
z^{n+1} & = & 2 z^n +S^{3n+2}+n & (\text{mod } 2). \\
\end{array}
\right.$$
So the $n^{th}$ bit of the encrypted and encoded binary message is inserted into the $z^n$ least significant bit of the pixel in position $(x^n, y^n)$ of Lena, to obtain the watermarked Lena in Figure \[fig:water\](a). In Figure \[fig:water\](b) the differences are shown between the original Lena and the watermarked Lena. This image illustrates the fact that LSCs to be replaced are chaotically chosen and uniformly distributed [@guyeux10].
![Watermarked Lenas (scale reduced).[]{data-label="fig:water"}](lena_marque){width="3.5cm"}
\(a) Watermarked Lena.
![Watermarked Lenas (scale reduced).[]{data-label="fig:water"}](lena_diff){width="3.5cm"}
\(b) Differences with Lena.
![Watermarked Lenas (scale reduced).[]{data-label="fig:water"}](lena_marque_attaque){width="3.5cm"}
\(c) Attacked Lena.
In [@guyeux10ter], our scheme has been defined without RS codes and its robustness has been evaluated. It is established that the watermark can resist rotation, cropping, JPEG compression, and gaussian noise attacks. However, the extracted watermark is slightly different from the original one and this difference increases with the number of attacks. These errors, which are undesirable in a social web search context, can be corrected by the use of RS codes.
To illustrate, the watermarked Lena is zeroed: a square of $40\times 40$ pixels is removed from the image, as in Figure \[fig:water\](c). So the message is extracted from the watermarked and attacked Lena: the strategy $S$ is regenerated from a logistic map with the same parameters as above. Then the sequences $x^n, y^n$ and $z^n$ can be regenerated too, and the embedded bits can thus be extracted. These bits are decoded in the reverse process: (24,16)-RS decoding, 3-way de-interlacing, and (32,24)-RS decoding codes. Lastly, the resulting bits sequence is decrypted, bits are grouped 7 by 7, and converted into characters with the ASCII table, to obtain the following message:
`Lena (Soderberg), a standard test image`\
`originally cropped from the November `\
`1972 issue of Playboy magazine.`
Improving robustness against frequency domain attacks {#frequency}
=====================================================
In this section, the way to use our scheme in frequency DWT domain is explained. Due to its robustness against frequency attacks such as JPEG compression, this scheme can be used to insert a copyright into a media (digital rights management context).
Stages and detail
-----------------
The carrier image and watermark are the same as in Section \[Geometric\], but Lena is now constituted by $512 \times 512$ pixels. The embedding domain is the discrete wavelets domain (DWT). In this paper, the Daubechies family of wavelets is chosen: Lena is converted into its Daubechies-1 DWT coefficients, which are altered by chaotic iterations. The watermark is encrypted by chaotic iterations before its embedding, with the same procedure as above.
Each example below depends on a decomposition level and a coefficient matrix (Figure \[fig:DWTs\]): $LL$ means approximation coefficient, when $HH,LH,HL$ denote respectively diagonal, vertical and horizontal detail coefficients. For example, the DWT coefficient HH2 is the matrix equal to the diagonal detail coefficient of the second level of decomposition of Lena.
![Wavelets coefficients.[]{data-label="fig:DWTs"}](DWTs){width="3.7cm"}
To embed the encrypted watermark, LSCs are obtained from the coefficients defined above. The system to iterate is the boolean vector of size $256^2$, constituted by these LSCs of Lena. Iterate function is the vectorial boolean negation, and chaotic strategy $U$ is defined as follows: $$\left\{
\begin{array}{lll}
U^{0} & = & S^{0} \\
U^{n+1} & = & S^{n+1}+2\times U^{n}+n ~ (\textrm{mod } \textsf{M}).
\end{array}%
\right.
\label{equation_strategie}$$
where $S$ denotes the strategy used in the encryption stage (see [@guyeux10ter]). Thus, bits of the LSCs are switched, not replaced: the whole embedding process satisfies Devaney’s chaos property [@guyeux10]. However, for this reason, the watermark cannot be extracted: contrary to Section \[Geometric\], we are not in a steganographic framework, but in a pure non-blind watermarking scheme used for digital rights management. To know if a given image $I'$ is the watermarked version of another image $I$:
- the whole process is applied to $I'$, with the same parameters (LSCs, watermark, *etc.*), to obtain $I''$,
- $I''$ is compared to the original $I$.
To evaluate the differences, the RMS value defined by $\bar{x} = \sqrt{\frac{1}{\mathsf{M}}\sum_{i=1}^\mathsf{M}{(I-I'')_i^2}}$ is computed. The probability that the image has been watermarked increases when the RMS decreases. Indeed, each bit of the LSCs of $I''$ has been switched an even number of times (the RMS is nonzero because of computational errors).
First example: coefficient HH2
------------------------------
### Embedding
In this first experiment, the watermark is inserted into the diagonal coefficient HH2 (a real matrix of size $128 \times 128$). LSCs are the second least significant bit of each integral value of HH2. To do the insertion, chaotic iterations are made. The system to iterate is the boolean vector of size $128^2$, constituted by the LSCs of Lena. Iterate function is the vectorial boolean negation and chaotic strategy is defined as in Equation \[equation\_strategie\], with $U^0 = 1$ and $\mathsf{M} = 256^2$.
![Data hiding in DWT domain[]{data-label="fig:DWT"}](lena512){width="3.5cm"}
\(a) Original Lena.
![Data hiding in DWT domain[]{data-label="fig:DWT"}](lena512marqueDWT){width="3.5cm"}
\(b) Watermarked Lena.
In this situation, PSNR = 53.45 dB. Pixel values have been modified by at most of one level of gray. The mean value of differences is equal to 0.294, when RMS = 0.542. The alteration can thus be considered as indistinguishable.
### Extraction
The system to iterate is constituted by the second least significant bit of each integral value of HH2, the approximation coefficient of the first decomposition level of the watermarked Lena. The iterate function is the vectorial boolean negation and the chaotic strategy is computed as above. Thus, the result is compared to the coefficient HH2 of the original Lena. The RMS is equal to 0.129. As a comparison, Table \[Tab:RMS\] gives the RMS values resulting on a bad extraction (wrong parameters, *etc.*) Symbol ‘-’ means that the value of the considered parameter is unchanged. We show that the least RMS is obtained for an extraction with the same parameters as the embedding. Let us notice that if the extraction is attempted to the original Lena, RMS is twice greater than 0.127.
------------ ------------- ------- ------------ ----------------- -------
$\mu$ $U^0$ Iterations Authentication RMS
(r)[2-6]{} 3.99987 - - - 1.131
- 0.64 - - 1.129
Encryption - - 19950 - 0.796
- - - MSB = \[5,6,7\] 1.122
Coefficient $S^0$ LSB RMS
(r)[2-5]{} HH1 - - 253.65
Embedding - 2 - 0.653
- - \[1\] 0.983
------------ ------------- ------- ------------ ----------------- -------
: RMS values for a HH2 embedding[]{data-label="Tab:RMS"}
Second example: coefficient LL1
-------------------------------
### Embedding
In this paragraph, the watermark is inserted into the approximation coefficient LL1 of Lena (a real matrix of size $256 \times 256$) and LSCs are the second least significant bit of each integral value of LL1.
To realize the embedding, chaotic iterations are realized as before. The system to iterate is the boolean vector, of size $256^2$, constituted by the LSCs of Lena. Iterate function is the vectorial boolean negation, chaotic strategy is defined as in Equation \[equation\_strategie\] with $U^0 = 1$, and $\mathsf{M} = 256^2$. In this situation, PSNR = 60.06 dB. Pixel values have been modified by at most two levels of gray. The mean value of differences is 0.063, when the RMS is equal to 0.245. For all of these reasons, the alteration can be considered again as indistinguishable.
### Extraction
The system to iterate is constituted by the second least significant bit of each integral value of LL1. Iterate function is the vectorial boolean negation and chaotic strategy is computed as above. Thus, the result is compared to the coefficient LL1 of the original Lena. In our example, we obtain RMS = 0.127. As a comparison, Table \[Tab:RMS2\] gives the RMS values resulting in a bad extraction (wrong parameters, *etc.*) Symbol ‘-’ means that the value of the considered parameter is unchanged. We show that the least RMS is obtained for an extraction with the same parameters as the embedding. Let us remark that if the extraction is tried on the original Lena, then RMS is twice greater than $0.127$.
------------ ------------- ------- ------------ ----------------- -------
$\mu$ $U^0$ Iterations Authentication RMS
(r)[2-6]{} 3.99987 - - - 0.669
- 0.64 - - 0.670
Encryption - - 19950 - 0.443
- - - MSB = \[5,6,7\] 0.667
Coefficient $S^0$ LSB RMS
(r)[2-5]{} HH1 - - 223.737
Embedding - 2 - 0.135
- - \[1\] 0.548
------------ ------------- ------- ------------ ----------------- -------
: RMS values for a LL1 embedding[]{data-label="Tab:RMS2"}
Discussion and future work
==========================
In this paper, the robustness of the data hiding scheme proposed in [@guyeux10ter] is improved to achieve properties required in Internet applications of data hiding techniques. This scheme depends on a general description of the carrier medium to watermark, in terms of the significance of some coefficients we called MSCs and LSCs. The encryption of the watermark and the selection of coefficients to alter are based on chaotic iterations, which generate topological chaos in the sense of Devaney [@guyeux10]. Thus, the proposed scheme has a sufficient level of security for Internet applications, such as digital rights management or social web search.
We have proposed in this paper to enlarge the relevance of our scheme in these contexts by using Reed-Solomon error correcting codes and wavelets domain. The first improvement is relevant in a social web search domain, in which the tags of an image must be recovered exactly, even though the image has faced geometric operations. The use of wavelets domain is linked more to digital rights management. This domain is known to present good results against frequency attacks, which can occur when someone tries to remove some DRM. It can be noticed that these two improvements can be realized together.
The schemes have been evaluated through attacks and results have been experimentally obtained. Choices that have been made in this first study are simple: spatial and Daubechies domains for the embedding, negation function as iteration function, *etc.* The aim was not to find the best watermarking method generated by our general scheme, but to explain how to improve robustness for Internet applications.
In future work, other choices of iteration functions and chaotic strategies will be explored and compared in order to increase authentication and robustness to attacks. In addition, new frequency domain representations will be used to select the MSCs and LSCs. Properties induced by topological chaos, such as entropy, will be explored and their role in Internet applications will be explained.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Energetic feedback from star clusters plays a pivotal role in shaping the dynamical evolution of giant molecular clouds (GMCs). To study the effects of stellar feedback on the star formation efficiency of the clouds and the dynamical response of embedded star clusters, we perform a suite of isolated GMC simulations with star formation and momentum feedback subgrid models using the moving-mesh hydrodynamics code <span style="font-variant:small-caps;">Arepo</span>. The properties of our simulated GMCs span a wide range of initial mass, radius, and velocity configurations. We find that the ratio of the final stellar mass to the total cloud mass, ${\epsilon_{\mathrm{int}}}$, scales strongly with the initial cloud surface density and momentum feedback strength. This correlation is explained by an analytic model that considers force balancing between gravity and momentum feedback. For all simulated GMCs, the stellar density profiles are systematically steeper than that of the gas at the epochs of the peaks of star formation, suggesting a centrally concentrated stellar distribution. We also find that star clusters are always in a sub-virial state with a virial parameter $\sim0.6$ prior to gas expulsion. Both the sub-virial dynamical state and steeper stellar density profiles prevent clusters from dispersal during the gas removal phase of their evolution. The final cluster bound fraction is a continuously increasing function of ${\epsilon_{\mathrm{int}}}$. GMCs with star formation efficiency smaller than 0.5 are still able to form clusters with large bound fractions.'
author:
- |
Hui Li$^{1}$[^1][](https://orcid.org/0000-0002-1253-2763), Mark Vogelsberger$^{1}$, Federico Marinacci$^{3,2,1}$[](https://orcid.org/0000-0003-3816-7028) Oleg Y. Gnedin$^{4}$[](https://orcid.org/0000-0001-9852-9954)\
$^{1}$Department of Physics, Kavli Institute for Astrophysics and Space Research, Massachusetts Institute of Technology, Cambridge, MA 02139, USA\
$^{2}$Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA\
$^{3}$Department of Physics & Astronomy, University of Bologna, via Gobetti 93/2, 40129 Bologna, Italy\
$^{4}$Department of Astronomy, University of Michigan, Ann Arbor, MI 48109, USA
bibliography:
- 'references.bib'
date: 'Accepted XXX. Received YYY; in original form ZZZ'
title: Disruption of giant molecular clouds and formation of bound star clusters under the influence of momentum stellar feedback
---
\[firstpage\]
methods: numerical – stars: formation – stars: kinematics and dynamics – galaxies: star clusters: general
Introduction {#sec:intro}
============
Most, if not all, stars are formed in clusters [@lada_lada03], which emerged from giant molecular clouds [GMC; @shu_etal87; @scoville_good89; @mckee_ostriker07; @krumholz_etal18]. Due to the complex interplay of gravity, supersonic turbulence and stellar feedback from massive stars, the dynamical evolution and cluster formation activities in GMCs are still highly debatable [e.g. @krumholz_mckee05; @ballesteros-paredes_hartmann07; @heitsch_etal09; @hennebelle_chabrier11; @padoan_nordlund11; @federrath_klessen12; @burkert_hartmann13; @traficante_etal18]. One of the key observables that can be used to constrain various physical process is the star formation efficiency (SFE) in star-forming regions.
It is well known that star formation is inefficient on galactic scales. The observed linear correlation between molecular gas surface density and star formation rate (SFR) surface density in normal star-forming galaxies suggests a nearly constant gas depletion time-scale around $\sim$ 2 Gyr, much longer than the dynamical time-scale of galactic disks [@kennicutt89; @bigiel_etal08; @saintonge_etal11; @leroy_etal13; @genzel_etal15; @tacconi_etal18]. In contrast, the SFE on GMC scales shows a large variation ranging from less than a few percent to nearly unity [@zuckerman_evans74; @krumholz_tan07; @wu_etal10; @evans_etal14; @heyer_etal16; @lee_etal16; @vutisalchavakul_etal16]. The origin of this large scatter is usually explained as a combination of the time variability of the SFR during the course of cloud evolution and intrinsic scatter of SFEs due to the diversity of GMC properties [@feldmann_gnedin11; @kruijssen_longmore14; @lee_etal16; @grudic_etal18b; @kruijssen_etal18b]. For example, recent theoretical models and high-resolution magneto-hydrodynamics simulations suggest that the SFE depends on the local virial parameters of the cloud controlled by large-scale turbulence [e.g. @krumholz_mckee05; @padoan_etal12]. However, it has recently been recognized that large-scale turbulence can only account for an $\sim$ 0.3 dex scatter, which is not sufficient to explain the observed SFE variations [@lee_etal16]. Another source of variation comes from different stellar feedback channels that alter the dynamical states of the GMCs [@fall_etal10; @murray_etal10; @dale_etal14; @myers_etal14; @raskutti_etal16; @kim_etal17; @grudic_etal18]. Previous studies found that GMC simulations adopting different stellar feedback mechanisms (stellar winds, ionizing radiation, or supernovae) lead to dramatically different final SFEs. The problem has recently been recognized to be more subtle than previously thought, since even small differences in numerical treatments, such as different radiative transfer schemes, massive star sampling, and momentum and energy deposition algorithms, can lead to drastic changes for the final SFE [@dale_etal05; @roskar_etal14; @raskutti_etal16; @grudic_etal18; @kim_etal18]. Therefore, how the SFE depends on GMC properties and the strength of stellar feedback remains an open question.
Stellar feedback not only changes the efficiency of star formation within GMCs, but also alters the dynamical state of star clusters by dispersing the cloud. The process of star cluster disruption due to rapid gas expulsion shortly after the cluster emerges from its natal cloud is believed to be the main culprit of the “infant mortality” of star clusters – a sharp decrease in the number of young star clusters with the increase of cluster age in local star-forming regions [e.g. @lada_lada03]. A simple virial analysis suggests that a star cluster that is initially in virial equilibrium will dissociate if more than half of the mass is instantaneously lost [@hills80; @mathieu83]. However, this statement does not take into account the highly non-linear star formation and stellar feedback process in realistic self-gravitating turbulent environments. For example, embedded clusters in star-forming regions are not necessarily in virial equilibrium. Recent hydrodynamical simulations suggest that the stellar velocity dispersions are in general much smaller than that of the gas, suggesting a sub-virial dynamical state of star clusters within GMCs [e.g. @offner_etal09]. Moreover, stars are usually not well mixed with gas but instead formed in the densest part of the cloud. The difference between the gas and stellar distribution can strongly affect the dynamical response of star clusters to gas dispersal [e.g. @kruijssen_etal12b; @shukirgaliyev_etal18]. Most importantly, GMCs are highly substructured. Stars are formed at the intersections of gas filaments and assembled into different subclusters hierarchically. Previous works have explored some of the above complications using different physical and numerical methods, including analytical models [@hills80; @mathieu83; @adams00; @boily_kroupa03a; @kruijssen12; @parmentier_pfalzner13], pure *N*-body simulations [@tutukov78; @lada_etal84; @boily_kroupa03b; @goodwin_bastian06; @baumgardt_kroupa07; @smith_etal11; @farias_etal18], and hydrodynamic simulations [@bonnell_etal11; @girichidis_etal12; @moeckel_etal12; @fujii_portegies_zwart_2016; @gavagnin_etal17]. Recent efforts have been made to include various relevant physical processes in hydrodynamic simulations [e.g. @parker_etal15; @gavagnin_etal17], however, due to the high computational costs, they usually focus on a handful of less massive GMCs.
In this paper, we perform a suite of hydrodynamic simulations of turbulent GMCs employed with a simple star formation and stellar feedback models in the moving-mesh code <span style="font-variant:small-caps;">AREPO</span>. We survey GMCs with a broad range of mass, size, and velocity configurations to investigate the physical origin of the intrinsic variations of SFEs and the properties of surviving star clusters. The structure of this paper is as follows. In Section \[sec:methods\], we describe the simulation setup, star formation and momentum stellar feedback implementations, and the design of initial conditions. In Section \[sec:sfe\], we examine the dependence of integrated SFE of GMCs on cloud mass, size, and momentum feedback intensity. In Section \[sec:fbound\], we describe the subsequent dynamical evolution of star clusters after the residual gas is completely expelled and investigate the relationship between SFE and cluster bound fraction. We summarize our conclusions in Section \[sec:summary\].
Methods {#sec:methods}
=======
Simulation setup {#sec:method-setup}
----------------
The simulations in this work are performed with <span style="font-variant:small-caps;">Arepo</span> [@springel10arepo], a moving-mesh, finite-volume hydrodynamic code employing a second-order unsplit Godunov scheme. The control volumes are discretized by a Voronoi tessellation, which is generated from its dual Delaunay tessellation determined by a set of mesh-generating points. These points can move freely within the simulation domain and follow gas flows in a quasi-Lagrangian fashion. Therefore, <span style="font-variant:small-caps;">Arepo</span> captures the advantages of both grid- and particle-based hydrodynamic methods and has already been applied to various astrophysical problems [e.g. @keres_etal12; @torrey_etal12; @vogelsberger_etal12; @vogelsberger_etal14; @springel_etal18]. Our simulations include hydrodynamics, self-gravity, radiative cooling, star formation, and momentum feedback from stellar winds.
We use an adaptive softening scheme for gas cells so that the gravitational forces are resolved all the way down to the size of each cell. We employ a quasi-Lagrangian refinement scheme that keeps the mass of gas cells close to a target mass determined by initial conditions. In addition, we refine a cell if its volume is more than 32 times larger than the minimum volume of all its face-touching neighbours. This volume-limited refinement scheme prevents large volume contrast between adjacent cells and helps to better resolve the regions that experience fast expansion due to stellar feedback.
Since radiative cooling is responsible for gas fragmentation and subsequent star formation in GMCs, it is important to follow the cooling process explicitly over a large range of temperatures. Instead of adopting isothermal or effective equations of state, which have been used in many previous GMC simulations [e.g. @dale_etal05; @raskutti_etal16; @kim_etal18], we explicitly include radiative cooling from multiple channels: a network implementing hydrogen and helium cooling and heating processes due to collisions, recombinations, free-free emission and photoionization from UV background radiation; high-temperature ($T > 10^{4}\, {\rm K}$) metal-line cooling that is added to the hydrogen and helium network following @vogelsberger_etal13; low-temperature ($T < 10^{4}\, {\rm K}$), metal-line, fine-structure and molecular cooling implemented as a fitting function, depending on temperature, density and gas metallicity, to <span style="font-variant:small-caps;">CLOUDY</span> calculations presented in @hopkins_etal18 and Marinacci et al. (in prep.). We set the metallicity of the GMCs to solar abundance when evaluating the cooling rates from metals. One caveat is that the adopted cooling curves are based on the <span style="font-variant:small-caps;">CLOUDY</span> model calculations under the spatially uniform UV background used for galaxy formation simulations. The cooling rate calculated based on this uniform UV background may not be accurate for GMC simulations, especially in close proximity of massive stars.
Star formation {#sec:method-sf}
--------------
During each simulation time-step, we identify all star-forming cells and convert them to stellar particles probabilistically. Star-forming cells are defined as gas cells that are cold ($T_{\rm cell}<100$ K), contracting ($\nabla\cdot \bm{v}<0$), and self-gravitating ($|\nabla\cdot\bm{v}|^2+|\nabla\times\bm{v}|^2<2G\rho$), where $\bm{v}$ and $\rho$ are the velocity and density of the cells, respectively. We also employ a density threshold for star formation, $n_{\rm cell} > 10^5{\rm cm^{-3}}$, to avoid rare situations where some self-gravitating clumps are formed in the very low density outskirt of the cloud.
A given star-forming cell is converted to stellar particles with a constant probability $p=\Delta t/{\tau_{\mathrm{ff}}}(\rho)$ at a given time-step $\Delta t$, where ${\tau_{\mathrm{ff}}}=(3\pi/32G\rho)^{1/2}$ is the free-fall time of the cell. The cells that are converted to stellar particles are removed and the volume of these cells is claimed by their neighbours. The mass, position, and velocity of the newly formed stellar particles are inherited from their parent gas cells. Therefore, the mass distribution of stars is similar to that of the gas particles, which is around the target mass of the simulations. After the stellar particles are created, they are treated as collisionless particles with a Plummer-equivalent softening length fixed to $10^{-4}$ of the initial diameter of the GMC.
Momentum stellar feedback {#sec:method-feedback}
-------------------------
The overall evolution of GMCs depends strongly on the strength of stellar feedback. Unfortunately, the exact amount of feedback that is associated with massive stars in the simulations is still debated. As has already been noticed in previous studies, GMC simulations with different stellar feedback sources (stellar winds, ionizing radiation, or supernovae) show dramatically different gas evolution and star formation efficiencies [@dale_etal05; @roskar_etal14; @raskutti_etal16; @grudic_etal18c; @kim_etal18]. Moreover, it has recently been recognized that, even some small changes in numerical implementations, such as radiation hydrodynamic methods [@kim_etal18], sampling of massive star formation [@grudic_etal18b], and momentum/energy deposition algorithms [@hopkins_etal18], can contribute noticeable variation to the star formation efficiencies of the clouds. Since exploring accurate feedback implementation from various sources is not the main focus of this paper, we simply treat stellar feedback by depositing mass and momentum fluxes from stellar particles to their neighbouring gas cells.
We set the fiducial mass-loss and momentum deposition rate to the initial mass function (IMF)-averaged values of stellar winds from a single stellar population with a Kroupa initial mass function [@kroupa01]. Following @hopkins_etal18, the mass-loss rate per unit stellar mass is $$\label{eq:wind-mass-loss}
\frac{\dot{m}_{\rm w}}{{\rm Gyr^{-1}}} = \begin{cases}
4.763(0.01+Z/Z_\odot) & t_6<1\\
4.763(0.01+Z/Z_\odot)t_6^{1.45+0.8\ln{(Z/Z_\odot)}} & 1<t_6<3.5\\
29.4(t_6/3.5)^{-3.25} + 0.0042 & 3.5<t_6<100
\end{cases}$$ where $Z$ is the metallicity and $t_6$ is the age of stellar particles in unit of Myr. The kinetic luminosity of winds per unit stellar mass is $$\label{eq:wind-lum}
l_{\rm w} = \left[\frac{5.94\times10^4}{1+(t_6/2.5)^{1.4}+(t_6/10)^5}+4.83\right]\times10^{12}\dot{m}_{\rm w}~{\rm erg/g},$$ for $t_6<100$. Winds from stellar particles older than $t_6>100$ are irrelevant here since the dynamical time-scales of our model GMCs are much shorter than 100 Myr.
The mass-loss and wind momentum are deposited to the gas cells around each stellar particle in the following way: for a given stellar particle of mass $m_*$ at time-step $\Delta t$, the total mass-loss is $\Delta m=\dot{m}_{\rm w}m_*\Delta t$ and wind momentum is $\Delta p=\sqrt{2l_{\rm w}\dot{m}_{\rm w}}{f_{\mathrm{boost}}}m_*\Delta t$, where ${f_{\mathrm{boost}}}$ is a boosting factor to the fiducial wind momentum to mimic the feedback intensity from different feedback sources. The mass and momentum fluxes from a stellar particle are distributed to its nearest 32 neighbouring gas cells in a weighted fashion so that cell $i$ with weight $w_i$ receives mass $\Delta m_i=(w_i/\Sigma_j{w_j})\Delta m$ and momentum $\Delta \bm{p}_i=(w_i/\Sigma_j{w_j})\Delta p\bm{r}_i/|\bm{r}_i|$, where $\bm{r}_i$ is the vector from the position of the stellar particle to the mesh-generating point of cell $i$. The weight can be chosen to be any physical quantities of the cells, such as volume, mass, or solid angle opened to the stellar particle. To test the robustness of the feedback implementation, we perform a series of numerical tests of wind-blowing bubbles created by a stellar particle with constant mass-loss rate $\dot{m}=10^{-5}{\,M_{\odot}}/$yr and wind velocity $v_{\rm w}=500$ km/s (see Appendix \[sec:appendix-wind\]). We find that the expansion history and the internal structure of the bubble are consistent with analytical solutions in @weaver_etal77. We also test the sensitivity of the star formation history of one GMC using different weighting methods (see Appendix \[sec:appendix-wind-gmc\]). To make consistent investigation across all GMCs, in the rest of the paper, we use solid angle as the weight to deposit mass and momentum.
---------- ------------------------------- ---------- -------------- ----------------------- --------------------------------- ------------------------------ -- -- --
Name $M_{\rm GMC} ({\,M_{\odot}})$ $R$ (pc) $\alpha_0^R$ $\tau_{\rm ff}$ (Myr) $m_{\rm res}$ (${\,M_{\odot}}$) $l_{\rm soft}$ ($10^{-3}$pc)
(i) (ii) (iii) (iv) (v) (vi) (vii)
RHO5 T/R $1.25\times10^4$ 5 0.1/0.9 1.6 $6.0\times10^{-3}$ 1
RHO10 $10^5$ 10 0.1/0.9 1.6 $4.8\times10^{-2}$ 2
RHO20 $8\times10^5$ 20 0.1/0.9 1.6 $0.38$ 4
RHO40 $6.4\times10^6$ 40 0.1/0.9 1.6 $3.1$ 8
RHO80 $5.12\times10^7$ 80 0.1/0.9 1.6 $24.4$ 16
SIGMA5 $2.5\times10^4$ 5 0.1/0.9 1.13 $1.2\times10^{-2}$ 1
SIGMA20 $4\times10^5$ 20 0.1/0.9 2.26 $0.19$ 4
SIGMA40 $1.6\times10^6$ 40 0.1/0.9 3.2 $0.76$ 8
---------- ------------------------------- ---------- -------------- ----------------------- --------------------------------- ------------------------------ -- -- --
\
**Note.** Column information: (i) model name, (ii) initial GMC mass, (iii) initial GMC radius, (iv) initial virial parameter for the rotational components: $\alpha_0^R=2E_{\rm rot}/|E_G|$, see in details, (v) initial free-fall time, (vi) target mass for gas cells, (vii) gravitational softening length of stellar particles.\
\[tab:run-parameter\]
Initial conditions {#sec:method-IC}
------------------
We set up the initial condition of GMCs as gas spheres of uniform density with initial turbulent velocity fields. The mass, radius, and other physical parameters of the initial conditions are listed in Table \[tab:run-parameter\]. We choose the initial mass (${M_{\mathrm{GMC}}}$) and radius (${R_{\mathrm{GMC}}}$) of the GMCs so that all “RHO” runs have the same initial volume density ($\rho_0={M_{\mathrm{GMC}}}/(\frac{4}{3}\pi{R_{\mathrm{GMC}}}^3)\approx24{\,M_{\odot}}{\rm pc^{-3}}$) and all “SIGMA” runs have the same initial surface density ($\Sigma_0={M_{\mathrm{GMC}}}/\pi{R_{\mathrm{GMC}}}^2\approx318{\,M_{\odot}}{\rm pc^{-2}}$). The goal of this experimental design is to determine whether SFE depends on volume density or surface density.
The initial velocity field is initialized as a combination of turbulent motions and rigid rotation along the $z$-axis. We assign the rotational velocity field as $$v_x^R(x,y,z)=-\Omega_iy; \; v_y^R(x,y,z)=\Omega_ix; \; v_z^R(x,y,z)=0,$$ where $v_x^R, v_y^R, v_z^R$ are the three components of the rotation velocities with circular frequency $\Omega_i$. For each run listed in , we construct two separate initial conditions: rotation-supported (“R”) and turbulent-supported (“T”) runs. The virial parameter contributed from rotation $\alpha_0^R\equiv2E_{\rm rot}/|E_G|=0.1/0.9$ is used to calculate $\Omega_i$ for the corresponding “T”/“R” runs, where $E_{\rm rot}$ and $E_G$ are the rotational energy and gravitational energy of the cloud.
{width="0.985\columnwidth"} {width="0.985\columnwidth"} {width="0.985\columnwidth"} {width="0.985\columnwidth"}
The turbulent velocity field is first initialized as a Gaussian random field in the Fourier space with the variance of the field determined by a given power spectrum, $P(k)$. Each dimension of the turbulent velocity field is treated independently and the result is a natural mixture of solenoidal and compressive turbulence. In order to rearrange the turbulent field into arbitrary solenoidal and compressive components, we perform a Helmholtz decomposition in $k$-space by applying the projection operator to the field [@federrath_etal10] $$\bm{\mathcal{H}}_{ij}(\mathbf{k})=\eta_{\rm comp}\bm{\mathcal{H}}_{ij}^{\parallel}(\mathbf{k})+(1-\eta_{\rm comp})\bm{\mathcal{H}}_{ij}^{\bot}(\mathbf{k}),$$ where $\bm{\mathcal{H}}_{ij}^{\parallel}(\mathbf{k})=k_ik_j/k^2$ and $\bm{\mathcal{H}}_{ij}^{\bot}(\mathbf{k})=\bm\delta_{ij}-k_ik_j/k^2$ are the compressive and solenoidal operators, respectively, and $\bm\delta_{ij}$ is the Kronecker delta function. $\eta_{\rm comp}$ is the contribution from the compressive mode ranging from 0 to 1. $\eta_{\rm comp}=1$ means a purely compressive turbulence field and $\eta_{\rm comp}=0$ means a purely solenoidal. Varying $\eta_{\rm comp}$ would lead to a different density structure of the clouds. For simplicity, we use $\eta_{\rm comp}=1/3$ so that the ratio of kinetic energy of the compressive mode to that of the solenoid mode is 2:1 to mimic the natural mixture of the two turbulent modes. After projection, the turbulence field in $k$ space is Fourier-transformed to real space and is then interpolated to the position of gas cells within the sphere. The field is renormalized so that the virial parameter due to turbulence is $\alpha_0^T\equiv2E_{\rm turb}/|E_G|$. We adopt a power-law power spectrum $P(k)\propto k^{-4}$, which is similar to the turbulence properties of GMCs [e.g. @dobbs_etal14]. We assume the cloud is initially in virial equilibrium so that the virial parameter of the cloud $\alpha_0=\alpha_0^R+\alpha_0^T=1$.
The gas temperature is initialized to 10 K, which is commonly used for GMC simulations [@dale_etal14; @raskutti_etal16]. The choice of the initial temperature does not change the evolution of the gas and star of the GMCs because of the short cooling time-scale compared to the dynamical time-scale of the clouds. For each initial condition, we perform five simulations with different momentum boosting factor, ${f_{\mathrm{boost}}}=$ 0.5, 1, 2, 4, 10, to test the effects of the strength of momentum feedback on the global evolution of GMCs and star clusters. GMCs are initially resolved by $128^3$ equal-mass gas cells, which sets the target mass $m_{\rm res}\approx{M_{\mathrm{GMC}}}/128^3$. We perform a convergence test for the RHO20T run with ${f_{\mathrm{boost}}}=2$ by varying the number of resolution elements from $64^3$ to $256^3$ and find that the star formation histories are not sensitive to mass resolutions. The final star formation efficiencies in runs with different resolutions vary by only a few percent.
\[fig:time-evolution\] {width="\columnwidth"}
Caveats of the sub-grid models
------------------------------
The sub-grid model used in this paper samples star particles probabilistically with star formation criteria described in . This is different from the more realistic sink particle approach that follows the accretion history of individual stars. We adopt this simplified star formation algorithm since the goal of this paper is to investigate the global properties of GMCs and star clusters but not to study the origin of the IMF or the detailed formation of single stellar objects. We interpret each star particle as a single stellar population, whose feedback intensity is estimated in an IMF-averaged fashion, see . For the most massive GMCs in our simulations, this IMF-averaged approach captures the overall energy budget of stellar feedback [see also @grudic_etal18]. However, it unavoidably underestimates the large variation of star formation efficiency in low-mass clouds, where a few massive stars can dominate the feedback process. In addition, the star particles in our simulations with a fixed gravitational softening length only trace the overall mass distribution of star clusters. Therefore, the detailed dynamical evolution could in principle depend on the choice of softening length [e.g. @bate_etal03; @bate12]. To fully capture the collisional process between star particles requires simulations that resolve the formation of individual stars over the whole mass spectrum and a more accurate gravity integrator, such as NBODY6 [@aarseth99]. Recent efforts have been made towards this direction [e.g. @parker_etal15; @gavagnin_etal17; @wall_etal19], however, due to the high computational cost, these simulations mainly focused on less massive GMCs and cannot explore the dynamical evolution of GMCs over a large parameter space.
Another main caveat is that we only take into account the momentum feedback from stellar winds, whose intensity is controlled by $f_{\rm boost}$. A larger $f_{\rm boost}$ is used to mimic stellar feedback from multiple feedback sources, such as stellar winds, ionizing radiation, and supernovae. However, in reality, different feedback mechanisms operate on different time-scales, at different locations, and through different physical processes. For example, ionizing radiation from massive stars strongly alters the ionizing state of the gas around massive stars and deposits both internal and kinetic energy to the surrounding medium. Moreover, for some of our simulations with very long $t_{\rm ff}$, for example SIGMA40, at the late stage of the cloud evolution when the age of the massive stars is longer than their main-sequence lifetime, core-collapse supernovae can deposit an enormous energy to the cloud and violently disrupt it. The combined effects of various feedback mechanisms will be investigated in an upcoming paper.
Integrated star formation efficiency (SFE) {#sec:sfe}
==========================================
In total, we have performed 80 GMC simulations with different masses, radii, velocity configurations, and feedback boosting factors. For all runs, we stop the hydrodynamical simulations when 99% of the gas mass is expelled from the initial spherical regions by momentum feedback. Although different GMCs show quantitative different star formation histories and final efficiencies, the general evolutionary stages of the clouds are very similar. Here we use the RHO20T run with ${f_{\mathrm{boost}}}=2$ as an example to describe the general pattern of GMC evolution.
shows the gas density projection of this run at four different epochs. The cloud evolution is initially governed by the turbulent velocity field which creates complex filamentary structures (upper left). After $\sim0.3{\tau_{\mathrm{ff}}}$, a roughly log-normal density distribution is established due to the supersonic turbulence. As turbulent energy dissipates by supersonic compression, the cloud starts to experience global contraction under self-gravity. Many subclusters are formed at the intersection of the filaments where dense gas clumps experience local runaway collapse (upper right). These subclusters move along the filaments, merge with each other frequently, and eventually form more massive subclusters. Due to momentum feedback, some gas mass is channelled outwards through low-density regions. In contrast, the high-density regions are compressed further and form young stars subsequently (lower left). When the central star cluster is massive enough so that its momentum feedback is able to counteract gravitational contraction, the majority of the gas mass is expelled from the cloud centre, causing the formation of giant wind-blowing bubbles (lower right).
Time evolution of cloud properties {#sec:sfe-evolution}
----------------------------------
In , we quantify the time evolution of various physical quantities of the cloud for the same run shown in . Panel (a) shows the star formation history of the cloud until it is fully disrupted. After the first group of stars forms at $\sim0.3{\tau_{\mathrm{ff}}}$, the SFR rises dramatically and peaks at around $t_{50}\sim{\tau_{\mathrm{ff}}}$. As momentum feedback from stellar particles clears some gas mass from the cloud centre, the SFR drops gradually. Although the whole star formation activity spans over $\sim2{\tau_{\mathrm{ff}}}$, the majority of the stellar mass is formed around the epoch of the star formation peak. The central 80% of the stellar mass is assembled within 0.6-1.4 ${\tau_{\mathrm{ff}}}$.
The cumulative version of panel (a), the stellar mass growth history, is shown in panel (b), together with the evolution of gas mass. We split all gas mass into bound and outflow components. The bound gas is defined as the total mass of gas cells with negative (kinetic+potential) energy, while the outflow is defined as the unbound gas cells that are outside twice the initial GMC radius. It is clear that the decrease of bound gas mass is partly due to gas consumption by star formation and partly due to the increase of feedback-driven outflows. The final stellar mass reaches $\sim60$% of the initial cloud mass, which is defined as the integrated SFE, ${\epsilon_{\mathrm{int}}}$.
Panel (c) shows the evolution of virial parameters for both gas ($\alpha_{\rm gas}$) and stars ($\alpha_{\rm star}$). By construction, gas is initially in virial equilibrium with $\alpha_0=1$. As the turbulent energy dissipates, the kinetic energy of the cloud decreases and the system collapses, which leads to a slight decrease of $\alpha_{\rm gas}$. The momentum feedback from stars adds kinetic energy to the gas cells and helps to increase $\alpha_{\rm gas}$ after $t\sim1.3{\tau_{\mathrm{ff}}}$. Eventually, the virial parameter comes back to unity. Yet momentum feedback cannot keep the cloud in a quasi-equilibrium state. $\alpha_{\rm gas}$ keeps rising and becomes much larger than unity very quickly until the majority of the gas mass is removed from the central region of the cloud. Interestingly, the virial parameter of stars, $\alpha_{\rm star}$, is always smaller than unity, suggesting that the model star cluster is sub-virial. As we will show later, this sub-virial dynamical state before gas expulsion has a dramatic effect on the formation of bound clusters in GMCs.
In panel (d), we present the evolution of the half-mass radius of the gas and stellar components of the GMC. The evolution of the half-mass radius of the gas tightly follows the evolution of its dynamical state as is described in panel (c). The size of the gas cloud first shrinks slightly due to the dissipation of the initial turbulent energy until stellar momentum feedback puffs it up. The evolution of the half-mass radius of the star cluster is more complicated. At first, stars are formed in dense gas clumps distributed over a large volume of the cloud, which leads to a relatively large initial stellar radius. As the cloud contracts, star formation activities concentrate more towards the central region and the stellar half-mass radius decreases until $t_{50}$. Later, as gas removal shallows the overall gravitational potential, the star cluster expands dynamically to reach a new equilibrium state.
Many theoretical works on gas expulsion and the formation of bound fraction suggest that, rather than ${\epsilon_{\mathrm{int}}}$, the local stellar fraction (LSF) is considered as an effective SFE to better probe the bound fraction of the cluster after gas expulsion [@goodwin09; @smith_etal11; @smith_etal13; @farias_etal18]. The LSF is defined as the mass fraction of stellar mass within the stellar half-mass radius right before the gas expulsion: $${\rm LSF}=\frac{M_*(<r_{\rm h})}{M_*(<r_{\rm h})+M_{\rm gas}(<r_{\rm h})}.$$ In panel (e), we show the evolution of the LSF in the simulations. Since neither the formation of stars nor the dispersion of gas happens instantaneously, the LSF changes dramatically during the course of GMC evolution. By definition, the LSF is initially zero. As star formation continues, the gas is gradually consumed and expelled from the central region, causing the increase of stellar mass and decrease of gas mass. Therefore, the LSF increases monotonically as a function of time until it reaches unity.
![Star formation histories of RHO20T runs with five different ${f_{\mathrm{boost}}}$: 0.5, 1, 2, 4, 10. The SFR is not affected by the choice of ${f_{\mathrm{boost}}}$ during the early phase of GMC evolution. During this phase, the SFR increases linearly with time, which is shown as the dashed line: ${\rm SFR}\propto t$. Clouds start to be disrupted by stellar feedback after the SFR reaches its peak. The epoch of the peaks of star formation as well as the final SFE decreases with increasing ${f_{\mathrm{boost}}}$.[]{data-label="fig:sfh_boost"}](figure/sfh_boost.pdf){width="\columnwidth"}
Effects of ${f_{\mathrm{boost}}}$ on the star formation history {#sec:sfe-fboost}
---------------------------------------------------------------
shows the star formation histories of the RHO20T run with different ${f_{\mathrm{boost}}}$. During the early stages, the total mass of young stars is so small that momentum feedback is not enough to affect the dynamical state of the cloud. Therefore, the effects of ${f_{\mathrm{boost}}}$ on the SFR is not visible and all lines overlap with each other until the SFR reaches its peak. During this period of time, the SFR presents a linear increase with time, ${\rm SFR}\propto t$. This linear time dependence is consistent with previous theoretical and numerical prediction of turbulent self-gravitating cloud with virial parameters close to unity [@lee_etal15; @murray_chang15; @murray_etal17].
When a sufficient fraction of gas mass is converted to stars, the momentum feedback is able to alter the overall dynamical state of the cloud (see ), and eventually reverts the increasing trend of the SFR. The exact epoch of the turning point and, in turn, the final ${\epsilon_{\mathrm{int}}}$, is determined by the balance between the feedback intensity and gravitational contraction, which will be discussed in the next section.
Surface density-dependent SFE {#sec:sfe-epsint}
-----------------------------
For all 80 runs, we obtain the total stellar mass at the end of the hydrodynamical simulations and calculate the integrated SFE, ${\epsilon_{\mathrm{int}}}$. shows ${\epsilon_{\mathrm{int}}}$ as a function of the cloud initial surface density, $\Sigma_0$, for all 80 runs. We find a positive correlation between ${\epsilon_{\mathrm{int}}}$ and $\Sigma_0$, for a given value of ${f_{\mathrm{boost}}}$. In contrast, we do not find clear correlations of ${\epsilon_{\mathrm{int}}}$ with either the initial mass, radius, or volume density of the clouds. For the same GMC, runs with larger ${f_{\mathrm{boost}}}$ produce less stars and smaller ${\epsilon_{\mathrm{int}}}$, consistent with the results described in . Moreover, we find that rotation-supported (“R”) runs in general show a slightly higher ${\epsilon_{\mathrm{int}}}$ than the corresponding turbulence-supported (“T”) runs, especially when a large ${f_{\mathrm{boost}}}$ is employed. This can be explained as follows. Because of the initial rotational velocity, GMCs in the “R” runs first collapse to a disk-like structure whose scale height is typically smaller than the cloud radius. The formation of the thin disk allows momentum feedback to escape easier than that of the “T” runs, where the spherical shape of cloud is roughly maintained. This geometric effect leads to a difference of ${\epsilon_{\mathrm{int}}}$ by about 10-20%.
We next build an analytical model to explain the correlation between ${\epsilon_{\mathrm{int}}}$ and $\Sigma_0$ by considering the force balance between gravitational contraction and gas expulsion by momentum feedback. We assume, when the balance is achieved, the residual gas forms a thin spherical shell with a radius $r_s$. For gravitational forces, we consider the contribution from both the central star cluster of mass $M_*$ and self-gravity of the gas shell of mass $M_{\rm sh}={M_{\mathrm{GMC}}}-M_*$. The gravitational force per unit area of the shell from the cluster is evaluated as $$F_{\rm sh,*}= \frac{GM_*M_{\rm sh}}{r_s^2 A_s}= \frac{GM_*\Sigma_{\rm sh}}{r_s^2},$$ while self-gravity of the gas shell is $$F_{\rm sh} = \frac{\beta GM_{\rm sh}^2}{r_s^2A_s} = \frac{\beta GM_{\rm sh}\Sigma_{\rm sh}}{r_s^2},$$ where $A_s=4\pi r_s^2$ is the surface area of the shell, $\Sigma_{\rm sh}=M_{\rm sh}/4\pi r_s^2$, and $\beta$ is the geometric factor that takes into account the anisotropic distribution of the gas shell. Note that $\Sigma_{\rm sh}$ is the surface density of the spherical shell seen from the central cluster, different from $\Sigma_0$, the cloud column density, by a factor of 4: $\Sigma_0=4\Sigma_{\rm sh}$. For a uniform density gas distribution, $\beta=0.5$. The expel force per unit area exerted onto the gas shell by momentum feedback is $$\label{eq:force-mom}
F_p = \frac{M_*\dot{p}}{4\pi r_s^2},$$ where $\dot{p}$ is the momentum deposition rate per unit stellar mass. As described in , we use an IMF-averaged wind injection as the default setup for feedback with a boosting factor, ${f_{\mathrm{boost}}}$. In , the deposition rate evolves as the stellar population ages. For simplicity, in this analytical model we assume a constant momentum deposition rate per unit mass $\dot{p}={f_{\mathrm{boost}}}\dot{p}_{\rm w}$ where $\dot{p}_{\rm w}$ is the IMF- and time-averaged value.
![Integrated SFE, ${\epsilon_{\mathrm{int}}}$, of all 80 runs as a function of initial gas surface densities, $\Sigma_0$. The solid points represent all 40 rotation-supported “R” runs while the crosses represent all 40 turbulent-supported “T” runs. Different colours represent runs using different ${f_{\mathrm{boost}}}$ (see legend for details). In general, we find that ${\epsilon_{\mathrm{int}}}$ increases with increasing $\Sigma_0$ and decreasing ${f_{\mathrm{boost}}}$. This trend is explained by a physical model that considers force balancing between gravitational collapse and momentum feedback. The solid lines represent the result of the physical model with best-fit parameters $\beta=1.83$ and $\dot{p}_{\rm w}=3.32\times10^{-9}{\rm cm/s^2}$ using .[]{data-label="fig:epsint"}](figure/epsint.pdf){width="\columnwidth"}
Force balancing between gravitational collapse and the momentum-driven wind, $F_{\rm sh, *}+F_{\rm sh}=F_p$, gives $$\frac{M_*\dot{p}}{4\pi r_s^2} = \frac{GM_*\Sigma_{\rm sh}}{r_s^2} + \frac{\beta GM_{\rm sh}\Sigma_{\rm sh}}{r_s^2}.$$ By defining variables $\Sigma_{\rm crit}\equiv\dot{p}/\pi G$ and $\Gamma\equiv\Sigma_{\rm sh}/\Sigma_{\rm crit}=\pi G \Sigma_0/4{f_{\mathrm{boost}}}\dot{p}_{\rm w}$ and assuming $r_s={R_{\mathrm{GMC}}}$, the above equation can be simplified to $$(1-{\epsilon_{\mathrm{int}}})\left(\frac{\beta}{{\epsilon_{\mathrm{int}}}}-\beta+1\right)=\frac{1}{\Gamma}.$$ Finally, ${\epsilon_{\mathrm{int}}}$ can be solved as $$\label{eq:epsint}
{\epsilon_{\mathrm{int}}}= \frac{\sqrt{\Gamma^2+(4\beta-2)\Gamma+1}-(2\beta-1)\Gamma-1}{2(1-\beta)\Gamma}.$$ For a uniform density gas shell, reduces to $${\epsilon_{\mathrm{int}}}= \frac{\sqrt{\Gamma^2+1}-1}{\Gamma}.$$
For clouds with high surface density, $\Gamma\gg1$, is reduced to ${\epsilon_{\mathrm{int}}}\approx1-1/\Gamma$, which suggests that almost all gas mass is converted into stars before the gravitational collapse is balanced by momentum feedback. Since the mass of the gas shell is much smaller than the mass of the central star cluster, self-gravity of the gas shell is negligible and therefore ${\epsilon_{\mathrm{int}}}$ is independent of $\beta$. For $\Gamma\ll1$, on the other hand, the gravitational force is dominated by the self-gravity of the shell and can be simplified to ${\epsilon_{\mathrm{int}}}\approx\beta\Gamma$, which shows a clear $\beta$ dependence. In this case, ${\epsilon_{\mathrm{int}}}$ depends linearly on the cloud surface density divided by the momentum deposition rate, $\Sigma_0/{f_{\mathrm{boost}}}\dot{p}_{\rm w}$.
We fit the value of ${\epsilon_{\mathrm{int}}}$ for all 80 GMC simulations using , and obtain the best-fit parameters with $1\sigma$ uncertainty: $\beta=1.83\pm0.89$ and $\dot{p}_{\rm w}=(3.32\pm0.64)\times10^{-9}{\rm cm/s^2}$. As can be seen in , the analytical model is in good agreement with the simulated ${\epsilon_{\mathrm{int}}}$ over a large range of $\Sigma_0$ and ${f_{\mathrm{boost}}}$. We notice that the analytical model overestimates ${\epsilon_{\mathrm{int}}}$ for runs with ${f_{\mathrm{boost}}}=10$. This is possibly because clouds are disrupted earlier in ${f_{\mathrm{boost}}}=10$ runs than other runs and the time-averaged $\dot{p}$ is systematically higher due to the decreasing wind kinetic luminosity used in the simulations, see .
Star formation time-scales {#sec:sfe-time-scale}
--------------------------
As described in and \[sec:sfe-epsint\], stronger momentum feedback changes the epoch of the peak of star formation to earlier times and reduces the final SFE. How important is the strength of feedback to the overall star formation time-scales? Can feedback be the main energy source to support the cloud and maintain a quasi-equilibrium state? We investigate these questions here by defining several relevant time-scales that characterize the star formation activities for the simulated GMCs. First, we define the initial free-fall time of the cloud as
$${\tau_{\mathrm{ff}}}= \sqrt{\frac{3\pi}{32G\rho_0}} \approx 1.6 {\rm Myr} \left(\frac{M}{10^5{\,M_{\odot}}}\right)^{-1/2} \left(\frac{{R_{\mathrm{GMC}}}}{10{\rm pc}}\right)^{3/2}.$$
where $\rho_0=3{M_{\mathrm{GMC}}}/4\pi{R_{\mathrm{GMC}}}^3$ is the initial volume density of the GMC. The free-fall times of all GMCs are listed in . We define the star formation duration as the time-scale during which the clouds form the central 80% of their stars: ${\tau_{\mathrm{dur}}}=t_{90} - t_{10}$. Following @li_etal18 [@grudic_etal18], we also define an age spread of star cluster as the ratio between the final stellar mass $M_*$ and the mass-weighted SFR: $$\tau_{\rm spread}\equiv\frac{M_*}{<\dot{M}>}=\frac{M_*}{\int \dot{M}_*^2dt/M_*},$$ where $\dot{M}_*$ is the instantaneous SFR. For a Gaussian-like star formation history with standard deviation $\sigma_*$, the age spread is approximated $\tau_{\rm spread}\approx2\sqrt{\pi}\sigma_*$.
\[fig:time-scales\] {width="1\columnwidth"}
In , we show the central star formation epoch ($t_{50}$), star formation duration ($\tau_{\rm dur}$), and age spread ($\tau_{\rm spread}$) as a function of initial free-fall time (${\tau_{\mathrm{ff}}}$) for all 80 GMCs. In the top panel, we find a clear linear correlation between $t_{50}$ and ${\tau_{\mathrm{ff}}}$. We know that $t_{50}$ roughly represents the epoch of the peak of star formation because of the Gaussian-like shape of the star formation history, see . In fact, $t_{50}$ is close to ${\tau_{\mathrm{ff}}}$ for clouds that are turbulence-supported (“T” runs). For “T” runs, the initial turbulent energy dissipates within the turbulence crossing time, which is shorter than the free-fall time of the cloud. The peak of star formation is determined by gravitational collapse of the whole cloud and is therefore similar to the free-fall time. The “R” runs, on the other hand, show a systematically larger $t_{50}$ than the corresponding “T” runs. This is because the rotation-supported cloud first collapses along the rotational axis and form a gaseous disk. The rotating disk contains more coherent motions whose kinetic energy dissipates over a longer time-scale than turbulent motions. Interestingly, as shown in the middle and bottom panels, the star formation duration and age spread for “T” and “R” runs do not show clear difference, which suggests that once the runaway collapse starts, the details of the initial configuration of the gas motion do not affect the subsequent star formation process.
Similar to $t_{50}$, ${\tau_{\mathrm{dur}}}$ and ${\tau_{\mathrm{spread}}}$ also correlate linearly with ${\tau_{\mathrm{ff}}}$. We perform a linear fit to the correlation between ${\tau_{\mathrm{ff}}}$ and ${\tau_{\mathrm{dur}}}$ and find that runs with different ${f_{\mathrm{boost}}}$ show similar scalings but with different normalizations. Although the normalization of the relations shows an anticorrelation to ${f_{\mathrm{boost}}}$, increasing ${f_{\mathrm{boost}}}$ by a factor of 20 from 0.5 to 10 only shortens the time-scale by a factor of 3 to 4. Quantitatively, this weak dependence of the momentum feedback intensity on the star formation duration can be understood as follows. In , we find a linearly increasing SFR from $t_{10}$ to $\sim t_{50}$ regardless the choice of ${f_{\mathrm{boost}}}$. Here we define ${\rm SFR}=A(t-t_{10})$, where $A$ is an arbitrary normalization. Therefore, the approximated total stellar mass is $M_{*, \rm{linear}}=5/2\int_{t_{10}}^{t_{50}}A(t-t_{10})dt=5/4A(t_{50}-t_{10})^2\sim5/16A{\tau_{\mathrm{dur}}}^2\propto {\tau_{\mathrm{dur}}}^2$, assuming $\tau_{\rm dur}=2(t_{50}-t_{10})$. We also obtained a correlation between ${\epsilon_{\mathrm{int}}}$ and $\Sigma_0$ from , which gives the final stellar mass as $M_*={\epsilon_{\mathrm{int}}}M_0$. For clouds with $\Sigma_0<<\Sigma_{\rm crit}$, the above expression can be simplified as $M_* \approx \beta\Gamma{M_{\mathrm{GMC}}}= \pi G \beta\Sigma_0 {M_{\mathrm{GMC}}}/ 8{f_{\mathrm{boost}}}\dot{p}_{\rm wind}\propto{f_{\mathrm{boost}}}^{-1}$. Equating $M_{*, \rm{linear}}$ and $M_*$ gives a scaling ${\tau_{\mathrm{dur}}}\propto{f_{\mathrm{boost}}}^{-1/2}$. For high surface density clouds when $\Sigma_0>>\Sigma_{\rm crit}$, ${\epsilon_{\mathrm{int}}}\approx1-1/\Gamma$ and therefore $\tau_{\rm dur}$ is almost independent of ${f_{\mathrm{boost}}}$. Indeed, we find that the correlation between $\tau_{\rm dur}$ and ${f_{\mathrm{boost}}}$ in the simulations scales between ${\tau_{\mathrm{dur}}}\propto{f_{\mathrm{boost}}}^{-1/2}$ and ${\tau_{\mathrm{dur}}}\propto{f_{\mathrm{boost}}}^0$, with a median power-law slope around -1/4. The weak dependence of star formation duration to the strength of momentum feedback suggests that the cluster formation time-scale is mainly determined by gravitational runaway collapse. Keep in mind that the turbulent velocity fields used in our simulations are initialized at the very beginning of the simulations. No subsequent turbulence driving is applied to feed in the kinetic energy after turbulence dissipation. Understanding how turbulent motions cascade from large-scale environments to the local star-forming regions and affect the long-term star formation activities requires simulations of GMCs in realistic galactic environments, which will be investigated in a future work.
Bound fraction of model clusters {#sec:fbound}
================================
As discussed in previous sections, momentum feedback from young stars disrupts star-forming regions, reduces the SFE, and shortens star formation time-scales. The gas expulsion and cloud disruption flatten the gravitational potential and inevitably leave some imprint on the dynamical state of the star clusters formed at the centre of the clouds. Previous works have studied extensively the effects of gas removal on the dynamical evolution of star clusters. A simple virial analysis shows that if more than 50% of the mass is instantaneously removed from a virialized system, the remaining mass will become gravitationally unbound [@hills80; @mathieu83]. This conclusion is based on several assumptions: (1) the system is initially in virial equilibrium; (2) the removed mass is initially well mixed with the residual mass; (3) the mass-loss time-scale is much shorter than the dynamical time-scale of the system. In realistic star-forming environments, all of the above assumptions are not strictly applicable. The star-forming regions are not necessarily in virial equilibrium [e.g. @offner_etal09]. Stars are not randomly distributed within the GMCs, but are formed hierarchically within the densest molecular cores at the intersections of the gas filaments [e.g. @smith_etal11; @smith_etal13; @farias_etal15; @lee_goodwin16; @farias_etal18]. The non-star-forming gas is expelled outward gradually rather than instantaneously [@geyer_burkert01; @smith_etal13] and is preferentially channelled through low-density holes and tunnels rather than being removed homogeneously. Here, we explore these factors in our simulations and investigate how gas expulsion affects the bound fraction of star clusters.
![Compilation of gas (blue) and stellar (yellow) density profiles for all 80 runs. The profiles are normalized to the central stellar density. Individual profiles are shown as background transparent lines, while the median and 25-75% interquartile range are shown as solid lines and shaded regions, respectively. The inset figure shows the distribution of the best-fit power-law slopes of the gas and stellar density profiles. The vertical dashed lines are the median of distribution of the slopes for gas and stellar profiles, respectively.[]{data-label="fig:profile"}](figure/profiles.pdf){width="\columnwidth"}
Stellar and gas distribution at $t_{50}$ {#sec:fbound-profile}
----------------------------------------
shows the density profiles for both gas and stars at $t_{50}$ for all 80 runs. The profiles are centred at the location of the deepest gravitational potential, which is usually the centre of the central star clusters. Both the gas and stellar profiles are normalized to the central stellar density of the corresponding run. We find that, for most of the runs, the stellar central density is systematically higher than that of the gas, suggesting that a large fraction of gas mass in the central region of the GMCs has already been converted to stars at $t_{50}$. As a result, the gravitational potential is dominated by stars rather than gas within $\sim0.1{R_{\mathrm{GMC}}}$. We fit both the gas and stellar density profiles with a power-law shape, $\rho(r)\propto r^{-\gamma}$, and show the distribution of the power-law slopes $\gamma$ in the inset of the figure. The gas density profiles shows a roughly isothermal profile with a median power-law slope $\gamma\sim1.9$. It is interesting that the gas distribution for all runs converges to quasi-isothermal profiles regardless of the fact that the initial conditions are uniform density spheres. This isothermal gas density profile is consistent with the observed radial profiles of star-forming molecular clumps [@mueller_etal02; @palau_etal14; @wyrowski_etal16; @csengeri_etal17] and is thought to be a natural consequence of scale-free gravitational collapse [e.g. @larson69; @penston69; @naranjo-romero_etal15; @donkov_stefanov18; @li18]. The stellar density profiles, on the other hand, are systematically steeper than that of the gas with slopes centred around 2.8 with a large variation. The steeper stellar density profiles imply more centrally concentrated star formation activities.
Virial state of star clusters at the peak of star formation {#sec:fbound-virial}
-----------------------------------------------------------
In panel (c) of , we showed the evolution of the virial parameter of stars in the RHO20T run with ${f_{\mathrm{boost}}}=2$. We found that the model star cluster is in a sub-virial state ($\alpha_{\rm vir,*}$<1) during the course of gas expulsion from $t_{50}$ to $t_{\rm exp}$. Here we calculate $\alpha_{\rm vir,*}$ at $t_{50}$ for all 80 GMCs in order to quantify the dynamical state of the star clusters before gas expulsion. We find that $\alpha_{\rm vir,*}$ has a median value around 0.61 with a 25-75% interquartile range 0.55-0.65, which suggests a systematic sub-virial dynamical state. This finding is qualitatively consistent with previous simulations, such as @offner_etal09, who suggest a sub-virial stellar velocity dispersion even in virialized GMCs. This is possibly because stars are preferentially formed in the densest molecular cores which on average have less velocity dispersion than the rest of the cloud. Moreover, as discussed in , the stellar distribution is much more compact than that of the gas, which also helps the central cluster to remain gravitationally bound. As will be shown later, the sub-virial state has a dramatic effect on the final boundness of star clusters after gas expulsion.
Gas expulsion time-scales vs dynamical time-scales {#sec:fbound-time-scale}
--------------------------------------------------
The dynamical response of star clusters to gas expulsion is a competition between the flattening of the gas potential that happened over gas expulsion time-scale and the energy exchange among stars that happened over the dynamical time-scale of the clusters. Previous $N$-body simulations mimic the gas expulsion process by gradually reducing the background gas potential over a given period of time [e.g. @geyer_burkert01; @baumgardt_kroupa07]. However, in realistic star-forming regions, star formation and gas expulsion happen at the same time and there is no clear separation between the two processes.
Here we define the gas expulsion time-scale as the duration from the peak of star formation to the epoch when the contribution of the gravitational potential energy from gas mass within twice the half-mass radius of the star cluster is less than 10%, $\tau_{\rm exp}=t_{\rm exp}-t_{50}$. We also modified the potential energy threshold from 1 to 10% and find the expulsion time-scale is not sensitive to the choice of this value. Similar to the star formation duration, we find that $t_{\rm exp}$ depends strongly on the initial free-fall time of the GMC, suggesting that gas expulsion associates well with the end of star formation. It also suggests that gas expulsion happens neither instantaneously nor much longer than the dynamical time-scale of the clouds.
![Evolution of Lagrangian radius of star clusters formed with GMCs during gas expulsion for RHO20T run with ${f_{\mathrm{boost}}}=2$. $R_{\rm lag}$ of different mass fractions are shown as lines with different colours. Lines with less transparency are for mass fractions at every 10%. $R_{\rm lag}$ for mass fraction that corresponds to the bound fraction estimated by the iterative method described in \[sec:fbound-methods\] is shown as thick dashed line.[]{data-label="fig:star-contour"}](figure/lagr_RHO20T_boost2.pdf){width="\columnwidth"}
Calculating the bound fraction of star clusters {#sec:fbound-methods}
-----------------------------------------------
The final efficiency of star formation and all relevant time-scales investigated above are determined once the majority of the gas mass is removed from the central region of the clouds. Right after gas removal, the dynamical state of star clusters will readjust according to the changes of gravitational potential for the next couple of dynamical time-scales. Therefore, after momentum feedback expels more than 99% of the gas mass out of twice the stellar half-mass radius, we stop the hydro runs, remove residual gas cells, and continue evolving the star clusters in a gravity-only mode with the same softening length of stellar particles as the corresponding hydro runs. The simulations keep running for another two free-fall times of the GMC, ${\tau_{\mathrm{ff}}}$. We analyze the cluster bound fraction from $N$-body snapshots at different epochs and find that the bound fraction usually becomes stable after only $\sim0.5{\tau_{\mathrm{ff}}}$.
We adopt two methods to estimate the bound fraction from the last snapshot of the $N$-body runs. The first and simplest way is to calculate the mass fraction of all stellar particles with negative energies. This method gives accurate results for clusters that have large bound fractions. However, for clusters with low bound fraction, simply summing up stellar particles with negative energy overestimates the bound fraction. Stars with negative energy are not guaranteed to be bound to the cluster since removing all stars with positive energy shallows the gravitational potential. Therefore, we design a new method that removes stellar particles with positive energies and updates gravitational energies for the remaining stars iteratively. The iteration stops when all remaining stars have negative energies and the bound mass fraction of the clusters, ${f_{\mathrm{bound}}}$, is obtained.
The second method is to use the “Lagrangian radius”, $R_{\rm lag}$, defined as a series of radii within which the star cluster contains a sequence of fractions of stellar mass. @brinkmann_etal17 suggest to use the evolution of the Lagrangian radii to determine the structural changes of the star cluster during gas expulsion. The bound fraction after gas expulsion is determined by the outermost Lagrange radius that shows a core collapse. Figure \[fig:star-contour\] shows the time evolution of the Lagrangian radii for the RHO20T run with ${f_{\mathrm{boost}}}=2$. We find that the Lagrangian radii of all mass fractions decrease during the first couple of ${\tau_{\mathrm{ff}}}$ due to the same reason of the decrease of the stellar half-mass radius as mentioned in . After the majority of stars are formed after ${\tau_{\mathrm{ff}}}$, $R_{\rm lag}$ starts to increase in response to gas removal and the decrease of the total gravitational potential. The $R_{\rm lag}$ of small mass fractions shows a turnover when the central component recollapses to form the central bound clusters. The evolution of $R_{\rm lag}$ for mass fraction the same as the bound fraction determined by the iterative method is highlighted in the same figure. We find that the evolution of this $R_{\rm lag}$ is indeed approximately the outermost Lagrange radius that shows a turnover. We have performed the same analysis for all 80 runs and find that the iterative and “Lagrangian radius” methods give consistent results, confirming the eligibility of both methods. Because the outermost Lagrangian radius with turnover is determined somewhat subjectively while the iterative method always converges to an accurate result, we will only report the bound fraction that is determined by iterative methods in later sections.
Bound fraction as a function of ${\epsilon_{\mathrm{int}}}$ {#sec:fbound-epsint}
-----------------------------------------------------------
![Bound fraction as a function of integrated SFE for all 80 GMCs. The colour and marker styles are the same as those used in . The solid line is the best-fit model with $f_{\rm sat}=0.95$ and $\alpha_*=0.48$ using . For reference, the relationship between ${\epsilon_{\mathrm{int}}}$ and ${f_{\mathrm{bound}}}$ derived from the semi-analytical model in @adams00 is shown by the dotted line. The sub-grid star cluster formation model used to estimate bound cluster mass in cosmological hydrodynamical simulations in @li_etal18 is shown by the dotted-dashed line.[]{data-label="fig:eps-fbound"}](figure/epsint_fbound.pdf){width="\columnwidth"}
shows a compilation of bound fractions for all 80 GMCs as a function of ${\epsilon_{\mathrm{int}}}$. The bound fractions are calculated from the last output of the $N$-body runs using the iterative method described in . We find that there exists a broad range of ${f_{\mathrm{bound}}}$ from almost completely bound to almost completely disruptive. There is an increasing trend of ${f_{\mathrm{bound}}}$ as a function of ${\epsilon_{\mathrm{int}}}$. Interestingly, several runs with ${\epsilon_{\mathrm{int}}}<0.5$ show large ${f_{\mathrm{bound}}}$, deviating from the simple virial analysis. For some runs with ${\epsilon_{\mathrm{int}}}\sim0.2$, they still form star clusters with large ${f_{\mathrm{bound}}}>0.5$. The emergence of bound clusters in low-SFE clouds is actually in line with the findings in previous hydrodynamic simulations of GMCs [e.g. @parker_etal15; @gavagnin_etal17; @farias_etal18], although the detailed relationship between ${\epsilon_{\mathrm{int}}}$ and ${f_{\mathrm{bound}}}$ shows subtle differences.
Here, we present a simple one-zone cluster model to explain the relationship between ${\epsilon_{\mathrm{int}}}$ and ${f_{\mathrm{bound}}}$. Assuming stars in clusters always follow a Maxwellian velocity distribution $$f(v)dv \equiv f(x)dx = \sqrt{\frac{2}{\pi}} x^2\exp{(-x^2/2)}dx,$$ where $x\equiv\sqrt{3}v/{v_{\mathrm{rms}}}$, ${v_{\mathrm{rms}}}=\sqrt{3kT/m}$, and $m$ and $T$ are the average mass and “temperature” of the cluster. Before gas expulsion, a cluster with mass $M_*$ and radius $r_*$ has a virial parameter $\alpha_*=-2T_{*,0}/\Omega_{*,0}$, where the kinetic energy of stars $T_{*,0}=M_*{v_{\mathrm{rms}}}^2/2$ and $\Omega_{*,0}=-G{M_{\mathrm{GMC}}}M_*/r_*$. Instantaneous gas expulsion flattens the gravitational potential but does not change the kinetic energy of the stars. Therefore, the potential energy of stars after gas expulsion drops according to the SFE, $\Omega_{*,1}=-GM_*M_*/r_*={\epsilon_{\mathrm{int}}}\Omega_{*,0}$, while the kinetic energy does not change, $T_{*,1}=T_{*,0}=M_*{v_{\mathrm{rms}}}^2/2$. The escape velocity of the cluster after gas expulsion is $$v_{\rm esc} = \sqrt{-\frac{2\Omega_{*,1}}{M_*}}=\sqrt{\frac{2{\epsilon_{\mathrm{int}}}}{\alpha_*}}{v_{\mathrm{rms}}}.$$
Assuming stars keep their Maxwellian distribution after gas expulsion, the bound fraction ${f_{\mathrm{bound}}}$ can be estimated as the fraction of stars that have velocities below $v_{\rm esc}$: $$\begin{aligned}
\label{eq:fbound-epsint}
{f_{\mathrm{bound}}}&=& f_{\rm sat}\int_0^{v_{\rm esc}}f(v)dv \nonumber \\
&=&f_{\rm sat}\int_0^{\sqrt{3}v_{\rm esc}/{v_{\mathrm{rms}}}}f(x)dx \\
&=&\left[{\rm erf}\left(\sqrt{\frac{3{\epsilon_{\mathrm{int}}}}{\alpha_*}}\right)-\sqrt{\frac{12{\epsilon_{\mathrm{int}}}}{\pi\alpha_*}}\exp{\left(-\frac{3{\epsilon_{\mathrm{int}}}}{\alpha_*}\right)}\right]f_{\rm sat},\nonumber\end{aligned}$$ where $f_{\rm sat}$ is the saturation of bound fractions. This saturation is probably due to some small fraction of substructures that are formed with low local star formation efficiencies and do not merge into the central cluster.
We fit the simulated ${f_{\mathrm{bound}}}$ using the above relationship with two parameters, $\alpha_*$ and $f_{\rm sat}$. The best-fit values and their corresponding 1-$\sigma$ confidence intervals are $\alpha_*=0.48\pm0.02$ and $f_{\rm sat}=0.94\pm0.03$. The best-fit $\alpha_*$ is consistent with the measured virial parameters at $t_{50}$ (see ), suggesting that the sub-virial dynamical state of model clusters is the main driver to prevent clusters from being dissociated by gas expulsion.
{width="49.00000%"} {width="49.00000%"}\
{width="49.00000%"} {width="49.00000%"}
We compare our best-fit relation with the semi-analytical model developed by @adams00, who evaluates the dynamical response of star clusters to instantaneous gas expulsion based on an equilibrium cluster model with different stellar and gas density profiles. The analytical fit to his isotropic velocity distribution model gives ${f_{\mathrm{bound}}}=(2\tilde{\epsilon}-\tilde{\epsilon}^2)^{2/3}$, where $\tilde{\epsilon}\equiv(10{\epsilon_{\mathrm{int}}}-1)/9$. We find that this model is roughly consistent with our simulation result but slightly underestimates ${f_{\mathrm{bound}}}$ for ${\epsilon_{\mathrm{int}}}<0.5$. Moreover, we also show the ${\epsilon_{\mathrm{int}}}$-${f_{\mathrm{bound}}}$ relation used in our previous cosmological hydrodynamic simulations of a Milky Way-sized galaxy. In these simulations, a continuous cluster formation prescription is used to model the formation of individual star clusters from dense clumps resolved by parsec-scale resolutions [@li_etal17]. The adopted ${\epsilon_{\mathrm{int}}}$-${f_{\mathrm{bound}}}$ relation, ${f_{\mathrm{bound}}}={\rm min}(2{\epsilon_{\mathrm{int}}},1)$, determines the final bound mass of model clusters and in turn affects the cluster initial mass function, cluster formation efficiency, and the properties of evolved cluster populations [@li_etal18; @li_gnedin18]. The piecewise function used in those cosmological simulations is in general agreement with the relationship found here.
Note that the bound fractions obtained above take into account the total bound stellar mass from not only the central star cluster but also all other surrounding substructures, which are not necessarily bound to the central cluster. In the following section, we will identify individual subclusters in the simulations, estimate the bound mass of central star clusters, and quantify its relationship to the total bound mass.
Properties of substructures {#sec:fbound-sub}
---------------------------
Previous studies on the bound fraction of star clusters after gas expulsion usually assume an initial spherical gas and stellar distribution. However, recent observational and theoretical works suggest a hierarchical star formation scenario due to the turbulent nature of GMCs [e.g. @elmegreen_elmegreen01; @bonnell_etal03; @allen_etal07; @bate09; @gutermuth_etal09; @bressert_etal10; @girichidis_etal11; @maury_etal11]. In , we show the stellar particle distributions in the $x$-$y$ plane for the RHO20T run with ${f_{\mathrm{boost}}}=2$ at four different epochs. At $t={\tau_{\mathrm{ff}}}$, the stellar distribution follows well with the gas distribution and subclusters are distributed along the filamentary structures. At $t=2{\tau_{\mathrm{ff}}}$ when the majority of the gas mass has already been pushed out of the central region, we find that some subclusters spiral into the centre of the GMCs due to gravitational attraction and, at the same time, merge with each other, and form more massive subclusters. Some of the most massive subclusters show a non-spherical shape because of recent mergers. At $t=$ 3-4 ${\tau_{\mathrm{ff}}}$, dynamical evolution after gas expulsion and violent relaxation during mergers erase the memory of the turbulent configurations of the gas cloud and help circularize the central star clusters. Some small substructures with high bulk velocities escape from the central region and never return to the central cluster.
To fully analyze the behavior of subclusters and quantify the central cluster properties, we identify bound substructures in the $N$-body simulations by adopting the SUBFIND algorithm [@springel_etal01]. The star particles are first linked into friends-of-friends (FOF) groups with separation less than 0.17 of the mean particle separation. For each FOF group, the SUBFIND algorithm is applied to identify all bound subclusters. We report bound substructures that are resolved by at least 200 stellar particles and assign the most massive bound subcluster as the central cluster. In most cases, the central cluster sits very close to the centre of the GMC and is surrounded by other less massive subclusters.
![Mass fraction of central cluster, $f_{\rm central}$, as a function of bound fraction, ${f_{\mathrm{bound}}}$, for all 80 simulated GMCs. $f_{\rm central}$ is defined as the ratio of the mass of the most massive subcluster to the total stellar mass in the GMC. The colour and marker styles are the same as those used in . The solid and dashed lines show the $1:1$ and $2:1$ ratio between ${f_{\mathrm{bound}}}$ and $f_{\rm central}$.[]{data-label="fig:fbound-sub"}](figure/central_cluster.pdf){width="\columnwidth"}
We find that the number of subclusters in the “R” runs is systematically larger than that in the corresponding “T” runs, although the total bound stellar masses are similar. In , we examine the relationship between the bound fraction, $f_{\rm bound}$, and the ratio of the mass of the central cluster to the total stellar mass, $f_{\rm central}$. In most cases, the central clusters dominate the total bound mass of the system. We also find that $f_{\rm central}$ is systematically higher for “T” runs than for “R” runs. Statistically, more than 50% of the bound stellar mass is contributed by the central clusters for “T” runs. This fraction is smaller for “R” runs but it exhibits a large scatter. As we discussed in the previous sections, GMCs in the “R” runs first collapse along the z-axis and form a gas disk. Therefore, many subclusters, that are formed in the dense clumps within the disk, obtain a similar bulk rotational velocity and are less likely to merge with each other than those formed in the “T” runs. The large number of subclusters in the “R” runs also implies that the stellar mass is distributed more broadly across different subclusters. Therefore, the central clusters in the “R” runs are less massive than those in the corresponding “T” runs. The exact hierarchical structure of the subclusters in the “R” runs depends strongly on the initial setup of the turbulent velocity fields, making the prediction of the mass of central star clusters less promising. This difficulty is reflected in the very large scatter of $f_{\rm central}$ for clouds with ${f_{\mathrm{bound}}}>0.8$.
It should be noted that recent observations of cluster formation efficiency, defined as the fraction of stars formed in bound clusters, identify centrally-concentrated clusters of spherical shape as bound clusters [e.g. @adamo_etal15]. This cluster selection method suggests that only the central clusters (and maybe other most massive subclusters) formed in GMCs are identified as bound star clusters when estimating cluster formation efficiencies from given galaxy patches. When interpreting the relevant observational correlations, such as the positive correlation between the SFR surface density and the cluster formation efficiency, through physics models [@kruijssen_etal12; @li_etal18; @li_gnedin18], the effect of substructures that are not bound to the central star clusters needs to be considered and taken with caution.
Summary {#sec:summary}
=======
We have performed a suite of three-dimensional hydrodynamic simulations of turbulent GMCs using the moving-mesh code <span style="font-variant:small-caps;">Arepo</span> with self-gravity, explicit cooling, star formation and momentum stellar feedback. We survey a large range of GMC masses and radii, and investigate the physical origin of the large variation of intrinsic SFE, ${\epsilon_{\mathrm{int}}}$. After the gas clouds are fully disrupted by stellar feedback, we follow the subsequent dynamical evolution of star clusters formed within the GMCs with $N$-body simulations. Below, we summarize our key conclusions:
- All simulated GMCs follow an initial linear increasing SFR before stellar momentum feedback disperses the whole cloud. The accelerating star formation activity leads to a superlinear stellar mass growth with time, $M_*\propto t^2$. This superlinearity is consistent with previous theoretical expectations of the gravitational runaway collapse of turbulent clouds.
- Momentum feedback from stellar particles adds kinetic energy to their ambient gas cells, inflates the virial parameters and radius of the clouds, drives outflows through low-density regions, and finally creates a large cavity at the centre of the clouds. The peak epoch and final efficiencies of star formation decrease with increasing strength of momentum feedback.
- ${\epsilon_{\mathrm{int}}}$ does not depend on the initial mass or radii of the clouds, but depends strongly on initial cloud surface density. This dependence is successfully explained by an analytical model that considers force balancing between gravitational collapse and momentum output from stellar particles. The model predicts ${\epsilon_{\mathrm{int}}}\approx1-\Sigma_{\rm crit}/\Sigma_0$ for clouds with high surface density while ${\epsilon_{\mathrm{int}}}\propto\Sigma_0/{f_{\mathrm{boost}}}\dot{p}_{\rm w}$ for clouds with low surface density.
- The duration of star formation in simulated GMCs is close to the initial free-fall time of the clouds, suggesting that the cluster formation time-scale is mainly determined by gravitational runaway collapse. The duration decreases with increasing feedback intensity, although the dependence is weak: $\tau_{\rm dur}\propto {f_{\mathrm{boost}}}^{-1/4}$.
- The model star clusters are assembled hierarchically. Subclusters are formed at the many density peaks across the GMCs controlled by the initial turbulence configuration. The subclusters move along the filamentary structures and merge with each other frequently. About 50% of the mass of subclusters is merged to form the most massive central clusters, but there are always a small fraction of subclusters that are unbound to the system and fly apart from the central clusters.
- The gas density distribution rearranges from an initial uniform density sphere to an isothermal profile with a power-law slope $\gamma\sim2$. At the peak of star formation, the stellar density profiles are systematically steeper than that of the gas with a power-law slope $\gamma\sim2.8$. The steeper stellar profiles suggest a fast conversion of gas to stars and gas expulsion by stellar feedback at the centre of GMCs.
- The model star clusters are always in a sub-virial state with a gradually increasing virial parameter as star formation continues. Interestingly, right before gas expulsion, the virial parameters of all simulated star clusters show a consistent value around $\alpha_{\rm vir,*}\approx0.6$.
- Due to the steep density profiles and sub-virial dynamical state of model clusters, clouds with low ${\epsilon_{\mathrm{int}}}$ (0.2-0.4) are still able to form clusters with relatively high bound fractions (0.3-0.8). The bound fraction of model clusters, ${f_{\mathrm{bound}}}$, is a continuously increasing function of the integrated SFE, ${\epsilon_{\mathrm{int}}}$. This relation is explained by a physical model that takes into account the mass fraction of stars with velocities below the escape velocity of a sub-virial system that obeys Maxwellian velocity distribution. The best-fit virial parameter of this model is around 0.5, consistent with the values obtained directly from the simulations.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous referee for detailed comments and suggestions. We thank Volker Springel for giving us access to <span style="font-variant:small-caps;">Arepo</span>. We are grateful to Peter Behroozi, Andreas Burkert, John Forbes, Nick Gnedin, Mike Grudic, Lee Hartmann, Jens Kauffmann, and Vadim Semenov for insightful comments and suggestions. MV acknowledges support through an MIT RSC award, a Kavli Research Investment Fund, NASA ATP grant NNX17AG29G, and NSF grants AST-1814053 and AST-1814259. FM is supported by the Program “Rita Levi Montalcini” of the Italian MIUR. OG acknowledges supported through NSF grant 1412144. The simulations of this work were run on the Harvard Odyssey clusters and the Comet HPC resource at San Diego Supercomputer Center as part of XSEDE through TG-AST170042 and TG-AST180025.
Tests of wind-blowing bubbles {#sec:appendix-wind}
=============================
We test the momentum deposition algorithm used in this paper by performing idealized simulations of wind-blowing bubbles. In this test, a stellar particle is located at the centre of a uniform density box with size 20 pc and total gas mass 320 ${\,M_{\odot}}$. The gas mass is initially resolved by $128^3$ gas cells. The central star deposits its wind material with a constant mass-loss rate $\dot{m}_{\rm w}=10^{-5}{\,M_{\odot}}/$yr at a fixed velocity $v_{\rm w}=500$ km/s. No self-gravity or cooling is used in this test in order to compare to analytical solutions of the evolution and internal structure of the wind-blowing bubble derived by @weaver_etal77.
The bubble first experiences free expansion with constant wind velocity until the mass of the swept-up material, $\frac{4}{3}\pi(v_{\rm w}t)^3\rho_0$, is comparable to the mass of the wind ejecta, $\dot{m}_{\rm w}t$. The initial free expansion phase only lasts for several hundred years and is followed by an adiabatic expansion phase. The time evolution of the bubble is $R_{\rm bubble}\approx0.88(\dot{m}_{\rm v}v_{\rm w}^2/2\rho_0)^{1/5}$, see Equation (5) in [@weaver_etal77]. shows the time evolution of the wind-blowing bubble from the numerical test. The edge of the bubble is identified as the densest gas shell surrounding the stellar particles. The radius of the bubble is calculated as the mass-weighted shell radius. In addition to the time evolution of the gas shell, we also examine the internal structure of the bubble. shows the density, velocity, and pressure profiles of the shocked interstellar gas at $t=0.2$ Myr. All profiles are normalized to the values at the outer shock of radius $R_2$. The analytical solution of the self-similar adiabatic flow within the shocked medium is obtained by numerically solving Eq. (6)-(8) in @weaver_etal77. We confirm that the momentum deposition algorithm used in our simulations reproduces the analytical time evolution and internal structure of the wind-blowing bubble with high precision.
![Time evolution of the radius of wind-blowing bubble from the test simulation (red dots). The analytical solution derived from @weaver_etal77 is overplotted as blue solid line for comparison.[]{data-label="fig:wind-radius"}](figure/wind_bubble_radius_evolve.pdf){width="1\columnwidth"}
![Velocity (black), density (red) and pressure (blue) profiles within the wind-blowing bubble. All physical quantities, $v, \rho, P$, are normalized to the values at the shock front for convenience. The x-axis is normalized to the outer shock radius $R_2$. The simulation result is shown as dotted lines, while the analytical expression from @weaver_etal77 is shown in solid lines for comparison. []{data-label="fig:wind-profile"}](figure/wind_bubble_profile.pdf){width="1\columnwidth"}
Convergence to momentum deposition methods {#sec:appendix-wind-gmc}
==========================================
In , we describe various algorithms for wind deposition to the ambient medium. Here, we test how different algorithms affect the star formation activities of the GMCs. The test runs use an identical simulation setup as the main GMC simulations described in except wind feedback. We use three different weights to deposit momentum flux to neighbouring cells: volume, mass, and solid angle from star particles. We also test an alternative algorithm that injects stellar winds in the form of pure thermal energy, rather than momentum.
In the upper panel of , we show the star formation histories of the GMCs using different wind deposition algorithms from the same initial condition RHO20T with ${f_{\mathrm{boost}}}=2$. We find that wind feedback in pure energy form is not able to disrupt the cloud and the star formation history is almost the same as that of the run without wind feedback. The failure of this energy deposition algorithm seems inconsistent with the result found in @rogers_pittard13, who used thermal energy injection to simulate wind feedback and obtained significant gas outflows from the central turbulent cloud. Note that, in their simulations, self-gravity is not included and the turbulent cloud never collapses to higher density. The maximum number density reached in their simulations is about $\sim10^4{\rm cm}^{-3}$. In contrast, our simulations track gravitational runaway collapse of dense clumps until they are converted to stars. In fact, our simulations always form much higher density clumps, $>10^{10} {\rm cm}^{-3}$. The thermal energies deposited into these dense environments suffer significant radiative cooling and therefore make the energy deposition inefficient. To capture the correct thermal dynamics of the adiabatic phase of the wind-blowing bubbles, the cooling radius needs to be resolved: $R_{\rm cool}\propto n^{-3/7}l_{\rm w}^{2/7}$ [@cioffi_etal88; @thornton_etal98]. Since the highest density in our simulations is about six orders of magnitudes higher than that in @rogers_pittard13, our simulations require about three orders of magnitude finer spatial resolution than in @rogers_pittard13 to resolve the wind energy feedback, which is computationally prohibitive. We conclude that under the current simulation setup, wind feedback through thermal energy deposition is not appropriate. Alternatively, wind feedback through momentum deposition can efficiently shut off star formation activities within 2 ${\tau_{\mathrm{ff}}}$. We find that using volume- or solid angle-weighted schemes for momentum deposition leads to faster cloud disruption than using a mass-weighted scheme. The reason is that the majority of the momentum is deposited to the densest clumps in a mass-weighted scheme and is therefore not able to channel gas out through low-density regions [see also @smith_etal18]. In reality, wind momentum should be deposited isotropically around star particles. It is more physically plausible to use volume and solid angle as weights.
![**Upper panel**: SFH of GMCs with different wind feedback mechanisms: without feedback (blue); energy feedback (purple); and momentum feedback with weights of solid angle (orange), cell volume (green), and cell mass (red). **Lower panel**: The SFH of GMCs with different mass resolutions (lower). The blue, orange, and green lines show the simulations with the number of resolution elements $64^3$, $128^3$, and $256^3$, respectively.[]{data-label="fig:sfh-wind"}](figure/numerics.pdf){width="1\columnwidth"}
Convergence to numerical resolutions {#sec:appendix-res}
====================================
Hydrodynamic simulations discretize continuous space into a finite number of resolution elements. The choice of optimal numerical resolution is to achieve a balance between scientific accuracy and computational costs. To study the convergence of the simulation outcomes to numerical resolutions, we perform GMC simulations of RHO20T runs with ${f_{\mathrm{boost}}}=2$ with different numbers of initial gas elements, $64^3$, $128^3$, and $256^3$, corresponding to target cell masses around $0.191$, $2.38\times10^{-2}$, and $2.98\times10^{-3}{\,M_{\odot}}$. The test runs are performed following the same physics as the production runs in the main text, see .
We find that the integral star formation efficiencies for the $64^3$, $128^3$, and $256^3$ runs are ${\epsilon_{\mathrm{int}}}=$ 0.612, 0.610, and 0.615, respectively. This consistency suggests that ${\epsilon_{\mathrm{int}}}$ is not sensitive to mass resolutions but is solely controlled by the force balance between gravitational collapse and gas expulsion by momentum stellar feedback. As shown in the lower panel of , the star formation histories of the three runs are also in general agreement with each other. The star formation first rises dramatically and peaks at around the free-fall time of the cloud. The only noticeable difference comes from the $64^3$ run. This run shows a narrower star formation history than the other two runs, suggesting delayed star formation at the beginning and an earlier gas removal process after a majority of the stellar mass is formed. The $128^3$ and $256^3$ runs present a more consistent star formation history across the course of GMC evolution. Therefore, for all production runs present in the main text, we choose $128^3$ as the default number of resolution elements.
\[lastpage\]
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'SSRN\_main.bib'
---
Introduction {#sec:intro}
============
In recent years, there has been a proliferation of “gig economy” online marketplaces providing temporary workers for jobs ranging from cleaning and dog-walking to tutoring and volunteering. At the center of each marketplace is a platform via which workers and jobs repeatedly form matches. Many of these platforms are decentralized, and they rely on the workers and job-posters to find compatible matches. However, user engagement and growth frequently depend on whether past interactions with the platform resulted in successful matches. Hence many gig economy platforms aim to increase user engagement by improving match efficiency.
One natural tool for improving match efficiency is to encourage workers to “adopt” compatible jobs. In an adopted match, a worker commits to repeatedly completing the same job in each period. As there is randomness in job compatibility, this improves match efficiency by reducing uncertainty about whether jobs can be matched. Adoption can also lead to an increase in engagement from both workers and jobs in the adopted matches. However, increasing adoption does not guarantee an increase in long-term user engagement and growth. High adoption levels deplete matching options for the remaining workers and jobs, which can reduce their future engagement, leading to a fundamental trade-off between options and adoption.
We find evidence of this trade-off in the operations of a nonprofit platform that matches volunteer workers to jobs in markets across the U.S. On this platform, [jobs]{} repeat weekly, and [workers]{} can either adopt jobs on a weekly recurring basis or sign up for jobs as a non-adopter on a one-off basis.
false
Our hypothesis is supported by our empirical analysis, which shows that a marginal decrease in the number of available [jobs]{} increases the dropout probability of non-adopters by around 1.3%.[^1] On the other hand, the number of available options does not impact the dropout rate of adopters.
[0.48]{} ![The evolution of the number of [jobs]{} on a nonprofit platform in two different locations over the course of 30 weeks. We hypothesize that the lack of options in location (a) between weeks 1 and 15 leads to more incomplete jobs when the number of total jobs increases. In contrast, because there are consistently more options in location (b), job growth can occur without a corresponding increase in incomplete jobs.[]{data-label="fig:FRUSlocations"}](loc_1_visual_v2.pdf "fig:"){width="\textwidth"}
[0.48]{}
To optimally resolve the trade-off between options and adoption, we first develop a model for a repeated matching market with random compatibility that captures the heterogeneity in workers’ future engagement based on their match type. Workers and jobs arrive in discrete periods, and are compatible with users on the other side of the market independently with probability. Each match is one of two types: User engagement and growth in each period is endogenously governed by the volume of matches in the previous period. Motivated by evidence from the volunteer matching platform, in our model worker engagement also depends on whether the previous-period match was an adoption or a non-adoption, whereas job growth depends on the volume but not the types of matches formed in the previous period.[^2] In a finite horizon setting, we use this framework to optimize the level of adoption in each period with the goal of maximizing the total discounted number of completed jobs.
A key ingredient that enables us to study the trade-off between options and adoption is our development of a new matching function that approximates greedy matching in two-sided random markets. We prove that the random matching process converges to a deterministic differential equation which admits a simple closed-form solution (see Proposition \[prop:matching\]).
We show that the optimal policy has an “all-or-nothing” form: either (1) full adoption is optimal, when worker growth due to adoption is larger than worker growth from non-adoption or when the number of jobs is small; or (2) no adoption is optimal (Theorem \[thm:myopic\]). Our proof relies on showing that the volume of matches in the next period is quasiconvex in the adoption level.
More generally, our analysis provides insight as to how to use commitment levels to influence the growth of a two-sided matching platform.
Our work is motivated by a collaboration with a nonprofit platform that operates a volunteer labor market. Non-monetary tools for influencing match efficiency, such as adoption, are especially critical in volunteer labor markets, as financial resources are limited and volunteers respond negatively to underutilization.
Related Literature
------------------
This paper contributes to the growing literature on designing dynamic matching platforms. A number of works study the trade-off between current matches and future match efficiency, such as in ridesharing [@ozkan2017dynamic; @ma2018spatio], kidney exchange [@akbarpour2017thickness; @anderson2017efficient; @ashlagi2018maximizing; @ashlagi2019matching], and other markets [@loertscher2016optimal; @baccara2018optimal]. We consider a similar trade-off in a setting where current matches affect future matches through user growth. Previous literature finds that limiting information and/or available actions on two-sided platforms can improve welfare, primarily through reducing search frictions [@halaburda2017competing; @kanoria2017facilitating; @arnosti2018managing]. Here, we focus on the lever of *commitment* in a repeated matching market and find that providing restrictions on different match types can increase the long-run number of matches. Most of the previous work on dynamic matching focuses on improving the matching of heterogeneous agents and items, e.g. in housing allocation [@bloch2012optimal; @kurino2014house; @arnosti2017design; @leshno2017dynamic], kidney exchange [@zenios2002optimal; @su2005patient; @dickerson2012dynamic], and other application areas [@feigenbaum2017dynamic; @hu2018dynamic; @chen2019pricing]. We consider a repeated setting with ex ante homogeneous jobs and workers, and we focus on how commitment to matches affects match volume through user engagement and growth.
Within the literature on two-sided marketplace design, the most closely related paper is [@lian2019optimal], which provides a theoretical framework for optimal growth in a two-sided revenue-maximizing marketplace. Their theory relies on a Cobb-Douglas matching function, as well as price levers for both sides of the market; this formulation allows them to characterize the optimal market balance and pricing policy to promote growth and maximize expected discounted profit. Similarly to [@lian2019optimal], we provide a theoretical formulation where user growth depends endogenously on the volume of past matches. However the structure of our optimal policies differ substantially, as we consider a setting without prices where the only lever is the type of matching. In order to capture uncertainty in matching and study the effects of adopted matches, we develop a novel framework using a different matching function.
Our study also contributes to the literature on improving labor management. In the nonprofit space, [@ata2016dynamic] consider the problem of volunteer management in gleaning operations, and use dynamic control techniques to provide optimal staffing policies for maximizing the volume of food gleaned in the presence of food and labor uncertainties. Similarly [@green2013nursevendor] and [@dong2017managing] study how to maximize the volume of completed jobs in call centers and hospitals when managerial decisions affect labor participation rates. There is also a growing body of work on how volunteer management should differ from traditional labor management given the unique characteristics of volunteer labor and retention (see e.g. [@locke2003hold] and [@sampson2006optimization] for overviews of these areas). Our setting is partially motivated by the volunteer labor market, and we similarly consider how to manage volunteers to improve operational and allocative efficiency.
Model Description {#sec:model}
=================
In this section, we formally introduce our model of a repeated matching market as well as the design problem that the platform faces when attempting to optimally use [workers]{} to complete [jobs]{} over a finite time horizon. In Section \[subsec:dynamics\], we describe the dynamics of our model and show that, properly scaled, our system converges to a deterministic dynamic. Then, in Section \[subsec:opt\] we formalize the platform’s design problem.
Model Dynamics {#subsec:dynamics}
--------------
The market starts at an initial state and evolves over $T$ periods. At the beginning of period $t \in [T]$, there are $D_t$ [jobs]{} that require one [worker]{}, and on the other side of the market, there are $V_t$ [workers]{} who are willing to complete one job.[^3] Recall that the platform allows for *adoption*, which implies that some [jobs]{} and [workers]{} present at time $t$ have already been matched. In particular, $K_t$ pairs of [workers]{} and [jobs]{} begin period $t$ as part of an adopted match, where $K_t \leq \min \{V_t, D_t \}$.
[**Matching Process:**]{} We now describe how [workers]{} and [jobs]{} match on the platform. In a given period, all [workers]{} who are not already committed to a [job]{} via adoption would like to match with one [job]{}; however, [workers]{} are not always compatible with the time or location of a particular [job]{}. We assume that [jobs]{} and [workers]{} are compatible with some identical probability, independently across pairs and time periods. During each period, some [workers]{} and [jobs]{} are already part of an adopted match. All others become part of that period’s *spot market*, where [workers]{} match with [jobs]{} in a decentralized manner. We model the matching process as sequentially greedy. Specifically, (1) [workers]{} arrive to the platform in a random order, (2) upon arrival, a [worker]{} checks the available [jobs]{} for compatibility, and (3) the [worker]{} matches if she finds a compatible [job]{} (breaking ties at random). At the conclusion of the matching process in period $t$, a total of $M_t$ pairs of [workers]{} and [jobs]{} have been matched, either from prior fixed matches or from the spot market in period $t$, where $M_t \leq \min \{V_t, D_t \}$.
[**Platform’s Match Type Decision:**]{} Even though the platform has no control over the decentralized matching process, it controls the *types* of the matches formed. In particular, the platform determines the fraction of matches in period $t$ which will be adopted going forward. We denote this fraction by $z_t$. We make the simplifying assumption that the platform can vary $z_t$ between $0$ and $1$, meaning that the platform can terminate previous fixed matches (e.g. by making the job generally available) or can require that all new matches are adoptions.[^4]
[**Job Dynamics:**]{} As discussed in Section \[sec:intro\], we consider settings where [jobs]{} recur, so the number of [jobs]{} in the market in one period is closely related to the number of [jobs]{} in the subsequent period. If a [job]{} is not completed, it may be removed from the platform, e.g. an alternate source of labor may be used to complete the [job]{} in the future. On the other hand, if the [job]{} is completed—whether as part of an adopted match or a [one-time]{} match—it is more likely that additional [jobs]{} will be added to the market. To reflect this in the model, we assume that [jobs]{} will leave the market independently with probability $\beta$ if they do not get completed in the previous period. However, if a [job]{} is completed in period $t$, then with probability $\beta'-\beta$, an additional [job]{} enters the market in the subsequent period, where $\beta' > \beta$. Thus, the expected number of [jobs]{} in period $t+1$ is given by $$\begin{aligned}
E[D_{t+1}|D_t, M_t] &= (1-\beta)(D_t - M_t) + (1+\beta' -\beta)M_t \nonumber \\
&= (1-\beta)D_t + \beta'M_t\end{aligned}$$
In this model, departures occur at a fixed rate while growth is proportional to the the number of matches. Since there is no money in this setting, the number of matches can be viewed as a measure of the surplus earned by the [job]{} side of the market. As a consequence, in our model the growth of the [job]{} side of the market is proportional to its surplus. Modeling growth in such a manner is broadly consistent with the matching market literature (see e.g. [@lian2019optimal]).
[**Worker Dynamics:**]{} We now turn our attention to the evolution of the number of [workers]{}. As discussed in Section \[sec:intro\], the future engagement of [workers]{} with the platform depends not only on whether or not the worker is matched but also on the type of her match. To capture this in our model, we consider three different dropout probabilities for the three different possible [worker]{} outcomes. (1) Unmatched workers drop out with probability $\alpha$. (2) Workers who are part of an adopted match drop out with probability $\gamma$.[^5] Naturally, we expect $\alpha > \gamma$. (3) Workers who are part of a [one-time]{} match drop out with probability $(\alpha - \alpha')$. The parameter $\alpha'$ captures the impact of getting matched on the worker’s future engagement with the platform.[^6] In the context of volunteer labor, previous work confirms such behavior: if a volunteer feels underutilized, they will be less engaged in the future [@sampson2006optimization]. Though we expect $\alpha > \gamma$, the relationship between $\alpha - \alpha'$ and $\gamma$ is unclear, and as we will see in Section \[sec:results\], it plays an important role in determining the optimal platform design. Similar to how we modeled [job]{} growth, we assume that as a result of each match, a new [worker]{} joins the market with probability $\gamma'$, i.e., external growth is proportional to the surplus of the [worker]{} side of the market.
Based on these modeling choices, we can specify the expected number of both adopted matches and [workers]{} in each period. As a result of the matching in period $t$, there are $z_tM_t$ adopted matches and $(1-z_t)M_t$ [one-time]{} matches. Since adopters (i.e., workers who are part of an adopted match) will drop out with probability $\gamma$, the expected number of adopted matches at the start of period $t+1$ is given by $$\begin{aligned}
E[K_{t+1}|M_t, z_t] = (1-\gamma)z_tM_t\end{aligned}$$
Similarly, based on the dynamics described above for the three different possible [worker]{} outcomes, the expected number of [workers]{} at the start of period $t+1$ is given by: $$\begin{aligned}
E[V_{t+1}|V_t, M_t, z_t] =& (1-\alpha)(V_t - M_t) & \tag{\text{unmatched workers}}\\ & + (1 + \gamma'-\gamma)z_tM_t& \tag{\text{adopters}} \\ &+ (1+\gamma'+\alpha'-\alpha)(1-z_t)M_t& \tag{\text{matched non-adopters}} \\
=& (1-\alpha)V_t + (\alpha'+\gamma')M_t + (\alpha - \alpha' - \gamma)z_tM_t&\end{aligned}$$
Note that the last line is the more compact (but less interpretable) representation of the dynamic. We highlight that the expected number of [workers]{} in period $t+1$ is increasing (decreasing) in the prior fraction of adopted matches $z_t$ if $\alpha-\alpha'$ is greater than (less than) $\gamma$.
In Proposition \[prop:convergence\], we show that the state variables $\{D_t, K_t, V_t, : t \in [T]\}$ and the matching outcomes $\{M_t: t \in [T]\}$ are concentrated around their expectations when the market size is large. Such concentration results enable us to approximate the multi-dimensional stochastic process by simple deterministic dynamics. This in turn provides the tractability needed to analyze the impact of the platform’s decisions on the dynamics of the system. We note that such an approach is common when studying design questions in complex stochastic systems (see e.g. [@crapis2017monopoly; @afeche2018ride])
\[prop:convergence\] For all $n > 1$, some fixed $c > 0$, and all $t \in [T]$, suppose that $(D^n_t, K^n_t, V^n_t)$ represents the state of the system in period $t$ assuming the initial conditions are given by $(nd_0, nk_0, nv_0) \in \mathbb{R}_+^3$, the platform decisions are given by $\mathbf{z}= \{z_t: z_t \in [0,1] \ \forall t \in [T]\}$, and the match probability is given by $\frac{c}{n}$.[^7] In addition, suppose that $M^n_t$ represents the number of completed [jobs]{} in period $t$. Then for all $t \in [T]$, the following limits hold almost surely as $n$ approaches infinity: $$\begin{aligned}
\frac{M_t^n}{n} \ \rightarrow \ m_t \ &= {s}(d_t-k_t, v_t-k_t) + k_t \label{eq:mt}\\
\frac{D_{t+1}^n}{n} \rightarrow d_{t+1} &= (1-\beta)d_t + \beta'm_t \label{eq:dt} \\
\frac{K_{t+1}^n}{n} \rightarrow k_{t+1}&= (1-\gamma)z_tm_t \label{eq:kt} \\
\frac{V_{t+1}^n}{n} \rightarrow v_{t+1} &= (1-\alpha)v_t + (\alpha'+\gamma')m_t + (\alpha - \alpha' - \gamma)z_tm_t \label{eq:vt}
\end{aligned}$$ where the matching function ${s}(\cdot, \cdot)$ is defined in .[^8]
The parameter $c$ represents the expected number of available and compatible [jobs]{} for each worker when there are exactly $n$ available [jobs]{}, or equivalently, the expected number of available and compatible [workers]{} for each [job]{} when there are exactly $n$ available [workers]{}. A proof of Proposition \[prop:convergence\] can be found in the appendix, while a characterization of the matching function ${s}(\cdot, \cdot)$ can be found in Section \[subsec:matchingprocess\].
Platform Optimization Problem {#subsec:opt}
-----------------------------
The platform wants to maximize the number of completed [jobs]{} between periods $t=1$ and $t = T$. Since different platforms may place different relative weights on current and future success, we allow for the value of a completed [job]{} to be discounted at a rate $\delta \in [0,1]$.
In each period $t$, the platform decides the fraction of adopted matches, i.e., the fraction that are fixed going into the next period, which we denote as $z_t \in [0,1]$. In extreme cases, by setting $z_t = 1$, the platform can require that each matched pair in period $t$ is part of an adopted match. On the other hand, by setting $z_t=0$, the platform can dissolve all matchesincluding adopted matches from the prior periodand make all [jobs]{} available in the spot market of period $t+1$ (or anything in between).[^9]
Given that the system follows the deterministic dynamics described in Proposition \[prop:convergence\], for any initial condition $(d, k, v)$, the [platform design problem]{} (PD) is to decide on the level of adoption in each period in order to maximize the total discounted number of completed jobs. Table \[table:platformproblem\] summarizes (PD).
$$\begin{array}{|ccclrr|}
\hline
&&&&& \\
\max & & & \sum_{t=1}^T \delta^{t-1} m_t&&(PD) \\
^{\{z_t, m_t, d_t, k_t, v_t : t \in [T]\}}&&&&& \\
\text{s.t.} & m_t & = & k_t + {s}(d_t-k_t, v_t-k_t) & \forall t \in [T]& \\
&d_{t+1} &=& (1-\beta)d_t + \beta'm_t &\forall t\in [T-1]& \\
&k_{t+1}&=& (1-\gamma)z_tm_t &\forall t\in [T-1]&\\
&v_{t+1} &=& (1-\alpha)v_t + (\alpha'+\gamma')m_t + (\alpha - \alpha' - \gamma)z_tm_t &\forall t\in [T-1]& \\
&\multicolumn{3}{l}{d_0 = d, \qquad k_0 = k, \qquad v_0 = v, \qquad z_t \in [0,1]}&\forall t \in [T]& \\
&&&&& \\
\hline
\end{array}$$ \[table:platformproblem\]
[Main Results]{} \[sec:results\] We now present our analysis of the [platform design problem]{}. In Section \[subsec:matchingprocess\], we derive a deterministic approximation for the matching process described in Section \[subsec:dynamics\]. Then, in Section \[subsec:myopic\], we characterize the platform’s optimal policy in a single-period version of the problem. Based on insights from the single-period model, in Sections \[subsec:adoptiondominance\] and \[subsec:optiondominance\] we separately study the [platform design problem]{} for two regimes depending on the relationship between $\gamma$ and $(\alpha-\alpha')$, i.e. the asymmetric dropout rates for [workers]{} in adopted matches and [workers]{} in [one-time]{} matches. We show a stark difference in the optimal policy for the two regimes.
Characterizing the Matching Function {#subsec:matchingprocess}
------------------------------------
As described in Section \[subsec:dynamics\], we model the random matching process as sequentially greedy, i.e., [workers]{} arrive in a random order and match with any available, compatible [job]{} (breaking ties randomly). Because [workers]{} and [jobs]{} are both assumed to be homogeneous populations, neither the arrival order nor the individual outcomes affect the process. The following proposition establishes that in a large market, the size of the matching is well concentrated around its expected size, which is itself the unique solution to a simple differential equation.
\[prop:matching\] Given two values $(a, b) \in \mathbb{R}_+^2$, suppose that for all $n>1$, $S^n_t$ represents the number of matches in a market with sides of size $na$ and $nb$ where the match probability is given by $\frac{c}{n}$ and the matching is governed by a sequentially greedy process. Then the following limit holds almost surely: $$\frac{S_t^n}{n} \rightarrow {s}(a, b) = a+b - \frac{1}{c} \log(e^{ca}+e^{cb}-1)
\label{eq:matchingfunc}$$
Proposition \[prop:matching\] provides a simple matching function for a two-sided market when matches are formed greedily and compatibilities are random and homogeneous. As such, it may have application in other studies of matching markets.
Two other commonly used matching functions are Cobb-Douglas with constant returns to scale, i.e., $\mu_1(a,b) = a^\rho b^{1-\rho}$ for some $\rho \in (0,1)$, and fully-efficient matching, i.e., $\mu_2(a,b) = \min\{a,b\}$. Like both of these alternatives, the matching function ${s}(a, b)$ is (1) increasing and concave in the size of each side of the market, and (2) if one side of the market is empty, no matching occurs.
However, our model diverges from these alternatives in two key ways. (1) If there is randomness in compatibility, then increasing both sides of the market by the same small constant $\epsilon$ should increase the size of the matching by less than $\epsilon$. Our matching function captures this, i.e., $\frac{d }{d \epsilon} {s}(a+\epsilon, b+\epsilon) < 1$, but the Cobb-Douglas and fully-efficient matching functions do not satisfy this property ($\frac{d }{d \epsilon} \mu_1(a+\epsilon, b+\epsilon) \geq 1$, and $\frac{d }{d \epsilon} \mu_2(a+\epsilon, b+\epsilon) = 1$).[^10] (2) If thickness improves the match efficiency, then the matching should exhibit increasing returns to scale. Our matching function satisfies this property, while $\mu_1(a,b)$ and $\mu_2(a,b)$ both have constant returns to scale. Though our matching function exhibits increasing returns to scale, it does so at *at a diminishing rate*, and it approaches the fully-efficient matching function in the limit: $\lim_{a,b \rightarrow \infty}{s}(a,b) = \min\{a,b\}$. Thus, this matching function is consistent with a market that gets more efficient as it grows while also respecting the physical limitation that the size of the matching cannot exceed the size of either side. We also remark that the size of the matching is increasing in the compatibility probability.
The proof of Proposition \[prop:matching\] is a generalization of Theorem 3 in [@mastin2013greedy] and makes use of stochastic differential approximation techniques from [@wormald1999models]. In Section \[sec:numerics\], we numerically show the accuracy of this approximation.
[Proof of Proposition \[prop:matching\]]{} Based on the greedy random matching process, if there are $na$ available [jobs]{} when a [worker]{} arrives, she will form a match with probability $1-(1-\frac{c}{n})^{na}$. As a consequence, the expected number of available [jobs]{} when the next [worker]{} arrives is given by $na -1+(1-\frac{c}{n})^{na}$. This insight enables us to write a stochastic differential equation for the number of matched [jobs]{}. We then show that the matching converges to the solution of a deterministic differential equation as the market size grows, holding fixed the expected number of [jobs]{} compatible with each worker.
Recall that the market has sides of size $na$ (which we will refer to as [jobs]{}) and $nb$ (which we will refer to as [workers]{}).
Let $Y(Z)$ be the number of matched [jobs]{} right before the $Z^{th}$ [worker]{} arrives in the spot market. According to the matching process, we have $Y(0) = 0$ and $${{\mathbb{E}}}[Y(Z+1)|Y(Z)] = Y(Z) + 1 - \left(1-\frac{c}{n}\right)^{n(a - \frac{Y(Z)}{n})}$$ If we define $y(z) = \frac{Y(nz)}{n}$, then for large $n$, $$\frac{{{\mathbb{E}}}[y(z + \frac{1}{n}) - y(z)|y(z)]}{1/n} = 1 - e^{-c(a-y(z))} + o(1) \footnote{For two functions $d, l: \mathbb{N}\rightarrow \mathbb{R}, l(n) =o(d(n))$ if $\lim_{n\rightarrow \infty} \frac{l(n)}{d(n)} = 0$.}$$ As $n \rightarrow \infty$, this corresponds to the differential equation $\frac{d y}{d z} = 1 - e^{-c(a-y(z))}$. Given the initial condition $y(0) = 0$, this differential equation has the unique solution $$y(z) = a + z - \frac{1}{c}\log(e^{ca} + e^{cz} - 1) \label{eq:y}$$ To show convergence, we apply techniques from [@wormald1999models], deferring the details to Appendix \[proof:prop:matching\]. To find the number of matches after $bn$ workers have arrived, we plug $z= b$ into , which yields $$\frac{S_t^n}{n} \ \ \substack{a.s.\\\rightarrow} \ \ a + b - \frac{1}{c}\log(e^{ca} + e^{cb} - 1) = {s}(a,b)$$
[Optimal Policy for a Single-Period Problem]{} \[subsec:myopic\] In this section, we consider an instance of the problem where $T = 1$, or equivalently an instance where the discount rate equals $0$. In such a setting, the platform chooses $z_0$ with the intention of maximizing $m_{1}(z_0)$.[^11]
Before presenting the structure of the optimal policy, we provide some intuition. First note that $z_0$ does not impact the number of completed jobs in period $0$ (accordingly, $d_1$ as defined in is independent of $z_0$). However, it impacts $m_1$ in three ways. (1) The match type influences the growth of the [worker]{} pool. (2) Adoption replaces the randomness in matching with certainty. However, (3) [one-time]{} matches ensure that the spot market in the subsequent period will be large. To compare the relative impact of effects (2) and (3), note that with certainty, each such adoption is worth $1$ match; however, increasing both sides of the spot market by $1$ in the next period is worth less than $1$ additional match due to randomness in matching. Therefore if adopted matches are preferable for growing the [worker]{} pool (i.e., if $\gamma \leq \alpha - \alpha'$), we would expect a policy of only adoption ($z_0 = 1$) to be optimal. On the other hand, when [one-time]{} matches lead to more growth in the [worker]{} pool (i.e., $\gamma > \alpha' - \alpha$) then no adoption can be preferable if the spot market is efficient enough, i.e., if increasing both sides of the spot market by $1$ is worth close to $1$ additional match.[^12]
As discussed after Proposition \[prop:matching\], the match efficiency in the spot market of period $1$ depends on its thickness. If both sides are large, then the matching is efficient (recall that in the extreme case, $\lim_{a,b \rightarrow \infty} {s}(a,b) = \min\{a,b\}$). Replacing an adopted match with a [one-time]{} match increases the size of the spot market in the following period, which implies that the marginal value of [one-time]{} matches should get larger as the adoption fraction decreases. Put another way, we would expect $m_1$, the number of matches in the next period, to be quasiconvex in $z_0$, the fraction of adopted matches, leading to an ‘all-or-nothing’ optimal policy in the single-period problem.
In the left panel of Figure \[fig:m1z0\], we show an example where the spot market is not thick enough and exhibits poor match efficiency. In that example, the optimal single-period policy is to have only adoption ($z_0 = 1$). In the right panel, we consider the same model primitives but with larger initial sides of the market, which leads to having a thicker spot market in period $1$. Since match efficiency is higher, the optimal single-period policy is to have no adoption ($z_0 = 0$).
In both examples, an ‘all-or-nothing’ policy is indeed optimal. In the following theorem, we confirm that in any single-period problem, the platform either wants as many adopted matches as possible, or the platform wants to create as thick of a spot market in the next period as possible by eliminating adoption.
\[thm:myopic\] Given an initial state $(d_0, k_0, v_0)$, the optimal single-period policy is either full adoption or no adoption. In particular, $$z_0^* = \begin{cases}1, &\text{if} \quad \gamma\leq \alpha-\alpha' \quad \text{or} \quad d_1 \leq \frac{1}{c}\log\left(\frac{ e^{c(1+\alpha'-\alpha)m_0} -1}{e^{c( \gamma + \alpha' - \alpha)m_0} - 1}\right) \\
0, &\text{otherwise}
\end{cases} \label{eq:optmyopic}$$
[0.45]{}\[1.0\][ ![Plotting $m_1(z_0)$ when when $(\alpha, \alpha', \gamma, \gamma', \beta, \beta') = (0.2, 0.15, 0.06, 0.2, 0.05, 0.2)$ and $c=10$. *Left:* For initial state $(d_0, k_0, v_0) = (0.8, 0, 0.8)$, full adoption maximizes $m_1$. *Right:* For initial state $(d_0, k_0, v_0) = (2.8, 0, 2.8)$, eliminating adoption maximizes $m_1$[]{data-label="fig:m1z0"}](m1_z0_ex1_v2.pdf "fig:"){width="\textwidth"}]{}
[0.45]{} \[1.0\][ ]{}
If $\gamma\leq \alpha-\alpha'$, the [worker]{} pool grows more per adopted match than per [one-time]{} match. Since this means that adoption provides both growth and certainty (i.e., there is no trade-off between growth and certainty), the platform prefers adopted matches in this regime regardless of the state of the system. In Section \[subsec:adoptiondominance\], we show that in this regime, the static policy of full adoption is also optimal in the multi-period version.
When $\gamma > \alpha-\alpha'$, adoption is optimal in the single-period problem only if the number of [jobs]{} is below a threshold, either because the spot market is not thick enough to be efficient or because there is an abundance of [workers]{}. However, once the number of [jobs]{} is sufficiently large, no adoption proves superior because the spot market is thick (and thus, efficient).[^13] To prove Theorem \[thm:myopic\], we first show that $m_{1}(z_0)$ is quasiconvex. This implies that either no adopted matches ($z_0=0$) or all adopted matches ($z_0 = 1$) is the single-period optimal solution. We then show that $m_{1}(1) \geq m_{1}(0)$ when either $\gamma \leq \alpha-\alpha'$ or $d_1\leq \frac{1}{c}\log\left(\frac{e^{c(1-\alpha'-\alpha)m_0}-1}{e^{c(\gamma-\alpha'-\alpha)m_0}-1}\right)$. A complete proof can be found in Appendix \[proof:thm:myopic\]. In Section \[sec:numerics\], we numerically show that the structure of the optimal single-period policy holds even when we consider the actual expected size of the matching (as opposed to its deterministic approximation).
[Optimal Policy in the Adoption Growth Dominance Regime]{} \[subsec:adoptiondominance\] In this section, we focus on instances where $\gamma \leq \alpha - \alpha' $, which we will call the Adoption Growth Dominance (AGD) regime. In this regime, having more adoption will increase the number of [workers]{} in the subsequent period, i.e., $v_{t+1}(z_t)$ as defined in is increasing in $z_t$. In addition, Theorem \[thm:myopic\] implies that $m_{t+1}(z_t)$ is maximized at $z_t = 1$. Thus, in the [AGD regime]{} there is no trade-off between long-term market growth and short-term certainty in matching. Increasing the adoption fraction will improve both objectives. We formalize this in the following theorem:
\[thm:optregimea\] In the [AGD regime]{} (i.e., $\gamma \leq \alpha - \alpha'$), the optimal policy is $z^*_t = 1$ for all $t \in [T]$.
Theorem \[thm:optregimea\] establishes that a policy of only adoption is optimal in the [AGD regime]{}. This intuitive result suggests that unless [one-time]{} matches provide significant gains in [worker]{} engagement (and therefore growth), the platform should promote adoption.
To prove Theorem \[thm:optregimea\], we use Theorem \[thm:myopic\] to show that in the [AGD regime]{}, the optimal single-period policy (i.e., the optimal myopic policy) is always $z_t = 1$. Thus, it is sufficient to show that when the optimal myopic policy also increases the [worker]{} pool, then it is the optimal policy in the long-run as well. We use the following lemma to establish this result.
\[lemma:myopicisbest\] Let $z_\tau'$ be the optimal myopic policy given state $(d_\tau, k_\tau, v_\tau)$. Then $z_\tau'$ is the optimal policy in period $\tau$ if $z_\tau' \in \text{argmax}_{z_\tau \in [0,1]} v_{\tau+1}(z_\tau)$.
[Proof of Lemma \[lemma:myopicisbest\]]{} Consider two policies $\mathbf{\hat{z}} = \{\hat{z}_t : t \in [T]\} $ and $\mathbf{\tilde{z}}= \{\tilde{z}_t : t \in [T]\}$ that differ only at time $\tau$, where $\hat{z}_\tau = z_\tau'$ and $\tilde{z}_\tau \neq z_\tau'$. Using identical notation to denote the state variables when following each policy, we have $\hat{m}_{\tau+1} \geq \tilde{m}_{\tau+1}$. In addition, when the condition in Lemma \[lemma:myopicisbest\] holds, we have $\hat{v}_{\tau+1} \geq \tilde{v}_{\tau+1}$, and by the dynamics described in , $\hat{d}_{\tau+1} \geq \tilde{d}_{\tau+1}$. To complete the proof, we must show that these conditions imply $\hat{m}_{t} \geq \tilde{m}_{t}$ for all $t \in [T] \setminus [\tau]$. We use the following claim to establish this result:
\[claim:statecoupling\] Suppose the platform’s decisions $z_t$ are fixed for all $t \in [T']$ where $T' \leq T$. In that case, for any two initial states $(d_0, k_0, v_0)$ and $(d_0', k_0', v_0')$, if $d_0 \geq d_0'$, $v_0 \geq v_0'$, and $m_0 \geq m_0'$, then for all $t \in [T']$, $d_t \geq d_t'$, $v_t \geq v_t'$, and $m_t \geq m_t'$.
We defer the proof of this claim to Appendix \[proof:lemma:myopicisbest\]. Since $\hat{z}_t = \tilde{z}_t$ for all $t \in [T] \setminus [\tau]$, we can apply Claim \[claim:statecoupling\] to the initial states $(\hat{d}_{\tau+1}\, \hat{k}_{\tau+1}, \hat{v}_{\tau+1})$ and $(\tilde{d}_{\tau+1}, \tilde{k}_{\tau+1}, \tilde{v}_{\tau+1})$ for $T' = T - \tau-1$ to show that $\hat{m}_{t} \geq \tilde{m}_{t}$ for all $t \in [T] \setminus [\tau]$. Therefore, by replacing the decision in period $\tau$ with $z_\tau'$, the platform can only increase the total number of completed [jobs]{}. Using a contradiction argument, no policy with $z_\tau \neq z_\tau'$ can be strictly optimal. This completes the proof that $z_\tau'$ (the optimal myopic policy) is the optimal policy in period $\tau$.[^14]
[Optimal Policy in the Adoption Growth Non-Dominance Regime]{} \[subsec:optiondominance\] As discussed above, in the [AGD regime]{}, there is no trade-off between short-term certainty in matching and long-term market growth. We now turn our attention to the regime where [one-time]{} matches are better for long-term market growth, i.e., when $\gamma > \alpha - \alpha'$. We call this the Adoption Growth Non-Dominance (AGN) regime. In this regime, the market can choose to promote the growth of the [worker]{} side of the market by reducing adoption and correspondingly increasing the size of the spot market in the subsequent period.
As a consequence of Lemma \[lemma:myopicisbest\], when $z_t = 0$ is myopically optimal in the [AGN regime]{}, no adoption is the optimal policy in period $t$. As established in Theorem \[thm:myopic\], $z_t = 0$ is optimal in a single-period problem if $\gamma > \alpha - \alpha'$ and the number of [jobs]{} exceeds a threshold. In other words, no adoption is optimal in the [AGN regime]{} if the spot market is thick enough that the benefits of short-term certainty are immediately offset by the benefits of [worker]{} growth.
If the number of [jobs]{} continues to grow beyond that threshold, a no-adoption policy will remain optimal. Based on this insight, we can characterize the optimal policy in the [AGN regime]{} when the market is sufficiently large.
\[thm:optregimeB\] In the [AGN regime]{} (i.e., $\gamma > \alpha -\alpha'$), if $d_\tau \geq \bar{d}$, and $v_\tau \geq \bar{v}$, then $z_t = 0$ is optimal for all $t \in [T] \setminus [\tau-1]$, where $\bar{d}$ and $\bar{v}$ are the unique solutions to the *minimum market size* program defined in .
$$\begin{aligned}
\min_{d,v} \quad \qquad &\ d \tag{MMS} \label{eq:mms} \\
s.t. \ \ \quad d \geq & \frac{1}{c(1-\beta)}\log\left(\frac{1+\alpha'-\alpha}{\gamma +\alpha' - \alpha}\right) \label{eq:mmscon1.5} \\
d \geq& \frac{1}{c(1-\beta)} \log \left(\frac{e^{c(1+\alpha' -\alpha)d}-1}{e^{c(\gamma+\alpha'-\alpha )d}-1} \right) - \frac{\beta'}{1-\beta}d \label{eq:mmscon1}\\
d \geq & \frac{1}{c} \log \left(\frac{e^{cv}-1}{e^{cv(1-\frac{\alpha}{\alpha'+\gamma'})}-1} \right) \label{eq:mmscon2}\\
v \geq & \frac{1}{c} \log \left(\frac{e^{cd}-1}{e^{cd(1-\frac{\beta}{\beta'})}-1} \right) \label{eq:mmscon3}
\end{aligned}$$
false
Theorem \[thm:optregimeB\] shows that in the [AGN regime]{}, once the market gets sufficiently thick, a static no-adoption policy is optimal for the remainder of the time horizon. In Figure \[fig:arrow\_diagram\] we provide a visual example of the thresholds $\bar{d}$ and $\bar{v}$ using dashed lines. Note that the model primitives for this example belong to the [AGN regime]{}. Figure \[fig:arrow\_diagram\] also provides a visualization for how the myopically optimal policy compares to the simulated optimal policy.
Specifically, in Figure \[fig:arrow\_diagram\], (1) the solid curve represents the threshold at which the myopically optimal policy changes. To the left of the curve, the resulting spot market is thin so $z_0 = 1$ is myopically optimal, while to the right, $z_0 = 0$ is myopically optimal. (2) The color of the arrow at a point $(d,v)$ represents the long-run optimal decision at the state $(d, 0, v)$. A blue arrow indicates an optimal policy of $z_0^* = 1$, while a red arrow indicates that $z_0^* = 0$. We remark all arrows to the right of the curve are red, as are some arrows to the left of the curve. In other words, even if adoption is myopically optimal, it can be sub-optimal when considering a longer time horizon due to its negative impact on [worker]{} growth. (3) The direction of the arrow indicates how the size of the [worker]{} and [job]{} pools will change in the subsequent period when following the optimal policy $z_0^*$. (4) The dashed lines represent the thresholds $\bar{d}$ and $\bar{v}$ defined in Theorem \[thm:optregimeB\], which outline an absorbing state where $z_t^* = 0$ is optimal for all $t \in [T]$. We remark that all arrows are red beyond those thresholds, and all arrows along the boundary point into the region.
The proof of Theorem \[thm:optregimeB\] relies on appropriately interpreting the constraints in . Constraints and are sufficient conditions to ensure [one-time]{} matches are myopically optimal. Constraint (resp. ) ensures that the number of [workers]{} (resp. [jobs]{}) will increase in the subsequent period assuming there is no adoption. We use these constraints to show that $ \{(d, v) : d \geq \bar{d}, v \geq \bar{v} \}$ is an absorbing state. When applied iteratively, this proves that $z_t = 0$ is optimal for all $t \in [T] \setminus [\tau-1]$. In Appendix \[proof:thm:optregimeB\], we present a rigorous proof of Theorem \[thm:optregimeB\] which also demonstrates that the unique solution to can be found via linear search.
If the market size is below the threshold specified by Theorem \[thm:optregimeB\], the optimal decision proves difficult to characterize, since the problem is not convex in the decision variables $\mathbf{z}$. However, our numerical analysis indicates that the optimal decisions are always $z_t = 0$ or $z_t=1$, resembling the ‘all-or-nothing’ optimal myopic policy. Further, if the horizon is long enough, we numerically observe that the optimal policy is either no adoption (i.e., $z_t^* = 0$ for all $t \in [T]$) or it is initially adoption before switching to no adoption (i.e., for some $\tau$, the optimal policy is $z_t^* = 1$ for all $t \in [\tau]$ and $z_{t}^* = 0$ for all $t \in [T] \setminus [\tau]$). In these cases, even though we cannot theoretically characterize the optimal policy, we can offer a guarantee on the performance of a static policy of no adoption.
\[thm:approxregimeB\] Let $r =\delta \max\{1-\beta+\beta', 1-\alpha+\alpha'+\gamma'\}$. In the [AGN regime]{} (i.e., $\gamma > \alpha -\alpha'$), the policy $z^*_t = 0$ for all $t \in [T]$ is $1-\kappa$ optimal, where $$\kappa = \min\{\frac{ \frac{1}{1-r^T} + \frac{\max\{ \beta', \alpha'+\gamma'\}}{1-r} }{c \min\{d_0, v_0\}} \log 2, 1\}. \label{eq:kappa}$$
Theorem \[thm:approxregimeB\] establishes a ratio between the performance of a no-adoption policy and the optimal policy. The ratio approaches $1$ as the market size $c\min\{d_0,v_0\}$ grows, and the ratio also improves the more the platform discounts the future. If $r \leq 1 - \frac{1+\max\{ \beta', \alpha'+\gamma'\}}{c \min\{v_0, d_0\}} \log 2$, then our approximation factor is non-trivial. In the example shown in Figure \[fig:arrow\_diagram\], a policy of no adoption at the initial condition $(d_0, v_0) = (0.5, 0.5)$ achieves an approximation ratio of $0.704$.
To prove Theorem \[thm:approxregimeB\], we first upper bound the achievable number of completed [jobs]{} by considering a setting where the matching process is perfectly efficient, i.e., the number of matches is the minimum of the two sides of the spot market. Using this matching function in the [AGN regime]{}, we can exactly characterize the optimal policy as $z_t^* = 0$ for all $t \in [T]$. We then directly compare this upper bound on the number of completed [jobs]{} to the number of completed [jobs]{} when the matching process is governed by and the platform follows the same no-adoption policy. We show that the matching outcome given by is within $\log(2)$ of the minimum side of the market. Propogating this gap over the time horizon yields the ratio in Theorem \[thm:approxregimeB\]. The full proof can be found in Appendix \[proof:thm:approxregimeB\].
[Matching Function Approximation Accuracy]{} \[sec:numerics\] Our insights in the previous section are based on deterministic approximations to the system dynamics. In this section, we numerically demonstrate the accuracy of our approximations for reasonable system sizes, and we discuss how some of our main structural results continue to apply.
Numerical analysis shows that the matching function ${s}(a,b)$ defined in is always an underestimate of the actual expected matching, but it becomes increasingly accurate as $n$ gets large. This is evident in Figure \[fig:numerics1\] in which we show the gap between the expected size of the matching and its corresponding deterministic approximation (defined in ) normalized by the matching size. We observe that the normalized gap is always positive, and it is below $2 \%$ when the size of both sides is at least $30$.
In addition, the structure of the optimal single-period solution appears unchanged; numerical results indicate that an ‘all-or-nothing’ policy continues to be optimal. In Figure \[fig:numerics2\], we use simulation to find the expected number of matches for various levels of adoption, and we compare those results to our deterministic approximation. We make the following observations from the plot: (1) the shape of $m_1(z_0)$, i.e., the deterministic approximation, is broadly consistent with the simulated results. However, (2) the deterministic matching function is always an underestimate, and (3) the gap is smallest at $z_0 = 1$, when most of the matching is pre-determined due to adoption. The latter two observations imply that the deterministic approximation relatively overvalues adoption. This insight leads us to conjecture that if the optimal single-period policy according to the deterministic approximation is $z_t = 0$ (i.e., if neither condition in holds), then the optimal single-period policy is $z_t = 0$ when considering the actual expected matching.
[0.48]{} ![*Left*: Convergence to the deterministic approximation of the matching function when both sides of the market are of size $n$ and the compatibility probability is $c/n$. *Right*: Comparing the expected value of $m_1(z_0)$ to its deterministic approximation in an identical setting to Figure \[fig:m1z0ex1\], except with initial condition $(d_0, k_0, v_0) = (0.8, 0.8, 0.8)$ to ensure the integrality of $nm_0$.[]{data-label="fig:numerics"}](matching_convergence.pdf "fig:"){width="\textwidth"}
[0.48]{}
Discussion {#sec:discussion}
==========
We conclude with some discussion and directions for future work.
**Benefits of Non-Adoption.** We have shown that although adoption increases match , in some settings the policy of no adoption can maximize the total number of discounted matches. the growth benefits of non-adoption can outweigh the loss in match . We remark that a number of our modeling choices have overvalued the of adoption relative to non-adoption. We assume that the probability of compatibility between a given job and worker is redrawn every period, unless the ; in the repeated matching setting, it is reasonable to assume that a worker matched with a job in period $t$ will likely still be compatible with the job in period $t+1$ . Our numerical analysis in Section \[sec:numerics\] suggests that our deterministic matching function is an underestimate of the true greedy matching, and that the error is larger as a percentage of the total number of completed jobs for small $z$ (low adoption levels). the case for the optimality of non-adoption , meaning that may be optimal for settings beyond those that we have identified.
**Benefits of Adoption.** We have shown that in the [AGD regime]{}, where adopter growth dominates non-adopter growth, a static policy of full adoption is optimal. This finding holds true beyond the specific assumptions of our model and the specific form of our deterministic matching function. For example, when using the actual expected matching instead of the deterministic approximation, we can directly prove that adoption is optimal in the [AGD regime]{} by extending Theorem \[thm:optregimea\]. Said differently, the proof of this result does not rely on the explicit functional form of ${s}(a,b)$ but rather on the structural properties of the underlying matching process, and so the result holds without appealing to the deterministic approximation. Adjusting the proof of Theorem \[thm:optregimea\] to a more general setting would involve an inductive coupling argument showing that in a single-period dynamic, the number of jobs, workers, and matches are largest under a policy of full adoption. We omit the details for the sake of brevity.
**Commitment as a Design Lever.** The literature on designing two-sided matching markets is rich and varied, and there have been a multitude of papers studying how to optimize centralized matching as well as how to design levers such as pricing, search, and information. To our knowledge, our work is one of the first to suggest that the level of *commitment* to a match is also a design lever that can be influenced by the platform. Such a lever is potentially useful for a wide range of settings where matchings are repeated over time, and our results provide a framework for evaluating the trade-offs. For example, in managing gig economy workers, a platform may consider creating a blended workforce where some workers lock in their shifts; our results suggest this may reduce options available for flexible workers, which in turn may reduce their engagement. Looked at from another perspective, a consumer-focused platform that provides repeated service by gig economy workers may want to increase customer goodwill by allowing customers to repeatedly match with their favorite service provider; however this may reduce the options available for other consumers and harm platform growth. There are also many models of commitment beyond our specific formulation that can be used to study a wider range of applications.
[Proof of Proposition \[prop:convergence\]]{} \[proof:prop:convergence\]
To aid in this proof, we define constants $\Delta_t$ and $\Phi_t$, where $\Delta_t$ is defined such that with probability $1-O(n^{1/4}e^{-c^6n^{1/4}})$, $$\left|\frac{S^n(n(d_t-k_t), n(v_t-k_t))}{n} - {s}(d_t-k_t, v_t-k_t)\right| \leq \Delta_t n^{-1/4} \label{eq:proof:app1}$$ We defer a proof that such a choice of $\Delta_t$ is possible until we prove Proposition \[prop:matching\] in Appendix \[proof:prop:matching\]. In addition, $\Phi_{-1} = 0$ and for all $t \in [T]$, $\Phi_t := \sum_{\tau = 0}^t (3+\alpha +\alpha'+\beta'+\gamma')^{t-\tau}\Delta_\tau$.
We now prove by induction that for all $t \in [T]$, with probability $1-O((t+1)n^{1/4}e^{-c^6n^{1/4}})$, $$\begin{aligned}
\left|\frac{K_{t}^n}{n}-k_t\right| &\leq \Phi_{t-1} n^{-1/4} \label{proof:eq:kconverge} \\
\left|\frac{D_t^n}{n}-d_t\right| &\leq (1+\beta')\Phi_{t-1} n^{-1/4} \label{proof:eq:dconverge} \\
\left|\frac{V_t^n}{n}-v_t\right| &\leq (1 + \alpha +\alpha'+\gamma') \Phi_{t-1} n^{-1/4} \label{proof:eq:vconverge} \\
\left|\frac{M_t^n}{n}-m_t\right| &\leq \Phi_{t} n^{-1/4} \label{proof:eq:mconverge}\end{aligned}$$ We note that the final inequality will be proved using the previous three inequalities as part of its inductive hypothesis.
The base cases for equations - are straightforward, since both sides of each inequality equals $0$ by definition. The base case for follows directly from the definition of $\Phi_0$ (above) and the proof of Proposition \[prop:matching\] (deferred to Appendix \[proof:prop:matching\]).
We now assume that this holds for $t = \tau$. We will prove that the statement also holds for $t = \tau+1$. We proceed in order: $$\begin{aligned}
\left|\frac{K_{\tau+1}^n}{n}-k_{\tau+1}\right|
&\leq \left|\frac{K_{\tau+1}^n}{n}-{{\mathbb{E}}}\left[\frac{K_{\tau+1}^n}{n}\mid M_\tau^n\right]\right|
+ \left|{{\mathbb{E}}}\left[\frac{K_{\tau+1}^n}{n}\mid M_\tau\right] - k_{\tau+1}\right| \\
&= \left|\frac{K_{\tau+1}^n}{n}-(1-\gamma)z_\tau \frac{M_\tau^n}{n}\right|
+ \left|(1-\gamma)z_\tau \frac{M_\tau^n}{n} - (1-\gamma)z_\tau m_\tau\right|
\\ &= \frac{1}{n}\left|K_{\tau+1}^n-(1-\gamma)z_\tau M_\tau^n\right|
+ (1-\gamma)z_\tau\left|\frac{M_\tau^n}{n} - m_\tau\right|
\end{aligned}$$ By the inductive hypothesis, with probability $1-O((\tau+1)n^{1/4}e^{-c^6n^{1/4}})$, we can bound the second term by $(1-\gamma)z_\tau \Phi_{\tau} n^{-1/4}$. We will use a Chernoff bound on the first term, recalling that $K_{\tau+1}^n$ is a binomial random variable. Thus, with probability $1-e^{n^{-1/2}}$, $$\frac{1}{n}|K_{\tau+1}^n-(1-\gamma)z_\tau M_\tau^n| \leq n^{-5/4} \sqrt{2(1-\gamma)z_\tau M_\tau^n}$$ Assuming the bound in the second term holds, $M_\tau^n$ is $O(n)$, which means that for large enough $n$, $n^{-5/4} \sqrt{2(1-\gamma)z_\tau M_\tau^n} \leq \gamma z_\tau \Phi_{\tau} n^{-1/4}$. Thus, taking a union bound, we have with probability $1-O\left((\tau+1)n^{1/4}e^{-c^6n^{1/4}}\right)$, $$\left|\frac{K_{\tau+1}^n}{n}-k_{\tau+1}\right| \leq \Phi_{\tau} n^{-1/4}$$ This proves by induction that holds for all $t \in [T]$.
We repeat a nearly identical process for $\left|\frac{D_{\tau+1}^n}{n}-d_{\tau+1}\right|$ and $\left|\frac{V_{\tau+1}^n}{n}-v_{\tau+1}\right|$, using Chernoff bounds two and three times, respectively, due to the additional binomial processes.
For the final step, we use techniques from [@wormald1999models] to prove the convergence of the matching function conditional on the current state, and then we leverage the sublinearity of the matching function to bound the distance between the matching resulting from the actual state and the deterministic approximation of the matching in period $\tau$.
$$\begin{aligned}
\left|\frac{M_{\tau+1}^n}{n}-m_{\tau+1}\right| \leq& \left|\frac{M_{\tau+1}^n}{n}-\frac{K_{\tau+1}^n}{n}-{s}\left(\frac{D_{\tau+1}^n-K_{\tau+1}^n}{n}, \frac{V_{\tau+1}^n-K_{\tau+1}^n}{n}\right)\right| \nonumber \\&+
\left|\frac{K_{\tau+1}^n}{n}+{s}\left(\frac{D_{\tau+1}^n-K_{\tau+1}^n}{n}, \frac{V_{\tau+1}^n-K_{\tau+1}^n}{n}\right) - m_{\tau+1}\right| \\
\leq&
\left|\frac{S_{\tau+1}^n}{n}-{s}\left(\frac{D_{\tau+1}^n-K_{\tau+1}^n}{n}, \frac{V_{\tau+1}^n-K_{\tau+1}^n}{n}\right)\right| \nonumber \\&
+\left|\frac{D_{\tau+1}^n}{n} - d_{\tau+1}\right| +\left|\frac{K_{\tau+1}^n}{n} - k_{\tau+1}\right| +\left|\frac{V_{\tau+1}^n}{n} - v_{\tau+1}\right|\end{aligned}$$
In the first term, we cancel out the certain adopted matches. We then break the second term into three terms, utilizing the fact that $\frac{ \partial m_t}{\partial d_t} \in [0,1]$, as are $\frac{ \partial m_t}{\partial k_t}$ and $\frac{ \partial m_t}{\partial v_t}$ (we defer presentation of these partial derivatives to Appendix \[proof:lemma:myopicisbest\]).
With probability $1-O\left(n^{1/4}e^{-c^6n^{1/4}}\right)$, the first term is less than $\Delta_{\tau+1}n^{-1/4}$ (this comes from the definition of $\Delta_t$, based on the proof of Proposition \[prop:matching\] in Appendix \[proof:prop:matching\]). Taking a union bound over three of our inductive hypotheses, with probability $1-O((\tau +1)n^{1/4}e^{-c^6n^{1/4}})$, $$\left|\frac{D_{\tau+1}^n}{n} - d_{\tau+1}\right| +\left|\frac{K_{\tau+1}^n}{n} - k_{\tau+1}\right| +\left|\frac{V_{\tau+1}^n}{n} - v_{\tau+1}\right| \leq (3+\alpha+\alpha'+\beta'+\gamma')\Phi_{\tau}n^{-1/4}$$ Combining the bounds on these two terms via one union bound, we see that with probability $1-O((\tau +2)n^{1/4}e^{-c^6n^{1/4}})$, $$\left|\frac{M_{\tau+1}^n}{n}-m_{\tau+1}\right| \leq
(3+\alpha+\alpha'+\beta'+\gamma')\Phi_{\tau}n^{-1/4} + \Delta_{\tau+1}n^{-1/4} = \Phi_{\tau+1}n^{-1/4}$$ This completes the proof by induction. To show that this convergence result holds for all $t \in [T]$, we can take another union bound over the $4T$ state variables. Since $T = O(1)$, with probability $1-O(n^{1/4}e^{-c^6n^{1/4}})$, for all $t \in [T]$, $$\begin{aligned}
\left|\frac{K_{t}^n}{n}-k_t\right| &\leq O\left(n^{-1/4}\right) \\
\left|\frac{D_t^n}{n}-d_t\right| &\leq O\left(n^{-1/4}\right) \\
\left|\frac{V_t^n}{n}-v_t\right| &\leq O\left(n^{-1/4}\right) \\
\left|\frac{M_t^n}{n}-m_t\right| &\leq O\left(n^{-1/4}\right) \end{aligned}$$ This completes the proof of almost sure convergence.
false We will prove this proposition in two parts. First, we will show that if $|\frac{D_t^n}{n}-d_t|$, $|\frac{K_t^n}{n}-k_t|$, and $|\frac{V_t^n}{n}-v_t|$ are all less than $O(t n^{-1/4})$ with probability $blank$, then $|\frac{M_t^n}{n} - m_t| \leq O(n^{-1/4})$ with probability $blank$.
Then we will prove via induction that for a series of constants $\{\Phi_t : t \in [T]\}$, $|\frac{D_t^n}{n}-d_t| \leq \Phi_t t n^{-1/4}$, $|\frac{K_t^n}{n}-k_t| \leq \Phi_t t n^{-1/4}$, $|\frac{V_t^n}{n}-v_t| \leq \Phi_t t n^{-1/4}$, and $|\frac{M_t^n}{n} - m_t| \leq \Phi_t (t+1) n^{-1/4}$ with probability $blank$.
Applying these steps inductively starting from $t=0$ to $t = T$, we then take a union bound to show that with probability $blank$, for all $t \in [T]$ $$\begin{aligned}
|\frac{D_t^n}{n}-d_t| \leq O(n^{-1/4}) \\
|\frac{K_t^n}{n}-k_t| \leq O(n^{-1/4}) \\
|\frac{V_t^n}{n}-v_t| \leq O(n^{-1/4}) \\
|\frac{M_t^n}{n}-m_t| \leq O(n^{-1/4})\end{aligned}$$ which completes the proof. Note that we require $T$ as well as $d_t, k_t,$ and $v_t$ to be of order $O(1)$ with respect to $n$.
We begin by showing the convergence of the matching process at any given period $\tau$. Using the proof of Proposition \[prop:matching\], we know that conditional on $D^n_\tau, K_\tau^n,$ and $V^n_\tau$, then with probability $1-O(n^{1/4}e^{-c^6n^{1/4}})$, $$|\frac{M^n_\tau}{n} - \frac{K_\tau^n}{n} - {s}(1/n(D^n_\tau - K_\tau^n), 1/n (V^n_\tau - K_\tau^n))| = O(n^{-1/4})$$ Since $|\frac{\partial {s}(a,b)}{\partial a}| \leq 1$ (and equivalently for $b$), with probability $1-O(n^{1/4}e^{-c^6n^{1/4}})$ $$|\frac{M^n_\tau}{n} - k_\tau - {s}(d_\tau - k_\tau, v_\tau - k_\tau)| \leq |d_\tau - \frac{D^n_\tau}{n}| + 3|k_\tau - \frac{K^n_\tau}{n}|+ |v_\tau - \frac{V^n_\tau}{n}|+ O(n^{-1/4})$$ By assumption, with probability $blank$, $|d_\tau - \frac{D^n_\tau}{n}| \leq \Phi_\tau \tau n^{-1/4}$ (similarly for $k_\tau$ and $v_\tau$). Thus, using a union bound, we have with probability $1-O(n^{1/4}e^{-c^6n^{1/4}})$, $$|\frac{M^n_\tau}{n} - m_\tau| \leq (5 \Phi_\tau + 1)$$ This concludes the first part of the proof.
We now show that for a series of constants $\Phi_t$, $|\frac{D_t^n}{n}-d_t| \leq \Phi_t t n^{-1/4}$, $|\frac{K_t^n}{n}-k_t| \leq \Phi_t t n^{-1/4}$, $|\frac{V_t^n}{n}-v_t| \leq \Phi_t t n^{-1/4}$, and $|\frac{M_t^n}{n} - m_t| \leq \Phi_t (t+1) n^{-1/4}$ with probability $blank$. We proceed by induction. Clearly, this is true for $t=0$, since the first three results hold by equality, and the last was proven above.
By the dynamics described in , $$D^n_{t+1} = D_t^n + \text{Binomial}(M_t^n, \beta') - \text{Binomial}(D_t^n - M_t^n, \beta)$$ Using a Chernoff bound, we have that with probability $1-O(e^{\frac{-n^(1/2)\beta'}{2}})$, $$\frac{1}{n}|\text{Binomial}(M_t^n, \beta')-\beta' M_t^n| \leq n^{-1/4} \beta' (\frac{M_t^n}{n})^{1/2}$$ Similarly, with probability $1-O(e^{\frac{-n^{1/2}(1-\beta)}{2}})$, $$\frac{1}{n}|\text{Binomial}(D_t^n-M_t^n, 1-\beta)-(1-\beta)(D_t^n M_t^n)| \leq n^{-1/4} (1-\beta)(\frac{D_t^n}{n}-\frac{M_t^n}{n})^{1/2}$$ Combining the two using a union bound, for some constant $\omega_1$, we have that with probability $1-O(e^{\frac{-n^(1/2)}{\omega_1}})$, $$\begin{aligned}
\frac{1}{n}|D^n_{t+1} - (1-\beta)D_t^n - (\beta'-\beta)M_t^n | &\leq n^{-1/4} ((\frac{M_t^n}{n})^{1/2} + (\frac{D_t^n}{n}-\frac{M_t^n}{n})^{1/2}) \\
&\leq n^{-1/4} ((m_t + t\Phi_t n^{-1/4} )^{1/2} + (d_t - m_t + 2t\Phi_t n^{-1/4})^{1/2}) \\
&\leq \Phi_{t+1} n^{-1/4}\end{aligned}$$ By the triangle inequality, this implies that with probability $blank$ $$\begin{aligned}
|\frac{D^n_{t+1}}{n} - (1-\beta)d_n^t - (\beta'-\beta)d_t^n | &\leq O(n^{-1/4}) + (1-\beta)|\frac{D_t^n}{n}-d_t| + (\beta'-\beta)|\frac{M_t^n}{n}-m_t| \\
|\frac{D^n_{t+1}}{n} - d_{t+1}| &\leq \Phi_{t+1} (t+1) n^{-1/4}\end{aligned}$$ Identical arguments using the dynamics in and show that with probability $blank$ $$\begin{aligned}
|\frac{K^n_{t+1}}{n} - k_{t+1}| &\leq \Phi_{t+1} (t+1) n^{-1/4} \\
|\frac{V^n_{t+1}}{n} - v_{t+1}| &\leq \Phi_{t+1} (t+1) n^{-1/4}\end{aligned}$$ Given these three results, we can the first step of this proof implies that with probability $blank$, $|\frac{M^n_{t+1}}{n} - m_{t+1}| \leq O(n^{-1/4})$. By the union bound, all three concentration results must hold with probability $blank$. As previously noted, repeating these steps from $t=0$ to $t = T$ and taking a union bound yields the result in Proposition \[prop:convergence\].
[Missing Proofs from Section \[sec:results\]]{}
[Remaining Details in the Proof of Proposition \[prop:matching\]]{} \[proof:prop:matching\]
In Section \[subsec:matchingprocess\], we showed that $${{\mathbb{E}}}\left[Y(Z+1) - Y(Z)\mid Y(Z)\right] = 1 - \left(1-\frac{c}{n}\right)^{n\left(a - \frac{Y(Z)}{n}\right)}$$
To complete the proof of Proposition \[prop:matching\], we apply Theorem 5.1 from [@wormald1999models]. Using identical notation to the statement of that theorem, we use a domain $D$ defined by $z \in (-\epsilon, b+\epsilon)$ and $y \in (-\delta, a+\delta)$.
To satisfy the boundedness hypothesis, we set $\beta = 1$ and $\gamma =0$ because with probability $1$, we have ${{\mathbb{E}}}[Y(Z+1) - Y(Z)|Y(Z)] \leq 1$. Turning our attention to the trend hypothesis, we first note for any $x \geq 0$, $e^{-x-x^2} \leq 1-x \leq e^{-x}$. Setting $x = \frac{c}{n}$ and raising each term to the $n\left(a - \frac{Y(Z)}{n}\right)$ power yields $$\begin{aligned}
e^{\left(-c-\frac{c^2}{n}\right)\left(a - \frac{Y(Z)}{n}\right)} &\leq \left(1-\frac{c}{n}\right)^{n\left(a - \frac{Y(Z)}{n}\right)} \leq e^{-c\left(a - \frac{Y(Z)}{n}\right)} \\
e^{(-c)(a - \frac{Y(Z)}{n})}\left(1-\frac{c^2(a - \frac{Y(Z)}{n})}{n}\right) &\leq \left(1-\frac{c}{n}\right)^{n\left(a - \frac{Y(Z)}{n}\right)} \leq e^{-c\left(a - \frac{Y(Z)}{n}\right)}\\
e^{(-c)\left(a - \frac{Y(Z)}{n}\right)}(1-\frac{c^2}{n}) &\leq \left(1-\frac{c}{n}\right)^{n\left(a - \frac{Y(Z)}{n}\right)} \leq e^{-c\left(a - \frac{Y(Z)}{n}\right)}
\end{aligned}$$ Since $\frac{c^2}{n}e^{(-c)\left(a - \frac{Y(Z)}{n}\right)} \leq \frac{c^2}{n}$, we can set $\lambda_1 = \frac{c^2}{n}$ to satisfy the trend hypothesis.
Finally, we note that the constant $L=c^2e^{c\delta}$ satisfies the Lipschitz condition for the function $1-e^{-c(a-y(z))}$ over domain $D$.
Given that the conditions of Theorem 5.1 in [@wormald1999models] are met, we define $\lambda = c^2n^{-1/4}$ to yield that for all $z \in [0,1]$, with probability $1-O(n^{1/4}e^{-c^6n^{1/4}})$, $$Y(nz) = ny(z) + O\left(n^{3/4}\right)$$ Plugging in $z = b$ completes that proof that the expected matching converges almost surely to $a+b - \frac{1}{c}\log(e^{ca}+e^{cb}-1)$. In addition, this implies that there is some constant $\Delta(a,b)$ whose magnitude depends on the initial conditions such that $|\frac{S^n_t}{n}-s_t| \leq \Delta n^{-1/4}$. The constants $\Delta_t$ in Line of Appendix \[proof:prop:convergence\] are defined to equal $\Delta(d_t-k_t, v_t-k_t)$.
[Proof of Theorem \[thm:myopic\]]{} \[proof:thm:myopic\] We begin by showing that $m_1(z_0)$ does not attain a local maximum. From -, we have $\frac{\partial d_{1}}{\partial z_0} = 0$, $\frac{\partial k_{1}}{\partial z_0} = (1-\gamma)m_0$, and $\frac{\partial v_1}{\partial z_0} = (\alpha - \alpha' - \gamma)m_0$. Combining these derivatives with and , we have $$\begin{aligned}
m_1 \ \ &= d_1 - k_1 + v_1 - \frac{1}{c} \log(e^{c(d_1-k_1)}+e^{c(v_1 - k_1)}-1) \\
\frac{\partial m_1}{\partial z_0} \ &= \left(\frac{(\alpha - \alpha'-\gamma)e^{c(d_1-k_1)} + (1 +\alpha' - \alpha) }{e^{c(d_1-k_1)}+e^{c(v_1 - k_1)}-1}\right)m_0 \label{eq:dmdz}\\
\frac{\partial^2 m_1}{\partial z_0^2} &= \left( \frac{(1-\gamma)^2e^{c(d_1-k_1)} + (1+\alpha'-\alpha)^2e^{c(v_1-k_1)}-(\alpha - \alpha' - \gamma)^2e^{c(d_1+v_1-2k_1)}}{(e^{c(d_1-k_1)}+e^{c(v_1 - k_1)}-1)^2}\right)m_0^2 \label{eq:d2mdz2}
\end{aligned}$$ The first-order condition prescribed by is equivalent to $e^{c(d_1-k_1)} = \frac{1 +\alpha' - \alpha}{\alpha - \alpha'-\gamma}$. When the FOC holds, the numerator of reduces to $(1-\gamma)^2e^{c(d_1-k_1)} + (1-\gamma)(1+\alpha'-\alpha)e^{c(v_1-k_1)}$, which must be positive. Thus, $m_1$ can have no local maxima as a function of $z_0$, so the optimal solution must be at a boundary, i.e., $z_0 = 0$ or $z_0 =1$.
Using the definitions in -, we see that $m_1(0) \geq m_1(1)$ if and only if $$d_1(0) + v_1(0) -k_1(0) - \frac{1}{c} \log(e^{c(d_1(0)-k_1(0))}+e^{c(v_1(0)-k_1(0))}-1) \geq$$ $$d_1(1) + v_1(1) -k_1(1) - \frac{1}{c} \log(e^{c(d_1(1)-k_1(1))}+e^{c(v_1(1)-k_1(1))}-1)$$ Note that with a slight abuse of notation, we augment $(d_1, k_1, v_1)$ by $z_t$ to compare them for $z_t = 0$ and $z_t = 1$. We have $d_1(0) = d_1(1) := d_1$ and $k_1(0) = 0$, so this condition reduces to $$\begin{aligned}
0 &\leq e^{c(v_1(0) - v_1(1) + k_1(1))} - \frac{e^{cv_1(0)}+e^{cd_1} -1}{e^{c(v_1(1)-k_1(1))}+e^{c(d_1-k_1(1))} -1} \\
0&\leq e^{c(1-\alpha+\alpha')m_0}(e^{cd_1}e^{-c(1-\gamma)m_0} -1) - e^{cd_1} +1 \label{eq:proof:myopic1} \\
0&\leq e^{cd_1}(e^{c(\gamma + \alpha' -\alpha)} - 1) - (e^{c(1-\alpha +\alpha')}-1)
\end{aligned}$$ Line comes from multiplying by the denominator and using the equalities $v_1(0) - v_1(1) + k_1(1) = (1-\alpha+\alpha')m_0$ and $k_1(1) = (1-\gamma)m_0$. Algebraically, this is equivalent to the two conditions in .
[Proof of Claim \[claim:statecoupling\]]{} \[proof:lemma:myopicisbest\] We proceed via induction. We have as a base case $d_0 \geq d_0'$, $v_0 \geq v_0'$, and $m_0 \geq m_0'$. Suppose this holds for all periods $t \leq \tau$. We will show it holds for $t = \tau+1$. Based on the dynamics described in -, since the platform’s policy decision in period $\tau$ is assumed to be fixed, we have $$\begin{aligned}
d_{\tau+1} &= (1-\beta)d_\tau + \beta'm_\tau \quad \quad \geq \quad (1-\beta)d_\tau' + \beta'm_\tau' &\geq d_{\tau+1}' \\
k_{\tau+1} &= (1-\gamma)z_\tau m_\tau \qquad \qquad \geq \quad (1-\gamma)z_\tau m_\tau' &\geq k_{\tau+1}' \\
v_{\tau+1} &= (1-\alpha)v_\tau + (\alpha'+\gamma')m_\tau + (\alpha - \alpha' - \gamma)z_\tau m_\tau & \nonumber \\ &\geq (1-\alpha)v_\tau' + (\alpha'+\gamma')m_\tau' + (\alpha - \alpha' - \gamma)z_\tau m_\tau' &\geq v_{\tau+1}'
\end{aligned}$$
To complete the proof of the claim, we need to show that $m_{\tau+1}$ is non-decreasing in $d_\tau, k_\tau$, and $m_\tau$. By the chain rule, it is sufficient to show that $m_{\tau+1}$ as defined in is increasing in $d_{\tau+1}$, $v_{\tau+1}$, and $k_{\tau+1}$. We establish this directly, using the matching function defined in . $$\begin{aligned}
\frac{\partial m_{\tau+1}}{\partial d_{\tau+1}} &= \ \ 1 - \frac{e^{c(d_{\tau+1}-k_{\tau+1})}}{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}-1} &= \frac{e^{c(v_{\tau+1}-k_{\tau+1})}-1}{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}-1} \in [0,1] \\
\frac{\partial m_{\tau+1}}{\partial v_{\tau+1}} &= \ \ 1 - \frac{e^{c(v_{\tau+1}-k_{\tau+1})}}{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}-1} &= \frac{e^{c(d_{\tau+1}-k_{\tau+1})}-1}{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}-1} \in [0,1] \\
\frac{\partial m_{\tau+1}}{\partial k_{\tau+1}} &= -1 + \frac{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}}{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}-1} &= \frac{1}{e^{c(d_{\tau+1}-k_{\tau+1})}+e^{c(v_{\tau+1}-k_{\tau+1})}-1} \in [0,1]
\end{aligned}$$ This shows that $m_{\tau + 1} \geq m_{\tau+1}'$, which completes the proof by induction.
[Proof of Theorem \[thm:optregimeB\]]{} \[proof:thm:optregimeB\] To prove this theorem, we first show that has a unique solution in the [AGN regime]{}. Then we show that if $d_{\tau} \geq \bar{d}$, $z_{\tau} = 0$ is optimal. Finally, we show that the thresholds define an absorbing state, i.e., if $d_{\tau} \geq \bar{d}$ and $v_{\tau} \geq \bar{v}$, then $d_{\tau+1} \geq \bar{d}$ and $v_{\tau+1} \geq \bar{v}$. This implies that $d_t \geq \bar{d}$ for all $t \in [T] \setminus [\tau-1]$, so in those periods, $z_t = 0$ must be optimal.
To show that has a unique solution, it is sufficient to prove that the right hand sides of and have slope less than $1$ and that the right hand sides of and have slopes in $[0,1]$ with respect to the decision variables. In that case, must be tight at any optimal solution. Applying the chain rule, the right hand sides of - must have slopes less than $1$ as a function of $d$. Thus, a linear search for the first value of $d$ to satisfy all three constraints solves the problem.
Clearly the right hand side of is increasing with slope less than $1$ (it is a constant). For the other three constraints, we can exploit their common structure. We will prove that the general function $f(x) = \frac{1}{c}\log(\frac{e^{c\zeta_1x}-1}{e^{c\zeta_2x} - 1})$ satisfies $f'(x) < 1$ when $0 < \zeta_2 < \zeta_1 < 1+ \zeta_2$. $$\begin{aligned}
\frac{d f}{dx} &= \frac{\zeta_1 e^{c\zeta_1x}}{e^{c\zeta_1x} - 1} - \frac{\zeta_2 e^{c\zeta_2x}}{e^{c\zeta_2x} - 1} \label{eq:proof:mms:0} \\
&= \zeta_1 - \zeta_2 + \frac{\zeta_1}{e^{c\zeta_1x} - 1} - \frac{\zeta_2}{e^{c\zeta_2x} - 1} \\
& \leq \zeta_1 - \zeta_2 \label{eq:proof:mms}\\
&< 1
\end{aligned}$$ Line comes from noting that $\frac{\zeta}{e^{cx \zeta}-1}$ is decreasing in $\zeta$, so $\frac{\zeta_1}{e^{c\zeta_1x} - 1} - \frac{\zeta_2}{e^{c\zeta_2x} - 1} \leq 0$. This immediately shows that the right hand sides of and both have slopes less than $1$. Similarly, the right hand side of must have a slope less than $\frac{1-\beta'}{1-\beta} < 1$.
We also note that $\zeta + \frac{\zeta}{e^{cx \zeta}-1}$ is increasing in $\zeta$, which means that according to , $\frac{d f}{dx} \geq 0$. This implies that and are weakly increasing in both decision variables. This completes the proof that there is a unique solution to $\eqref{eq:mms}$ which can be found via a linear search on $d$.
We now show that if $d_\tau \geq \bar{d}$, $z_\tau = 0$ is optimal. By appealing to Lemma \[lemma:myopicisbest\] in the [AGN regime]{}, it is sufficient to show that $d_{\tau+1} \geq \frac{1}{c}\log\left(\frac{ e^{c(1+\alpha'-\alpha)m_\tau} -1}{e^{c( \gamma + \alpha' - \alpha)m_\tau} - 1}\right)$. Using to rewrite $d_{\tau+1}$ in terms of $d_\tau$ and $m_\tau$, we find that the above condition is equivalent to $$d_\tau \geq \frac{1}{c(1-\beta)}\log\left(\frac{ e^{c(1+\alpha'-\alpha)m_\tau} -1}{e^{c( \gamma + \alpha' - \alpha)m_\tau} - 1}\right) - \frac{\beta'}{1-\beta}m_\tau$$ This threshold is convex in $m_\tau$, so its maximum value over its domain must occur when $m_\tau = d_\tau$ or when $m_\tau = 0$. Constraint ensures that $d_\tau$ exceeds the threshold when $m_\tau = d_\tau$, while constraint ensures that $d_\tau$ exceeds the threshold when $m_\tau = 0$ (recall that we defined the threshold to equal its limiting value when $m_\tau = 0$). This means that if these two constraints are satisfied, the optimal single-period policy is $z_\tau = 0$. Since we are in the [AGN regime]{}, applying Lemma \[lemma:myopicisbest\] proves that the optimal policy is $z_\tau = 0$.
We now show that $\bar{d}$ and $\bar{v}$ define an absorbing state when the platform follows a policy of no adoption. We define $$\bar{m} := {s}(\bar{d}, \bar{v}) = \bar{d}+\bar{v} - \frac{1}{c} \log(e^{c\bar{d}}+e^{c\bar{v}}-1) \label{eq:proof:mms1.5}$$ Since the matching function in is increasing in each side of the market, $\bar{m}$ represents the minimum number of matches when $d_\tau \geq \bar{d}$ and $v_\tau \geq \bar{v}$.
By constraint , $\bar{d} \geq \frac{1}{c} \log \left(\frac{e^{p\bar{v}}-1}{e^{p\bar{v}(1-\frac{\alpha}{\alpha'+\gamma'})}-1} \right)$. Rearranging terms, we have $$\begin{aligned}
e^{c(\bar{d} + \bar{v}(1-\frac{\alpha}{\alpha'+\gamma'}))}
&\geq e^{c\bar{d}} + e^{c\bar{v}} - 1 \\
&\geq e^{c(\bar{d} + \bar{v} - \bar{m})} \label{eq:proof:mms2}
\end{aligned}$$ Line comes from rearranging the definition of $\bar{m}$ in . This inequality is equivalent to $\bar{d} + \bar{v}(1-\frac{\alpha}{\alpha'+\gamma'} \geq \bar{d} + \bar{v} - \bar{m}$. Rearranging terms, we see that $$\begin{aligned}
\bar{v} &\leq (1-\alpha)\bar{v} + (\alpha'+\gamma')\bar{m}
\end{aligned}$$ If $v_\tau \geq \bar{v}$ and $d_\tau \geq \bar{v}$, we must have $m_\tau \geq \bar{m}$ since the matching function is increasing in both sides of the matching market. Thus, $$\bar{v} \leq (1-\alpha)v_\tau + (\alpha'+\gamma')m_\tau = v_{\tau+1}(0)$$ An identical proof using constraint shows that $\bar{d} \leq d_{\tau+1}(0)$ when $v_\tau \geq \bar{v}$ and $d_\tau \geq \bar{v}$. Thus, when following a policy of no adoption, $\bar{v}$ and $\bar{d}$ define an absorbing state where no adoption remains the optimal policy.
[Proof of Theorem \[thm:approxregimeB\]]{} \[proof:thm:approxregimeB\] For ease of notation throughout the proof, we define constants $A_1:= \max\{1-\beta+\beta', 1-\alpha+\alpha'+\gamma'\}$, $A_2:= \min\{1-\beta+\beta', 1-\alpha+\alpha'+\gamma'\}$ and $A_3:=\max\{\beta', \alpha'+\gamma'\}$.
We first establish an upper bound on the value of the [platform design problem]{}. Then, for each period, we upper bound the difference between the number of completed [jobs]{} when following a no-adoption policy and the number of [jobs]{} completed in the upper bound. Ultimately, this allows us to lower bound the performance of a static no-adoption policy as a function of the upper bound on the value of the [platform design problem]{}.
Consider a modified setting where the matching in each period is equal to the small side in the matching market, e.g., $\hat{m_t}(d_t, k_t, v_t) = \min\{d_t, v_t\}$. We will call this the *[match-min]{}* setting, and we will use consistent notation to indicate the state variables in the [match-min]{} setting, e.g. $\hat{d}_{t+1} = (1-\beta)\hat{d_t} + \beta'\hat{m}_t$. In this setting, the platform’s decision at time $t$ only impacts $\hat{m}_{t+1}$ through its impact on $\hat{v}_{t+1}$. Thus, in the [AGN regime]{}, the platform’s optimal policy is to set $z_t = 0$ for all $t \in [T]$, which maximizes the growth of the [worker]{} side of the market.
We remark that the the size of the matching as given in is less than the minimum side of the market: $$\begin{aligned}
m_t(d_t, k_t, v_t) &= d_t + v_t - k_t - \frac{1}{c}\log(e^{c(d_t-k_t)} + e^{c(v_t-k_t)} -1) \\
&= \min\{d_t, v_t\} -\frac{1}{c}\log(1 + e^{c(\min\{d_t, v_t\}-\max\{d_t, v_t\})} - e^{-c(\max\{d_t, v_t\}-k_t)}) \label{eq:proof:approx1}\\
&\leq \min\{d_t, v_t\}
\end{aligned}$$ Thus, given the same initial state $(d_0, k_0, v_0)$ and the same policy choice $z_0$, we have $\hat{m}_1 \geq m_1$, $\hat{d}_1 \geq d_1$, and $\hat{v}_1 \geq v_1$. Using a similar argument as in the proof of Lemma \[lemma:myopicisbest\], these inequalities will continue to hold for all $t \in [T]$ as long as the policies $z_t$ are identical. As a consequence, the value of completed [jobs]{} in the [match-min]{} setting when $z_t = 0$ for all $t \in [T]$ is an upper bound on the value of the [platform design problem]{}. The value of this upper bound is given by $\sum_{t=1}^T \delta^{t-1}\hat{m}_t$.
We now place a lower bound on the ratio $\frac{\sum_{t=1}^T \delta^{t-1}m_t}{\sum_{t=1}^T \delta^{t-1}\hat{m}_t} = 1- \frac{\sum_{t=1}^T \delta^{t-1}(\hat{m}_t-m_t)}{\sum_{t=1}^T \delta^{t-1}\hat{m}_t}$, where $m_t$ represents the number of matches in period $t$ when the platform follows a static policy of no adoption and the matching output is determined by . We start by lower-bounding $\sum_{t=1}^T \delta^{t-1}\hat{m}_t$.
$$\begin{aligned}
\sum_{t=1}^T \delta^{t-1}\hat{m}_t = \sum_{t=1}^T \delta^{t-1}\min\{\hat{d}_t, \hat{v}_t\} &\geq \sum_{t=1}^T \delta^{t-1}(A_2)^t\min\{d_0, v_0\} =A_2\min\{d_0, v_0\} \frac{1-(A_2\delta)^{T}}{1-A_2\delta} \label{eq:proof:approxub}
\end{aligned}$$
Now we must place an upper bound on $\sum_{t=1}^T \delta^{t-1}(\hat{m}_t-m_t)$. From line , we see that $m_t \geq \min\{d_t, v_t\} - \frac{1}{c}\log(2)$. This allows us to upper bound the difference between $\hat{m}_t$ and $m_t$ in any period $t \in [T]$: $$\begin{aligned}
\hat{m}_t - m_t &\leq \min\{\hat{d}_t, \hat{v}_t\} - \min\{d_t, v_t\} + \frac{1}{c} \log(2) \\
&\leq \max\{\hat{d}_{t} - d_{t}, \hat{v}_{t} -v_{t}\} + \frac{1}{c} \log(2) \label{eq:proof:matchdiff}
\end{aligned}$$ To place an upper bound on $\max\{\hat{d}_{t} - d_{t}, \hat{v}_{t} -v_{t}\}$, note that when the platform follows a static policy of no adoption, the [worker]{} dynamics are given by $$\begin{aligned}
v_{t+1} &=(1-\alpha)v_t +(\alpha'+\gamma')m_t \\
&\geq (1-\alpha)v_t +(\alpha'+\gamma')(\min\{d_t, v_t\} - \frac{1}{c}\log(2))\\
&= (1-\alpha)\hat{v}_t +(\alpha'+\gamma')(\min\{\hat{d}_t, \hat{v}_t\})
-(1-\alpha)(\hat{v}_t - v_t) - \nonumber \\& \quad (\alpha'+\gamma')(\min\{\hat{d}_t, \hat{v}_t\} - \min\{d_t,v_t\}) - \frac{\alpha' + \gamma'}{c}\log(2) \\
&\geq \hat{v}_{t+1} -(1-\alpha)(\hat{v}_t - v_t)- (\alpha'+\gamma')(\max\{\hat{d}_t - d_t, \hat{v}_t -v_t\}) - \frac{\alpha' + \gamma'}{c}\log(2) \\
&\geq \hat{v}_{t+1} - (1-\alpha+\alpha'+\gamma')(\max\{\hat{d}_t - d_t, \hat{v}_t -v_t\}) - \frac{\alpha' + \gamma'}{c}\log(2)
\end{aligned}$$ Similarly, $$\begin{aligned}
d_{t+1} &\geq (1-\beta)d_t +(\beta')(\min\{d_t, v_t\} - \frac{1}{c}\log(2))\\
&= (1-\beta)\hat{d}_t +(\beta')(\min\{\hat{d}_t, \hat{v}_t\})
-(1-\beta)(\hat{d}_t - d_t) - \nonumber \\& \quad (\beta')(\min\{\hat{d}_t, \hat{v}_t\} - \min\{d_t,v_t\}) - \frac{\beta'}{c}\log(2) \\
&\geq \hat{d}_{t+1} -(1-\beta)(\hat{d}_t - d_t)- (\beta')(\max\{\hat{d}_t - d_t, \hat{v}_t -v_t\}) - \frac{\beta'}{c}\log(2) \\
&\geq \hat{d}_{t+1} - (1-\beta+\beta')(\max\{\hat{d}_t - d_t, \hat{v}_t -v_t\}) - \frac{\beta'}{c}\log(2)
\end{aligned}$$ We can combine these two inequalities to yield $$\begin{aligned}
\max\{\hat{d}_{t+1} - d_{t+1}, \hat{v}_{t+1} -v_{t+1}\} &\leq \max\{(1-\beta + \beta'), (1-\alpha +\alpha' +\gamma')\}\max\{\hat{d}_t - d_t, \hat{v}_t -v_t\} \nonumber \\ &\quad + \frac{\max\{\beta', \alpha'+\gamma'\}}{c} \log (2) \\
&=A_1\max\{\hat{d}_t - d_t, \hat{v}_t -v_t\}+\frac{A_3}{c}\log(2)
\end{aligned}$$
Given initial condition $\max\{\hat{d}_{0} - d_{0}, \hat{v}_{0} -v_{0}\}=0$, we have $$\begin{aligned}
\max\{\hat{d}_{t} - d_{t}, \hat{v}_{t} -v_{t}\} &\leq \frac{A_3\log(2)}{c} \sum_{\tau=0}^{t-1} A_1^\tau \\
&= \frac{A_3\log(2) (A_1^t-1)}{c(A_1-1)}
\end{aligned}$$ Taking a discounted sum of over all $t \in [T]$, the difference in the total value of completed jobs is given by $$\begin{aligned}
\sum_{t=1}^T \delta^{t-1}(\hat{m}_t - m_t) &\leq \sum_{t=1}^T \delta^{t-1}\left(\frac{\log(2)}{c} + \frac{A_3\log(2) (A_1^t-1)}{c(A_1-1)}\right) \\
&= \frac{\log(2)}{c(A_1-1)}\sum_{t=1}^T \delta^{t-1}(A_1 - 1 - A_3 +A_3A_1^t)) \\
&= \frac{\log(2)}{c(A_1-1)}\left(\frac{(A_1-1-A_3 )(1-\delta^{T})}{1-\delta} + \frac{A_1A_3(1-(\delta A_1)^{T})}{1-\delta A_1} \right) \\
&\leq \frac{\log(2)}{c(1-\delta)} + \frac{A_3 \log(2) (1-(\delta A_1)^{T})}{c(1-\delta)(1-\delta A_1)}
\end{aligned}$$
Using this result and the upper bound from , we can place a lower bound on the approximation factor of a no-adoption policy:
$$\begin{aligned}
\frac{\sum_{t=1}^T \delta^{t-1}(m_t)}{\sum_{t=1}^T \delta^{t-1}(\hat{m}_t)} & = 1 - \frac{\sum_{t=1}^T \delta^{t-1}(\hat{m_t} - m_t)}{\sum_{t=1}^T \delta^{t-1}(\hat{m}_t)} \\
&\geq 1 - \frac{\frac{\log(2)}{c(1-\delta)} + \frac{A_3 \log(2) (1-(\delta A_1)^{T})}{c(1-\delta)(1-\delta A_1)}}{A_2\min\{d_0, v_0\} \frac{1-(A_2\delta)^{T}}{1-A_2\delta}} \\
&\geq 1 - \frac{\log(2)(\frac{1}{1-(\delta A_1)^T} + \frac{A_3}{1-\delta A_1})}{c\min\{d_0, v_0\}}
\end{aligned}$$
In these steps, we require $\delta A_1 < 1$ for the lower bounds to hold. This is equivalent to the approximation factor reported in Theorem \[thm:approxregimeB\].
false $$\begin{aligned}
\sum_{t=1}^T \delta^{t-1}(\hat{m}_t - m_t) &\leq \sum_{t=1}^T \delta^{t-1}\left(\frac{\log(2)}{c} + \frac{A_3\log(2)}{c} \sum_{\tau=0}^{t-1} A_1^\tau\right) \\
&\leq \frac{\log(2)}{c(1-\delta)} +\frac{A_3\log(2)}{c} \sum_{t=1}^\infty \delta^{t-1} \sum_{\tau=0}^{t-1} A_1^\tau \\
&= \frac{\log(2)}{c(1-\delta)} + \frac{A_3\log(2)}{c(1-\delta)(1-\delta A_1)}\end{aligned}$$
$$\begin{aligned}
\frac{\sum_{t=1}^T \delta^{t-1}(m_t)}{\sum_{t=1}^T \delta^{t-1}(\hat{m}_t)} & = 1 - \frac{\sum_{t=1}^T \delta^{t-1}(\hat{m_t} - m_t)}{\sum_{t=1}^T \delta^{t-1}(\hat{m}_t)} \\
&\geq 1 - \frac{\frac{\log(2)}{c(1-\delta)} + \frac{A_3\log(2)}{c(1-\delta)(1-\delta A_1)}}{\frac{A_2\min\{d_0, v_0\}}{1-A_2\delta}} \\
&\geq 1 - \frac{\log(2)(1 + \frac{A_3}{1-\delta A_1})}{c\min\{d_0, v_0\}}\end{aligned}$$
[^1]: We estimate this probability via a regression analysis, using fixed effects and controls for confounding variables. The $p$-value is $0.053$. The effect size and significance level would increase if the controls were removed. Using the same procedure for adopters, the effect of options is minimal and not statistically significant.
[^2]: We use the terms “[one-time]{} matches” and “non-adoptions” interchangeably.
[^3]: For ease of notation, for any $a \in \mathbb{N}$, we use $[a]$ to refer to the set $\{0, 1, 2, \dots, a\}$.
[^4]: In the setting we consider, the platform can implement this level of control by making minor adjustments to its website. In settings where this level of control is impractical, the structure of our results is maintained.
[^5]: If an adopter leaves the platform, we assume its corresponding [job]{} will enter the spot market in the subsequent period. This simplifying assumption is reasonable in the setting we consider, where workers are homogeneous.
[^6]: We highlight that $\alpha'$ can be thought of as a decrease in workers’ dropout rate or as an increase in the rate at which workers sign up for jobs.
[^7]: We note that for some $n$, it may not be possible to have a system with these initial conditions, e.g. $nd_0$ is not an integer. Thus a more accurate statement is to consider the subsequence $n_1, n_2, \dots$ such that the initial conditions are all integers and $\frac{c}{n_1} \leq 1$. However, for the sake of brevity we do not mention this in our statements.
[^8]: In the statement of this proposition, we use capital (resp. lower case) letters for our state variables (resp. approximations) to highlight that they are random (resp. deterministic) variables.
[^9]: Forcing each new match to be an adopted match may not be achievable in practice. We remark that all of our results go through with minor adjustments when restricting the domain of $z_t$ to $[\underline{z},\bar{z}]$, though we omit the details for the sake of brevity.
[^10]: This critical property ensures that there is randomness in matching, or equivalently, that the certainty provided by an adopted match is valuable relative to a corresponding increase in the size of the spot market.
[^11]: With a slight abuse of notation, we augment $m_1$ by $z_0$ to highlight the role of the decision variable.
[^12]: We expect the exact threshold to depend on the gap $\gamma-\alpha'+\alpha$.
[^13]: We remark that the threshold depends on the model primitives (including the gap between $\gamma$ and $\alpha - \alpha'$) as well as $m_0$, which is uniquely determined by the initial state.
[^14]: Note that in the presence of multiple optimal policies, we follow the convention of choosing the myopically optimal one.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Gianluca Grignani\
Dipartimento di Fisica and Sezione I.N.F.N., Università di Perugia, Via A. Pascoli I-06123, Perugia, Italia.
- |
Marta Orselli\
NORDITA, Blegdamsvej 17, DK-2100 Copenhagen, Denmark.
- |
Bojan Ramadanovic, Gordon W. Semenoff and Donovan Young\
Department of Physics and Astronomy, University of British Columbia, 6224 Agricultural Road, Vancouver, British Columbia V6T 1Z1 Canada.
title: 'AdS/CFT v.s. String Loops'
---
Ł § ł ø Ø v
Introduction and Conclusions
============================
The AdS/CFT correspondence [@Juan; @Gubser:1998bc; @Witten:1998qj] has provided one explicit example of the long conjectured duality between gauge fields and strings. One of the most important testing grounds for this correspondence is string theory in the pp-wave geometry and its mapping to the BMN limit of Yang-Mills theory.
The pp-wave geometry is produced by taking the Penrose limit of AdS$_5\times$S$^5$ [@Blau:2001ne; @Blau:2002dy]. On that geometry, non-interacting IIB string theory is explicitly solvable and the complete spectrum of free strings can be found [@Metsaev]. The corresponding BMN limit of $\mathcal{N}=4$ super-Yang-Mills theory can be taken by identifying the appropriate operators [@Berenstein:2002jq] and taking a large quantum number limit. The planar limit of Yang-Mills theory corresponds to non-interacting strings and the planar spectrum of the Yang-Mills dilatation operator, which is dual to the string Hamiltonian, can be computed perturbatively [@Berenstein:2002jq]-[@Santambrogio:2002sb]. As far as these computations have been done, the result shows beautiful agreement between planar Yang-Mills and non-interacting strings. This agreement has been extended to scenarios beyond the BMN limit [@Gubser:2002tv; @Callan:2003xr] and to the non-perturbative sector [@Green:2005pg; @Green:2005rh] and has led to many promising insights.
One of those insights has been the recognition that the problem of computing dimensions of composite operators in $\mathcal{N}=4$ super-Yang-Mills theory can be mapped onto integrable spin chains [@Minahan:2002ve]-[@Beisert:2003xu]. The string theory sigma model on $AdS_5\times S^5$ also has an integrable structure [@Bena:2003wd] and much progress has been made to the point that a complete matching of the precise details of planar Yang-Mills and and non-interacting strings on the full $AdS_5\times
S^5$ background is a possibility that is sometimes contemplated [@Beisert:2005bm]-[@Rej:2005qt].
However, in spite of this optimistic outlook, beyond the planar limit of Yang-Mills theory and non-interacting string theory, there has been very little success in checking the AdS/CFT correspondence, even in the BMN limit. For example, the Yang-Mills prediction for the string-loop corrections to energies of 2-impurity BMN states were computed early on [@Kristjansen:2002bb; @Constable:2002hw; @Beisert:2002bb; @Constable:2002vq]. The gauge theory prediction for the BMN energy of a 2-impurity state is [^1]
$$\label{gauge}
\Delta-J =
2\left(1+{\small\frac{1}{2}}\lambda'n^2-{\small\frac{1}{8}}{\lambda'}^2n^4+\ldots
\right)+\frac{g_2^2}{4\pi^2} \left(\frac{1}{12}+\frac{35}{32\pi^2
n^2}\right)\left( \lambda'-\frac{1}{2}{\lambda'}^2 n^2\right)+\ldots$$
Attempts to produce a result which matches this one using string theory have spawned a large literature [@Spradlin:2002ar]-[@Dobashi:2006fu] the best available published computation using light-cone string field theory is due to Gutjahr and Pankiewicz [@Gutjahr:2004dv] (their Eq. (4.17)) $$\begin{aligned}
\label{string}
\frac{p^-}{\mu}=2\sqrt{1+\lambda'n^2}+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\nonumber
\\ +\frac{g_2^2\lambda'}{4\pi^2}\left( \left(\frac{1}{24}+\frac{65}{64\pi^2
n^2}\right) -\frac{3}{16\pi^2}{\lambda'}^{\small\frac{1}{2}}
-\frac{n^2}{2}\left(\frac{1}{24}+\frac{89}{64\pi^2n^2}\right){\lambda'}
+\frac{9n^2}{32\pi^2}{\lambda'}^{\small\frac{3}{2}}+\ldots\right)\end{aligned}$$ This computation, as did those which preceded it, uses an unjustified truncation of the string vertex to the 2-impurity channel. It clearly does not match the gauge theory result (\[gauge\]). The three-string vertex actually has an arbitrary pre-factor, the choice of which gives an arbitrary re-scaling of the entire expression in Eq. (\[string\]). The pre-factor can thus be chosen so that either the first or the second term in the leading order $\lambda'$ contribution agrees with gauge theory, but not both. There are other differences in the terms beyond the leading order.
On the other hand, in spite of its shortcomings, the formula in (\[string\]) has some remarkable features. The natural expansion parameter on the string side is ${\lambda'}^{\small\frac{1}{2}}$. In (\[string\]), the naive leading term that one would expect from power-counting, $\sim {\lambda'}^{\small\frac{1}{2}}$, is absent. It was argued that this is generally so in Ref. [@Grignani:2005yv]. The leading non-zero term, of order $\lambda'$, has contributions of the same functional form in $n$ as the gauge theory result, it is only the coefficients that are wrong. The bigger problem begins with the order ${\lambda'}^{\small\frac{3}{2}}$ term which is clearly absent in the gauge theory, where the expansion parameter is in integer powers of $\lambda'$. One might argue that such a fractional power is generated non-perturbatively, by re-summing logarithmic divergent diagrams for example, and it could appear in principle. This does happen elsewhere, for example in the expansion of the free energy of Yang-Mills theory at finite temperature. However, the gauge theory result seems to be free of infrared problems, this has been checked explicitly to at least order ${\lambda'}^2$, so it is difficult to see how a term of order ${\lambda'}^{\small\frac{3}{2}}$ could occur.
In the present paper, we will repeat the light-cone string field theory computation that led to (\[string\]), using the same truncation to the 2-impurity channel, and a modified form of the pp-wave background string vertex which was suggested by Dobashi and Yoneya in Ref. [@Dobashi:2004nm]. Other details of the computation are identical to those in Ref. [@Gutjahr:2004dv] which led to (\[string\]).[^2] Our result will be $$\begin{aligned}
\label{ours}
\frac{p^-}{\mu}&=&2\sqrt{1+\lambda'n^2}+ \nonumber
\\ &+&\frac{g_2^2\lambda'}{4\pi^2}|f|^2 \left(\frac{1}{12}+\frac{35}{32\pi^2
n^2}\right)\left( \frac{3}{4}-\frac{n^2}{2}\lambda'+{\cal
O}({\lambda'}^2)\right)\end{aligned}$$ where $f$ is the unknown pre-factor of the vertex. Note that now, remarkably, if we set the pre-factor $|f|^2=\frac{4}{3}$, the order $\lambda'$ term agrees with gauge theory. The order ${\lambda'}^2$ term, however, does not. Further to this, the fractional powers of $\lambda'$ are absent, at least up to order $7/2$.
The essential new aspect of this computation is the use of the Dobashi-Yoneya vertex. Unlike the case of Minkowski space, on the pp-wave background there are competing proposals for the three string vertex. The original one [@Spradlin:2002ar; @Spradlin:2002rv; @Pankiewicz:2002gs; @Pankiewicz:2002tg; @Pankiewicz:2003kj] (which we will call the SVPS vertex) was fixed using the supersymmetry algebra up to a pre-factor function of the light-cone momentum (which is $f$ in Eq. (\[ours\])). Another vertex was proposed in Ref. [@DiVecchia:2003yp] and we will call it the DVPPRT vertex. The DVPPRT vertex solves the supersymmetry algebra in the simplest possible way, by acting upon the oscillator representation of the Dirac delta function which enforces world-sheet locality by the [*quadratic*]{} Hamiltonian and supercharge. This vertex is trivial in Minkowski space, but is non-trivial in the pp-wave background.
Then, in Ref. [@Dobashi:2004nm], Dobashi and Yoneya proposed another form for the cubic Hamiltonian and supercharge based on consistency with the AdS/CFT holographic relations for three-point functions [@Gubser:1998bc; @Witten:1998qj] and their comparison with supergravity. This “holographic” vertex, which we shall call the DY vertex, is an equal-weighted average of the original SVPS vertex and the DVPPRT vertex: DY=${\small\frac{1}{2}}$SVPS+${\small\frac{1}{2}}$DVPPRT. It, and the four-string contact term that is generated using the supersymmetry algebra, are the vertices that are used in deriving Eq. (\[ours\]). A correction to the DY vertex based on the supersymmetry algebra was suggested in Ref. [@Lee:2004cq]. Because of the truncation to the 2-impurity channel, this modification does not influence the computations in the present paper.
Though the result (\[ours\]) is a big improvement on the previous one, it is still not in complete agreement with the gauge theory computation. It disagrees at order ${\lambda'}^2$. There might (or might not) be a simple reason for this disagreement. We have not performed computations beyond the 2-impurity channel. It could be that higher impurity channels contribute only to orders ${\lambda'}^2$ or higher, but do not influence the order $\lambda'$ contribution. This would require a miraculous cancelation of a number of orders in the small ${\lambda'}^{\small\frac{1}{2}}$ expansion. After all, from power counting and the generic structure of the amplitude, one would expect that higher impurities begin to contribute at order ${\lambda'}^{\small\frac{1}{2}}$. In previous work, we have shown that this leading order cancels [@Grignani:2005yv]. There, it was associated with cancelation of divergences, which were also generic, and supersymmetry played an important role. Examining whether this cancelation could also occur at orders $\lambda'$ and ${\lambda'}^{\small\frac{3}{2}}$ is a challenge that has not been addressed yet. A careful check of this possibility would be very interesting.
There is another possibility for discrepancy. In all computations to date, the contact term with the supercharge $g_2^2Q_4$ has been assumed to not contribute. Indeed, the supersymmetry algebra shows that the contact term is (schematically) $g_2^2H_4=g_2^2\{Q_3,Q_3\}
+ g_2^2\{Q_2,Q_4\} + g_2^2\{Q_4,Q_2\}$ and only the first term on the right-hand side has been used in all computations. Generally, these contact terms are needed to cancel divergences arising from iterations of lower order vertices [@Greensite:1987hm; @Green:1987qu]. In principle, $Q_4$ could be determined by finding multi-string matrix elements of the supersymmetry algebra. To our knowledge, this has not been attempted on the pp-wave background. We only observe that $Q_4$ is not needed to cancel divergences in any of the quantities that we compute. (This was also found on the Minkowski background in Ref. [@Green:1987qu].) However, we cannot rule out its having a non-zero finite contribution that would affect our results.
One further observation that we can make is that, we could consider any linear combination of the SVPS and DVPPRT vertices: $\alpha$SVPS+$\beta$DVPPRT. In this case, it would seem that, by using the supersymmetry algebra, one could consistently construct higher order contact terms in the Hamiltonian and supercharges, so that this is also a viable possibility for the vertex. In particular, we will see that divergences in the energy shifts of 2-impurity states cancel for any values of $\alpha$ and $\beta$. However, if we use this vertex to compute the energy shift, we find a result that agrees with gauge theory (\[gauge\]) to the leading order $\lambda'$ only for the particular combination in the DY vertex, that is only when $\alpha=\beta={\small\frac{1}{2}}$.
There is another intriguing and unexplained feature of these results, which was observed in Ref. [@Roiban:2002xr]. Consider the expansion of the string field theory Hamiltonian into free (quadratic) and interacting – cubic, quartic, etc. terms, $H=H_2+g_2H_3+g_2^2H_4+\ldots$ and the expression for second order quantum mechanical perturbation theory which is used to compute (\[string\]), $$\label{hhhhh} \delta E^{(2)} ~=~ g_2^2<\psi_0|
H_3\frac{1}{E_0-H_2}H_3|\psi_0>+g_2^2<\psi_0|H_4|\psi_0>$$ If, in the computation which arrives at (\[string\]), we change the terms on the right-hand-side of (\[hhhhh\]) by a relative factor of 2, either multiplying the first term by $\frac{1}{2}$ or the second term by $2$, then the order $\lambda'$ term would be different from that quoted in (\[string\]) and in that case the pre-factor could be chosen so that the order $\lambda'$ term agrees with gauge theory. Here, we observe that this interesting fact persists in (\[string\]) to higher orders. In that case, with factor of 2 and the same choice of prefactor the order $\l'^2$ term also agrees with gauge theory, and the $\l'^{3/2}$ term vanishes. In addition, this intriguing fact persists in the computation of (\[ours\]), if one inserts a relative factor of 2 in (\[hhhhh\]), (\[ours\]) is modified so that it agrees with gauge theory up to and including order ${\lambda'}^2$ and the coefficients of the fractional powers with exponents $3/2$ or $5/2$ still vanish. At this point, we have no explanation for this fact. Inserting the factor of 2 is definitely not mathematically correct here. Aside from the violence it would do to quantum mechanical perturbation theory, it would upset the divergence cancelation that was found in Ref. [@Grignani:2005yv], for example. The reason, if any, for this numerological coincidence remains a mystery.
In the remainder of this Paper, we will outline the computation that leads to Eq. (\[ours\]). The notation and techniques are identical to those used in Ref. [@Gutjahr:2004dv] and Ref. [@Grignani:2005yv] and we defer to them for the details.
Divergence Cancelation
======================
The light-cone energy of the two-oscillator free string state on the pp-wave background is $p^-=2\mu\sqrt{1+\lambda'n^2}$. This matches the energy of the two impurity BMN operator in planar Yang-Mills theory. The energy-shift of these states due to string loop corrections is calculated in second order quantum mechanical perturbation theory using the formula in Eq. (\[hhhhh\]). We will call the first term in (\[hhhhh\]) the “$H_3$ term” and the second the “contact term”.
In our previous paper [@Grignani:2005yv] we showed that, in the computation of the energy-shifts of some two-impurity states using SVPS vertex, the $H_3$ and contact terms individually contain logarithmically divergent sums over intermediate state mode numbers. These divergences were shown to always cancel, leaving a finite result which leads as $g_2^2\l'$. This behavior was shown to be generic, and to exist at arbitrary order in intermediate state impurities. This was important because, of course it is necessary to obtain finite amplitudes. In addition, it is also the mechanism whereby the leading order ${\lambda'}^{\small\frac{1}{2}}$ contributions cancel.
We shall now show that this mechanism is at play for the DVPPRT vertices, and that any linear combination of the SVPS and DVPPRT vertices will similarly be divergence free. The special choice of an equal weighted average - the DY vertex - is thus well behaved.
The simplest method to understand the divergence cancelation is to consider the energy shift of the two-impurity trace state |\[[**1**]{}, [**1**]{}\]. = \^[i]{}\_n \^[i ]{}\_[-n]{} |\[singlet\]restricted to the impurity conserving channel. For details of this computation we refer the reader to [@Grignani:2005yv] and for details of definitions and notation to Ref. [@Gutjahr:2004dv] and other literature quoted there. The DVPPRT vertex is given by the following expressions [@DiVecchia:2003yp],
|H\^D\_3=&&- g\_2f(\_3,) |V,|Q\^D\_[3\_1\_2]{}=&& g\_2f(\_3, ) (Z\_[\_1\_2]{}K\_[\_1]{}\^[\_1]{}- i Y\_[\_1\_2]{}K\_[2]{}\^[\_2]{})|V,|Q\^D\_[3\_1\_2]{}=&& g\_2|f(\_3, ) (Y\_[\_1\_2]{}K\_[\_1]{}\^[\_1]{}-iZ\_[\_1\_2]{}K\_[\_2]{}\^[\_2]{})|V. \[divecchia\]
Unlike the SVPS case, the $H_3$ divergence does not stem from the two-bosonic-impurity intermediate state. This can be traced to the substitution of $K^2 + \wt K^2$ for $K\,\wt K$ in the $H_3$ prefactor. There is, however, another divergence that was not present in the SVPS case. It is due to the contribution coming from matrix elements with two fermionic impurities in the intermediate state. In particular, the relevant matrix elements are given by
\[ferm\] && \_3 | \^i\_[n]{} \^i\_[-n]{} \_2 | \_1 | \_[p(1)]{}\^[\_1\_2]{} \_[-p(1)\_1\_2]{} |H\^D\_3 =&& 4g\_2 r(1-r)( + )([Q]{}\_[-pp]{}\^[11]{}. .- [Q]{}\_[p-p]{}\^[11]{})[N]{}\_[-nn]{}\^[33]{}\_[\_1]{}\^[\_1]{} \_[\_2]{}\^[\_2]{}
and similarly for the intermediate state with dotted indices. The divergent contribution to the energy shift coming from these matrix elements is found by taking the large $p$ limits of the summands in (\[hhhhh\]). One finds
$$\label{diveccH3div}
\delta E^{\mbox{div}}_{H^D_3} \sim -\frac{1}{2}\int_0^1 dr\,
\,\frac{g_2^2\,r(1-r)}{r\,|\alpha_3|\,\pi^2} \left( {\widetilde
N}^{3\,3}_{n\,-n} \right)^2 \, \sum_p \frac{1}{|p|}$$
The contribution from the contact term stems from the following matrix element,
\[contactmed\] &&( g\_2 )\^[-1]{} \_3 | \^i\_[n]{} \^i\_[-n]{} \_2 | \_1 | \^[K(1)]{}\_[p]{} \^[(1) \_1\_2]{}\_[-p]{} |Q\^D\_[3 \_1 \_2]{} =&& 2( G\_[|p|]{}\^[(1)]{} K\_[-n]{}\^[(3)]{} [N]{}\^[31]{}\_[np]{} + G\_[|p|]{}\^[(1)]{} K\_[n]{}\^[(3)]{} [N]{}\^[31]{}\_[-np]{} )( \^k)\^\_[ \_1 ]{} \^[[\_2]{}]{}\_[[\_2]{}]{} +8 G\_[|p|]{}\^[(1)]{} K\_[p]{}\^[(1)]{} [N]{}\^[33]{}\_[n-n]{} ( \^K)\^\_[ ]{} \^\_.
The divergent contribution to the energy shift is found to be,
$$\delta E^{\mbox{div}}_{H^D_4} \sim +\int_0^1 dr\,
\,\frac{g_2^2\,r(1-r)}{r\,|\alpha_3|\,\pi^2} \left( {\widetilde
N}^{3\,3}_{n\,-n} \right)^2 \, \sum_{p>0} \frac{1}{p}$$
Noting that in the $H^D_3$ contribution the divergence is found for both positive and negative $p$, while in the $H^D_4$ contribution the divergence occurs only for negative $p$, and hence a relative factor of 2 is induced in the $H^D_3$ term, one sees that the logarithmically divergent sums cancel identically between the $H^D_3$ and contact terms, leaving a convergent sum. This result can be generalized to arbitrary impurity channels, as was done for the SVPS case in [@Grignani:2005yv].
We now show that an arbitrary linear combination of the SVPS and DVPPRT vertices,
H\^N\_3 = H\^S\_3 + H\^D\_3\
Q\^N\_3 = Q\^S\_3 + Q\^D\_3
similarly yields a finite energy shift. The divergence stemming from the $H_3$ term is simply $\a^2$ times the SVPS $H_3$ divergence plus $\b^2$ times (\[diveccH3div\]). The reason is simple - the SVPS divergence stems from an entirely bosonic intermediate state, while (\[diveccH3div\]) results from an entirely fermionic one. This precludes any divergences arising from cross terms. Referring the reader to equation (2.7) of [@Grignani:2005yv], we note that the SVPS divergence is exactly equal to (\[diveccH3div\]), therefore we have,
E\^\_[H\^N\_3]{} \~-(\^2+\^2) \_0\^1 dr ( [N]{}\^[33]{}\_[n-n]{} )\^2 \_p
The pieces of the SVPS $Q_3$ relevant to a two-impurity channel calculation are exactly $Q_3^D$ with $K \lr \wt K$, see again [@Grignani:2005yv].
\[contactmey\] &&( g\_2 )\^[-1]{} \_3 | \^i\_[n]{} \^i\_[-n]{} \_2 | \_1 | \^[K(1)]{}\_[p]{} \^[(1) \_1\_2]{}\_[-p]{} |Q\^Y\_[3 \_1 \_2]{} =&& 2 G\_[|p|]{}\^[(1)]{} ( ( \^k)\^\_[ \_1 ]{} \^[[\_2]{}]{}\_[[ \_2]{}]{}&& +4 (K\_[p]{}\^[(1)]{}+K\_[-p]{}\^[(1)]{}) [N]{}\^[33]{}\_[n-n]{} ( \^K)\^\_[ ]{} \^\_)
The last term in (\[contactmey\]) gives rise to a log-divergent sum, the large-$p$ behaviour of which is:
$$\delta E^{\mbox{div}}_{H^N_4} \sim +(\a^2+\b^2)\int_0^1 dr\,
\,\frac{g_2^2\,r(1-r)}{r\,|\alpha_3|\,\pi^2} \left( {\widetilde
N}^{3\,3}_{n\,-n} \right)^2 \, \sum_{p>0} \frac{1}{p}$$
Thus the energy shift is finite for arbitrary $\a$ and $\b$. The DY vertex uses $\a=\b=1/2$, and this combination exclusively gives rise to the agreement with gauge theory discussed in the introduction. The generalization of these arguments to the impurity non-conserving channels is a straightforward application of the treatment given in [@Grignani:2005yv].
Results
=======
The calculations undertaken in this Paper are practically identical to those in [@Gutjahr:2004dv], using the DVPPRT and DY vertices in place of the SVPS vertices used there. One small difference in the case of the SVPS vertex is that the half-integer powers of $\l'$ calculated in Ref. [@Gutjahr:2004dv] and quoted vertabim in our Eq. (\[string\]) are incomplete and suffer from a sign error, and are correctly given below. We refer the reader to this reference for details, and simply give results below.
The external state for which we are calculating the energy shift is
$$\left. |[{\bf 9}, {\bf
1}]\ra^{(ij)}\right. =
\frac{1}{\sqrt{2}}\left(\a^{\dag\,i}_n\a^{\dag\,j}_{-n}+\a^{\dag\,j}_n\a^{\dag\,i}_{-n}
-\frac{1}{2}\d^{ij}\a^{\dag\,k}_n\a^{\dag\,k}_{-n}\right)|3\ra.$$
For this particular state, individual $H_3$ and contact terms are not divergent in the two impurity approximation. It should be further noted that for this state, and for the impurity conserving channel, we shall find that use of the DY vertex, rather than the SVPS vertex, is equivalent to making the replacements of the quantities $(K,\wt K)$ as $K \rightarrow (K+\wt K)/2$ and $\wt K
\rightarrow (K+\wt K)/2$ in the SVPS vertex. This is the simplest way of reproducing our results.
The separate $H_3$ and contact term contributions to the energy shift for each of the three vertices are given below. We find that the DY energy shift agrees with gauge theory only at the leading order, while also enjoying the vanishing of the $3/2$ and $5/2$ powers of $\l'$. The order-$\l'^2$ term is of the correct form, but suffers from an overall factor of $4/3$. The SVPS and DVPPRT results do not agree with gauge theory at the leading order. By multiplying the contact terms by two (an unjustified operation), one can recover the correct gauge theory result up to $\l'^2$ order with the SVPS (including vanishing of its $\l'^{3/2}$ term) and DY vertices. Further, this operation does not spoil the vanishing $3/2$ and $5/2$ powers of $\l'$ for the DY result.
$H_3$ terms
-----------
E\^\_[H\_3]{} &=&
E\^\_[H\_3]{} &=&
E\^\_[H\_3]{} &=&
Contact terms
-------------
E\^\_[H\_4]{} &=&
E\^\_[H\_4]{} = E\^\_[H\_4]{}
E\^\_[H\_4]{} &=&
Energy shifts
-------------
The results for the complete energy shifts are as follows,
E\^ &=&
E\^ &=&
\[N4final\] E\^ &=&
Recall that the leading $3/4$ is irrelevant and can be scaled away by fixing the overall $f$ factor which multiplies the vertices (and which has not been written in the above formulae, where it would appear in each as an overall factor of $|f|^2$). We see that the gauge theory result (\[gauge\]) is matched only by the DY result, and only at leading order in $\l'$, with the $\l'^2$ term being of the correct form but with an overall factor of $4/3$. We also see the miraculous absence of the $\l'^{3/2}$ and $\l'^{5/2}$ terms which are clearly generic in the string field theory. The result (\[N4final\]) represents the best matching of this quantity to gauge theory so far, and thus is an indication that the DY vertex is an improvement over its predecessors.
Mysteriously, if the contact terms are scaled by a factor of 2, the agreement with gauge theory is enhanced for both the SVPS and DY results,
E\^\_[2H\_4]{} &=&
E\^\_[2H\_4]{} &=&
however, the DY result is still superior in that the $\l'^{5/2}$ power is absent.
Acknowledgements: {#acknowledgements .unnumbered}
=================
The work of B. Ramadanovic, G. Semenoff and D. Young is supported by the Natural Sciences and Engineering Research Council of Canada and by University of British Columbia Graduate Fellowships. G. Semenoff acknowledges the hospitality of the University of Perugia where some of this work was done. The work of M. Orselli is supported in part by the European Community’s Human Potential Programme under contract MRTN-CT-2004-005104 ‘Constituents, fundamental forces and symmetries of the universe’. The work of G.Grignani is supported by the I.N.F.N. and M.U.I.R. of Italy. G. Grignani and M. Orselli acknowledge the hospitality of the Pacific Institute for Theoretical Physics and the University of British Columbia where parts of this work were done.
[0]{}
J. Maldacena, “The Large N limit of superconformal field theories and supergravity,” [*Adv. Theor. Math. Phys.*]{} [**2**]{} (1998) 231, \[arXiv:hep-th/9711200\]. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, “Gauge theory correlators from non-critical string theory,” Phys. Lett. B [**428**]{}, 105 (1998) \[arXiv:hep-th/9802109\]. E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math. Phys. [**2**]{}, 253 (1998) \[arXiv:hep-th/9802150\]. M. Blau, J. Figueroa-O’Farrill, C. Hull and G. Papadopoulos, “A new maximally supersymmetric background of IIB superstring theory,” JHEP [**0201**]{}, 047 (2002) \[arXiv:hep-th/0110242\]. M. Blau, J. Figueroa-O’Farrill, C. Hull and G. Papadopoulos, “Penrose limits and maximal supersymmetry,” Class. Quant. Grav. [**19**]{}, L87 (2002) \[arXiv:hep-th/0201081\]. R. R. Metsaev, “Type IIB Green-Schwarz superstring in plane wave Ramond-Ramond background,” Nucl. Phys. B [**625**]{}, 70 (2002) \[arXiv:hep-th/0112044\]. D. Berenstein, J. M. Maldacena and H. Nastase, “Strings in flat space and pp waves from N = 4 super Yang Mills,” JHEP [**0204**]{}, 013 (2002) \[arXiv:hep-th/0202021\]. C. Kristjansen, J. Plefka, G. W. Semenoff and M. Staudacher, “A new double-scaling limit of N = 4 super Yang-Mills theory and PP-wave strings,” Nucl. Phys. B [**643**]{}, 3 (2002) \[arXiv:hep-th/0205033\]. N. R. Constable, D. Z. Freedman, M. Headrick, S. Minwalla, L. Motl, A. Postnikov and W. Skiba, “PP-wave string interactions from perturbative Yang-Mills theory,” JHEP [**0207**]{}, 017 (2002) \[arXiv:hep-th/0205089\]. D. J. Gross, A. Mikhailov and R. Roiban, “Operators with large R charge in N = 4 Yang-Mills theory,” Annals Phys. [**301**]{}, 31 (2002) \[arXiv:hep-th/0205066\]. A. Santambrogio and D. Zanon, “Exact anomalous dimensions of N = 4 Yang-Mills operators with large R charge,” Phys. Lett. B [**545**]{}, 425 (2002) \[arXiv:hep-th/0206079\]. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, “A semi-classical limit of the gauge/string correspondence,” Nucl. Phys. B [**636**]{}, 99 (2002) \[arXiv:hep-th/0204051\]. C. G. . Callan, H. K. Lee, T. McLoughlin, J. H. Schwarz, I. Swanson and X. Wu, “Quantizing string theory in AdS(5) x S\*\*5: Beyond the pp-wave,” Nucl. Phys. B [**673**]{}, 3 (2003) \[arXiv:hep-th/0307032\]. M. B. Green, A. Sinha and S. Kovacs, “Non-perturbative effects in the BMN limit of N = 4 supersymmetric Yang-Mills,” arXiv:hep-th/0506200. M. B. Green, S. Kovacs and A. Sinha, “Non-perturbative contributions to the plane-wave string mass matrix,” JHEP [**0505**]{}, 055 (2005) \[arXiv:hep-th/0503077\].
J. A. Minahan and K. Zarembo, “The Bethe-ansatz for N = 4 super Yang-Mills,” JHEP [**0303**]{}, 013 (2003) \[arXiv:hep-th/0212208\]. N. Beisert, C. Kristjansen and M. Staudacher, “The dilatation operator of N = 4 super Yang-Mills theory,” Nucl. Phys. B [**664**]{}, 131 (2003) \[arXiv:hep-th/0303060\]. N. Beisert, J. A. Minahan, M. Staudacher and K. Zarembo, “Stringing spins and spinning strings,” JHEP [**0309**]{}, 010 (2003) \[arXiv:hep-th/0306139\]. I. Bena, J. Polchinski and R. Roiban, “Hidden symmetries of the AdS(5) x S\*\*5 superstring,” Phys. Rev. D [**69**]{}, 046002 (2004) \[arXiv:hep-th/0305116\]. N. Beisert, V. A. Kazakov, K. Sakai and K. Zarembo, “The algebraic curve of classical superstrings on AdS(5) x S\*\*5,” Commun. Math. Phys. [**263**]{}, 659 (2006) \[arXiv:hep-th/0502226\]. N. Beisert, V. A. Kazakov, K. Sakai and K. Zarembo, “Complete spectrum of long operators in N = 4 SYM at one loop,” JHEP [**0507**]{}, 030 (2005) \[arXiv:hep-th/0503200\]. N. Beisert and A. A. Tseytlin, “On quantum corrections to spinning strings and Bethe equations,” Phys. Lett. B [**629**]{}, 102 (2005) \[arXiv:hep-th/0509084\]. A. Rej, D. Serban and M. Staudacher, “Planar N = 4 gauge theory and the Hubbard model,” JHEP [**0603**]{}, 018 (2006) \[arXiv:hep-th/0512077\]. N. Beisert, C. Kristjansen, J. Plefka, G. W. Semenoff and M. Staudacher, “BMN correlators and operator mixing in N = 4 super Yang-Mills theory,” Nucl. Phys. B [**650**]{}, 125 (2003) \[arXiv:hep-th/0208178\]. N. R. Constable, D. Z. Freedman, M. Headrick and S. Minwalla, “Operator mixing and the BMN correspondence,” JHEP [**0210**]{}, 068 (2002) \[arXiv:hep-th/0209002\]. M. Spradlin and A. Volovich, “Superstring interactions in a pp-wave background,” Phys. Rev. D [**66**]{}, 086004 (2002) \[arXiv:hep-th/0204146\]. D. Berenstein and H. Nastase, “On lightcone string field theory from super Yang-Mills and holography,” arXiv:hep-th/0205048. R. Gopakumar, “String interactions in PP-waves,” Phys. Rev. Lett. [**89**]{}, 171601 (2002) \[arXiv:hep-th/0205174\]. Y. j. Kiem, Y. b. Kim, S. m. Lee and J. m. Park, “pp-wave / Yang-Mills correspondence: An explicit check,” Nucl. Phys. B [**642**]{}, 389 (2002) \[arXiv:hep-th/0205279\]. M. x. Huang, “Three point functions of N = 4 super Yang Mills from light cone string field theory in pp-wave,” Phys. Lett. B [**542**]{}, 255 (2002) \[arXiv:hep-th/0205311\]. M. Bianchi, B. Eden, G. Rossi and Y. S. Stanev, “On operator mixing in N = 4 SYM,” Nucl. Phys. B [**646**]{}, 69 (2002) \[arXiv:hep-th/0205321\].
C. S. Chu, V. V. Khoze and G. Travaglini, “Three-point functions in N = 4 Yang-Mills theory and pp-waves,” JHEP [**0206**]{}, 011 (2002) \[arXiv:hep-th/0206005\]. H. L. Verlinde, “Bits, matrices and 1/N,” JHEP [**0312**]{}, 052 (2003) \[arXiv:hep-th/0206059\]. P. Lee, S. Moriyama and J. w. Park, “Cubic interactions in pp-wave light cone string field theory,” Phys. Rev. D [**66**]{}, 085021 (2002) \[arXiv:hep-th/0206065\]. M. Spradlin and A. Volovich, “Superstring interactions in a pp-wave background. II,” JHEP [**0301**]{}, 036 (2003) \[arXiv:hep-th/0206073\]. C. S. Chu, V. V. Khoze and G. Travaglini, “pp-wave string interactions from n-point correlators of BMN operators,” JHEP [**0209**]{}, 054 (2002) \[arXiv:hep-th/0206167\]. I. R. Klebanov, M. Spradlin and A. Volovich, “New effects in gauge theory from pp-wave superstrings,” Phys. Lett. B [**548**]{}, 111 (2002) \[arXiv:hep-th/0206221\]. C. S. Chu, V. V. Khoze, M. Petrini, R. Russo and A. Tanzini, “A note on string interaction on the pp-wave background,” Class. Quant. Grav. [**21**]{}, 1999 (2004) \[arXiv:hep-th/0208148\]. A. Pankiewicz, “More comments on superstring interactions in the pp-wave background,” JHEP [**0209**]{}, 056 (2002) \[arXiv:hep-th/0208209\]. D. Vaman and H. L. Verlinde, “Bit strings from N = 4 gauge theory,” JHEP [**0311**]{}, 041 (2003) \[arXiv:hep-th/0209215\]. B. Eynard and C. Kristjansen, “BMN correlators by loop equations,” JHEP [**0210**]{}, 027 (2002) \[arXiv:hep-th/0209244\]. S. Dobashi, H. Shimada and T. Yoneya, “Holographic reformulation of string theory on AdS(5) x S\*\*5 background in Nucl. Phys. B [**665**]{}, 94 (2003) \[arXiv:hep-th/0209251\]. R. A. Janik, “BMN operators and string field theory,” Phys. Lett. B [**549**]{}, 237 (2002) \[arXiv:hep-th/0209263\]. J. Pearson, M. Spradlin, D. Vaman, H. L. Verlinde and A. Volovich, “Tracing the string: BMN correspondence at finite J\*\*2/N,” JHEP [**0305**]{}, 022 (2003) \[arXiv:hep-th/0210102\]. J. Gomis, S. Moriyama and J. w. Park, “SYM description of SFT Hamiltonian in a pp-wave background,” Nucl. Phys. B [**659**]{}, 179 (2003) \[arXiv:hep-th/0210153\]. A. Pankiewicz and B. . J. Stefanski, “pp-wave light-cone superstring field theory,” Nucl. Phys. B [**657**]{}, 79 (2003) \[arXiv:hep-th/0210246\]. C. S. Chu, M. Petrini, R. Russo and A. Tanzini, “String interactions and discrete symmetries of the pp-wave background,” Class. Quant. Grav. [**20**]{}, S457 (2003) \[arXiv:hep-th/0211188\]. Y. H. He, J. H. Schwarz, M. Spradlin and A. Volovich, “Explicit formulas for Neumann coefficients in the plane-wave geometry,” Phys. Rev. D [**67**]{}, 086005 (2003) \[arXiv:hep-th/0211198\]. R. Roiban, M. Spradlin and A. Volovich, “On light-cone SFT contact terms in a plane wave,” JHEP [**0310**]{}, 055 (2003) \[arXiv:hep-th/0211220\]. U. Gursoy, “Predictions for pp-wave string amplitudes from perturbative SYM,” JHEP [**0310**]{}, 027 (2003) \[arXiv:hep-th/0212118\]. N. Beisert, C. Kristjansen, J. Plefka and M. Staudacher, “BMN gauge theory as a quantum mechanical system,” Phys. Lett. B [**558**]{}, 229 (2003) \[arXiv:hep-th/0212269\]. J. Gomis, S. Moriyama and J. w. Park, “SYM description of pp-wave string interactions: Singlet sector and arbitrary impurities,” Nucl. Phys. B [**665**]{}, 49 (2003) \[arXiv:hep-th/0301250\]. P. Di Vecchia, J. L. Petersen, M. Petrini, R. Russo and A. Tanzini, “The 3-string vertex and the AdS/CFT duality in the pp-wave limit,” Class. Quant. Grav. [**21**]{}, 2221 (2004) \[arXiv:hep-th/0304025\]. B. . J. Stefanski, “Open string plane-wave light-cone superstring field theory,” arXiv:hep-th/0304114. A. Pankiewicz, “An alternative formulation of light-cone string field theory on the plane wave,” JHEP [**0306**]{}, 047 (2003) \[arXiv:hep-th/0304232\]. J. C. Plefka, “Lectures on the plane-wave string / gauge theory duality,” Fortsch. Phys. [**52**]{}, 264 (2004) \[arXiv:hep-th/0307101\]. M. Spradlin and A. Volovich, “Light-cone string field theory in a plane wave,” arXiv:hep-th/0310033. J. Lucietti, S. Schafer-Nameki and A. Sinha, “On the exact open-closed vertex in plane-wave light-cone string field theory,” Phys. Rev. D [**69**]{}, 086005 (2004) \[arXiv:hep-th/0311231\]. D. Sadri and M. M. Sheikh-Jabbari, “The plane-wave / super Yang-Mills duality,” Rev. Mod. Phys. [**76**]{}, 853 (2004) \[arXiv:hep-th/0310119\]. M. Bianchi, G. Rossi and Y. S. Stanev, “Surprises from the resolution of operator mixing in N = 4 SYM,” Nucl. Phys. B [**685**]{}, 65 (2004) \[arXiv:hep-th/0312228\]. M. Bianchi, G. D’Appollonio, E. Kiritsis and O. Zapata, “String amplitudes in the Hpp-wave limit of AdS(3) x S\*\*3,” JHEP [**0404**]{}, 074 (2004) \[arXiv:hep-th/0402004\]. J. Lucietti, S. Schafer-Nameki and A. Sinha, “On the plane-wave cubic vertex,” Phys. Rev. D [**70**]{}, 026005 (2004) \[arXiv:hep-th/0402185\]. P. Gutjahr and J. Plefka, “Decay widths of three-impurity states in the BMN correspondence,” Nucl. Phys. B [**692**]{}, 110 (2004) \[arXiv:hep-th/0402211\]. S. Dobashi and T. Yoneya, Nucl. Phys. B [**711**]{}, 3 (2005) \[arXiv:hep-th/0406225\]. P. Gutjahr and A. Pankiewicz, “New aspects of the BMN correspondence beyond the planar limit,” Nucl. Phys. B [**704**]{}, 583 (2005) \[arXiv:hep-th/0407098\]. S. Dobashi and T. Yoneya, “Impurity non-preserving 3-point correlators of BMN operators from pp-wave Nucl. Phys. B [**711**]{}, 54 (2005) \[arXiv:hep-th/0409058\]. S. Lee and R. Russo, “Holographic cubic vertex in the pp-wave,” Nucl. Phys. B [**705**]{}, 296 (2005) \[arXiv:hep-th/0409261\]. G. Grignani, M. Orselli, B. Ramadanovic, G. W. Semenoff and D. Young, JHEP [**0512**]{}, 017 (2005) \[arXiv:hep-th/0508126\]. S. Dobashi, “Impurity non-preserving 3-point correlators of BMN operators from PP-wave holography. II: Fermionic excitations,” arXiv:hep-th/0604082. J. Greensite and F. R. Klinkhamer, “Superstring Amplitudes And Contact Interactions,” Nucl. Phys. B [**304**]{}, 108 (1988).
M. B. Green and N. Seiberg, “Contact Interactions In Superstring Theory,” Nucl. Phys. B [**299**]{}, 559 (1988).
[^1]: We remind the reader that the string light-cone momenta are related to Yang-Mills conformal dimension $\Delta$ and R-charge $J$ as $$p^-=\mu\left( \Delta-J\right) ~~,~~ p^+=\frac{
\Delta+J }{2 \mu \sqrt{ g_{YM}^2N } \alpha' }$$ where in the BMN limit $N,\Delta,J\to\infty$ so that $(p^+,p^-)$ remain finite. Two convenient couplings are $$\frac{1}{\left(\mu\alpha'p^+\right)^2}=\frac{g_{YM}^2N}{J^2}
\equiv\lambda' ~~,~~ 4\pi g_s\left(\mu\alpha'p^+\right)^2 =
\frac{J^2}{N} \equiv g_2~~,~~N,J\to\infty$$ $\lambda'$ is proportional to the string tension. $g_2$ is the string coupling which weights the genus of the string world-sheet.
[^2]: There are a few minor corrections which affect the fractional powers in (\[string\]), but they remain non-zero in the corrected (\[string\]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The goal of this paper is to develop a general approach to solution of ill-posed nonlinear problems in a Hilbert space based on continuous processes with a regularization procedure. To avoid the ill-posed inversion of the Fréchet derivative operator a regularizing one-parametric family of operators is introduced. Under certain assumptions on the regularizing family a general convergence theorem is proved. The proof is based on a lemma describing asymptotic behavior of solutions of a new nonlinear integral inequality. Then the applicability of the theorem to the continuous analogs of the Newton, Gauss-Newton and simple iteration methods is demonstrated.
[**AMS subject classification:**]{} Primary: 47H17. Secondary: 65J15, 58C15.
---
\[section\] \[theorem\][Lemma]{}
\[theorem\][Definition]{} \[theorem\][Proposition]{} \[theorem\][Exercise]{}
\[theorem\][Remark]{}
¶ ł Ł
[**Continuous methods for solving nonlinear ill-posed problems**]{}\
Ruben G. Airapetyan\
E-mail: [email protected]\
Department of Mathematics\
Kansas State University\
Manhattan, Kansas 66506-2602, U.S.A.\
Alexander G. Ramm\
E-mail: [email protected]\
Department of Mathematics\
Kansas State University\
Manhattan, Kansas 66506-2602, U.S.A.\
Alexandra B. Smirnova\
E-mail: [email protected]\
Department of Mathematics\
Kansas State University\
Manhattan, Kansas 66506-2602, U.S.A.\
Introduction
============
Let us consider a nonlinear operator equation $$\label{noneq}
F(z)=0,\q F: H\to H,$$ in a real Hilbert space $H$ (equation (\[noneq\]) in a complex Hilbert space can be treated similarly).
Assume that (\[noneq\]) is solvable (not necessarily uniquely). If the Fréchet derivative of the operator $F$ has nontrivial null-space at the solution to (\[noneq\]), then one can use the classical Newton method for solution to (\[noneq\]) only under some strong assumptions on the operator $F$ (see [@dkk; @ag]). Otherwise in order to construct a numerical method for solution to (\[noneq\]) one needs some regularization procedure.
In the theory of ill-posed problems many different discrete methods based on a regularization are known. Many different convergence theorems for such schemes describe the efficiency of the numerical algorithms for solving various nonlinear problems, and give existence results (see, for example, [@bns; @ehn; @va]). However it is quite difficult to navigate in the sea of the discrete schemes and corresponding convergence theorems. Proofs of these theorems are usually based on the contraction mapping principle and are sometimes rather complicated.
On the other hand an analysis of continuous processes is based on the investigation of the asymptotical behavior of nonlinear dynamical systems in Banach and Hilbert spaces. If a convergence theorem is proved for a continuous method, one can construct various discrete schemes generated by this continuous process. Thus construction of a discrete numerical scheme is split into two parts: construction of the continuous process and numerical integration of the corresponding nonlinear operator differential equation. Convergence theorems for regularized continuous Newton-like methods are established in [@ars; @al; @r].
The goal of this paper is to develop a general approach to continuous analogs of discrete methods and to establish fairly general convergence theorem. This approach is based on an analysis of the solution to the Cauchy problem for a nonlinear differential equation in a Hilbert space. Such an analysis was done for well-posed problems in [@a], where it was based on a usage of an integral inequality. It is more difficult to study nonlinear ill-posed problems. In this case one has to use more complicated new integral inequality (Lemma \[ric\]). Based on Lemma \[ric\] the general theorem establishing convergence of a regularized continuous process is proved (Theorem \[t1\]). Applying this theorem to the regularized Newton’s and simple iteration methods (for monotone operators) and to Gauss-Newton-type methods (for non-monotone operators) convergence theorems are obtained under less restrictive conditions on the equation than the theorems known for the corresponding discrete methods. According to these theorems one can choose a regularizing operator depending on the “degree of degeneracy” of the original nonlinear problem and estimate the rate of convergence of the regularized process.
The paper is organized as follows. In Section 2 Lemma \[ric\] about the solution of a new integral inequality and Theorem \[t1\] about the convergence of a regularized continuous process are formulated and proved. In Section 3 this theorem is applied to continuous Newton’s and simple iteration methods and in Section 4 to the Gauss-Newton-type methods. A practically interesting numerical example is considered in Section 5. A lemma about nonlinear differential inequality is proved in the Appendix.
Regularization procedure
========================
In the well-posed case (the Fréchet derivative $F'$ of the operator $F$ is a bijection in a neighborhood of the solution of equation (\[noneq\])) in order to solve equation (\[noneq\]) one can use the following continuous processes:
- simple iteration method: $$\label{esim2}
\dot{z}(t)=-F(z(t)),\q z(0)=z_0\in H,$$
- Newton’s method: $$\label{enm2}
\dot{z}(t)=-[F'(z(t))]^{-1}F(z(t)),\q z(0)=z_0\in H,$$
- Gauss-Newton’s method: $$\label{egnm}
\dot{z}(t)=-[F^{\pr *}(z(t))F^{\pr}(z(t))]^{-1}F^{\pr *}(z(t))F(z(t)),\q
z(0)=z_0\in H,$$
or some of their modifications (see [@a; @g]).
However if $F'$ is not continuously invertible (ill-posed case) one has to replace equations (\[esim2\]) -(\[egnm\]) by the corresponding regularized equations:
- regularized simple iteration method: $$\label{esim1}
\dot{z}(t)=-[F(z(t))+\ep(t)(z(t)-z_0)],\q z(0)=z_0\in H,$$
- regularized Newton’s method: $$\label{enm1}
\dot{z}(t)=-[F'(z(t))+\ep(t)I]^{-1}[F(z(t))+\ep(t)(z(t)-z_0)],\q z(0)=z_0\in H,$$
- regularized Gauss-Newton’s methods: equation (\[caupr\]) with the function $\P$ defined in (\[P1\]) or in (\[P2\]),
with an appropriate choice of the function $\ep(t)$ and the point $z_0$. Here $I$ is the identity operator.
The goal of this paper is to develop a uniform approach to such regularized methods. Let us consider the Cauchy problem: $$\label{caupr}
\dot{z}(t)=\P(z(t),t),\q z(0)=z_0\in H,$$ with an operator $\P:H\times[0,\infty)\to H$. The choice of $\P$ yields the corresponding continuous process.
In this section a general convergence theorem (Theorem \[t1\]) is established. In the next two sections the convergence theorems for the processes mentioned above are derived from this general theorem. In the proof of the general theorem the technique of integral inequalities is used.
The following lemma is known. It is a version of some results concerning integral inequalities (see e.g. Theorem 22.1 in [@JS]). For convenience of the reader and to make the presentation essentially self-contained we include a proof.
Let $f(t,w)$, $g(t,u)$ be continuous on region $[0,T)\times D$ ($D\subset
R$, $T\le\infty$) and $f(t,w)\leq g(t,u)$ if $w\leq u$, $t\in (0,T)$, $w,u\in D$. Assume that $g(t,u)$ is such that the Cauchy problem $$\label{Cauchy}
\dot u = g(t,u), \q u(0)=u_0,\q u_0\in D$$ has a unique solution. If $$\label{inequal}
\dot w\leq f(t,w), \q w(0)=w_0\leq u_0,\q w_0\in D,$$ then $u(t)\geq w(t)$ for all $t$ for which $u(t)$ and $w(t)$ are defined. \[ineq\]
[**Proof**]{} Suppose first $f(t,w)<g(t,u)$, if $w\leq u$. Since $w_0\leq u_0$ and $\dot w (0)\leq f(t,w_0)<g(t,u_0)=\dot u (0)$, there exists $\delta
>0$ such that $u(t)> w(t)$ on $(0,\delta]$. Assume that for some $t_1> \delta$ one has $u(t_1)< w(t_1).$ Then for some $ t_2<t_1$ one has $$u(t_2)=w(t_2) \q\t{and}\q
u(t)<w(t) \q\t{for}\q t\in (t_2,t_1].$$ One gets $$\dot w(t_2)\geq \dot u(t_2)=g(t,u(t_2))>f(t,w(t_2))\geq \dot w(t_2).$$ This contradiction proves that there is no point $t_2$ such that $u(t_2)=w(t_2)$.
Now consider the case $f(t,w)\leq g(t,u)$, if $w\leq u$. Define $$\dot u_n = g(t,u_n)+\ep_n, \q u_n(0)=u_0,\q \ep_n>0,\q n=0,1,...,$$ where $\ep_n$ tends monotonically to zero. Then $$\dot w\leq f(t,w)\leq g(t,u)<g(t,u)+\ep_n, \q w\leq u.$$ By Step 1 $u_n(t)\geq w(t)$, $n=0,1,...\, $. Fix an arbitrary compact set $[0,T_1]$, $0<T_1<T.$ $$\label{un}
u_n(t)=u_0+\il_0^tg(\tau,u_n(\tau))d\tau + \ep_n t.$$ Since $g(t,u)$ is continuous, the sequence $\{u_n\}$ is uniformly bounded and equicontinuous on $[0,T_1]$. Therefore there exists a subsequence $\{u_{n_k}\}$ which converges uniformly to a continuous function $u(t)$. By continuity of $g(t,u)$ we can pass to the limit in (\[un\]) and get $$\label{un1a}
u(t)=u_0+\il_0^tg(\tau,u(\tau))d\tau,\q t\in [0,T_1].$$ Since $T_1$ is arbitrary (\[un1a\]) is equivalent to the initial Cauchy problem that has a unique solution. The inequality $u_{n_k}(t)\geq w(t),$ $k=0,1,...$ implies $u(t)\geq w(t).$ If the solution to the Cauchy problem (\[Cauchy\]) is not unique, the inequality $w(t)\leq u(t)$ holds for the maximal solution to (\[Cauchy\]). $\Box$
Our second lemma is a key to the basic result of this section, namely to Theorem \[t1\].
Let $\g(t), \s(t), \b(t)\in C[0,\infty)$. If there exists a positive function $\mu(t)\in C^1[0,\infty)$ such that $$\label{ric1}
0\le\s(t)\le\f{\mu(t)}{2}\left(\g(t)-\f{\dot{\mu}(t)}{\mu(t)}\right),\q
\b(t)\le\f{1}{2\mu(t)}\left(\g(t)-\f{\dot{\mu}(t)}{\mu(t)}\right),\q
\mu(0)v(0)<1,$$ then a nonnegative solution to the following inequality $$\label{ric4}
\dot{v}(t)\le -\g(t) v(t)+\s(t)v^2(t)+\b(t).$$ satisfies the estimate: $$\label{ric5}
v(t)\,<\, \f{1}{\mu(t)}.$$ \[ric\]
Without loss of generality one can assume $\b(t)\geq 0$.
In [@al1] (see also Appendix) a differential inequality $\dot{u}\le -a(t)\psi(u(t))+b(t)$ was studied under assumptions 1) - 3) of Lemma \[apineq\] of the Appendix. These assumptions alone, as we show in the Appendix, do not imply the desired conclusion (\[descon\]). We have added assumption 4) in order to prove conclusion (\[descon\]). In Lemma \[ric\] the term $-\g(t)v(t)+\s(t)v^2(t)$ (which is analogous to some extent to the term $-a(t)\psi(u(t))$) can change sign. Our Lemma \[ric\] is not covered by the result in [@al1]. In particular, in Lemma \[ric\] an analog of $\psi(u)$, for the case $\g(t)=\s(t)=a(t)$, is the function $\psi(u):=u-u^2$. This function goes to $-\infty$ as $u$ goes to $+\infty$, so it does not satisfy the positivity condition imposed in [@al1].
Unlike in the case of Bihari integral inequality ([@bb]) one cannot separate variables in the right hand side of inequality (\[ric4\]) and estimate $v(t)$ by a solution of the Cauchy problem for a differential equation with separating variables. The proof below is based on a special choice of the solution to the Riccati equation majorizing a solution of integral inequality (\[ric4\]). \[bihari\]
[**Proof**]{}[**of Lemma \[ric\]**]{} Denote: $$\label{rc1}
w(t):=v(t)e^{\int_0^t \g(s)ds},$$ then (\[ric4\]) implies: $$\label{rc2}
\dot{w}(t)\le a(t)w^2(t)+b(t),\q w(0)=v(0),$$ where $$a(t)=\s(t)e^{-\int_0^t \g(s)ds},\q b(t)=\b(t)e^{\int_0^t \g(s)ds}.$$ Consider Riccati’s equation: $$\label{rc3}
\dot{u}(t)=\f{\dot{f}(t)}{g(t)}u^2(t)-\f{\dot{g}(t)}{f(t)}.$$ One can check by a direct calculation that the the solution to problem (\[rc3\]) is given by the following formula [@kam eq. 1.33]: $$\label{rc5}
u(t)=-\f{g(t)}{f(t)}+\left[f^2(t)\left(C-
\int_0^t\f{\dot{f}(s)}{g(s)f^2(s)}ds\right)\right]^{-1}.$$ Define $f$ and $g$ as follows: $$\label{rc4}
f(t):=\mu^{\f{1}{2}}(t)e^{-\f{1}{2}\int_0^t \g(s)ds},\q
g(t):=-\mu^{-\f{1}{2}}(t)e^{\f{1}{2}\int_0^t \g(s)ds},$$ and consider the Cauchy problem for equation (\[rc3\]) with the initial condition $u(0)=v(0)$. Then $C$ in (\[rc5\]) takes the form: $$C=\f{1}{\mu(0)v(0)-1}.$$ From (\[ric1\]) one gets $$a(t)\le \f{\dot{f}(t)}{g(t)},\q b(t)\le -\f{\dot{g}(t)}{f(t)}.$$ Since $fg=-1$ one has: $$\int_0^t\f{\dot{f}(s)}{g(s)f^2(s)}ds=-\int_0^t\f{\dot{f}(s)}{f(s)}ds=
\f{1}{2}\int_0^t\left(\g(s)-\f{\dot{\mu}(s)}{\mu(s)}\right)ds.$$ Thus $$\label{rc6}
u(t)=\f{e^{\int_0^t \g(s)ds}}{\mu(t)}\left[1-\left(\f{1}{1-\mu(0)v(0)}+
\f{1}{2}\int_0^t\left(\g(s)-\f{\dot{\mu}(s)}{\mu(s)}\right)ds\right)^{-1}\right].$$ It follows from condition (\[ric1\]) that the solution to problem (\[rc3\]) exists for all $t\in[0,\infty)$ and the following inequality holds: $$\label{estimate}
1\, >\, 1-\left(\f{1}{1-\mu(0)v(0)}+
\f{1}{2}\int_0^t\left(\g(s)-\f{\dot{\mu}(s)}{\mu(s)}\right)ds\right)^{-1}
\ge\mu(0)v(0).$$ From Lemma \[ineq\] and from formula (\[rc6\]) one gets: $$\label{rc7}
v(t)e^{\int_0^t \g(s)ds}:=
w(t)\le u(t)<\f{1}{\mu(t)}e^{\int_0^t \g(s)ds},$$ and thus estimate (\[ric5\]) is proved. $\Box$
[**Examples.**]{} To illustrate conditions (\[ric1\]) of Lemma \[ric\] consider the following examples of functions $\g$, $\s$, $\b$, satisfying (\[ric1\]).
[**1.**]{} Let $$\label{rcex}
\g(t)=c_1(1+t)^{\nu_1},\q \s(t)=c_2(1+t)^{\nu_2},\q \b(t)=c_3(1+t)^{\nu_3},$$ where $c_2>0$, $c_3>0$. Choose $\mu(t):=c(1+t)^{\nu}$, $c>0$. From (\[ric1\]) one gets the following conditions $$c_2\le \f{cc_1}{2}(1+t)^{\nu+\nu_1-\nu_2}-\f{c\nu}{2}(1+t)^{\nu-1-\nu_2},\q$$ $$\label{rcex1}
c_3\le \f{c_1}{2c}(1+t)^{\nu_1-\nu-\nu_3}-\f{\nu}{2c}(1+t)^{-\nu-1-\nu_3},\q
cv(0)<1.$$ Thus one obtains the following conditions: $$\label{rcex2}
\nu_1\ge -1,\q \nu_2-\nu_1\le\nu\le\nu_1-\nu_3,$$ and $$\label{rcex3}
c_1>\nu,\q \f{2c_2}{c_1-\nu}\le c\le\f{c_1-\nu}{2c_3},\q
cv(0)<1.$$ Therefore for such $\g$, $\s$, $\b$ a function $\mu$ with the desired properties exists if $$\label{rcex4}
\nu_1\ge -1,\q \nu_2+\nu_3\le 2\nu_1,$$ and $$\label{rcex5}
c_1>\nu_2-\nu_1,\q 2\sqrt{c_2c_3}\le c_1+\nu_1-\nu_2,\q
2c_2v(0)<c_1+\nu_1-\nu_2.$$ In this case one can choose $\nu=\nu_2-\nu_1$, $c=\f{2c_2}{c_1+\nu_1-\nu_2}$. However in order to have $v(t)\to 0$ as $t\to+\infty$ (the case of interest in Theorem \[t1\]) one needs the following conditions: $$\label{rcex6}
\nu_1\ge -1,\q \nu_2+\nu_3\le 2\nu_1,\q \nu_1>\nu_3,$$ and $$\label{rcex7}
c_1>\nu_2-\nu_1,\q 2\sqrt{c_2c_3}\le c_1,\q 2c_2v(0)<c_1.$$
[**2.**]{} If $$\g(t)=\g_0,\q \s(t)=\s_0e^{\nu t},\q \b(t)=\b_0e^{-\nu t},\q \mu(t)=\mu_0e^{\nu t},$$ then conditions (\[ric1\]) are satisfied if $$0\le \s_0\le\f{\mu_0}{2}(\g_0-\nu),\q\b_0\le\f{1}{2\mu_0}(\g_0-\nu),\q\mu_0v(0)<1.$$
[**3.**]{} If $$\g(t)=\f{1}{\sqrt{\log(t+t_0)}},\q \mu(t)=c\log(t+t_0),$$ then conditions (\[ric1\]) are satisfied if $$0\le\s(t)\le\f{c}{2}\left(\sqrt{\log(t+t_0)}-\f{1}{t+t_0}\right),$$ $$\b(t)\le\f{1}{2c\log^2(t+t_0)}\left(\sqrt{\log(t+t_0)}-\f{1}{t+t_0}\right),\q
v(0)c\log t_0<1.$$
In all considered examples $\mu(t)$ can tend to infinity as $t\to +\infty$ and provide a decay of a nonnegative solution to integral inequality (\[ric4\]) even if $\s(t)$ tends to infinity. Moreover in the first and the third examples $v(t)$ tends to zero as $t\to
+\infty$ when $\g(t)\to 0$ and $\s(t)\to +\infty$.
Let $\P(h,t)$ be Fréchet differentiable with respect to $h$ and satisfy the following condition:
there exists a differentiable function $x(t)$, $x:[0,+\infty)\to H$, such that for any $h\in H,\, t\in[0,+\infty)$ $$\label{semon}
(\P(h,t),h-x(t))\le \a(t)||h-x(t)||-\g(t)||h-x(t)||^2+\s(t)||h-x(t)||^3,$$ where $\alpha(t)$ is a continuous function, $\alpha(t)\ge 0$, $\g(t)$ and $\s(t)$ satisfy conditions (\[ric1\]) of Lemma \[ric\] with $$\label{xdot}
\b(t):=||\dot{x}(t)||+\a(t),\q v(0):=||z_0-x(0)||,\q v(t):=||z(t)-x(t)||,$$ and $\mu(t)$ tends to $+\infty$ as $t\to +\infty$.
Then problem (\[caupr\]) has a unique solution $z(t)$ defined for all $t\in
[0,\infty)$, and $$\label{solut}
||z(t)-x(t)||<\frac 1 {\mu(t)},\q \lim_{t\to+\infty}||z(t)-x(t)||=0.$$ \[t1\]
One can choose the regularizing operator $\P(h,t)$ in (\[caupr\]) such that condition (\[semon\]) holds in the case when $F^{\pr *}(h)F'(h)$ is not boundedly invertible (see Sect. 3). \[nobound\]
[**Proof**]{} [**of Theorem \[t1\]**]{} Since $\P(h,t)$ is Fréchet differentiable with respect to $h$ there exists the solution to problem (\[caupr\]) on the maximal interval $[0,T_1)$ of the existence of the solution to (\[caupr\]). One has to show that $T_1=+\infty$. Assume $T_1<+\infty$. Since $H$ is a real Hilbert space one has: $$\label{eneq}
\f{1}{2}\f{d}{dt}||z(t)-x(t)||^2=(\dot{z}-\dot{x},z(t)-x(t))
=(\P(z(t),t),z(t)-x(t))-(\dot{x},z(t)-x(t)).$$ Therefore from (\[semon\]) and (\[xdot\]) one obtains $$\label{est1}
\f{1}{2}\f{d}{dt}||z(t)-x(t)||^2\le
-\g||z(t)-x(t)||^2+\s(t)||z(t)-x(t)||^3+\b(t)||z(t)-x(t)||.$$ Denote $$v(t):=||z(t)-x(t)||.$$ From (\[est1\]) one has: $$v(t)\dot{v}(t)\le -\g(t)v^2(t)+\s(t)v^3(t)+\b(t)v(t).$$ If $v>0$, one gets: $$\label{ineq11}
\dot{v}(t)\le -\g(t)v(t)+\s(t)v^2(t)+\b(t).$$ If $v=0$ on some interval, then inequality (\[ineq11\]) is satisfied trivially because $\beta(t)\ge 0$. Thus (\[ineq11\]) holds for all $t>0$.
By Lemma \[ric\] one obtains $$\label{appr}
||z(t)-x(t)||\le\f{1}{\mu(t)},\q \t{ for } \q t\in[0,T_1).$$ From (\[appr\]) one concludes that $z(t)$ does not leave the ball $B_1$ centered at $x(t)$ with radius $(\min_{t\in [0,T_1]}\mu(t))^{-1}$ $ > 0$. Since $\max_{0\leq t \leq T_1}||x(t)||<\infty$, one concludes that $\sup_{0\leq t < T_1}||z(t)||<\infty$. Therefore there exists a sequence $\{t_n\}\to T_1$ such that $\{z(t_n)\}$ converges weakly to some $\tilde{z}$. From equation (\[caupr\]) one derives the uniform boundedness of the norm $||\dot{z}(t)||$ on $[0,T_1)$. Thus there exists $\lim_{t\to T_1}||z(t)-\tilde{z}||=0$. Since the conditions for the uniqueness and local solvability of the Cauchy problem for equation (\[caupr\]) with initial condition $z(T_1)=\tilde{z}$ are satisfied, one can continue the solution to (\[caupr\]) through $T_1$. This contradicts the assumption of maximality of $T_1$, thus $T_1=+\infty$. Moreover, from (\[ric5\]) one gets: $$\label{est2}
\lim_{t\to+\infty}||z(t)-x(t)||\le \lim_{t\to+\infty}\f{1}{\mu(t)}=0.$$ $\Box$
Regularized Continuous Methods\
for Monotone Operators
===============================
In this section we apply the regularization procedure described in Sect. 2 to solve nonlinear operator equation (\[noneq\]). Assume that $F$ is Fréchet differentiable and $$\label{frder}
(F'(h)\xi,\xi)\ge 0\q\t{ for all }\q h,\xi\in H.$$ Under this assumption the operator $F'(h)+\ep(t)I$ is boundedly invertible. Define $\P$ as follows: $$\label{rchn}
\P(h,t):=-[F'(h)+\ep(t)I]^{-1}[F(h)+\ep(t)(h-z_0)],$$ where $z_0\in H$ is an initial approximation point and $\ep(t)$ is some positive function on the interval $[0,\infty)$. Some restrictions on $\ep(t)$ will be stated in Theorem \[t3\].
An outline of the convergence proof is the following. First one considers an auxiliary well-posed problem: $$\label{pr1}
F_\ep(x):=F(x)+\ep(x-z_0)=0, \q \ep>0,$$ and shows that the difference between its solution $x(t)$ and the solution $z(t)$ to problem (\[caupr\]) tends to zero as $t\to +\infty$. On the other hand one shows that $x(t)$ converges to the exact solution $y$ of equation (\[noneq\]). Thus one proves the convergence of $z(t)$ to $y$ as $t\to +\infty$.
We recall first some definitions from nonlinear functional analysis which are used below. The most essential restrictions on the operator $F$ imposed in this section are (\[frder\]) and [*w*]{}-continuity of $F$. In particular they imply monotonicity and hemicontinuity of $F$.
A mapping $\p$ is monotone in a Hilbert space $H$ if $$(\p(x_1)-\p(x_2),x_1-x_2)\ge 0,\q \forall x_1,x_2\in H.$$
A mapping $\p$ is hemicontinuous at $x_0\in H$ if the map $t\to (\p(x_0+th_1),h_2)$ is continuous in a neighborhood of $t=0$ for any $h_1,h_1\in H$.
A mapping $\p$ is strongly monotone in a Hilbert space $H$ if there exists a constant $k>0$ such that $$(\p(x_1)-\p(x_2),x_1-x_2)\ge k||x_1-x_2||^2,\q \forall x_1,x_2\in H.$$
If $F$ is monotone and hemicontinuous then the problem (\[pr1\]) is uniquely solvable. \[l1\]
[**Proof**]{} [**of Lemma \[l1\]**]{} According to [@deim p. 100] problem (\[pr1\]) is solvable if the operator $F_\ep$ is monotone, hemicontinuous and $||F_\ep(x)||\to\infty$ as $||x||\to\infty$. For sufficiently small $\delta>0$ from (\[pr1\]) one has: $$||F_\ep(x)||^2\ge
||F(x)||^2+\ep^2||x-z_0||^2-\f{1}{4\delta}||F(z_0)||^2-\delta||x-z_0||^2$$ $$\ge(\ep^2-\delta)||x-z_0||^2-C\to +\infty$$ as $||x||\to+\infty$. Here $C$ is a constant. Since $F_\ep$ is strongly monotone the solution to (\[pr1\]) is unique. Therefore Lemma \[l1\] is proved. $\Box$
The result given by Lemma \[l1\] is well known and its proof is given for the convenience of the reader.
Let $\rightharpoonup$ denote weak convergence in $H$.
We say that $F$ is [*w*]{}-continuous if $x\rightharpoonup \xi$ implies $F(x)
\rightharpoonup F(\xi)$.
Suppose that $F$ is w-continuous, all the assumptions of Lemma \[l1\] are satisfied, and there exists a unique solution $y$ to (\[noneq\]). Let $x(t)$ solve (\[pr1\]) for $\ep=\ep(t)$, and $\ep(t)$ tend to zero as $t\to +\infty$. Then $$\label{znorms}
\lim_{t\to+\infty}||x(t)-y||=0.$$ \[l2\]
[**Proof**]{} First let us show that $x(t)$ is bounded. Indeed, it follows from (\[pr1\]) that $$F(x(t))-F(y)+\ep(t)(x(t)-y)=\ep(t)(z_0-y).$$ Therefore $$\label{3.6}
(F(x(t))-F(y), x(t)-y)+\ep(t)||x(t)-y||^2=\ep(t)(z_0-y,x(t)-y).$$ This and (\[frder\]) imply $$\label{toobad}
||x(t)-y||\le||z_0-y||.$$ Thus there exists a sequence $\{x(t_n)\}$, $t_n\to\infty$ as $n\to\infty$, which converges weakly to some element $\y\in H$. Let us show that $\y$ is the (unique) solution to problem (\[noneq\]). Since $F$ is [*w*]{}-continuous, $F(x(t_n))\rightharpoonup F(\y)$. Because of the weak lower semicontinuity of the norm in a Hilbert space one has: $$\label{wsol}
||F(\y)||\le\liminf_{n\to\infty}||F(x(t_n))||=\liminf_{n\to\infty}\ep(t_n)||x(t_n)-z_0||
=0.$$ The conclusion $F(\y)=0$ follows from (\[wsol\]) and can also be derived directly from (\[pr1\]) with $\ep=\ep(t)\to 0$ as $t\to +\infty$. By the uniqueness of the solution to equation (\[noneq\]) one concludes that $\y=y$. Let us show that the sequence $\{x(t_n)\}$ converges strongly to $y$. Indeed, from (\[3.6\]), (\[frder\]) and the relation $x(t_n)\rightharpoonup y$, one gets: $$\label{ineq1}
||x(t_n)-y||^2\le (z_0-y,x(t_n)-y)\to 0\t{ as }n\to\infty.$$ Thus $$\label{lim1}
\lim_{n\to\infty}||x(t_n)-y||=0.$$ From (\[lim1\]) it follows by the standard argument that $x(t)\to y$ as $t\to\infty$. Lemma \[l2\] is proved. $\Box$
Assume that $F$ is continuously Fréchet differentiable, $\sup_{x\in H}||F'(x)||\le N_1$, and condition (\[frder\]) holds. If $\ep(t)$ is continuously differentiable, then the solution $x(t)$ to problem (\[pr1\]) with $\ep=\ep(t)$ is continuously differentiable in the strong sense and one has $$\label{est8}
||\dot{x}(t)||\le\f{|\dot{\ep}(t)|}{\ep(t)}||y-z_0||,\q t\in [0,+\infty).$$ \[l3\]
[**Proof**]{} Fréchet differentiability of $F$ implies hemicontinuity of $F$. Therefore problem (\[pr1\]) with $\ep=\ep(t)$ is uniquely solvable. The differentiability of $x(t)$ with respect to $t$ follows from the implicit function theorem \[2\]. To derive (\[est8\]) one differentiates equation (\[pr1\]) and uses the estimate $\left\Vert\left[F'(x(t))+\ep(t)I\right]^{-1}\right\Vert\le\f{1}{\ep(t)}$. The result is: $$\label{deriv}
||\dot{x}(t)||=|\dot{\ep}(t)|\cdot ||[F'(x(t))+\ep(t)I]^{-1}(x(t)-z_0)||\le
\f{|\dot\ep(t)|}{\ep(t)}||x(t)-z_0||.$$ Here we have used the estimate $$\label{3.12}
||x(t)-z_0||\leq ||y-z_0||,$$ which can be derived from (\[pr1\]) similarly to the derivation of (\[toobad\]). Thus estimate (\[est8\]) follows from (\[deriv\]) and (\[3.12\]). $\Box$
Assume that $\ep=\ep(t)>0$, $F$ is twice Fréchet differentiable, condition (\[frder\]) holds, and $$\label{N}
||F'(x)||\le N_1, \q ||F''(x)||\le N_2\q \forall x\in H.$$ Then for the operator $\P$ defined by (\[rchn\]) and $x(t)$, the solution to (\[pr1\]) with $\ep=\ep(t)$, estimate (\[semon\]) holds with $$\label{sigma}
\a(t)\equiv 0,\q \g(t)\equiv 1,\t{ and } \s(t):=\f{N_2}{2\ep(t)}.$$ \[l4\]
[**Proof**]{} Since $x(t)$ is the solution to (\[pr1\]) applying Taylor’s formula one gets: $$(\P(h,t),h-x(t))=-([F'(h)+\ep(t)I]^{-1}[F(h)-F(x(t))+\ep(t)(h-x(t))],h-x(t))$$ $$\leq - \Biggl([F'(h)+\ep(t)I]^{-1}[F'(h)(h-x(t))+\ep(t)(h-x(t))],h-x(t)\Biggl)+
\f{N_2||h-x(t)||^3}{2\ep(t)}$$ $$\label{comp}
= -||h-x(t)||^2+\f{N_2||h-x(t)||^3}{2\ep(t)}.$$ From (\[comp\]) and (\[semon\]) the conclusion of Lemma \[l4\] follows. $\Box$
Let us state the main result of this section.
Assume:
1. problem (\[noneq\]) has a unique solution $y$;
2. $F$ is w-continuous, twice Fréchet differentiable and inequalities (\[frder\]), (\[N\]) hold;
3. $\ep(t)>0$ is continuously differentiable and monotonically tends to $0$, $$\label{rhoc}
C_\ep:=\max_{t\in[0,\infty)}\f{\ep(0)|\dot{\ep}(t)|}{\ep^2(t)}<1;$$
4. $$\label{rho}
\ep(0)>\f{N_2||z_0-y||}{1-C_\ep}
\max\left\{1,\f{2C_\ep}{1-C_\ep}\right\};$$
5. $\P$ is defined by (\[rchn\]).
Then Cauchy problem (\[caupr\]) has a unique solution $z(t)$ for $t\in [0,+\infty)$ and $$\label{lap}
||z(t)-x(t)||\le\f{1-C_\ep}{N_2}\ep(t),\q \lim_{t\to+\infty}||z(t)-y||=0.$$ \[t3\]
First notice that Theorem \[t3\] establishes convergence for any initial approximation point $z_0$ if $\ep(t)$ is appropriately chosen. To make an appropriate choice of $\ep(t)$ one has to choose some function $\ep(t)$ satisfying condition (\[rhoc\]). Examples of such functions $\ep(t)$ are given below. One can observe that condition (\[rhoc\]) is invariant with respect to a multiplication $\ep(t)$ by a constant. Therefore one can choose $\ep(t)$ satisfying condition (\[rho\]) by a multiplication of the original $\ep(t)$ by a sufficiently large constant. If $\f{|\dot{\ep}(t)|}{\ep^2(t)}$ is not increasing, then in condition (\[rho\]) $C_\ep:=\max_{t\in[0,\infty)}\f{\ep(0)|\dot{\ep}(t)|}{\ep^2(t)}$ can be replaced by $C_\ep:=\f{|\dot{\ep}(0)|}{\ep(0)}$. \[noincon\]
In order to get an estimate of the convergence rate for $||x(t)-y||$ one has to make some additional assumptions either on $F(x)$ or on the choice of the initial approximation $z_0$. Without such assumptions one cannot give an estimate of the convergence rate. Indeed, as a simple example consider the scalar equation $F(x):=x^m=0$. Then one gets the following algebraic equation for $x(\ep)$: $$\label{algeq}
F_\ep(x):=x^m+\ep(x-z_0)=0.$$ Assume $m$ is a positive integer and $z_0>0$. It is known that the solution to this equation is an algebraic function which can be represented by the Puiseux series: $x=\sum_{j=1}^\infty c_j\ep^{\f{j}{p}}$ in some neighborhood of zero. Thus $x=c_1\ep^{\f{1}{p}}(1+O(\ep))$ as $\ep\to 0$. Now from (\[algeq\]) one gets: $$c_1^m\ep^{\f{m}{p}}(1+O(\ep))+c_1\ep^{1+\f{1}{p}}(1+O(\ep))=z_0\ep.$$ Thus $p=m$, $c_1=z_0^{\f{1}{m}}$ and $x(\ep)=z_0^{\f{1}{m}}\ep^{\f{1}{m}}(1+O(\ep))$. For $\ep=0$ one gets the solution $y=0$. Therefore $$\label{algeq1}
|x(\ep)-y|\sim \ep^{\f{1}{m}},\q \ep\to 0,$$ where $m$ can be chosen arbitrary large.
Below in Propositions \[prop1\] and \[prop2\] some sufficient conditions are given that allow one to obtain the estimates for $||x(t)-y||$. \[prem\]
[**Proof**]{} [**of Theorem \[t3\]**]{} Choosing $\mu(t)=\f{\lambda}{\ep(t)}$, where $\lambda$ is a constant, from conditions (\[ric1\]) and (\[xdot\]) one gets the following inequalities: $$\label{rho1}
N_2\le\lambda\left(1-\f{|\dot{\ep}(t)|}{\ep(t)}\right),$$ $$\label{rho2}
2||z_0-y||\f{|\dot{\ep}(t)|}{\ep^2(t)}\le\f{1}{\lambda}
\left(1-\f{|\dot{\ep}(t)|}{\ep(t)}\right),\q \lambda||z_0-x(0)||<\ep(0).$$ Choose $$\label{chlam}
\lambda:=\f{N_2}{1-C_\ep}.$$ It follows from (\[rhoc\]) that (\[rho1\]) holds. From (\[rho\]) one gets: $$\label{ad1}
\ep(0)>\f{N_2||z_0-y||}{1-C_\ep}.$$ On the other hand from (\[toobad\]) it follows that $$\label{ad2}
||z_0-x(0)||\le ||z_0-y||.$$ Thus from (\[ad1\]), (\[ad2\]) and (\[chlam\]) one obtains the second inequality in (\[rho2\]). Using (\[rho\]) once again, one gets: $$\label{ad3}
\ep(0)\geq \f{2N_2C_\ep||z_0-y||}{(1-C_\ep)^2}.$$ This inequality and (\[rhoc\]) imply: $$\label{ad4}
\lambda=\f{N_2}{1-C_\ep}\le
\f{1-C_\ep}{2||z_0-y||\f{|\dot{\ep}(t)|}{\ep^2(t)}}.$$ By (\[rhoc\]) one gets $$\label{ad5}
\f{N_2}{1-C_\ep}\le\f{1-\f{|\dot{\ep}(t)|}{\ep(t)}}{2||z_0-y||
\f{|\dot{\ep}(t)|}{\ep^2(t)}}.$$ The first inequality in (\[rho2\]) is equivalent to (\[ad5\]) for $\lambda$ chosen in (\[chlam\]). Therefore one gets inequality (\[lap\]) by applying Theorem \[t1\], while the second relation (\[lap\]) follows from (\[znorms\]), inequality (\[lap\]) and the triangle inequality: $$||z(t)-y||\le ||z(t)-x(t)||+||x(t)-y||.$$ Theorem \[t3\] is proved. $\Box$
[**Examples.**]{}
[**1.**]{} Let $\ep(t)=\ep_0(t_0+t)^{-\nu}$, $\ep_0$, $t_0$ and $\nu$ are positive constants. Then $C_\ep=\f{\nu}{t_0}$ and condition (\[rhoc\]) is satisfied if $\nu\in (0,1]$ and $t_0>\nu$.
[**2.**]{} If $\ep(t)=\frac {\ep_0}{\log(t_0+t)}$, then $C_\ep=\f{1}{t_0\log t_0}$ and condition (\[rhoc\]) is satisfied if $t_0\log t_0>1$.
Note that if $\ep(t)=\ep_0e^{-\nu t}$ then condition (\[rhoc\]) is not satisfied.
[*Let the assumptions of Theorem \[t3\] hold. Suppose also that the following inequality holds: $$\label{pr11}
(F(h),h-y)\ge c||h-y||^{1+a},\q a>0.$$ Then for the solution $z(t)$ to problem (\[caupr\]) the following estimate holds: $$\label{pr12}
||z(t)-y||=O\left(\ep^{\f{1}{a}}(t)\right).$$* ]{} \[prop1\]
[**Proof**]{} Denote $||x(t)-y||:=\varrho(t)$. Since $F(y)=0$, inequality (\[3.6\]) implies $$\label{pr13}
c\varrho^{1+a}(t)+\ep(t)\varrho^2(t)\le \ep(t)||z_0-y||\varrho(t)$$ and $\varrho(t) \to 0$ as $t\to +\infty$. This inequality can be reduced to $$\label{pr14}
c\varrho^{a}(t)+\ep(t)\varrho(t)\le \ep(t)||z_0-y||.$$ Thus $\varrho^{a}(t)\le \f{||z_0-y||}{c} \ep(t)$, and $$\label{pr15}
||x(t)-y||\le \left( \f{||z_0-y||}{c}\right)^{\f{1}{a}}\ep^{\f{1}{a}}(t).$$ Combining this estimate with estimate (\[lap\]) for $||z(t)-x(t)||$ one completes the proof. $\Box$
[**Example.**]{} In the case of a scalar function $f(h)$ and even integer $a>0$ the estimate $f(h)(h-y)\ge c|h-y|^{1+a}$ means that $f(h)=(h-y)^{a}g(h)$, where $g(h)\ge c>0$, and hence $y$ is a zero of the multiplicity $a$ for $f$.
[*Let all the assumptions of Theorem \[t3\] hold and there exists $v\in H$ such that $$\label{pr21}
z_0-y=F'(y)v,\q ||v||<\f{2}{N_2}.$$ Then for the solution $z(t)$ to problem (\[caupr\]) the following convergence rate estimate holds: $$\label{pr22}
||z(t)-y||\le \left [\f{1-C_\ep}{N_2}+\f{4||v||}{2-N_2||v||}\right ]\ep(t).$$* ]{} \[prop2\]
[**Proof**]{} From (\[pr1\]) for an arbitrary $\ep>0$ one gets $$F(x)-F(y)+\ep(x-y)=\ep(z_0-y).$$ Therefore by the Lagrange formula one has: $$\label{pr23}
\left\{\il_0^1(F'(y+s(x-y))ds+\ep I\right\}(x-y)=\ep(z_0-y).$$ Introduce the notation $Q_\ep(x):=\il_0^1(F'(y+s(x-y))ds+\ep I$. From (\[pr23\]) it follows that $$||x-y||=\ep||Q_\ep^{-1}(x)Q_0(y)v||\le \ep||Q_\ep^{-1}(x)(Q_0(y)-Q_\ep(x))v||
+\ep||Q_\ep^{-1}(x)Q_\ep(x)v||.$$ Since $Q_\ep(x)=Q_0(x)+\ep I$, one obtains $$||x-y||\le \ep||Q_\ep^{-1}(x)(Q_0(y)-Q_0(x))v||+\ep||Q_\ep^{-1}(x)\ep v||
+\ep||v||$$ $$\label{pr24}
\le\f{N_2}{2}||x-y||\,||v||+2\ep ||v||.$$ So, from (\[lap\]), (\[pr21\]) and (\[pr24\]) for $\ep=\ep(t)$ and correspondingly $x=x(t)$ satisfying the assumptions of Theorem \[t3\] one concludes that estimate (\[pr22\]) holds. $\Box$
Now we describe the simple iteration scheme for solving nonlinear equation (\[noneq\]). Define: $$\label{4.1}
\P(h,t):=-[F(h)+\ep(t)(h-z_0)],\q \ep(t)>0.$$
Assume that $F$ is monotone, $\P$ is defined by (\[4.1\]), and $x(t)$ is a solution to problem (\[pr1\]) with $\ep=\ep(t)>0,$ $t\in [0,+\infty)$. Then for the positive function $\g(t):=\ep(t)$ and for $\s(t)=\a(t)\equiv 0$ estimate (\[semon\]) holds. \[l9\]
[**Proof**]{} Since $x(t)$ is a solution to problem (\[pr1\]), by the monotonicity of $F$ one has: $$(\P(h,t),h-x(t))=-(F(h)-F(x(t),h-x(t))-\ep(t)(h-x(t),h-x(t))$$ $$\label{star1}
\le -\ep(t)||h-x(t)||^2.$$ Lemma \[l9\] is proved. $\Box$
Lemma \[l9\] together with Lemma \[stupid\] presented below allow one to formulate the convergence result concerning the simple iteration procedure (see Theorem \[t5\]).
Let $\nu(t)$ be integrable on $[0,+\infty)$. Suppose that there exists $T\geq 0$ such that $\nu(t)\in C^1[T,+\infty)$ and $$\label{1b}
\nu(t)>0,\q -\f{\dot{\nu}(t)}{\nu^2(t)}\leq C,\q \hbox{for}\q t\in [T,+\infty).$$ Then $$\label{2b}
\lim_{t\to+\infty}\il^t_0\nu(\tau)d\tau = +\infty.$$ \[stupid\]
[**Proof**]{} One can integrate (\[1b\]) $$-\il^t_T \f{\dot{\nu}(\tau)}
{\nu^2(\tau)}d\tau\leq \il^t_T Cd\tau,\q t\in [T,+\infty)$$ and get $$\f{1}{\nu(t)}\leq C(t-T) + \f{1}{\nu(T)}.$$ Without loss of generality we can assume that $C>0$, and then $$\nu(t)\geq\f{1}{ C(t-T) + \f{1}{\nu(T)}}.$$ Integrating this inequality one gets (\[2b\]) and completes the proof. $\Box$
Lemmas \[l1\] - \[l3\] and Lemmas \[l9\] - \[stupid\] imply the following result.
Assume that:
1. problem (\[noneq\]) has a unique solution $y$;
2. $F$ is w-continuous and monotone;
3. $F$ is continuously Fréchet differentiable and $$\label{N1}
||F'(x)||\le N_1, \q \forall x\in H;$$
4. $\ep(t)>0$ is continuously differentiable and tends to zero monotonically as $t\to
+\infty$, and $\lim_{t\to +\infty}\f{\dot{\ep}(t)}{\ep^2(t)}=0$.
Then, for $\P$ defined by (\[4.1\]), Cauchy problem (\[caupr\]) has a unique solution $z(t)$ all for $t\in [0,+\infty)$ and $$\lim_{t\to\infty}||z(t)-y||=0.$$ \[t5\]
[**Proof**]{} In order to verify the assumptions of Theorem \[t1\] we use estimate (\[star1\]) to conclude that $\a(t)=\s(t)=0$ and $\g(t)=\ep(t)$ in formula (\[semon\]). By (\[xdot\]) $\b(t)=||\dot{x}(t)||$ because $\a(t)=0$. By (\[est8\]) $$\
\b(t)=||\dot{x}(t)||\le\f{|\dot{\ep}(t)|}{\ep(t)}||y-z_0||.$$ To apply Theorem \[t1\] one has to find a function $\mu(t) \in C^1[0,+\infty)$ satisfying (\[ric1\]) that is $$\label{1c}
\f{|\dot\ep(t)|}{\ep(t)}||y-z_0||\leq \f{1}{2\mu(t)}\left(\ep(t)-\f{
\dot\mu(t)}{\mu(t)}\right),\q \mu(0)||x(0)-z_0||<1.$$ Such a function can be chosen as the solution to the differential equation $$\label{2c}
-\f{\dot{\mu}(t)}{\mu^2(t)}+\f{\ep(t)}{\mu(t)}=\f{A|\dot{\ep}(t)|}{\ep(t)},$$ where $A:=2||y-z_0||$. Denote $\rho(t):=\f{1}{\mu(t)}.$ Then $$\dot{\rho}(t)+\ep(t)\rho(t)=\f{A|\dot{\ep}(t)|}{\ep(t)}$$ and $$\label{3c}
\rho(t)=\left[A\il^t_0\f{|\dot{\ep}(s)|}{\ep(s)}e^{\il^s_0\ep(\tau)d\tau}ds
+\f{1}{\mu(0)}\right]
e^{-\il^t_0\ep(\tau)d\tau}.$$ Since by Lemma \[stupid\] $e^{\il^t_0\ep(\tau)d\tau}\to\infty$ as $t\to\infty$ one can apply L’Hôspital’s rule to obtain from (\[3c\]) and condition 4 that $$\lim_{t\to +\infty}\rho(t)=\lim_{t\to +\infty}\f{|\dot{\ep}(t)|}{\ep^2(t)}=0.$$ Therefore $\mu(t)$ tends to $+\infty$. To complete the proof one can take $\mu(0)$ sufficiently small for the second inequality in (\[1c\]) to hold. By Theorem \[t1\] one concludes that $||z(t)-x(t)||\to 0$ as $t\to +\infty$ and by Lemma \[l2\] that $||x(t)-y||\to 0$ as $t\to +\infty$. Therefore it follows from the estimate: $$||z(t)-y||\le ||z(t)-x(t)||+||x(t)-y||,$$ that $||z(t)-y||\to 0$ as $t\to +\infty$. $\Box$
From the proof it is clear that one can get the estimate $||z(t)-x(t)||\le\f{1}{\mu(t)}\to 0$ as $t\to +\infty$, and for the term $||x(t)-y||$ one can get the rate of convergence if some additional assumptions are made on $F$ or on $z_0$ (see Propositions \[prop1\], \[prop2\], and also Remark \[prem\]).
The result obtained in Theorem \[t5\] is similar to the result in [@al2]. The assumptions in [@al2] are slightly different. The method of investigation in [@al2] is based on a linear differential inequality which is a particular case of (\[ric4\]) with $\s(t)\equiv 0$. This linear differential inequality has been used often in the literature by many authors.
[**Examples.**]{}
[**1.**]{} Let $\ep(t)=\ep_0(1+t)^{-\nu}$, $\ep_0$ and $\nu$ are positive constants. Then the assumptions of Theorem \[t5\] are satisfied if $\nu\in (0,1)$.
[**2.**]{} If $\ep(t)=\f{\ep_0}{\log(1+t)}$, then the assumptions of Theorem \[t5\] are satisfied.
If $\ep(t)=\ep_0e^{-\nu t}$ then condition 4 of Theorem \[t5\] is not satisfied.
Regularized Methods for Non-monotone Operators
==============================================
In this section we discuss two approaches to the regularization of the Gauss-Newton-type schemes for nonlinear equations with non-monotone operators. To describe the first one, assume that $F$ in (\[noneq\]) is compact and Fréchet differentiable. Denote: $$\label{T}
T(h):=F^{\pr *}(h)F'(h),\q T_{\ep}(h):=T(h)+\ep I.$$ Then $T(h)$ is a nonnegative self-adjoint compact operator. Such an operator cannot be boundedly invertible if $H$ is infinite-dimensional. One has: $$\label{ep}
|| T^{-1}_{\ep}(h)||\leq \f{1}{\ep}$$ for any $\ep >0$. Define $\P$: $$\label{P1}
\P(h,t):=(P_{\ep(t)}(\xi)-I)(h-z_0)-P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(h)F^{\pr
*}(h)F(h),\q
\ep(t)>0.$$ Here $\xi\in H$ is a fixed element, which will be chosen so that inequality (\[sup\]) (see below) holds, and $$\label{proj}
P_{\ep}(\xi):=\il_{\ep}^{N_1^2}dE(s),$$ $E(s):=E(s,\xi)$ is the resolution of the identity of the self-adjoint operator $T(\xi)$, and $$\label{N12}
\sup_{h\in H}||F'(h)||\leq N_1,$$ so that $||T(\xi)||\le N_1^2$.
Assume that:
1. problem (\[noneq\]) has a unique solution $y$;
2. $\P$ is defined by (\[P1\]), $F$ is compact, twice Fréchet differentiable and inequalities (\[N\]) hold;
3. there exist $\xi\in H$ and $ \ep_0>0$ such that $$\label{sup}
C:=\sup_{\ep\in (0,\ep_0]} ||P_{\ep}(\xi)T^{-1}_{\ep}(\xi)
\{T(y)-T(\xi)\}||<\f{1}{2};$$
4. $0<\ep(t)\leq \ep_0$.
Then there exist positive functions $\a(t)$, $\g(t)$ and $\s(t)$, $t\in [0,+\infty)$, such that estimate (\[semon\]) holds with $x(t)$ replaced by $y$. \[l41\]
Condition (\[sup\]) contains a priori information about a nonlinear operator $F$. This condition allows one to get a convergence rate for ill-posed problem (\[noneq\]). It is always satisfied in a well-posed case (for a boundedly invertible operator $T$) if $\xi$ is sufficiently close to $y$. However it is not clear yet how restrictive this condition is, and how it is related to other conditions that one has to use in order to prove the convergence of the process in ill-posed cases. \[wrem\]
[**Proof**]{} [**of Lemma \[l41\]**]{} Using the polar decomposition $F'(h)=U(F^{\pr *}(h)F'(h))^{\f{1}{2}}$, where $U$ is a partial isometry, one gets $F^{\pr *}(h)=T^{\f{1}{2}}U^*$, and, since $||U^*||=1$, one obtains: $$\label{polar}
||T^{-1}_{\ep(t)}(h)F^{\pr *}(h)||\leq||T^{-1}_{\ep(t)}(h)T^{\f{1}{2}}(h)||\leq
\max_{0\le s<+\infty}\f{\sqrt{s}}{s+\ep(t)}=\f{1}{2\sqrt{\ep(t)}}.$$ Using (\[P1\]) and the relation $$\label{47a}
F(h)=F(h)-F(y)=F'(h)(h-y)+R(y,h),$$ where $||R(y,h)||\le\f{N_2}{2}||h-y||^2$, one gets $$(\P(h,t),h-y)=((P_{\ep(t)}(\xi)-I)(h-z_0),h-y)
-(P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(h)T(h)(h-y),h-y)$$ $$+\f{N_2}{4\sqrt{\ep(t)}}||h-y||^3\leq -||h-y||^2+||(P_{\ep(t)}(\xi)-I)(y-z_0)||
||h-y||$$ $$\label{4.7}
+\ep(t)(P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(h)(h-y),h-y)+\f{N_2}{4\sqrt{\ep(t)}}||h-y||^3.$$ Also one has the following estimates: $$\label{2ep}
||P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(\xi)||\le\max_{s\ge\ep(t)}\f{1}{s+\ep(t)}=
\f{1}{2\ep(t)},$$ and $$||T(h)-T(y)||\le ||F^{\pr *}(h)[F'(h)-F'(y)]+[F^{\pr *}(h)-F^{\pr *}(y)]F'(y)||$$ $$\label{TT}
\leq 2N_1N_2||h-y||.$$ Thus, using the identity $A^{-1}-B^{-1}=-B^{-1}(A-B)A^{-1}$ and inequalities (\[2ep\]), (\[TT\]), one obtains: $$\ep(t)(P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(h)(h-y),h-y)=
\ep(t)(P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(\xi)(h-y),h-y)$$ $$+\ep(t)(P_{\ep(t)}(\xi)
\{T^{-1}_{\ep(t)}(h)-T^{-1}_{\ep(t)}(\xi)\}(h-y),h-y)\leq \f{1}{2}||h-y||^2$$ $$+||P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(\xi)||\,||T(h)-T(y)||\,
||h-y||^2+||P_{\ep(t)}(\xi)T^{-1}_{\ep(t)}(\xi)\{T(y)-T(\xi)\}||\,||h-y||^2$$ $$\label{4.8}
\leq\left(\f{1}{2}+C\right)||h-y||^2+\f{N_1N_2}{\ep(t)}
||h-y||^3.$$ Define: $$\label{g}
\g(t)\equiv\g:=\f{1}{2}-C;$$ $$\label{s}
\sigma(t):=\f{N_1N_2}{\ep(t)}+\f{N_2}{4\sqrt{\ep(t)}};$$ $$\label{a}
\a(t):=||(P_{\ep(t)}(\xi)-I)(y-z_0)||.$$ These functions are positive if (\[sup\]) holds. If $\epsilon (t)\to 0$ then $\sigma(t)\to +\infty$ and $\alpha(t)\to 0$ as $t\to +\infty$. Comparing (\[4.7\]) - (\[a\]) with (\[semon\]) and applying Lemma \[ric\] one completes the proof. $\Box$
Suppose that the assumptions of Lemma \[l41\] are satisfied and:
1. $P(y-z_0)=0$, where $P$ is an orthonormal projector onto the null-space of $T(\xi)$;
2. $\ep(t)>0$ is continuously differentiable, monotonically tends to $0$, and $$\label{c0min}
C_0:=\min_{t\in [0,+\infty)}\left\{\g-\f{|\dot{\ep}(t)|}{\ep(t)}\right\}>0,$$ with $\g$ defined in (\[g\]);
3. $C_{\a}:=\ep(0)\max_{t\in [0,+\infty)}\left\{\f{\a(t)}{\ep(t)}\right\}
<\infty$;
4. $\ep(0)$ is chosen so that $$\label{sbg}
\f{4N_1N_2+N_2\sqrt{\ep(0)}}{2C_0}<\ep(0)\min\left\{\f{C_0}{2C_\a},
\f{1}{||y-z_0||}\right\}.$$
Then Cauchy problem (\[caupr\]) has a unique solution $z(t)$ for $t\in
[0,\infty)$ and $$\label{n}
||z(t)-y||\le \f{2C_0}{4N_1N_2+N_2\sqrt{\ep(0)}}\ep(t).$$ \[t41\]
If $T(\xi)$ is injective then Condition 1 is satisfied automatically. \[nlsp\]
From (\[a\]) one can see that $\a$ depends on $y$ and cannot be known a priori. However it follows from condition 1 of Theorem \[t41\] that $\a(t)\to 0$ as $t\to
+\infty$. Therefore in numerical applications of this scheme one should try different functions $\ep(t)$ (and different points $z_0$) to satisfy condition 3. Then, since $C_0$ and $C_\a$ are invariant with respect to multiplication of $\ep(t)$ by a positive constant, one can choose $\ep(0)$ sufficiently large in order to satisfy condition 4. Such a choice can be done for an arbitrary $z_0$. \[yz0\]
[**Proof**]{} [**of Theorem \[t41\]**]{} Since $x(t)\equiv y$ in our case, one gets $\beta(t)=\a(t)$. Let us choose $\mu(t):=\f{\lambda}{\ep(t)}$. Conditions of Theorem \[t1\] can be written as follows: $$\label{gnc1}
\f{N_1N_2}{\ep(t)}+\f{N_2}{4\sqrt{\ep(t)}}\le
\f{\lambda}{2\ep(t)}\left(\g-\f{|\dot{\ep}(t)|}{\ep(t)}\right),$$ $$\label{gnc2}
\a(t)\le\f{\ep(t)}{2\lambda}\left(\g-\f{|\dot{\ep}(t)|}{\ep(t)}\right),\q
\lambda||y-z_0||<\ep(0).$$ Inequality (\[gnc1\]) is equivalent to the following one: $$\label{gnc1a}
\lambda\ge \f{4N_1N_2+N_2\sqrt{\ep(t)}}{2\left(\g-\f{|\dot{\ep}(t)|}{\ep(t)}\right)}.$$ Take $$\label{lmd}
\lambda:=\f{4N_1N_2+N_2\sqrt{\ep(0)}}{2C_0}.$$ Then (\[gnc1a\]) follows from (\[lmd\]), from the monotonicity of $\ep(t)$ and from (\[c0min\]). Inequality (\[sbg\]) implies that $$\label{4.18c}
\f{4N_1N_2+N_2\sqrt{\ep(0)}}{2C_0}<\f{C_0}{2\max_{t\in
[0,+\infty)}\left\{\f{\a(t)}{\ep(t)}\right\}}$$ If $\lambda$ is defined by (\[lmd\]), then one has: $$\max_{t\in
[0,+\infty)}\left\{\f{\a(t)}{\ep(t)}\right\}<\f{C_0}{2\lambda}.$$ From (\[c0min\]) it follows that the first inequality in (\[gnc2\]) holds. Finally, one obtains from (\[sbg\]) that $$\f{4N_1N_2+N_2\sqrt{\ep(0)}}{2C_0}<\f{\ep(0)}{||y-z_0||},$$ which implies the second inequality in (\[gnc2\]). Since $x(t)\equiv y$ and $\mu(t):=\f{\lambda}{\ep(t)}$, by (\[est2\]) and (\[lmd\]) one concludes that (\[n\]) holds. $\Box$
Consider now a variant of Gauss-Newton continuous method with $\P$ defined as follows: $$\label{P2}
\P(h,t):=-T^{-1}_{\ep(t)}(h)\{F^{\pr *}(h)F(h)+\ep(t)(h-z_0)\}.$$ Here $F$ is not assumed compact. The following lemma is a consequence of Theorem 2.3 in [@ars] (see also Theorem 2.4 in [@bns]).
Assume that:
1. problem (\[noneq\]) has a unique solution $y$;
2. $\P$ is defined by (\[P2\]), $F$ is twice Fréchet differentiable and estimates (\[N\]) hold;
3. $z_0$ is chosen so that, for some $v\in H$, one has: $$\label{istok}
y-z_0=T^\zeta(y)v,\q \f{1}{2}\le\zeta\le 1;$$
4. $\ep(t)>0$, $$\left[\f{1}{2}\ep^{\zeta-\f{1}{2}}(t)N_1\zeta^\zeta(1-\zeta)^{1-\zeta}+
\f{N_1^{2(\zeta+1)}}{N_1^2+\ep(t)}\right]||v||<1,$$ where $$(1-\zeta)^{1-\zeta}|_{\zeta=1}:=\lim_{\zeta\to
1-0}(1-\zeta)^{1-\zeta}=1.$$
Then for $x(t)\equiv y$ estimate (\[semon\]) holds with $$\label{alsig}
\a(t):=\ep^\zeta(t)\zeta^\zeta(1-\zeta)^{1-\zeta}||v||,\q
\s(t):=\f{N_2}{4\sqrt{\ep(t)}},$$ $$\label{gam}
\g(t):=1-\f{1}{2}\ep^{\zeta-\f{1}{2}}(t)N_2\zeta^\zeta(1-\zeta)^{1-\zeta}||v||
-\f{N_1^{2(\zeta+\f{1}{2})}N_2||v||}{N_1^2+\ep(t)}>0.$$ \[l42\]
The convergence theorem in the case $\zeta=1$ for continuous Gauss-Newton method is proved in [@ars Theorem 2.3]. For the discrete Gauss-Newton method the case $0\le\zeta <\f{1}{2}$ is analyzed in [@bns] under some additional assumptions on the operator. For the noise-free case it is shown that the rate of convergence is $o(\ep^{\zeta})$. \[bnsrem\]
[**Proof**]{} [**of Lemma \[l42\]**]{} From (\[47a\]) and (\[istok\]) one gets $$(\Phi(h,t),h-y)=-(T_{\ep(t)}^{-1}(h)[F^{\pr *}(h)F'(h)+\ep(t)(h-y)+\ep(t)(y-z_0)],
h-y)$$ $$\le -||h-y||^2+\f{N_2}{4\sqrt{\ep(t)}}||h-y||^3-
\ep(t)(T_{\ep(t)}^{-1}(h)T^\zeta(y)v,h-y).$$ Following [@bns] one estimates the inner product: $$(T_{\ep(t)}^{-1}(h)T^\zeta(y)v,h-y)=(T_{\ep(t)}^{-1}(h)[T(y)-T(h)]T_{\ep(t)}^{-1}(y)
T^\zeta(y)v,h-y)+(T_{\ep(t)}^{-1}(y)T^\zeta(y)v,h-y).$$ From the spectral theorem for selfadjoint linear operator $T(\eta)$, and for any $\eta\in H$, one gets: $$\label{spectral}
||T_{\ep(t)}^{-1}(\eta)T^\zeta(\eta)||\le\max_{s\ge
0}\f{s^\zeta}{s+\ep(t)}=
\f{\zeta^\zeta(1-\zeta)^{1-\zeta}}{\ep^{1-\zeta}(t)}.$$ Since $$T(h)-T(y)=F^{\pr *}(h)[F'(h)-F'(y)]+[F^{\pr *}(h)-F^{\pr *}(y)]F'(y)$$ and $\f{1}{2}\leq\zeta\le 1$, from the polar decomposition one gets the estimate $$||T^{\f{1}{2}}(\eta)T_{\ep(t)}^{-1}(\eta)T^\zeta(\eta)||
\le\max_{N_1^2\ge s\ge
0}\f{s^{\zeta+\f{1}{2}}}{s+\ep(t)}
\le\f{N_1^{2(\zeta+\f{1}{2})}}{N_1^2+\ep(t)},$$ which implies $$||T_{\ep(t)}^{-1}(h)[T(y)-T(h)]T_{\ep(t)}^{-1}(y)T^\zeta(y)||
\le ||T_{\ep(t)}^{-1}(h)F^{\pr *}(h)(F'(h)-F'(y))T_{\ep(t)}^{-1}(y)T^\zeta(y)||$$ $$+||T_{\ep(t)}^{-1}(h)(F^{\pr *}(h)-
F^{\pr *}(y))F'(y)T_{\ep(t)}^{-1}(y)T^\zeta(y)||$$ $$\le
\left[\f{N_2}{2\sqrt{\ep(t)}}\f{\zeta^\zeta(1-\zeta)^{1-\zeta}}{\ep^{1-\zeta}(t)}+
\f{N_2}{\ep(t)}\f{N_1^{2(\zeta+\f{1}{2})}}{N_1^2+\ep(t)}\right]||h-y||.$$ Therefore one obtains: $$(\Phi(h,t),h-y)\le \ep^\zeta(t)\zeta^\zeta(1-\zeta)^{1-\zeta}||v||\,||h-y||$$ $$\label{es423}
-\left\{1-\f{1}{2}\ep^{\zeta-\f{1}{2}}(t)N_2\zeta^\zeta(1-\zeta)^{1-\zeta}||v||-
\f{N_1^{2(\zeta+\f{1}{2})}N_2||v||}{N_1^2+\ep(t)}\right\}||h-y||^2+\f{N_2}{4\sqrt{\ep(t)}}||h-y||^3.$$ $\Box$
Note that assumption (\[istok\]) is not algorithmically verifiable. However, practitioners may try different $z_0$ and choose the one for which the algorithm works better, that is convergence is more rapid and the algorithm is more stable.
Assumptions of the type (\[istok\]) (sourcewise representation) became popular recently, because they allow one to establish some error estimates for the approximate solution. But one has to remember that the results based on such assumptions are of limited value because one has no algorithm for choosing $z_0$ for which (\[istok\]) holds, and $y$ in (\[istok\]) is unknown.
If $T=T^*$ is compact and the null space $N(T)=\{0\}$, then the range $R(T)$ is dense in $H$, so in any neighborhood of $y$ there are points $z_0$ for which (\[istok\]) holds. On the other hand, since $R(T)$ is not closed in the same neighborhood there are also points $z_0$ for which (\[istok\]) fails to hold. This is why the methods for solving nonlinear ill-posed problems, based on the assumption (\[istok\]) or similar assumptions are not quite satisfactory although they might work in practice sometimes, for reasons which are yet not clear.
In general, in order to get a convergence theorem in an ill-posed case one needs some additional assumptions on the Fréchet derivative of the operator $F$, for example condition (\[sup\]), or (\[istok\]), or some other condition ( see e.g., [@des], condition (2.11)).
Let the assumptions of Lemma \[l42\] be satisfied and
1. $\ep(t)$ is continuously differentiable, monotonically tends to $0$, and $$\label{C0}
C_0:=\min_{t\in [0,+\infty)}\left\{\g(t)-\zeta\f{|\dot{\ep}(t)|}{\ep(t)}\right\}>0,$$ with $\g(t)$ defined in (\[gam\]);
2. $z_0$ and $\ep(0)$ are chosen so that $$\label{min}
\f{N_2\ep^{\zeta-\f{1}{2}}(0)}{2C_0}<\min
\left\{\f{C_0}{2\zeta^\zeta(1-\zeta)^{1-\zeta}||v||},
\f{\ep^\zeta(0)}{||y-z_0||}\right\}.$$
Then the Cauchy problem (\[caupr\]) has a unique solution $z(t)$ for $t\in [0,+\infty)$ and $$\label{nn}
||z(t)-y||\le \f{2C_0}{N_2\ep^{\zeta-\f{1}{2}}(0)}\ep^\zeta(t).$$ \[t42\]
[**Proof**]{} Choose $\mu(t):=\f{\lambda}{\ep^\zeta(t)}$. Conditions of Theorem \[t1\] can be rewritten as follows: $$\label{perv}
\f{N_2}{4\sqrt{\ep(t)}}\le\f{\lambda}{2\ep^\zeta(t)}
\left(\g(t)-\zeta\f{|\dot{\ep}(t)|}{\ep(t)}\right),$$ $$\label{vtor}
\zeta^\zeta(1-\zeta)^{1-\zeta}||v||\le\f{1}{2\lambda}
\left(\g(t)-\zeta\f{|\dot{\ep}(t)|}{\ep(t)}\right),\q
\f{\lambda||y-z_0||}{\ep^\zeta(0)}<1.$$ Inequality (\[perv\]) is equivalent to the following one: $$\label{4.27a}
\f{N_2\ep^{\zeta-\f{1}{2}}(t)}{2\left(\g(t)-\zeta\f{|\dot{\ep}(t)|}
{\ep(t)}\right)}\leq \lambda.$$ If one takes $$\label{4.27b}
\lambda:=\f{N_2\ep^{\zeta-\f{1}{2}}(0)}{2C_0},$$ then (\[4.27a\]) follows from (\[4.27b\]), from the monotonicity of $\ep(t)$ and from (\[C0\]). Inequality (\[min\]) implies that $$\label{4.27c}
\f{N_2\ep^{\zeta-\f{1}{2}}(0)}{2C_0}<
\f{C_0}{2\zeta^\zeta(1-\zeta)^{1-\zeta}||v||}.$$ For $\lambda$ defined by (\[4.27b\]) inequality (\[4.27c\]) can be written as $$\label{4.27d}
\zeta^\zeta(1-\zeta)^{1-\zeta}||v||<\f{C_0}{2\lambda}.$$ From (\[4.27d\]) one obtains the first inequality (\[vtor\]). Finally, from (\[min\]) one concludes that $$\f{N_2\ep^{\zeta-\f{1}{2}}(0)}{2C_0}<
\f{\ep^{\zeta}(0)}{||y-z_0||},$$ which is equivalent to the second inequality (\[vtor\]) for $\lambda$ defined by (\[4.27b\]). Since $x(t)\equiv y$ and $\mu(t):=\f{\lambda}{\ep^\zeta(t)}$, by (\[est2\]) and (\[4.27b\]) one gets estimate (\[nn\]). $\Box$
One can take $x(t)$ in Theorem \[t42\] as the minimizer of the problem $$\label{argmin}
||\,F(x)||^2+\ep(t)||x-z_0||^2=\inf,\,\, x\in H,\q 0<\ep(t)\to 0,\t{
as }t\to +\infty,$$ instead of taking $x(t)\equiv y$. Problem (\[argmin\]) is solvable for [*w*]{}-continuous operator $F$ and $\lim_{t\to\infty}||x(t)-y||=0$. If one can obtain the estimate: $$\label{st}
||F(x(t))||=O(\ep(t)),$$ then one can prove that $$\label{star}
||\dot{x}(t)||=O\left(\f{|\dot{\ep}(t)|}{\ep(t)}\right)$$ and estimate (\[semon\]) holds. However we do not have examples of nonlinear operators $F(x)$ satisfying estimate (\[st\]) and such that $F'(y)$ is not boundedly invertible.
Numerical Example
=================
The aim of this section is to illustrate the efficiency of scheme (\[caupr\]) with $\P$ defined in (\[rchn\]) and in (\[P2\]) for solving a practically interesting ill-posed nonlinear equation.
Consider the following iterative process $$x^{k+1}=f_{\mu,z}(x^k)$$ for a differentiable function $f_{\mu,z}$ depending on two parameters $\mu$ and $z$ with the only maximum at the point $\hat{x}$. This process is characterized by the Feigenbaum constants: $$\delta_z=\lim_{j\to\infty}\f{\mu_{z_j}-\mu_{z_{j-1}}}{\mu_{z_{j+1}}-\mu_{z_{j}}},\q
\alpha_z=\lim_{j\to\infty}\f{d_{z_j}}{d_{z_{j+1}}},$$ where $\mu_j$ are the critical values of parameter $\mu$, for which a doubling of the period of the function $f_{\mu,z}$ occurs (the appearance of $2^j$ cycle), and $d_{z_j}$ is the algebraic distance (could be negative) between zero and the nearest attractor (the limit point in $2^j$ cycle). The calculation of $\a_z$ is a problem of a practical interest because it is not known yet if they satisfy any algebraic relations or not. As it is shown in [@fig], $\alpha_z$ can be found from the following nonlinear functional equation: $$\label{feig}
g_z(x)=-\a_zg_z\left(g_z\left(\f{x}{\a_z}\right)\right)$$ with the unknown function $g_z$ and the initial condition $g(0)=1$. Then $$\label{univ1}
\a_z=-\f{1}{g_z(1)}.$$ Functional equation (\[feig\]) does not have in general a unique solution. In [@briggs] numerical results are given which suggest, according to [@briggs], that in certain restricted classes of analytic functions the solution to (\[feig\]) is unique.
In [@briggs] the constants $\alpha_z$ are computed with high accuracy on a class of even concave functions analytic on $[-1,1]$ for integer $z$, $2\le z\le 12$. Approximate solutions of (\[feig\]) are constructed as polynomial approximations: $$g_z(x)=1+\sum_{i=1}^nq_i|x|^{zi},$$ where $q_i$ are the solutions to the following nonlinear system: $$\label{fj}
F_j:=\left(1+\sum_{i=1}^nq_i\right)\left(1+\sum_{i=1}^nq_i|x_j|^{zi}\right)-
1-\sum_{i=1}^nq_i\left|1+\sum_{i=1}^nq_i\biggl|(1+\sum_{i=1}^nq_i)x_j\biggl|^{zi}
\right|^{zi}=0,$$ where $x_j$ are obtained from the partition of the segment $[0,1]$.
In [@briggs] classical Newton’s method is successfully applied to a numerical solution of system (\[fj\]) and computation of $\a_z$ for $z=2,\dots,12$. For $z>12$ constants $\a_z$ are not found in [@briggs].
The goal of our experiment is to calculate $\a_z$ for $z=2,\dots,12$ (and to compare with [@briggs]) and also for $z>12$ using schemes (\[caupr\]) – (\[rchn\]) and (\[caupr\]) – (\[P2\]). The function $g_z(x)$ is even, therefore it is sufficient to find it on $[0,1]$. Since uniform partition of $[0,1]$ works for small $z$ only ($z=2,3$), the nonlinear partition $x_j:=\left(\f{j}{n}\right)^{\f{1}{z}}$ is chosen.
The Jacobi matrix $$F':=\left[\f{\partial F_j}{\partial q_l}\right]_{j,l=\overline{1,n}}=
|x_j|^{zl}\left(1+\sum_{i=1}^nq_i\right)+\left(1+\sum_{i=1}^nq_i|x_j|^{zi}\right)
-\left|1+\sum_{i=1}^nq_i\biggl|(1+\sum_{i=1}^nq_i)x_j\biggl|^{zi}\right|^{zl}$$ $$-\sum_{i=1}^nq_izi\left|1+\sum_{i=1}^nq_i\biggl|(1+\sum_{i=1}^nq_i)x_j\biggl|^{zi}
\right|^{zi-1}\left|\biggl|\biggl(1+\sum_{i=1}^nq_i\biggl)x_j\biggl|^{zl}+
\sum_{i=1}^nq_izi\biggl|\biggl(1+\sum_{i=1}^nq_i\biggl)x_j\biggl|^{zi-1}\right|$$ is strictly ill-posed for $z\ge 2$ and $n\ge 2$. The condition number for any fixed $n$ increases about ten times as $z$ is replaced by $z+1$. Therefore solving (\[fj\]) for large $z$ and $n\ge 2$ is a very unstable problem to which the standard numerical methods are not applicable. However our methods, based on Theorems \[t1\], \[t3\] and \[t42\], do work and yield the Feigenbaum constants $\a_z$ for $z=2,...,26$.
For a more accurate approximation of $g_z$ one has to take $n$ large enough, but then the problem of the choice of an initial approximation occurs: for $z=2$, $n=2$ or $n=3$ system (\[fj\]) has many solutions. By this reason the scheme described in [@b] is used. First, system (\[fj\]) is solved for $n=1$, then the solution of (\[fj\]) with $n=1$ is taken as the initial guess for the case $n=2$, etc. When $z=2$, $n=1$, $x=1$ system (\[fj\]) is reduced to one algebraic equation with respect to $q_1$: $$q_1(q_1+1)(q_1^5+3q_1^4+3q_1^3+3q_1^2+2q_1-1)=0$$ and the two obvious solutions are $q_1^{(1)}=0$, $q_1^{(2)}=-1$. Since the function $g_z(x)$ is even and concave, the initial condition $g_z(0)=1$ implies $g_z(1)<1$, that is $1+q_1<1$, $q_1<0$. Therefore one has to find the negative roots of the equation: $$q_1^5+3q_1^4+3q_1^3+3q_1^2+2q_1-1=0.$$ Such roots are: $q_1^{(3)}=-1.8597174...,$ $q_1^{(4)}=-1.4021968...$. Thus for the system of two equations ($z=2$, $n=2$) the initial data are:
1. $q_1=-1, \q q_2=0$;
2. $q_1=-1.8597174..., \q q_2=0$;
3. $q_1=-1.4021968..., \q q_2=0$.
In the first two cases the solutions to (\[fj\]) ($z=2$, $n=2$) are not concave on $[-1,1]$. In the third case the graph of the polynomial is concave and $$g_2(x)\approx 1-1.5416948x^2+0.1439197x^4.$$ For the system of three equations ($z=2$, $n=3$) the initial data are: $$q_1=-1.5416948, \q q_2=0.1439197, \q q_3=0.$$ Then we continue this process. The maximum dimension we take is $n=12$. If $n=13$, the discrepancy is not less than for $n=12$, and after $n=14$ it grows.
For $z=3$ we begin the computation with one equation ($n=1$) also. As the initial approximation $q_1=-1.4021968$ is taken, that is the solution to (\[fj\]) with $z=2$, $n=1$. The dimension increases step by step till the discrepancy improves. For $z=3$, $n=1$ the solution to (\[fj\]) with $z=3$, $n=1$ is used, etc. In our experiment $\a_z$ for $z=2,\dots,26$ are found. For $z=2,\dots,12$ they coincide with $\a_z$ proposed in [@briggs]. Below the values of $\a_{13} - \a_{26}$ are presented (the values of $\a_2 - \a_{12}$ can be found in [@briggs]). As the exact digits the ones that were the same as the result of both regularized procedures were taken. $$\a_{13}=-1.22902, \q \a_{14}=-1.21391, \q \a_{15}=-1.20072, \a_{16}=-1.18910,$$ $$\a_{17}=-1.17879, \q \a_{18}=-1.16957, \q \a_{19}=-1.1612, \a_{20}=-1.1537,$$ $$\a_{21}=-1.1469, \q \a_{22}=-1.140, \q \a_{23}=-1.134, \a_{24}=-1.129,$$ $$\a_{25}=-1.124, \q \a_{26}=-1.12.$$
Our numerical results indeed demonstrate the efficiency of procedures (\[caupr\]) – (\[rchn\]) and (\[caupr\]) – (\[P2\]) and give Feigenbaum’s constants for much larger range than in [@briggs], which is of some practical interest. Contrary to the original conjecture [@fig] our numerical results confirm the conclusion of [@briggs], which says that the Feigenbaum constants in fact depend on the parameter $z$.
Appendix
========
Here we prove a lemma about nonlinear differential inequalities. As we have shown in the previous sections, such inequalities are very useful in applications.
Let $u\in C^1[0,+\infty)$, $u(t)\ge 0$ for $t>0$, and $$\label{ap1}
\dot{u}\le -a(t)f(u(t))+b(t),\q \t{ for }\q t>0,\q u(0)=u_0.$$ Assume:
1\) $a(t),b(t)\in C[0,+\infty)$, $a(t)>0$, $b(t)\ge 0$ for $t>0$,
2\) $\int^{+\infty}a(t)dt=+\infty$, $\f{b(t)}{a(t)}\to 0$ as $t\to +\infty$,
3\) $f\in C[0,+\infty)$, $f(0)=0$, $f(u)>0$ for $u>0$,
4\) there exists $c>0$ such that $f(u)\ge c$ for $u\ge 1$.
Under these assumptions (\[ap1\]) implies $$\label{descon}
u(t)\to 0\q\t{as}\q t\to +\infty.$$ \[apineq\]
This Lemma is essentially Lemma 1 from [@al1]. We have added condition 4) and changed the proof slightly. Condition 4) is omitted in [@al1]. Without condition 4 the conclusion of Lemma 1 in [@al1] is false as we show by a counterexample at the end of this Appendix. Condition 4) is equivalent to the condition $f(w)\to 0$ implies $w\to 0$.
Assumptions about smoothness of $a(t)$ and $b(t)$ in [@al1] are not formulated. In Lemma \[apineq\] we assume that these functions are continuous and $a(t)>0$. The continuity assumption can be relaxed, but in applications it is not restrictive, since we deal with the inequality.
[**Proof**]{} By assumptions 1), 2), the new variable $s=s(t):=\il_0^t a(\tau)d\tau$, maps $t\in [0,+\infty)$ onto $s\in [0,+\infty)$. Write (\[ap1\]) as $$\label{ap2}
\f{dw}{ds}\le -f(w(s))+\b(s),\q \t{ for }\q s>0,\q w(0)=u_0,$$ where $w(s):=u(t(s))$ and $\b(s)=\f{b(t(s))}{a(t(s))}\to 0$ as $s\to +\infty$.
The lemma is proved if one proves $$\label{ap3}
w(s)\to 0\q \t{ as }\q s\to +\infty.$$ Let $\kappa(s)\in C[0,+\infty)$ be an arbitrary function such that $\kappa(s)\ge 0$, $\kappa(s)\to 0$ as $s\to +\infty$ and $\int^{+\infty}\kappa(s)ds=+\infty$. For example one can take $\kappa(s)=\f{1}{s+1}$. Define subsets of ${\bf R}_+:=\{s:s\ge 0\}$ as follows: $$\label{ap4}
E:=\{s: s>0, f(w(s))-\b(s)\le\kappa(s)\},\q F:={\bf R}_+\setminus E.$$ [**Claim:**]{} $$\label{ap5}
\sup E=+\infty.$$ We prove (\[ap5\]) later.
Assuming (\[ap5\]), consider $s_1\in E$, $(s_1,s_2)\subset F$. Then $$\label{ap6}
-f(w(s))+\b(s)<-\kappa(s)\t{ for } s_1<s<s_2.$$ From (\[ap2\]) and (\[ap6\]) one gets $$\label{ap61}
\f{dw}{ds}< -\kappa(s)\t{ for } s_1<s<s_2.$$ Therefore for $s_1<s<s_2$ one has $$\label{ap7}
w(s)\le w(s_1)-\int_{s_1}^s\kappa(\tau)d\tau\le w(s_1)\t{ for } s_1<s<s_2,
\quad s\in F.$$ Since $s_1\in E$ one has $$\label{ap8}
f(w(s_1))\le \kappa(s_1)+\b(s_1)\to 0\t{ as } s_1\to +\infty, \quad s_1\in E.$$ Here we have used the assumptions $\kappa(s)\to 0$ and $\beta(s)\to 0$ as $ s\to +\infty$.
From (\[ap8\]) and assumption 4) it follows, that $$\label{ap9}
w(s_1)\to 0\q\t{ as }\q s_1\to +\infty,\q s_1\in E,$$ and from (\[ap7\]) and (\[ap9\]) it follows that $$\label{ap10}
w(s)\to 0\q\t{ as }\q s\to +\infty,\q s\in F.$$ Thus, to prove Lemma it is sufficient to prove (\[ap5\]).
Suppose (\[ap5\]) is false, that is $\sup E=s_3<+\infty$. Then $$\label{ap11}
f(w(s))-\b(s)>\kappa(s)\q\t{for}\q s>s_3.$$ From (\[ap2\]) and (\[ap11\]) one gets $$\label{ap12}
\f{dw}{ds}\le -\kappa(s)\q\t{for}\q s>s_3.$$ Thus $$\label{ap13}
w(s)\le w(s_3)-\il_{s_3}^s\kappa(\tau)d\tau\to -\infty\q\t{as}\q s\to
+\infty,$$ where we have used the assumption $\int^{+\infty}\kappa(s)ds=+\infty$. This contradicts the assumption $u\ge 0$ and proves Lemma \[apineq\]. $\Box$
The following example (which is a counterexample to Lemma 1 in [@al1]) shows that condition 4) of Lemma \[apineq\] is essential.
Take $$\label{ap14}
f(u)=\left\{\begin{array}{lcc}
u,\q\t{for}\q 0\le u\le 1,\\
\\
e^{1-u},\q\t{for}\q 1\le u<+\infty,
\end{array}\right.\q a(t)\equiv 1,\q b(t)=\f{3}{t+c},$$ where $c>e^{-1}$ is an arbitrary constant.
One can check immediately that $$\label{ap15}
u(t)=1+log(t+c)$$ satisfies inequality (\[ap1\]). The choice $c>e^{-1}$ guarantees that $u(t)>0$ for all $t\ge 0$. Clearly $u(t)\to +\infty$ as $t\to +\infty$, so that conclusion (\[descon\]) of Lemma \[apineq\] is false if condition 4) is omitted.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank Professor Ya. Alber for useful remarks and Professor V. Vasin for a discussion of the numerical example.
[99]{} Airapetyan, R.G. [*Continuous Newton method and its modification*]{}, Applicable Analysis, (to appear). Airapetyan, R.G., Ramm A.G. and Smirnova, A.B. \[1999\] [*Continuous analog of Gauss-Newton method*]{}, Math. Models and Meth. in Appl. Sci., [**9**]{}, N3. Alber, Ya.I. \[1975\] [*On a solution of operator equations of the first kind with accretive operators in Banach spaces,*]{} Diffferen. Uravneniya, [**11**]{}, N12, 2242–2248. Alber, Ya.I. \[1993\] [*The regularization method for variational inequalities with nonsmooth unbounded operators in Banach space,*]{} Appl. Math. Lett., [**6**]{}, N4, 63–68. Alber, Ya.I. \[1994\] [*A new approach to the investigation of evolution differential equations in Banach spaces,*]{} Nonlin. Anal., Theory, Methods & Appl., [**23**]{}, N9, 1115–1134. Argyros, I.K. \[1998\] [*Polynomial operator equations in abstract spaces and applications,*]{} CRC Press, Boca Raton. Babenko, K.I. \[1986\] [*Fundamentals of the numerical analysis,*]{} Nauka, Moscow. Beckenbach, E. and Bellman R. \[1961\] [*Inequalities,*]{} Springer-Verlag, Berlin. Blaschke, B., Neubauer, A. and Scherzer O. \[1997\] [*On convergence rates for the iteratively regularized Gauss-Newton method,*]{} IMA J. Num. Anal., [**17**]{}, 421–436. Briggs, K \[1991\] [*A precise calculation of the Feigenbaum constants,*]{} Mathematics of computations, [**57**]{}, N195, 435–439. Decker, D.W., Keller, H.B. and Kelley, C.T. \[1983\] [*Convergence rates for Newton’s method at singular points,*]{} SIAM J. Numer. Anal., [**20**]{}, N2, 296–314. Deimling, K. \[1985\] [*Nonlinear functional analysis,*]{} Springer-Verlag, New York. Deuflhard, P., Engl, H.W. and Scherzer, O. \[1998\] [*A convergence analysis of iterative methods for the solution of nonlinear ill-posed problems under affinely invariant conditions,*]{} Inv. Probl., [**14**]{}, 1081–1106. Engl, H.W., Hanke, M. and Neubauer, A. \[1996\] [*Regularization of inverse problems,*]{} Kluwer Acad. Publ. Group, Dordrecht. Feigenbaum, M.J. \[1978\] [*Quantitative of universality for a class of nonlinear transformations,*]{} J. Stat. Phys. [**19**]{}, N1, 25–52. New York. M.K. Gavurin, \[1958\] [*Nonlinear functional equations and continuous analogies of iterative methods,*]{} Izv.Vuzov.Ser.Matematika. [**5**]{} (1958), 18–31. Kamke, E. \[1974\] [*Differentialgleichungen. Lösungmethoden und Lösungen,*]{} Chelsea, New York. Ryazantseva, I.P. \[1994\] [*On some continuous regularization methods for monotone equations,*]{} Comput. Math. Math. Phys., [**34**]{}, N1, 1–7. Szarski, J. \[1967\] [*Differential inequalities,*]{} PWN, Warszawa. Vasin, V.V. and Ageev, A.L., \[1995\] [*Ill-posed problems with a priori information*]{}, VNU, Utrecht.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We test the statistical isotropy of the universe by analyzing the distribution of WISE extragalactic sources that were also observed by 2MASS. We pay particular attention to color cuts and foreground marginalization in order to cull a uniform sample of extragalactic objects and avoid stars. We detect a dipole gradient in the number-counts with an amplitude of $\sim$0.05, somewhat larger than expectations based on local structures corresponding to the depth and (independently measured) bias of our WISE-2MASS sources. The direction of the dipole, $(l,b)\simeq (310\,^{\circ}, -15\,^{\circ})$, is in reasonably good agreement with that found previously in the (shallower) 2MASS Extended Source Catalog alone. Interestingly, the dipole direction is not far from the direction of the dipolar modulation in the CMB found by Planck, and also fairly closely matches large-scale-structure bulk-flow directions found by various groups using galaxies and type Ia supernovae. It is difficult, however, to draw specific conclusions from the near-agreement of these directions.'
author:
- |
Mijin Yoon$^{1}$[^1], Dragan Huterer$^{1}$, Cameron Gibelyou$^{1}$, András Kovács$^{2}$, and István Szapudi$^{3}$\
${}^{1}$Department of Physics, University of Michigan, 450 Church St, Ann Arbor, MI 48109-1040\
${}^{2}$Institute of Physics, Eötvös Loránd University, 1117 Pázmány Péter sétány 1/A, Budapest, Hungary\
MTA-ELTE EIRSA “Lendület” Astrophysics Research Group, 1117 Pázmány Péter sétány 1/A Budapest, Hungary\
${}^{3}$Institute for Astronomy, University of Hawaii 2680 Woodlawn Drive, Honolulu, HI, 96822
bibliography:
- 'mnras\_wise\_dipole.bib'
title: 'Dipolar modulation in number counts of WISE-2MASS sources'
---
Introduction {#sec:intro}
============
Modern surveys of large-scale structure allow tests of some of the most fundamental properties of the universe – in particular, its statistical isotropy. One of the most fundamental such tests is measuring the dipole in the distribution of extragalactic sources. One expects a nonzero amplitude consistent with the fluctuations in structure due to the finite depth of the survey; this “local-structure dipole” in the nomenclature of @Gibelyou2012 is of order 0.1 for shallow surveys extending to $z_{\rm
max}\sim 0.1$, but significantly smaller ($A\lesssim 0.01$) for deeper surveys. The motion of our Galaxy through the cosmic microwave background (CMB) rest frame also contributes to the dipole, but only at the level of $v/c\simeq 0.001$; while this kinematic dipole was detected in the CMB a long time ago, and more recently even solely via its effects on the higher multipoles in the CMB fluctuations [@Aghanim:2013suk], it has not yet been seen in large-scale-structure (LSS) surveys.
Measurements of the dipole in LSS therefore represent consistency tests of the fundamental cosmological model, and have in the past been applied to the distribution of sources in NVSS [@blake2002detection; @Hirata2009; @Rubart:2013tx; @Fernandez-Cobos]. Detection of an anomalously large (or small) dipole in LSS could indicate new physics: for example, motion between the CMB and LSS rest frames, or the presence of superhorizon fluctuations [@zibin2008gauging; @itoh2010dipole]. Moreover, in recent years, measurements of the bulk motion of nearby structures have been conducted, out to several hundred megaparsecs, using CMB-LSS correlations [@kashlinsky2008measurement], or out to somewhat smaller distances, using peculiar velocities [@Watkins_2009; @Feld_Watk_Hudson_2010].
In this study, for the first time we test statistical isotropy using WISE (Wide-field Infrared Survey Explorer) [@2010AJ....140.1868W]. WISE is, at least at first glance, perfectly suited to tests of statistical isotropy since it is deep and covers nearly the full sky. Moreover, its selection functions have been increasingly well understood over the past few years based on its observations in four bands sensitive to 3.4, 4.6, 12, and 22 $\mu m $ wavelengths with resolution in the 6“-12” range [@Yan:2012yk; @Menard:2013aaa].
Culling of the WISE dataset {#sec:data}
===========================
Our measurement of the dipole relies on a suitable selection of a representative sample of sources. The most important goal is to exclude Galactic sources – mainly stars. Galactic sources are expected to be concentrated around the Galactic plane, with density falling off to the north and south. While they are therefore expected to look like a $Y_{20}$ [ *quadrupole*]{} in Galactic coordinates, the residual contamination of the dipole may still be significant. Hence, in what follows we pay particular attention to magnitude and color cuts applied to WISE in order to leave a trustworthy set of extragalactic sources.
{width="80.00000%"}
The Nov. 2013 release of WISE data includes 747 million objects in total. Individual objects were not identified in the raw data, so data selection is the key part of the analysis. We therefore apply carefully chosen criteria to define a map as uncontaminated by Galactic objects as possible. As argued in @Kovacs:2013cga, color cuts using only the WISE bands are not sufficient, so we have applied 2MASS[^2] magnitudes ($J_{\rm 2mass}$) to distinguish between stars and galaxies. In other words, every source we use is observed in both WISE and 2MASS, though we refer to our sample as “WISE” because using that survey is crucial to give our sample greater depth. To cull a uniform, extragalactic sample of sources, we adopt the following color cuts:
- $W1 <15.2$,\
- $J_{\rm 2mass} < 16.5$,\
- $W1 - J_{\rm 2mass} < -1.7$.\
Note that the first two criteria simply remove the faintest objects in the respective band. To account for the effects of extinction by dust, we correct the magnitudes for these two cuts using the SFD [@SFD] map[^3]. The third criterion above represents the color cut that serves to separate galaxies from stars. The detailed analysis on the data selection was described in [@Kovacs:2013cga]; the resulting WISE map is shown in Fig. \[fig:map\].
Unlike the previous studies that used WISE for cosmological tests [@Kovacs:2013rs; @Ferraro:2014msa], our map does not show obvious contamination in regions affected by the appearance of the Moon. Therefore, we do not need to make further (and typically severe) cuts that remove these regions. We do use the WMAP dust map [@Bennett:2012zja] to mask out the pixels with remaining contamination; these mostly fall within $\pm 15^{\circ}$ Galactic latitude. In addition, we cut out all pixels with $E(B-V)>0.5$ from the SFD map (most of these have already been excluded by the WMAP dust map). We also checked for any unusual gradients with Galactic latitude, especially around the Galactic plane, due to contamination from stars. These tests were consistent with zero gradient.
In the analysis, there are of order 2 million galaxies. We used the GAMA DR2 [@2008AAONw.114....3D] catalog to find sources in the WISE dataset that are within 3” of GAMA sources. We can thus determine the redshift distribution of our objects. In the 144 sq. deg. overlapping region on the sky, the matching rate is 96.9%. The redshift distribution of matched objects, $N(z)$, is shown in Figure \[fig:N(z)\]; the mean is $\bar{z}=0.139$. We use a smooth fit to the full distribution to obtain our theoretical expectation for the local-structure dipole below.
Methodology
===========
Dipole estimator
----------------
A robust and easy-to-implement dipole estimator was first suggested by @Hirata2009, who measured hemispherical anomalies of quasars, and later adopted by @Gibelyou2012 to measure the dipole in a variety of LSS surveys. The number of sources in direction $\hat{\mathbf{n}}$ can be written as $$N(\hat{\mathbf{n}})=[1+A\,\hat{\mathbf{d}}\cdot\hat{\mathbf{n}} ]{\bar N }+\epsilon (\hat{\mathbf{n}})$$ where $A$ and $ \hat{\mathbf{d}}$ are the amplitude and direction of the dipole, and $\epsilon$ is noise.
The modulation in number counts can be written as the sum of contributions from a dipole, fluctuations due to systematics, and a mean offset [@Hirata2009]. $$\delta N/\bar{N} = A\, {\bf \hat d} \cdot {\bf \hat n} + \sum_i k_i t_i({\bf \hat n}) + C.
\label{eq:N_with_templates}$$ Here $t_i({\bf \hat n})$ represent the systematics maps, while the coefficients $k_i$ give the amplitudes of the contributions of these systematics to the observed density field. The presence of the monopole term, $C$, allows us to account for covariance between the monopole and other estimated parameters, especially covariance between the monopole and any systematic templates. The best linear unbiased estimator of the combination ([**d**]{}, $k_i$, $C$), with corresponding errors, is obtained as follows. First, we rewrite the above equation as $\delta N/N = {\bf x} \cdot {\bf
T(\hat n)}$ where ${\bf x} = (d_x, d_y, d_z, k_1, ..., k_N, C)$, ${\bf
T(\hat n)} = (n_x, n_y, n_z, t_1(\hat n), ..., t_N(\hat n), 1)$, and $n_x^2
+ n_y^2 + n_z^2 = 1$. The best linear unbiased estimator of [**x**]{} is $${\bf \hat x} = F^{-1} g$$ where the components of the vector $g$ are $g_i = \int T_i(\hat n) \delta
N^{\Omega}(\hat n) d^2 \hat n$ and the Fisher matrix $F$ is given by $F_{ij} =
\bar N^{\Omega} \int T_i(\hat n) T_j(\hat n) d^2 \hat n$, where $N^{\Omega}
\equiv dN/d\Omega$ is the number of galaxies per steradian ($\Omega$ is a solid angle). The integrals from which the vector $g$ and the Fisher matrix $F$ are calculated are discretized in our survey. We adopt a HEALPix [@healpix] pixelization with [ NSIDE]{}=128, so that each pixel corresponds to about half a degree on a side and contains roughly 14 sources.
The formalism above returns the best-fit dipole components (first three elements of the vector ${\bf x}$), together with their covariance (inverse of the corresponding Fisher matrix). We are however most interested in the likelihood of the amplitude of the dipole, $A=(d_x^2+d_y^2+d_z^2)^{1/2}$. We can construct a marginalized likelihood function for the amplitude $A$ [@Hirata2009]: $$\mathcal{L}(A) \propto \int \exp \left \lbrack -\frac{1}{2}(A {{\bf \hat n}}-
{{\bf d_{best}}}) {\rm Cov}^{-1}(A {{\bf \hat n}}- {{\bf d_{best}}}) \right \rbrack d^2 {{\bf \hat n}}\label{eq:like_A}$$ where $d^2 {{\bf \hat n}}$ indicates integration over all possible directions on the sphere. Thus we readily obtain a full likelihood for the amplitude. In our results, we quote the $68\%$ region around the best-fit amplitude.
![Number counts of WISE sources as a function of redshift. We obtain redshift information by matching WISE sources to those from the GAMA DR2 catalog. As explained in the text, matching works very well.[]{data-label="fig:N(z)"}](N_of_Z){width="40.00000%"}
Foreground Templates and Estimator Validation
---------------------------------------------
Despite our carefully chosen magnitude and color cuts, it is likely that there is some star contamination to our extragalactic source map. Moreover, on a cut sky, the dipole is not completely decoupled from the monopole, quadrupole, and other multipoles, and hence we need to marginalize over some of them in order to get correct results. We therefore include several templates – maps $t_i({\bf \hat n})$ in the parlance of Eq. (\[eq:N\_with\_templates\]) – with amplitudes $k_i$ over which we marginalize:
- To deal with the remaining star contamination, we add a star map as a template. The star map was generated based on the Tycho 2 catalog , as suggested in [@Kovacs:2013rs]. The inclusion of this template affects the measured dipole negligibly, reinforcing our confidence that star contamination does not affect the result.
- To account for the other multipoles, we add the monopole (corresponding to the constant $C$ in Eq. (\[eq:N\_with\_templates\]) with no spatial dependence), as well as the quadrupole and octopole that include 5 and 7 extra parameters. We therefore marginalize over these 13 parameters in addition to the amplitude of the star map. We experimented with marginalization over a few more ($\ell\geq 4$) multipoles, but for small Galactic cuts ($b_{\rm cut}\lesssim 15^\circ$), the shift in the dipole direction and magnitude were small.
We validated our estimator by running simulations with an input dipole of a given amplitude assuming various sky cuts and marginalizing over templates. We verified that the input dipole is recovered within the error bars.
Theoretical expectation
-----------------------
We calculate the theoretical expectation for the local-structure dipole using standard methods (see e.g. Sec. 2.2 of @Gibelyou2012). We calculate the angular power spectrum of large-scale structure for the given source distribution $N(z)$, and evaluate it at the dipole ($C_\ell$ at $\ell=1$); this calculation does not assume the Limber approximation since the latter is inaccurate at these very large scales. The amplitude is then given as $A_{\rm
theory}=(9C_1/(4\pi))^{1/2}$ [@Gibelyou2012], while the theory error is given by cosmic variance for $\ell=1$: $\delta A_{\rm theory}/A_{\rm
theory}=(1/2)\sqrt{2/((2\ell+1){f_{\rm sky}})}=(6{f_{\rm sky}})^{-1/2}$. Evaluating the theoretically expected dipole for the source distribution shown in Fig. \[fig:N(z)\], we get $$A_{\rm theory} = (0.0233 \pm 0.0094{f_{\rm sky}}^{-1/2})\times
\left (\frac{{\rm bias}}{1.41}\right )
\label{eq:A_th}$$ Here we make explicit the dependence of the cosmic variance error on the fraction of the sky covered ${f_{\rm sky}}$, and also on the bias of WISE sources. To obtain the latter, we followed @Kovacs:2013cga, and estimated the bias of the galaxy catalog using [SpICE]{} [@spice] and the Python CosmoPy[^4] package. We note that the estimation of the bias is particularly sensitive to $\sigma_{8}$ because they both act to renormalize the angular power spectrum, and in linear theory $C^{gg}_{\ell} \propto (b\sigma_{8})^{2}$. We fix $\sigma_{8}=0.8$ in our measurements, finding $b=1.41\pm0.07$. This value is comparable to earlier findings [@RassatEtal2003] that measured a value of $b=1.40\pm0.03$ for a 2MASS selected galaxy sample.
![Theoretical prediction for the dipole amplitude (horizontal blue line), together with the measured values in WISE (green points). The two sets of error bars on the measurements correspond to 68% and 95% confidence; they have been calculated from the full likelihood in Eq. (\[eq:like\_A\]) and are rather symmetric around the maximum-likelihood value. The two large horizontal bands around the theory prediction correspond to 1- and 2-sigma cosmic variance error. []{data-label="fig:A_th_and_directions"}](theory_exp_value_plot){width="48.00000%"}
Results
=======
Our measurements of the dipole’s amplitude and direction, as a function of the (isolatitude) Galactic cut, are presented in Table \[tab:A\]. The best-fit direction of the dipole is also shown in Fig. \[fig:map\] for the $10^\circ$ and $20^\circ$ Galactic cut, the two cases roughly illustrating the dependence of the direction on the Galactic cut.
We first note a reasonably good consistency between the recovered directions, despite the fact that the number of sources decreases by a factor of $\sim$1.4 as we increase the Galactic cut in the range shown. We also note that the overall amplitude is roughly 1.5 - 2.7 times larger than the theoretically expected one, and is roughly 1-2$\sigma$ high, where $\sigma$ corresponds to cosmic variance since the measurement error is much smaller (see Table \[tab:A\]). Finally, we note that while the dipole amplitude does vary with $b_{\rm cut}$ more than its typical measurement errors, it is overall consistent at $A_{\rm WISE}\simeq 0.04$-$0.05$, which is rather robustly stable given the large decrease of the number of sources with increasing Galactic cut.
It is interesting to note that 2MASS Extended Source Catalog data, as analyzed in @Gibelyou2012 (redshift $0<z<0.2$, $N =3.8\times10^5$), give $A_{\rm 2MASS} = 0.104 \pm
0.004$, $(l,b) = (268.4^{\circ}, 0.0^{\circ})$ – amplitude higher than ours due to the greater contribution of the local-structure dipole for the shallower survey, direction not far. Relative to this previous work, we have therefore made progress by pushing down a factor of 2.5 in the dipole amplitude. This is a welcome development toward being able to probe the kinematic dipole due to our motion relative to the overall LSS rest frame, which will require reaching the level $A \sim 10^{-3}$, and therefore a deeper survey (or a deeper sample of WISE sources).
[|| c | c | c | c | c ||]{}
------------------------------------------------------------------------
$b_{\rm cut}$& ${f_{\rm sky}}$ & $A_{\rm WISE}$ & $A_{\rm theory}$& $\hat{\mathbf{d}}(l\,^{\circ},b\,^{\circ})$
------------------------------------------------------------------------
$10^{\circ}$& 0.65 & $0.035\pm 0.002$ & $0.023 \pm
0.012$ & $(326\pm3, \,-17\pm2)$\
------------------------------------------------------------------------
$15^{\circ}$ & 0.62 & $0.042\pm 0.002$ & $0.023 \pm
0.012$ & $(316\pm3, \,-15\pm2)$\
------------------------------------------------------------------------
$20^{\circ}$& 0.57 & $0.052\pm 0.002$ & $0.023 \pm
0.012$ & $(308\pm4, \,-14 \pm2)$\
------------------------------------------------------------------------
$25^{\circ}$& 0.51 & $0.062\pm 0.003$ & $0.023 \pm
0.013$ & $(315\pm6, \,-12 \pm2)$\
------------------------------------------------------------------------
$30^{\circ}$& 0.45 & $0.051\pm 0.004$ & $0.023 \pm
0.014$ & $(335\pm6, \,-18 \pm3)$\
Conclusions {#sec:concl}
===========
We measured the clustering dipole in the WISE survey, using a carefully culled sample that contains 2 million extragalactic sources with a known redshift distribution. The amplitude of the measured dipole is $A\simeq 0.05\pm
0.01$, where we quote the central value corresponding to the $20^{\circ}$ cut case and error that shows the dispersion of central values for $15^{\circ}\leq b_{\rm cut}\leq 25^{\circ}$. The amplitude is therefore roughly twice as large as the theoretical expectation; see Eq. (\[eq:A\_th\]). The direction of the dipole is $\simeq
(310^\circ\pm5, \,-15^\circ\pm2)$.
What could explain the excess dipole measured relative to theoretical expectation? The systematics, while an obvious first suspect, are not necessarily at fault given the rather extensive care we took to account for them: we carefully culled the dataset by imposing cuts based on WISE and 2MASS magnitudes; we included cuts based on Galactic latitude and on the WMAP dust map, and we further marginalized over a carefully derived star-map template as well as templates corresponding to the quadrupole and octopole.
Another possibility is that the excess signal is cosmological. For example, a large void might generate the excess observed here [@Rubart:2014lia]. Such a void was incidentally just detected in the analysis of the WISE data itself [@Szapudi:2014zha; @FinelliEtal2014]. At this time it is too early to tell whether the WISE void is contributing significantly to the excess dipole that we measured, though a rough comparison with numbers in @Rubart:2014lia appears to indicate that it is not.
It is also interesting to note that Planck found a best-fit modulation with both amplitude and direction roughly (within $\sim$3$\sigma$ of their errors) in agreement with ours [@Ade:2013nlj]: $A_{\rm Planck} = 0.078 \pm 0.021$, $(l,b) = (227^\circ, -15^\circ) \pm 19^\circ $. It is not clear at this time what, if any, significance to assign to the comparable-looking modulations in WISE and Planck since their sources are at vastly different redshifts ($z\sim
0.15$ and $\sim 1000$), and the agreement in amplitude and direction is only approximate. Finally, the direction we find is [*also*]{} close to the peculiar-velocity bulk-flow directions found using type Ia supernovae [@Dai_Kinney_Stojkovic; @Kalus_SN; @Rathaus_SN], galaxies [@Feld_Watk_Hudson_2010; @Turnbull; @Ma_Gordon_Feldman; @Ma_Pan], and the kinetic Sunyaev-Zeldovich effect [@Lavaux_kSZ]. While the agreement between the directions is suggestive, it is not immediately clear how our WISE dipole is related to these. For example, interpreting the excess dipole amplitude $\delta A\sim 0.03$ as a bulk motion is clearly out of the question, since it would correspond to a huge velocity of $v\simeq 0.015c = 4500$ km/s, an order of magnitude larger than what typical bulk-motion measurements indicate.
With recent measurements of the cross-correlation of its sources with the CMB and the detection of a large underdense void, WISE is finally making major contributions to cosmology. Its nearly all-sky coverage is a huge asset and gives the survey a big advantage on that front over most other LSS surveys. In this paper we have taken another step in testing fundamental cosmology with WISE by measuring the clustering dipole in the distribution of its extragalactic sources. We look forward to further investigations of this result, especially in conjunction with other related findings in the CMB and LSS.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Maciej Bilicki, Brice Ménard and Dominik Schwarz for comments on the manuscript, and Matt Ashby and Jeffrey Newman for discussion of color cuts. The work of MY and DH has been supported by the DOE and the NSF. AK acknowledges OTKA grant no. 101666, and support from the Campus Hungary fellowship program. IS acknowledges NASA grants NNX12AF83G and NNX10AD53G. This work made use of the HEALPix package [@healpix].
[^1]: [`m`[email protected]]{}
[^2]: Two Micron All Sky Survey [@skrutskie2006two].
[^3]: http://lambda.gsfc.nasa.gov/product/foreground/fg\_sfd\_get.cfm
[^4]: `http://www.ifa.hawaii.edu/cosmopy/`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A dynamic memory model is proposed in which an agent “learns” a new agent by means of recommendation. The agents can also “remember” and “forget”. The memory size is decreased while the population size is kept constant. “Fame” emerged as a few agents become very well known in expense of the majority being completely forgotten. The minimum and the maximum of fame change linearly with the relative memory size. The network properties of the who-knows-who graph, which represents the state of the system, are investigated.'
author:
- Haluk
bibliography:
- 'bingolFame.bib'
title: Fame Emerges as a Result of Small Memory
---
Motivation
==========
One of the observations of complex systems is that they are made out of many interacting agents. Usually, the number of agents is simply too big for an agent to be informed of all the others. Therefore, agents act based on limited information. Many real-life examples can be given: A consumer can only have access to a limited number of suppliers. A car can only encounter a small number of other cars in a traffic jam. In the brain, a neuron cannot be connected to all the other $10^{11}$ neurons [@Chialvo2004]. No web page can connect to all the other existing web pages. Similarly, no router can be connected to all other routers on the Internet. Even in many simple models, access to only the local information is a common property. In Bak’s Sandpile Model, a sand particle communicates only with the four sand particles in its neighborhood [@Bak1996]. Similarly, in Axelrod’s 2D Culture Model, an agent interacts with its four neighbors only [@Axelrod1997]. In Conway’s Game of Life a cell checks its eight neighbors in order to decide whether to live or die in the next cycle [@Berlekamp1982]. Although information exchange is relatively local and the rules of exchange are quite simple, these systems manage to become complex systems.
An individual cannot know the entire population but a small fraction of it. Consider the ratio of the number of people that one knows to the size of the population of the city or the country that she lives in. One expects that this ratio, which will be an important parameter of the model developed in this study, is a very small number [^1]. Another observation is that the people that we know constantly changes. We “learn” new people from many sources including people, books, newspapers, radio, television, e-mail, www, SMS. On the other hand, we do not “remember” all the people that we learn. We have a limited cognitive capacity. A mechanism enables us to “forget” people. Therefore, a model should deal with concepts such as population, memory, learning, remembering, forgetting and interaction of individuals that change their memory content. This paper mainly considers the human population in the development of the model but the findings are applicable to many systems.
Mobile phones are a good example which satisfy many properties of the model presented in this paper [^2]. They have a limited memory. When they receive a call, they try to store the caller number. They usually do not store their own phone number, that is, they do not “know” themselves. Another example would be routers in computer networks.
The Recommendation Model
========================
A reasonable question would be: What happens if an agent is allowed to interact with all the other agents, but remembers only a small fraction of them? In order to investigate this question a simple model is constructed. The dynamics of the system is investigated as the memory size is decreased.
A Static Memory Model
---------------------
Let $\bm{a_i}$ be an agent. Let $\bm{A} = \{ a_i \: | \: 1 \leq i \leq n \}$ be a *population* of $\bm{n}$ agents. Each agent $a_i$ has a memory $\bm{M_i}\subseteq A$. An agent $a_i$ *knows* agent $a_j$ if $a_j \in M_i$. The *knownness* $\bm{k_i}$ of agent $a_i$ is the number of agents that know $a_i$. Then $k_i = \left| \{ a_j \: | \: a_i \in M_j
\} \right|$. If everybody knows the agent, that is $k_i = n$, then the agent is called *perfectly known*. On the other hand, if nobody knows the agent, that is $k_i = 0$, then the agent is called *completely forgotten*. Knownness depends on $n$. For example, for $n=100$, one can be known by at most 100 people but for $n=1\,000$, knownness can be as high as $1\,000$. In order to compare populations of different sizes, a metric, independent of $n$, is needed. The *fame* $\bm{f_i}$ of agent $a_i$ is defined as the ratio of its knownness to the population size, that is $f_i=k_i/n$. Since $0 \leq k_i \leq n$, it is always the case that $0 \leq f_i \leq
1$. Hence fame is a normalized measure of knownness.
An agent *learns* an agent $a_i$ if it gets $a_i$ in its memory. An agent $a_i$ *remembers* agent $a_r$, if $a_i$ selects $a_r$ among the agents stored in its memory. An agent *forgets* agent $a_f$ if it removes $a_f$ from its memory.
An abstraction which simplifies the model is made. It is assumed that every agent has the same *memory size* $\bm{m}$, that is $ \forall i \, \left| M_i \right| = m$. Then the *total memory capacity* of the population is $n m$. The *memory ratio* $\bm{\rho}$ is defined to be the ratio of memory size to the population size, $ \rho = m / n$. We have $0 \leq \rho \leq 1$, since this work considers $m$ values in the range $0 \leq m \leq n$. Note that $\rho$ corresponds to the ratio of the number of people that one knows to the number of people one possibly know. As discussed in the “motivation” section, $\rho$ is expected to be a small value which corresponds to small memory sizes.
The *state* of an agent is the content of its memory. Similarly the *state* of the system is the memories of the all agents. The state of the system can be represented by an $n \times m$ matrix as in Fig. \[fig:nmMatrix\] where row $i$ corresponds to the memory $M_i$.
(120,120)(0,0) (15,5)[(1,0)[80]{}]{} (15,110)[(1,0)[80]{}]{} (15,5)[(0,1)[105]{}]{} (95,5)[(0,1)[105]{}]{}
(40,125)[memory]{} (5,100)[1]{} (15,78)[(1,0)[80]{}]{} (5,70)[g]{} (15,66)[(1,0)[80]{}]{} (15,43)[(1,0)[80]{}]{} (5,35)[t]{} (15,31)[(1,0)[80]{}]{} (5,6)[n]{}
(17,113)[1]{} (40,113)[r]{} (70,113)[f]{} (88,113)[m]{}
(50,66)[(0,1)[12]{}]{} (40,70)[$a_r$]{} (38,66)[(0,1)[12]{}]{} (80,31)[(0,1)[12]{}]{} (70,35)[$a_f$]{} (68,31)[(0,1)[12]{}]{}
A Dynamic Memory Model
----------------------
The system defined so far is a static one. In order to make it dynamic, interaction between agents is defined by means of recommendation. Agent $a_g$ *recommends* agent $a_r$ to agent $a_t$ as visualized in Fig. \[fig:nmMatrix\]. The agents $\bm{a_g}$, $\bm{a_r}$ and $\bm{a_t}$ are called the *giver*, the *recommended* and the *taker*, respectively. The steps of recommendation process are i) $a_g$ remembers $a_r$; ii) $a_g$ gives $a_r$ to $a_t$; iii) $a_t$ learns $a_r$ if $a_t$ does not already know $a_r$. “Remembering” and “forgetting” are primitive operations. On the other hand, “learning” is not a primitive operation for $m<n$, since there is no empty space in the memory for the recommended agent. So learning comprises of three basic operations: i) remember some agent $a_f$ ii) forget $a_f$ in order to obtain an empty slot iii) put the recommended agent $a_r$ to this slot.
Some remarks about the recommendation operation is needed:
i\) *Selections.* The recommendation operation is carefully defined so that it is open to extensions. There are four selections in every recommendation operation, namely selections of giver-taker ($a_g$, $a_t$) and recommended-forgotten ($a_r$, $a_f$) as illustrated in Fig. \[fig:nmMatrix\]. These correspond to selections of $g$ and $t$ from the set of $\{1, 2,
\cdots, n \}$ and memory positions $r$ and $f$ from $\{1, 2, \cdots, m \}$. Different specifications of selections would produce different results. In the simple recommendation model of this paper, all four selections are defined to be randomly chosen from a uniform distribution.
ii\) *Axelrod’s Culture Model.* Another selection criterion could be the case that both the giver and the taker should know the same people in order to interact. Restrict the selection of $a_g$ and $a_t$ in such a way that $| M_g \cap M_t | \geq k$, the case where two agents commonly know at least $k$ agents. For $k=1$, this leads to a model that is similar to Axelrod’s Culture Model where culture vector has only one feature and the corresponding set of traits is $\{1, 2,
\cdots, n \}$.
iii\) *Invariants.* The recommendation operation preserves some global values. Since there are $n m$ memory locations, the summation of the knownnesses of the system is given as $\sum_{i=1}^{n} k_i = n m$. This summation is invariant with respect to recommendation operation, since a recommendation increases the knownness of the recommended by one while decreases that of the forgotten by one.
iv\) *Completely forgotten.* If an agent becomes completely forgotten, then there is no way to be known again.
v\) *Perfectly known.* If an agent becomes perfectly known, it does not mean that it will stay this way unless the system is in one of its “absorbing states”.
vi\) *Recommending items.* Note that although the memory model is presented as agents recommending agents, it can be extended to a model for a general case of agents recommending any type of items, such as books, songs, to other agents. Concepts such as “completely forgotten” would be difficult to explain for human population since a person would know herself even if the rest forgets her. On the other hand, it is not hard to talk about a song, a book, a cultural tradition or even a language that is completely forgotten. In science, there are many examples of concepts discovered, forgotten and re-discovered. A few changes would be needed to extend the model. Let $\bm{B}$ be the set of items. Then, an agent $a_i \in A$ would have items $b_j \in B$ in her memory, that is $M_i \subseteq B$. The memory ratio, that is the ratio of the actual number of memorized to the number of possibly memorized, would be $\rho = m / |B|$. The recommendation operation would be defined as an agent $a_g$ recommends item $b_r$ to agent $a_t$. In the rest of the paper, we assume that agents recommend agents to agents, that is $B = A$.
A Simple Recommendation Model
-----------------------------
Many models can be built on these concepts. One of the simplest models, called the *Simple Recommendation Model*, is obtained by defining all the selection mechanisms as random selections. There are four random selections for each recommendation. The giver, $g$, and the taker, $t$, are selected randomly from the set of $\{1, 2, \cdots, n \}$. The giver $a_g$ selects the recommended agent $a_r$ from its memory by selecting $r$ from $\{1, 2, \cdots, m \}$ randomly. This is the remembering process. If the taker $a_t$ already knows $a_r$, then it does nothing. Otherwise it has to learn it. Learning calls for selecting a memory slot. This selection of $f$ is also done randomly from $\{1, 2, \cdots, m \}$.
This definition implies a number of properties. i) The selection rules do not prefer one agent to another. That is, the process is symmetric with respect to agents. ii) Any agent can get a recommendation from any other agent. Note that this may be an over simplification, since in real life examples an agent can get in touch with only a limited number of agents. On the other hand, increase in communication (e.g. via e-mail) may enable one to communicate with almost anybody.
Termination Conditions
----------------------
When to terminate a simulation is a difficult issue. Defined this way, the memories of the agents are kept changing as long as the recommendations continues. There are some special cases in which continuing recommendations cannot change the state of the system. In these cases, the simulation can be terminated.
**Absorbing States.** A state where every agent has exactly the same memory content, that is $\forall i,j \: M_i = M_j$, is called an *absorbing state*. In an absorbing state case, nobody can recommend anything new since everybody knows exactly the same $m$ agents and the remaining $n-m$ agents are completely forgotten. So there is no point continuing simulation. Therefore an absorbing state is a termination point. Since the system asymptotically converges into one of these absorbing states, absorbing states are theoretical terminations points. Note that $m=0$ and $n = m$ cases are special cases of absorbing states.
Simulations show that there are two regimes in the system [@Bingol2005]. In the beginning, the system tends to forget. This forgetting mechanism works so powerful that many agents become completely forgotten at the very early stages of the simulation. As simulations proceeds, the number of known people becomes much less than the population size. Then system reverses this behavior. This time it tries not to forget. This is an expected behavior since the system converges to an absorbing state asymptotically. In this paper the second regime is investigated. The *average recommendation per agent* $\bm{\nu}$ is defined to be the ratio of total number of recommendations to the population size $n$. Throughout this study $\nu=10^6$ is used.
Random Initial Memory
=====================
The initial configuration of the agent memory is important for the model. The memories of the agents are initially filled with randomly selected indexes from $\{1, 2, \cdots, n \}$. Repetitions are not allowed. The model is simulated for different values of $n$ and $m$. Population sizes of $n
= 10^2, 10^3, 10^4$ and memory ratios of $\rho = 0.50, 0.30, 0.20, 0.10, 0.05, 0.01$ are used.
![(Color online) Even distribution of fame degrades as $\rho$ decreases. The model is simulated for various $m$ values where $n=100$ and $\nu=10^6$. At the end of the simulation, memory dumps of agents provide who-knows-who information. Fame of each agent is calculated and for better visualization, the agents are sorted in decreasing order in fame. []{data-label="fig:fig_effectOfRho"}](fig_effectOfRho.eps){width="\columnwidth" height="\columnwidth"}
Effects of memory size
----------------------
For the same population size $n$, effects of changing memory size $m$ in the interval $0 < m < n$ is investigated. Since the memories of the agents are initially randomly filled, the initial fame of an agent is around the average value of $<\!\!f\!\!>\,= m/n$. As $m$ is decreased, some agents become more known than others, at the expense of others becoming less known. Further decrease of $m$ increases the degeneration further.
For $n=100$, $\rho$ is changed and the change in fame $f$ is observed. Fig. \[fig:fig\_effectOfRho\] is an example of various simulations which produce similar results. In this visualization, the agent number $1$ is the most famous one and the agent number $100$ the least famous. Note that the area under the curve is equal to the total memory capacity $n \times
m$. As $m$ decreases, agents on the right become completely forgotten, as a result the agents on the left become increasingly famous. Around $\rho = 0.5$, the knownness of some agents becomes very low. Completely forgotten agents starts to appear around $\rho = 0.35$. From that point on, decrease in $m$ increases the number of completely forgotten agents. Since the total memory capacity is fixed, a few agents become very well known as a result of this process. Hence, fame emerges as an effect of small memory size.
Eventually $m$ decreases to the extreme case of $m = 1$ where an agent can remember only one agent. In this $\rho=0.01$ case, the dynamics of the system goes to an extreme. All the agents become completely forgotten, except only one. That lucky agent is known by the all other agents. This is the expected absorbing state since the number of known agents is $m=1$. In order to check this finding, simulations with larger values of $n$ is done for $m = 1$. It is observed that as the population size gets larger; reaching an absorbing state becomes harder.
Regular Initial Memory
======================
One may suspect that these findings are due to small fluctuations of the random initial memory. Although random initialization does not prefer any agent systematically, it has some statistical variation. As a result of that, some agents could be slightly more known then others. This initial unbalance may affect the dynamics. In order to check this possibility, a perfectly symmetrical memory initialization scheme is used. In the regular initial memory scheme, each agent $a_i$ is allowed to know its $m$-neighbor, that is $M_i = \{ a_k \: | \: k \equiv i+j \pmod{n}$ for $1 \leq
j \leq m \}$, similar to the case of [@Watts1998]. In this way, it is guaranteed that the knownness of every agent is exactly $m$.
For regular initial memory, $n=100, \cdots, 1\,000$ range with 100 increments is simulated. For each $n$, $\rho=0.10, \cdots, 0.90$ range with 0.05 increments is studied. Additionally, $\rho=0.01, \cdots,0.05$ range with 0.01 increments are simulated in order to see the behavior at very small values of $\rho$. Every $n$ and $\rho$ combination is simulated 20 times. Interestingly, both random and regular initial memory strategies produce similar results.
[ ![(Color online) Averages of 20 simulations with various $n$ values where $\nu=10^6$ and $m$ changes as $\rho$ does. (a) Change of the minimum fame, (b) change of the percentage of the completely forgotten agents as $\rho$ changes.[]{data-label="fig:fig_rho_fmin_u"}](fig_rho_fmin.eps "fig:"){width="\columnwidth" height="\columnwidth"} \[fig:fig\_rho\_fminA\] ]{} [ ![(Color online) Averages of 20 simulations with various $n$ values where $\nu=10^6$ and $m$ changes as $\rho$ does. (a) Change of the minimum fame, (b) change of the percentage of the completely forgotten agents as $\rho$ changes.[]{data-label="fig:fig_rho_fmin_u"}](fig_rho_u.eps "fig:"){width="\columnwidth" height="\columnwidth"} \[fig:fig\_rho\_u\] ]{}
Minimum Fame
------------
The minimum fame $\bm{f_{min}}$ in the population for a range of parametric settings is investigated. As $\rho$ decreases, the minimum value of fame decreases as in Fig. \[fig:fig\_rho\_fmin\_u\](a). This decrease turns out to be linear. As $n$ increases the linear region becomes more visible and $n$ values $800, \cdots, 1\,000$ produce almost the same line. For $n=1\,000$, the line is given as $f_{min}\approx 1.1 \rho -0.12$. The minimum value of fame is $f =
0$ when the first agent becomes completely forgotten. Occurrence of the first $f=0$ case depends on $n$ and it has quite a dynamic range. The first $f = 0$ case occurs when $\rho$ is around 0.35 for $n=100$. As $n$ increases, the first $f=0$ case moves to smaller values of $\rho$. For $n=1\,000$, it happens at $\rho=0.1$.
Percentage of Forgotten Agents
------------------------------
As $\rho$ is decreased beyond the point where at least one agent is forgotten, the minimum fame does not provide any further information. For those values of $\rho$, the number of completely forgotten agents can be investigated.
The percentage $\bm{u}$ of the population that is completely forgotten is used and Fig. \[fig:fig\_rho\_fmin\_u\](b) is obtained. Note that the graphs in Fig. \[fig:fig\_rho\_fmin\_u\](a) and Fig. \[fig:fig\_rho\_fmin\_u\](b) complement each other for any particular value of $n$. As expected, for any values of $\rho$, one graph has non-zero values whenever the other graph has zeros. Here again, as $n$ increases, the curves converge to a line which is given as $u\approx -9.2 \rho + 1$ calculated for $n=1\,000$.
[ ![(Color online) Averages of 20 simulations with various $n$ values where $\nu=10^6$ and $m$ changes as $\rho$ does. (a) Change of the maximum fame, (b) change of the cumulative fame of the top 5 % as $\rho$ changes.[]{data-label="fig:fig_rho_fmax_ftop5"}](fig_rho_fmax.eps "fig:"){width="\columnwidth" height="\columnwidth"} \[fig:fig\_rho\_fmax\] ]{} [ ![(Color online) Averages of 20 simulations with various $n$ values where $\nu=10^6$ and $m$ changes as $\rho$ does. (a) Change of the maximum fame, (b) change of the cumulative fame of the top 5 % as $\rho$ changes.[]{data-label="fig:fig_rho_fmax_ftop5"}](fig_rho_ftop5.eps "fig:"){width="\columnwidth" height="\columnwidth"} \[fig:fig\_rho\_ftop5\] ]{}
Maximum Fame
------------
The maximum value of fame $\bm{f_{max}}$ has an interesting behavior as $\rho$ changes. For $n=100$, $f_{max}$ slowly decreases as $\rho$ decreases from 0.9 to 0.1. It reaches a minimum value around $\rho = 0.1$. Interestingly, further decrease of $\rho$ causes $f_{max}$ to increase. This pattern can be seen by tracing Fig. \[fig:fig\_rho\_fmax\_ftop5\](a) from right to left, where the emergence of fame can be observed as the relative memory size (indicated by $\rho$) is decreased. This unexpected behavior can be explained: When $\rho = 1$, every agent is known by everybody else so the fame is 1. As $\rho$ decreases, the memory size of the agents decreases. Since no one dominates the memories yet, people are almost evenly distributed in the memories. So the reduction of the maximum fame is due to the decrease of the memory size. But as $\rho$ keeps decreasing, after a certain point some people become completely forgotten and some others become the dominating ones. As $\rho$ approaches to the limit of 0, more people become completely forgotten and fewer people dominate the memories. Those that dominate take all the references. So the rapid increase of maximum fame in the vicinity of $\rho=0$ can be explained due to this positive feedback.
Another observation is that the decrease of maximum fame is also linear. As in the case of minimum fame, as $n$ gets larger, the linear pattern becomes more apparent. For $n=1\,000$, it is given as $f_{max} \approx 0.91 \rho + 0.11$.
Cumulative Fame
---------------
Maximum fame is a measure of the dominance of one agent. Dominance of a group of famous agents is investigated by means of the cumulative fame of the top $\bm{p}$ percent of agents ordered according to their fames. The top $p\%$ of the population is selected. The *cumulative fame* $\bm{f_{p\%}}$ is obtained by adding their fames. The maximum possible value for the cumulative fame is $np/100$ when all top $p\%$ are completely known, that is, each has fame of $f=1$. This value is used for normalization.
In this study, $p=5$ is used and Fig. \[fig:fig\_rho\_fmax\_ftop5\](b) is obtained. As $n$ increases, the curves converge to a line which is given as $f_{5 \% } \approx
0.95 \rho + 0.071$ for $n=1\,000$. Fig. \[fig:fig\_rho\_fmax\_ftop5\](a) and Fig. \[fig:fig\_rho\_fmax\_ftop5\](b) are quite similar as expected. On the other hand, $f_{5 \% }$ line decreases slightly faster than $f_{max}$ line as $\rho$ decreases. As $\rho$ decreases to $\rho=0.1$, the curves become saturated. They stay this way for awhile and then as $\rho$ approaches 0, they start to decrease again. This behavior near $\rho=0$ can be explained by the memory size. As $m$ decreases, at some point there is no space to keep 5% of the population. Whenever that happens, the cumulative fame of the top 5% starts to decrease towards 0. This final decrease is much sharper. This behavior can be seen more clearly for small values of $n$ in the figure. For example for $n=100$, top 5%, means 5 agents. If $m$ becomes less than 5, that is $\rho<0.05$, a sharp decrease is expected as in the figure.
Network Issues
==============
Who-knows-who graph is another representation of the state of the system. The directed graph $G(A,E)$ where $A$ is the set of agents and $E = \{ (a_i, a_j) \: | \: a_j \in M_i \}$ is called the *who-knows-who graph*. The graph is a directed graph, since the corresponding relation “to know” is not symmetric.
In this directed graph, out-degree is not interesting since all vertices have the same out-degree of $m$, independent of recommendations. On the other hand, in-degree of a vertex changes by the recommendations and has a dynamic range starting from 0 and it can be as large as $n$. For both random and regular initial memory cases, the initial in-degree distribution is uniform since every agent has the same knownness of $m$. As a result of recommendations, in-degrees of a few agents increase while the majority decreases to 0. So as a result of recommendations, uniform in-degree distribution degrades to the one with two peaks around 0 and $n$. At an absorbing state, there would be exactly two non-zero points in the in-degree distribution, namely 0 and $n$. There is a nucleus of $m$ vertices in which a vertex is connected to other $m-1$ vertices and itself. The remaining $n-m$ vertices are connected to this $m$-vertex nucleus. The $m$ vertices of the nucleus have in-degree of $n$ and $n-m$ vertices have 0.
The undirected graph underlined by the directed who-knows-who graph is topologically investigated. In the random case, the initial network is a random graph. In the regular case, the initial graph is regular. As the recommendation dynamics takes place and fame emerges, both initial graphs transform into one common graph structure. The few famous vertices, which are in the process of forming the nucleus, become hubs. The rest of the vertices are connected to these hubs. Giving this picture, the graph is more towards to star-connected rather than power-law degree distribution. Therefore, the average distance is very low. The clustering coefficient is also low. The recommendation is a transformation that uses local information only. There are some network growth models that also use local information only but they produce power-law degree distribution [@Krapivsky2001; @Rozenfeld2004].
At the absorbing states
=======================
The main focus of this work is the behavior of the system as it approaches but never reaches to an absorbing state. In this section a brief investigation of the system at the absorbing state is done and the rest is left as future work. The system simply takes too long time to reach an absorbing state for $n$ values considered so far. On the other hand, if one reduces $n$, absorbing states become obtainable within reasonable durations. Simulations are done for small values of $n$ and $m$ such as $n \in \{ 20, 30, 40, 50, 60\}$ and $m \in \{ 1, 2 \}$. As a measure of time, the number of simulation cycles required to reach an absorbing state is measured.
It is clear that the system reaches an absorbing state asymptotically. Therefore, near absorbing state, forgetting the next person becomes increasingly harder. Let $t_{i}$ be the time the $i$th person is forgotten. Then, define the *time needed to forget the next person* after the $i$th person as $\Delta t(i) = t_{i+1}-t_{i}$ where $i \in \{ 1, \cdots, n-m\}$. As expected, $\Delta t(i)$ rapidly increases as $i$ approaches to $n-m$.
In some systems, system size makes a big effect to the behavior. When the parameters are scaled with the systems sizes, then some regularities become visible [@Hinrichsen2000]. In this system $n$ turns out to be an important parameter. Fig. \[fig:fig\_absorbingStateA1B1\] provides the behavior of $\Delta t(i)$ as average of 40 simulation runs. In the X-axis the number of persons forgotten is scaled by $n$ which is the percentage of forgotten, that is $u$. In the Y-axis $\Delta t(i)$ is scaled as $\Delta t(i)/m$ for various values of $n$.
Interestingly, there are two family of curves. The upper family belongs to $m=2$. It starts with a slight decrease which corresponds to the initial trend of forgetting. Then the regime changes and forgetting becomes harder and harder as $u$ approaches to 1. The $m=1$ family does not have this pattern. Unfortunately, these $n$ and $m$ values are too small to observe the patterns that are focused in this work.
![(Color online) Time to forget next person scales with $n$ for small $n$ and $m$ values. Average of 40 simulations.[]{data-label="fig:fig_absorbingStateA1B1"}](fig_absorbingStateA1B1.eps){width="\columnwidth" height="\columnwidth"}
Related Work on Fame
====================
Recently some studies on fame has been done. The difficulty starts with the definition of fame. An innovative metric for fame is defined as the number of hits returned from a search of a person’s name on Google [@Simkin2006a; @Simkin2006b]. In this study, the fames and the achievements of WWI fighter pilots are examined. “Fame” $F$ is defined as the number of Google hits. “Achievement” is the number of opponent aircrafts destroyed. It is found that fame grows exponentially with achievement. The distribution of fame is given as $P(F)\propto F^{- \gamma}$ where $\gamma\approx 2$. A study on scientist using the number of Google hits for fame gives another distribution, $P(F)\propto e^{-\eta F}$ where $\eta=0.00102$ [@Bagrow2004; @Bagrow2005].
Scientific papers can be “famous” by getting cited. A study on scientific papers published in Physical Review D in 1975-1994 has been done [@Simkin2003a]. There are 24000 papers, 350000 citations, that is 15 citation per paper on the average. Yet, 44 papers are cited 500 times or more. It is found that copying from the list of references used in other papers has an impact. A paper that is already cited has more chance to get cited again.
Early results of the simple recommendation model such as the fast and slow forget regimes, asymptotic approach to absorbing states and degeneration of the distribution of knownness to fame as $\rho$ decreases were presented in [@Bingol2005].
Conclusions
===========
“Too many to remember” is quite the common case in many complex systems. A dynamic memory model is defined where agents interact by exchanging recommendations. A random-selection based model is described as the simplest instantiation of the general model. Although the model does not prefer any agent, some agents become increasingly famous as the memory gets smaller. This observation can be interpreted as the *emergence of fame*.
The model can be used in some practical applications. Suppose some agents are preferred in the recommendations. Then, their fame is expected to increase and last longer. This can be used to model the social dynamics of advertisement. Essential questions in advertisement such as how frequently advertise or how widely advertise can be better estimated. Voting or election results are studied in opinion dynamics [@SznajdWeron2000]. Emergence of fame can be considered as formation of an opinion through interactions of agents.
The general model will serve as a basis for building sophisticated models as different selection criteria are adopted and the agent interaction scheme is restricted with new assumptions. For example, it is possible to define selections so that only agents with a common friend can interact. This leads to a version of Axelrod’s Culture Model [@Axelrod1997]. The model can be modified so that the giver always recommends itself rather than some agent from its memory. Then it becomes very close to the small-world model presented in [@Watts1998]. Another possibility is to place the agents on the vertices of an interaction graph, possibly with small-word or scale-free properties in order to introduce real world flavor.
This work was partially supported by Bogazici University Research Projects under the grant number 07A105 and was partially based on the work performed in the framework of the FP6 project SEE-GRID-2, which is funded by the European Community (under contract number INFSO-RI-031775). The author would like to thank Sidney Redner, Alber Ali Salah, Cem Ersoy, Can Ozturan and Pinar Yolum for their useful comments. The author also thanks to Muhittin Mungan, Amac Herdagdelen for their invaluable support especially in scaling and to anonymous referees for their enriching comments and suggestions.
[^1]: For example, one may know $2 \times 10^3$ people while there are $15 \times 10^6$ in Istanbul that makes the ratio of $1.3 \times 10^{-4}$.
[^2]: Example of mobile phone is proposed by one of the anonymous referee to whom the author would like to thank.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A short quantum Markov chain is a tripartite state $\rho_{ABC}$ such that system $A$ can be recovered perfectly by acting on system $C$ of the reduced state $\rho_{BC}$. Such states have conditional mutual information $I(A;B|C)$ equal to zero and are the only states with this property. A quantum channel $\mathcal{N}$ is sufficient for two states $\rho$ and $\sigma$ if there exists a recovery channel using which one can perfectly recover $\rho$ from $\mathcal{N}(\rho)$ and $\sigma$ from $\mathcal{N}(\sigma)$. The relative entropy difference $D(\rho\Vert\sigma)-D(\mathcal{N}(\rho)\Vert\mathcal{N}(\sigma))$ is equal to zero if and only if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$. In this paper, we show that these properties extend to Rényi generalizations of these information measures which were proposed in \[Berta *et al.*, J. Math. Phys. **56**, 022205, (2015) and Seshadreesan *et al.*, J. Phys. A **48**, 395303, 2015\], thus providing an alternate characterization of short quantum Markov chains and sufficient quantum channels. These results give further support to these quantities as being legitimate Rényi generalizations of the conditional mutual information and the relative entropy difference. Along the way, we solve some open questions of Ruskai and Zhang, regarding the trace of particular matrices that arise in the study of monotonicity of relative entropy under quantum operations and strong subadditivity of the von Neumann entropy.'
author:
- 'Nilanjana Datta[^1]'
- 'Mark M. Wilde[^2]'
bibliography:
- 'Ref.bib'
title: '**Quantum Markov chains, sufficiency of quantum channels, and Rényi information measures**'
---
Introduction
============
Markov chains and sufficient statistics are two fundamental notions in probability [@Feller; @Norris] and statistics [@stats]. Three random variables $X$, $Y$, and $Z$ constitute a three-step Markov chain (denoted as $X-Y-Z$) if $X$ and $Z$ are independent when conditioned on $Y$. In particular, if $p_{XYZ}(x,y,z)$ is their joint probability distribution, then$$\begin{aligned}
p_{XYZ}(x,y,z) & =p_{X}(x)\ p_{Y|X}(y|x)\ p_{Z|Y}(z|y)\nonumber\\
& =p_{X|Y}(x|y)\ p_{Z|Y}(z|y)\ p_{Y}(y). \label{eq:MC}$$ In the information-theoretic framework, such a Markov chain corresponds to a *[recoverability condition]{}* in the following sense. Consider $X$, $Y$, and $Z$ to be the inputs and outputs of two channels (i.e., stochastic maps) $p_{Y|X}$ and $p_{Z|Y}$, as in the figure below.
If $X-Y-Z$ is a three-step Markov chain, then the input $X$, if lost, can be recovered from $Y$ alone (without any knowledge of $Z$) by the action of the stochastic map $p_{X|Y}$, as is evident from (\[eq:MC\]).
It is well known that such a Markov chain $X-Y-Z$ can be characterized by an information measure [@book1991cover], namely, the *[conditional mutual information]{}* $I(X;Z|Y)$. For any three random variables $X$, $Y$ and $Z$, it is defined as$$I(X;Z|Y)\equiv H(XY)+H(ZY)-H(Y)-H(XYZ),$$ where $H(W)\equiv-\sum_{w}p_{W}(w)\log p_{W}(w)$ is the Shannon entropy of a random variable $W\sim p_{W}(w)$. It is non-negative and equal to zero if and only if $X-Y-Z$ is a Markov chain.
In statistics, for a given sample of independent and identically distributed data conditioned on an unknown parameter $\theta$, a *[sufficient statistic]{}* is a function of the sample whose value contains all the information needed to compute any estimate of the parameter. One can extend this notion to that of a sufficient channel (or sufficient stochastic map), as discussed in [@Petz1986; @Petz1988; @Mosonyi2004; @M05]. A channel $T\equiv
T_{Y|X}$ is sufficient for two input distributions $p_{X}$ and $q_{X}$ if there exists another channel (a *[recovery channel]{}*) such that both these inputs can be recovered perfectly by sending the outputs of the channel $T_{Y|X}$ corresponding to them through it. This notion of channel sufficiency is likewise characterized by an information measure, namely, the relative entropy difference$$D(p_{X}\Vert q_{X})-D(T(p_{X})\Vert T(q_{X})), \label{eq:cl-rel-ent-diff}$$ where $T(p_{X})$ and $T(q_{X})$ are the distributions obtained after the action of the channel, and $D(p_{X}\Vert q_{X})$ denotes the relative entropy (or Kullback-Leibler divergence) [@book1991cover] between $p_{X}$ and $q_{X}$. It is defined as$$D(p_{X}\Vert q_{X})\equiv\sum_{x}p_{X}(x)\log\left( \frac{p_{X}(x)}{q_{X}(x)}\right) ,$$ if for all $x$, $q_{X}(x)\neq0$ if $p_{X}(x)\neq0$. It is equal to $+\infty$ otherwise. The notion of recoverability provides a connection between the notions of Markov chains and sufficient channels.
The generalization of the above ideas to quantum information theory has been a topic of continuing and increasing interest (see, e.g., [@HMPB11; @BSW14; @FR14] and references therein). In the quantum setting, density operators play a role analogous to that of probability distributions in the classical case, and in [@HJPW04], a quantum Markov chain $A-C-B$ was defined to be a tripartite density operator $\rho_{ABC}$ with conditional (quantum) mutual information $I(A;B|C)_{\rho}$ equal to zero, where$$I(A;B|C)_{\rho}\equiv H(AC)_{\rho}+H(BC)_{\rho}-H(C)_{\rho}-H(ABC)_{\rho},
\label{eq:CMI}$$ and $H(F)_{\sigma}\equiv-$Tr$\left\{ \sigma_{F}\log\sigma_{F}\right\} $ denotes the von Neumann entropy of a density operator $\sigma_{F}$. (We take the convention $A-C-B$ for a quantum Markov chain because we are often interested in quantum correlations between Alice ($A$) and Bob ($B$), which are potentially mediated by a third party, here labeled by $C$.) Strong subadditivity of the von Neumann entropy guarantees that the conditional mutual information $I(A;B|C)_{\rho}$ is non-negative for all density operators [@PhysRevLett.30.434; @LR73], and it is equal to zero if and only if there is a decomposition of $\mathcal{H}_{C}$ as$$\mathcal{H}_{C}=\bigoplus\limits_{j}\mathcal{H}_{C_{L_{j}}}\otimes
\mathcal{H}_{C_{R_{j}}} \label{eq:h-c-direct-sum}$$ such that$$\rho_{ABC}=\bigoplus\limits_{j}q(j)\rho_{AC_{L_{j}}}^{j}\otimes\rho_{C_{R_{j}}B}^{j}, \label{eq:direct-sum-structure}$$ for a probability distribution $\{q(j)\}$ and sets of density operators $\{\rho_{AC_{L_{j}}}^{j},\rho_{C_{R_{j}}B}^{j}\}$ [@HJPW04]. Following [@HJPW04], we call such states *short quantum Markov chains* $A-C-B$. In analogy with the classical case, $I(A;B|C)_{\rho}=0$ is equivalent to the full state $\rho_{ABC}$ being recoverable after the loss of system $A$ by the action of a quantum recovery channel $\mathcal{R}_{C\rightarrow AC}^{P}$ on system $C$ alone:$$I(A;B|C)_{\rho}=0\Leftrightarrow\rho_{ABC}=\mathcal{R}_{C\rightarrow AC}^{P}(\rho_{BC}),$$ where$$\mathcal{R}_{C\rightarrow AC}^{P}(\cdot)\equiv\rho_{AC}^{\frac{1}{2}}\rho
_{C}^{-\frac{1}{2}}(\cdot)\rho_{C}^{-\frac{1}{2}}\rho_{AC}^{\frac{1}{2}}
\label{eq:special-Petz}$$ is a special case of the so-called *Petz recovery channel* [@Petz1986; @Petz1988]. Note that this channel acts as the identity on system $B$ and the superscript $P$ refers to Petz.
As a generalization of the classical notion of a sufficient channel, several works have discussed and studied the notion of a sufficient quantum channel [@Petz1986; @Petz1988; @Mosonyi2004; @M05]. A definition of this concept is as follows:
\[Sufficiency of a quantum channel\]\[def:channel-suff\]Let $\rho$ and $\sigma$ be density operators acting on a Hilbert space $\mathcal{H}$, and let $\mathcal{N}$ be a quantum channel taking these density operators to density operators acting on a Hilbert space $\mathcal{K}$. Then the quantum channel $\mathcal{N}$ is sufficient for them if one can perfectly recover $\rho$ from $\mathcal{N(}\rho)$ and $\sigma$ from $\mathcal{N(}\sigma)$ by the action of a quantum recovery channel $\mathcal{R}$, i.e., if there exists an $\mathcal{R}$ such that$$\rho=\left( \mathcal{R\circ N}\right) (\rho),\ \ \ \ \ \ \ \ \sigma=\left(
\mathcal{R\circ N}\right) (\sigma). \label{eq:recover-rho}$$ We define sufficiency in the same way even if $\rho$ and $\sigma$ are arbitrary positive semi-definite operators.
If is true for some recovery channel $\mathcal{R}$, it is known that the following Petz recovery channel $\mathcal{R}_{\sigma
,\mathcal{N}}^{P}$ satisfies (\[eq:recover-rho\]) as well [@Petz1988]: $$\mathcal{R}_{\sigma,\mathcal{N}}^{P}(\omega)\equiv\sigma^{\frac{1}{2}}\mathcal{N}^{\dag}\left( [\mathcal{N}(\sigma)]^{-\frac{1}{2}}\omega
\lbrack\mathcal{N}(\sigma)]^{-\frac{1}{2}}\right) \sigma^{\frac{1}{2}}.
\label{eq:Petz-map}$$ (Note that the Petz recovery channel in (\[eq:special-Petz\]) is a special case of (\[eq:Petz-map\]) with $\sigma=\rho_{AC}$ and $\mathcal{N}=\operatorname{Tr}_{A}$.) As a generalization of the classical case, the sufficiency of a quantum channel is characterized by the following information measure, the relative entropy difference$$D(\rho\Vert\sigma)-D(\mathcal{N(}\rho)\Vert\mathcal{N(}\sigma)),
\label{eq:rel-ent-diff}$$ where $D(\rho\Vert\sigma)$ denotes the *[quantum relative entropy]{}* [@U62]. It is defined as$$D(\rho\Vert\sigma)\equiv\text{Tr}\left\{ \rho\left[ \log\rho-\log
\sigma\right] \right\} , \label{rel-entropy}$$ whenever the support of $\rho$ is contained in the support of $\sigma$ and it is equal to $+\infty$ otherwise. The relative entropy difference in (\[eq:rel-ent-diff\]) is non-negative due to the monotonicity of relative entropy under quantum channels [@Lindblad1975; @U77], and it is equal to zero if and only if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$, i.e., if the Petz recovery channel $\mathcal{R}_{\sigma,\mathcal{N}}^{P}$ satisfies (\[eq:recover-rho\]) [@Petz1986; @Petz1988]. Further, Mosonyi and Petz have shown that the relative entropy difference in (\[eq:rel-ent-diff\]) is equal to zero if and only if $\rho$, $\sigma$, and $\mathcal{N}$ have the explicit form recalled below in Theorem \[thm:mos-petz-decomp\] of Section \[sec:prelim\] [@Mosonyi2004; @M05], which generalizes the result stated in (\[eq:h-c-direct-sum\])-(\[eq:direct-sum-structure\]).
Due to its operational interpretation in the quantum Stein’s lemma [@HP91], the quantum relative entropy plays a central role in quantum information theory. In particular, fundamental limits on the performance of information-processing tasks in the so-called asymptotic, memoryless (or i.i.d.) setting are given in terms of quantities derived from the quantum relative entropy.
There are, however, other generalized relative entropies (or divergences) which are also of operational significance. Important among these are the *Rényi relative entropies* [@Renyi; @P86] and the more recently defined *sandwiched Rényi relative entropies* [@MDSFT13; @WWY13]. For $\alpha\in(0,1)$, the Rényi relative entropies arise in the quantum Chernoff bound [@chernoff], which characterizes the minimum probability of error in discriminating two different quantum states in the setting of asymptotically many copies. Moreover, in analogy with the operational interpretation of their classical counterparts, the Rényi relative entropies can be viewed as generalized cutoff rates in quantum binary state discrimination [@milan]. The sandwiched Rényi relative entropies find application in the strong converse domain of a number of settings dealing with hypothesis testing or channel capacity [@WWY13; @MO13; @GW13; @TWW14; @CMW14; @HT14].
This motivates the introduction of Rényi generalizations of the conditional mutual information. Two of these generalizations, defined in [@BSW14], are given as follows:$$\begin{aligned}
I_{\alpha}(A;B|C)_{\rho} & \equiv\frac{1}{\alpha-1}\log\text{Tr}\left\{
\rho_{ABC}^{\alpha}\rho_{AC}^{\frac{1-\alpha}{2}}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{BC}^{1-\alpha}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{AC}^{\frac{1-\alpha
}{2}}\right\} ,\label{eq:Renyi-CMI}\\
\widetilde{I}_{\alpha}(A;B|C)_{\rho} & \equiv\frac{2\alpha}{\alpha-1}\log\left\Vert \rho_{ABC}^{\frac{1}{2}}\rho_{AC}^{\frac{1-\alpha}{2\alpha}}\rho_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{BC}^{\frac{1-\alpha}{2\alpha}}\right\Vert _{2\alpha}, \label{eq:sand-Renyi-cmi}$$ where $\alpha\in\left( 0,1\right) \cup\left( 1,\infty\right) $ denotes the Rényi parameter. (Note that we use the notation $\left\Vert A\right\Vert
_{\alpha}\equiv\operatorname{Tr}\{(\sqrt{A^{\dag}A})^{\alpha}\}^{1/\alpha}$ even for $\alpha\in\left( 0,1\right) $, when it is not a norm.) Both these quantities converge to (\[eq:CMI\]) in the limit $\alpha\rightarrow1$, they are non-negative,$\,$ and obey several properties of the conditional mutual information defined in (\[eq:CMI\]), as shown in [@BSW14]. In [@SBW14], the authors proposed some definitions for Rényi generalizations of a relative entropy difference, two of which are as follows:$$\begin{aligned}
\Delta_{\alpha}(\rho,\sigma,\mathcal{N}) & \equiv\frac{1}{\alpha-1}\log\text{Tr}\left\{ \rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right)
\sigma^{\frac{1-\alpha}{2}}\right\} ,\label{eq:Renyi-rel-ent-diff}\\
\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N}) & \equiv\frac{\alpha
}{\alpha-1}\log\left\Vert \rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N(}\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N(}\sigma)^{\frac
{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac
{1}{2}}\right\Vert _{\alpha}. \label{eq:Renyi-rel-ent-diff-sand}$$ The quantities above converge to (\[eq:rel-ent-diff\]) in the limit $\alpha\rightarrow1$ [@SBW14].
The quantities defined in (\[eq:Renyi-CMI\])-(\[eq:Renyi-rel-ent-diff-sand\]) can be expressed in terms of Rényi relative entropies, namely the $\alpha$-Rényi relative entropy and the $\alpha$-sandwiched Rényi relative entropy defined in Section \[Sec-rel-entropies\] [@BSW14]. The corresponding expressions for the Rényi generalizations of the conditional mutual information are$$\begin{aligned}
I_{\alpha}(A;B|C)_{\rho} & =D_{\alpha}\left( \rho_{ABC}\middle\Vert\left(
\rho_{AC}^{\frac{1-\alpha}{2}}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{BC}^{1-\alpha}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{AC}^{\frac{1-\alpha}{2}}\right)
^{\frac{1}{1-\alpha}}\right) ,\\
\widetilde{I}_{\alpha}(A;B|C)_{\rho} & =\widetilde{D}_{\alpha}\left(
\rho_{ABC}\middle\Vert\left( \rho_{AC}^{\frac{1-\alpha}{2\alpha}}\rho
_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{BC}^{\frac{1-\alpha}{\alpha}}\rho
_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{AC}^{\frac{1-\alpha}{2\alpha}}\right)
^{\frac{\alpha}{1-\alpha}}\right) ,\end{aligned}$$ respectively, and those for the relative entropy difference are given in (\[eq:ren-rel-ent-diff-1\]) and (\[eq:ren-rel-ent-diff-2\]), respectively.
Before proceeding, we should note that the conditional mutual information is a special case of a relative entropy difference. Indeed, one can check that$$I(A;B|C)_{\rho}=D(\rho\Vert\sigma)-D(\mathcal{N}(\rho)\Vert\mathcal{N}(\sigma)) \label{eq:rel-ent-cmi}$$ for the choices$$\rho=\rho_{ABC},\ \ \ \ \ \ \ \ \sigma=\rho_{AC}\otimes I_{B},\ \ \ \ \ \ \ \ \mathcal{N}=\operatorname{Tr}_{A}. \label{eq:rel-ent-cmi-1}$$ This reduction extends as well to the Rényi quantities for all $\alpha
\in(0,1)\cup(1,\infty)$:$$\begin{aligned}
I_{\alpha}(A;B|C)_{\rho} & =\Delta_{\alpha}(\rho_{ABC},\rho_{AC}\otimes
I_{B},\operatorname{Tr}_{A}),\\
\widetilde{I}_{\alpha}(A;B|C)_{\rho} & =\widetilde{\Delta}_{\alpha}(\rho_{ABC},\rho_{AC}\otimes I_{B},\operatorname{Tr}_{A}),\end{aligned}$$ as pointed out in [@SBW14]. This realization is helpful in simplifying some of the arguments in this paper.
Summary of results
==================
As highlighted in the Introduction, an important property of the conditional (quantum) mutual information of a tripartite quatum state is that it is always non-negative and vanishes if and only if the state is a short quantum Markov chain. The relative entropy difference of a pair of quantum states and a quantum channel is also non-negative and vanishes if and only if the channel is sufficient for the pair of states. Consequently, it is reasonable to require that Rényi generalizations of these information measures are also non-negative and vanish under the same necessary and sufficient conditions as mentioned above.
In this paper, we prove these properties for the quantities $I_{\alpha}$ and $\widetilde{I}_{\alpha}$, and the quantities $\Delta_{\alpha}$ and $\widetilde{\Delta}_{\alpha}$ defined in the Introduction. This contributes further evidence that $I_{\alpha}( A;B|C) _{\rho}$ and $\widetilde{I}_{\alpha
}( A;B|C) _{\rho}$ are legitimate Rényi generalizations of the conditional mutual information $I( A;B|C) _{\rho}$ and that $\Delta_{\alpha}( \rho
,\sigma,\mathcal{N}) $ and $\widetilde{\Delta}_{\alpha}( \rho,\sigma
,\mathcal{N}) $ are legitimate Rényi generalizations of the relative entropy difference $D( \rho\Vert\sigma) -D( \mathcal{N}( \rho) \Vert
\mathcal{N}( \sigma) ) $. In particular, we prove the following:
1. $I_{\alpha}( A;B|C) _{\rho}=0$ for some $\alpha\in\left( 0,1\right)
\cup\left( 1,2\right) $ if and only if $\rho_{ABC}$ is a short quantum Markov chain with a decomposition as in (\[eq:direct-sum-structure\]).
2. $\widetilde{I}_{\alpha}( A;B|C) _{\rho}=0$ for some $\alpha\in\left(
1/2,1\right) \cup\left( 1,\infty\right) $ if and only if $\rho_{ABC}$ is a short quantum Markov chain with a decomposition as in (\[eq:direct-sum-structure\]).
3. $\Delta_{\alpha}( \rho,\sigma,\mathcal{N}) $ is non-negative for $\alpha\in\left( 0,1\right) \cup\left( 1,2\right) $ and $\widetilde
{\Delta}_{\alpha}( \rho,\sigma,\mathcal{N}) $ is non-negative for $\alpha
\in\left( 1/2,1\right) \cup\left( 1,\infty\right) $.
4. $\Delta_{\alpha}( \rho,\sigma,\mathcal{N}) =0$ for some $\alpha
\in\left( 0,1\right) \cup\left( 1,2\right) $ if and only if the quantum channel $\mathcal{N}$ is sufficient for states $\rho$ and $\sigma$, so that $\mathcal{N}$, $\rho$, and $\sigma$ decompose as in (\[eq:sufficiency-states-decompose\])-(\[eq:sufficiency-channel-decompose\]).
5. $\widetilde{\Delta}_{\alpha}( \rho,\sigma,\mathcal{N}) =0$ for some $\alpha\in\left( 1/2,1\right) \cup\left( 1,\infty\right) $ if and only if the quantum channel $\mathcal{N}$ is sufficient for states $\rho$ and $\sigma
$, so that $\mathcal{N}$, $\rho$, and $\sigma$ decompose as in (\[eq:sufficiency-states-decompose\])-(\[eq:sufficiency-channel-decompose\]).
6. Generalizations of the conditional mutual information and the relative entropy difference arising from the so-called min- and max-relative entropies [@min-max-ND; @Dupuis] (which play an important role in one-shot information theory) satisfy identical properties.
Along the way, we resolve some open questions stated in [@R02; @Z14b]. Let $\rho_{ABC}$ be a positive definite density operator. We prove that$$\operatorname{Tr}\left\{ \left( \rho_{AC}^{\frac{1-\alpha}{2}}\rho
_{C}^{\frac{\alpha-1}{2}}\rho_{BC}^{1-\alpha}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{AC}^{\frac{1-\alpha}{2}}\right) ^{\frac{1}{1-\alpha}}\right\} \leq1.$$ for all $\alpha\in\left( 0,1\right) \cup\left( 1,2\right) $, and$$\operatorname{Tr}\left\{ \left( \rho_{AC}^{\frac{1-\alpha}{2\alpha}}\rho
_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{BC}^{\frac{1-\alpha}{\alpha}}\rho
_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{AC}^{\frac{1-\alpha}{2\alpha}}\right)
^{\frac{\alpha}{1-\alpha}}\right\} \leq1,$$ for all $\alpha\in\left( 1/2,1\right) \cup\left( 1,\infty\right) $. Let $\rho$ and $\sigma$ be positive definite density operators and let $\mathcal{N}$ be a strict completely positive trace preserving (CPTP) map (that is, a CPTP map such that $\mathcal{N}(X)$ is positive definite whenever $X$ is positive definite). We prove that$$\operatorname{Tr}\left\{ \left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right)
\sigma^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}\right\} \leq1,
\label{eq:N-rho-sigma-leq-1}$$ for $\alpha\in\left( 0,1\right) \cup(1,2)$, and$$\operatorname{Tr}\left\{ \left[ \sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac
{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\right]
^{\frac{\alpha}{1-\alpha}}\right\} \leq1, \label{eq:2-N-rho-sigma-leq-1}$$ for $\alpha\in\left( 1/2,1\right) \cup(1,\infty)$. By taking the limit $\alpha\rightarrow1$, the inequalities in and imply that$$\operatorname{Tr}\left\{ \exp\left\{ \log\sigma+\mathcal{N}^{\dag}\left(
\log\mathcal{N}(\rho)-\log\mathcal{N}(\sigma)\right) \right\} \right\}
\leq1.$$
The rest of the paper is devoted to establishing these claims. We begin by recalling some mathematical preliminaries and known results, and follow by establishing the latter claims first and then move on to the former ones.
Preliminaries {#sec:prelim}
=============
Let $\mathcal{B}(\mathcal{H})$ denote the algebra of bounded linear operators acting on a Hilbert space $\mathcal{H}$. We restrict to finite-dimensional Hilbert spaces throughout this paper. We denote the support of an operator $A$ by supp$(A)$. For a Hermitian operator $A$, by $A^{-1}$ we mean the inverse restricted to supp$(A)$, so that $AA^{-1}=A^{-1}A$ is the orthogonal projection onto supp$(A)$. More generally, for a function $f$ and Hermitian operator $A$ with spectral decomposition $A=\sum_{i}\lambda_{i}|i\rangle\langle i|$, we define $f(A)$ to be $\sum_{i:\lambda_{i}\neq0}f(\lambda_{i})|i\rangle\langle i|$. Let $\mathcal{B}(\mathcal{H})_{+}$ denote the subset of positive semidefinite operators, and let $\mathcal{B}(\mathcal{H})_{++}$ denote the subset of positive definite operators. We also write $X\geq0$ if $X\in\mathcal{B}(\mathcal{H})_{+}$ and $X>0$ if $X\in\mathcal{B}(\mathcal{H})_{++}$. An operator $\rho$ is in the set $\mathcal{S}(\mathcal{H})$ of density operators (or states) if $\rho\in\mathcal{B}(\mathcal{H})_{+}$ and Tr$\{\rho\}=1$, and an operator $\rho$ is in the set $\mathcal{S}(\mathcal{H})_{++}$ of positive definite density operators if $\rho\in\mathcal{B}(\mathcal{H})_{++}$ and Tr$\{\rho\}=1$. Throughout much of the paper, for technical convenience and simplicity, we consider states in $\mathcal{S}(\mathcal{H})_{++}$. For $\alpha\geq1$, we define the $\alpha$-norm of an operator $X$ as$$\left\Vert X\right\Vert _{\alpha}\equiv\left[ \text{Tr}\{|X|^{\alpha
}\}\right] ^{1/\alpha}, \label{eq:a-norm}$$ where $|X|\equiv\sqrt{X^{\dag}X}$, and we use the same notation even for the case $\alpha\in(0,1)$, when it is not a norm. The fidelity $F(\rho,\sigma)$ of two states $\rho,\sigma\in\mathcal{S}(\mathcal{H})$ is defined as $$F(\rho,\sigma)\equiv\left\Vert \sqrt{\rho}\sqrt{\sigma}\right\Vert _{1}^{2}.$$
A quantum channel is given by a completely positive, trace-preserving (CPTP) map $\mathcal{N}:\mathcal{B}(\mathcal{H})\mapsto\mathcal{B}(\mathcal{K})$, with $\mathcal{H}$ and $\mathcal{K}$ being the input and output Hilbert spaces of the channel, respectively. Let $\left\langle C,D\right\rangle
\equiv\operatorname{Tr}\left\{ C^{\dag}D\right\} $ denote the Hilbert-Schmidt inner product of $C,D\in\mathcal{B}(\mathcal{H})$. The adjoint of the quantum channel $\mathcal{N}$ is a completely positive unital map $\mathcal{N}^{\dagger}:\mathcal{B}(\mathcal{K})\mapsto\mathcal{B}(\mathcal{H})$ defined through the following relation:$$\left\langle B,\mathcal{N}(A)\right\rangle =\langle\mathcal{N}^{\dagger
}(B),A\rangle,$$ for all $A\in\mathcal{B}(\mathcal{H)}$ and $B\in\mathcal{B}(\mathcal{K})$. A linear map is said to be a strict CPTP map if it is CPTP and if $\mathcal{N}(A)\in\mathcal{B}(\mathcal{K})_{++}$ for all $A\in\mathcal{B}(\mathcal{H})_{++}$. Note that a CPTP map is strict if and only if $\mathcal{N}(I)>0$ [@B07 Section 2.2]. We denote the identity channel as id but often suppress it for notational simplicity.
The set $\left\{ U^{i}\right\} $ of Heisenberg-Weyl unitaries acting on a finite-dimensional Hilbert space $\mathcal{H}$ of dimension $d$ has the property that$$\frac{1}{d^{2}}\sum_{i}U^{i}X\left( U^{i}\right) ^{\dag}=\text{Tr}\left\{
X\right\} \frac{I}{d},$$ for any operator $X$ acting on $\mathcal{H}$.
Generalized relative entropies {#Sec-rel-entropies}
------------------------------
The following relative entropies of a density operator $\rho$ with respect to a positive semidefinite operator $\sigma$ play an important role in this paper. In what follows, we restrict the definitions to the case in which $\rho$ and $\sigma$ satisfy the condition $\operatorname{supp}\rho\subseteq\operatorname{supp}\sigma$, with the understanding that they are equal to $+\infty$ when $\alpha>1$ and $\operatorname{supp}\rho\not \subseteq \operatorname{supp}\sigma$. The $\alpha$-Rényi relative entropy is defined for $\alpha\in\left(
0,1\right) \cup\left( 1,\infty\right) $ as follows [@P86]:$$D_{\alpha}(\rho\Vert\sigma)\equiv\frac{1}{\alpha-1}\log\text{Tr}\left\{
\rho^{\alpha}\sigma^{1-\alpha}\right\} .$$ The $\alpha$-sandwiched Rényi relative entropy is defined for $\alpha
\in\left( 0,1\right) \cup\left( 1,\infty\right) $ as follows [@MDSFT13; @WWY13]:$$\widetilde{D}_{\alpha}(\rho\Vert\sigma)\equiv\frac{1}{\alpha-1}\log
\text{Tr}\left\{ \left( \sigma^{\frac{1-\alpha}{2\alpha}}\rho\sigma
^{\frac{1-\alpha}{2\alpha}}\right) ^{\alpha}\right\} .$$ Both quantities above reduce to the quantum relative entropy in (\[rel-entropy\]) in the limit$~\alpha\rightarrow1$.
A fundamental property of the quantum relative entropy is that it is monotone with respect to quantum channels (also known as the *[data-processing inequality]{}*):$$D(\rho\Vert\sigma)\geq D\left( \mathcal{N}(\rho)\Vert\mathcal{N}(\sigma)\right) , \label{DPI}$$ where $\mathcal{N}$ is a quantum channel. It is known that the data-processing inequality is satisfied by the $\alpha$-Rényi relative entropy for $\alpha\in\lbrack0,1)\cup(1,2]$ [[@P86]]{}, and for the $\alpha$-sandwiched Rényi relative entropy for $\alpha\in\lbrack1/2,1)\cup(1,\infty]$ [@FL13; @B13; @MDSFT13; @WWY13; @MO13].
Two special cases of the $\alpha$-sandwiched Rényi relative entropy are of particular significance in one-shot information theory [@renato-thesis; @marco-thesis], namely the min-relative entropy [@Dupuis] and the max-relative entropy [@min-max-ND]. These are defined as follows:$$D_{\min}(\rho\Vert\sigma)\equiv\widetilde{D}_{1/2}(\rho\Vert\sigma)=-\log
F(\rho,\sigma),$$ and $$D_{\max}(\rho\Vert\sigma)\equiv\inf\{\lambda:\rho\leq2^{\lambda}\sigma
\}=\lim_{\alpha\rightarrow\infty}\widetilde{D}_{\alpha}(\rho\Vert\sigma).$$ The relative entropies defined above, satisfy the following lemma [@MDSFT13]:
\[lem:min-max-0\]For $\omega\in\mathcal{S}( \mathcal{H}) $ and $\tau
\in\mathcal{B}( \mathcal{H}) _{+} $, such that $\operatorname{Tr}\left\{
\omega\right\} \geq\operatorname{Tr}\left\{ \tau\right\} $,$$\begin{aligned}
D_{\alpha}(\omega\Vert\tau) & \geq0\text{ for }\alpha\in\left( 0,1\right)
\cup\left( 1,2\right) ,\\
\widetilde{D}_{\alpha}(\omega\Vert\tau) & \geq0\text{ for }\alpha\in\left(
1/2,1\right) \cup\left( 1,\infty\right) ,\\
D_{\min}(\omega\Vert\tau) & \geq0,\label{dmin-0}\\
D_{\max}(\omega\Vert\tau) & \geq0, \label{dmax-0}$$ with equalities holding if and only if $\omega=\tau$.
In proving our results, we also employ the notion of a *[quantum $f$-divergence]{}*, first introduced by Petz [@P86]. It can be conveniently expressed as follows [@FQAEP]:
For $A\in\mathcal{B}( \mathcal{H}) _{+}$, $B\in\mathcal{B}\left(
\mathcal{H}\right) _{++}$ and an operator convex function $f$ on $[0,\infty
)$, the $f$-divergence of $A$ with respect to $B$ is given by $$S_{f}(A\Vert B)=\langle\Gamma|\left( \sqrt{B}\otimes I\right) f\left(
B^{-1}\otimes A^{T}\right) \left( \sqrt{B}\otimes I\right) |\Gamma\rangle,$$ where $|\Gamma\rangle=\sum_{i}|i\rangle\otimes|i\rangle$, and $\{|i\rangle\}$ is an orthonormal basis of $\mathcal{H}$ with respect to which the transpose is defined.
\[rem:renyi-f-div\]Special cases of this are as follows:
1. The trace expression $\operatorname{Tr}\left\{ \rho^{\alpha}\sigma^{1-\alpha}\right\} $ of the $\alpha$-Rényi relative entropy, for the choice $f(x)=x^{\alpha}$ for $\alpha\in\left( 1,2\right) $, and $-\operatorname{Tr}\left\{ \rho^{\alpha}\sigma^{1-\alpha}\right\} $ for the choice $f(x)=-x^{\alpha}$ for $\alpha\in\left( 0,1\right) $.
2. The quantum relative entropy for the choice $f(x)=x\log x$.
The equivalence relations given in Lemma \[lem-equiv\] below follow directly from [@HMPB11 Theorem 5.1].
\[lem-equiv\] For $A,B\in\mathcal{B}(\mathcal{H})_{++}$ and a strict CPTP map $\mathcal{N}$ acting on $\mathcal{B}\left( \mathcal{H}\right) $, the following conditions are equivalent:[$$\begin{aligned}
S_{f}(\mathcal{N}(A)\Vert\mathcal{N}(B)) & =S_{f}(A\Vert B)\quad
{\hbox{for the functions in Remark~\ref{rem:renyi-f-div}}},\label{item1}\\
{\mathcal{N}}^{\dagger}\left[ \log\mathcal{N}(A)-\log\mathcal{N}(B)\right]
& =\log A-\log B,\label{item2}\\
& \mathcal{N}\text{ is sufficient for }A\text{ and }B. \label{eq:item3}$$ ]{}
\[lem-MC\] If $\rho_{ABC}$ is a positive definite density operator such that $$\label{eq-equal}\log\rho_{ABC}=\log\rho_{AC}+\log\rho_{BC}-\log\rho_{C},$$ then it is a short quantum Markov chain $A-C-B$.
The identity (\[eq-equal\]) is known from [@R02] to be a condition for the conditional mutual information $I(A;B|C)_{\rho}$ of $\rho_{ABC}$ to be equal to zero, which, by [@HJPW04], implies that $\rho_{ABC}$ is a short quantum Markov chain $A-C-B$.
\[[@Mosonyi2004; @M05]\]\[thm:mos-petz-decomp\]Let $\rho\in\mathcal{S}(
\mathcal{H}) _{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and $\mathcal{N}$ be a strict CPTP map. Then $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$ (as in ) if and only if the following conditions hold
1. There exist decompositions of $\mathcal{H}$ and $\mathcal{K}$ as follows:$$\mathcal{H}=\bigoplus\limits_{j}\mathcal{H}_{L_{j}}\otimes\mathcal{H}_{R_{j}},\ \ \ \ \ \ \ \ \mathcal{K}=\bigoplus\limits_{j}\mathcal{K}_{L_{j}}\otimes\mathcal{K}_{R_{j}}, \label{eq:suff-decomp}$$ where $\dim\left( \mathcal{H}_{L_{j}}\right) =\dim\left( \mathcal{K}_{L_{j}}\right) $ for all $j$.
2. With respect to the decomposition in , $\rho$ and $\sigma$ can be written as follows:$$\rho=\bigoplus\limits_{j}p( j) \rho_{L_{j}}^{j}\otimes\tau_{R_{j}}^{j},\ \ \ \ \ \ \ \ \sigma=\bigoplus\limits_{j}q( j) \sigma_{L_{j}}^{j}\otimes\tau_{R_{j}}^{j}, \label{eq:sufficiency-states-decompose}$$ for some probability distribution $\{p( j) \}$, positive reals $\{q( j) \}$, sets of states $\{\rho_{L_{j}}^{j}\}$ and $\{\tau_{R_{j}}^{j}\}$ and set of positive definite operators $\{\sigma_{L_{j}}^{j}\}$.
3. With respect to the decomposition in , the quantum channel $\mathcal{N}$ can be written as$$\mathcal{N}=\bigoplus\limits_{j}\mathcal{U}_{j}\otimes\mathcal{N}_{j}^{R},
\label{eq:sufficiency-channel-decompose}$$ where $\left\{ \mathcal{U}_{j}:\mathcal{B}(\mathcal{H}_{L_{j}})\rightarrow
\mathcal{B}(\mathcal{K}_{L_{j}})\right\} $ is a set of unitary channels and $\{\mathcal{N}_{j}^{R}:\mathcal{B}(\mathcal{H}_{R_{j}})\mapsto\mathcal{B}(\mathcal{K}_{R_{j}})\}$ is a set of quantum channels. Furthermore, with respect to the decomposition in , the adjoint of $\mathcal{N}$ acts as$$\mathcal{N}^{\dag}=\bigoplus\limits_{j}\mathcal{U}_{j}^{\dag}\otimes\left(
\mathcal{N}_{j}^{R}\right) ^{\dag}.$$
Trace inequalities
------------------
The following lemma is a consequence of [@epstein1973] (see also [@Carlen2008 Theorem 1.1]):
\[lem:concave-norm\]For $A\in\mathcal{B}( \mathcal{H}) $, $B\in
\mathcal{B}( \mathcal{H}) _{+}$, and $p\in\left( 0,1\right) $, the map $B\mapsto\operatorname{Tr}\{\left( AB^{p}A^{\dag}\right) ^{1/p}\}$ is concave. For invertible $A\in\mathcal{B}\left( \mathcal{H}\right) $, $B\in\mathcal{B}( \mathcal{H}) _{++}$, and $p\in\left( -1,0\right) $, the map $B\mapsto\operatorname{Tr}\{\left( AB^{p}A^{\dag}\right) ^{1/p}\}$ is concave.
\[lem:trace-ineq-gen\]Let $\rho\in\mathcal{S}(\mathcal{H})$, $\sigma
\in\mathcal{B}(\mathcal{H})_{+}$, and let $\mathcal{N}:\mathcal{B}(\mathcal{H})\mapsto\mathcal{B}(\mathcal{K})$ be a CPTP map. For $\alpha
\in\left( 0,1\right) $$$\operatorname{Tr}\left\{ \left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right)
\sigma^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}\right\} \leq1.
\label{eq:trace-ineq-rel-ent-diff}$$ For $\alpha\in\left( 1/2,1\right) $$$\operatorname{Tr}\left\{ \left[ \sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac
{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\right]
^{\frac{\alpha}{1-\alpha}}\right\} \leq1. \label{eq:other-trace-ineq-petz}$$ Let $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and let $\mathcal{N}:\mathcal{B}(\mathcal{H})\mapsto\mathcal{B}(\mathcal{K})$ be a CPTP map such that $\mathcal{N}(\rho)\in\mathcal{S}(\mathcal{K})_{++}$, $\mathcal{N}(\sigma)\in
\mathcal{B}(\mathcal{K})_{++}$. For these choices, the inequality in holds if $\alpha\in(1,2)$, and the inequality in holds if $\alpha\in(1,\infty)$.
We begin by proving (\[eq:trace-ineq-rel-ent-diff\]) for $\alpha\in\left(
0,1\right) $. By Stinespring’s dilation theorem [@S55], a given quantum channel $\mathcal{N}$ can be realized as$$\mathcal{N}(\omega)=\text{Tr}_{E^{\prime}}\left\{ U\left( \omega
\otimes|0\rangle\langle0|_{E}\right) U^{\dag}\right\} \quad\forall
\,\,\omega\in\mathcal{B}\left( \mathcal{H}\right) ,$$ for some unitary $U$ taking $\mathcal{H\otimes H}_{E}$ to $\mathcal{K\otimes
H}_{E^{\prime}}$ and fixed state $|0\rangle_{E}$ in an auxiliary Hilbert space$~{\mathcal{H}}_{E}$. Furthermore, it suffices to take $\dim\left(
\mathcal{H}_{E}\right) \leq\dim(\mathcal{H})\dim\left( \mathcal{K}\right) $ because the number of Kraus operators for the channel $\mathcal{N}$ can always be taken less than or equal to $\dim(\mathcal{H})\dim\left(
\mathcal{K}\right) $ and this is the dimension needed for an environment system $\mathcal{H}_{E}$. The adjoint of $\mathcal{N}$ is given by$$\mathcal{N}^{\dag}\left( \tau\right) =\langle0|_{E}U^{\dag}\left(
\tau\otimes I_{E^{\prime}}\right) U|0\rangle_{E}\quad\forall\,\,\tau
\in\mathcal{B}\left( \mathcal{K}\right) ,$$ Then$$\begin{aligned}
& \text{Tr}\left\{ \left( \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag
}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \sigma^{\frac{1-\alpha}{2}}\right) ^{\frac{1}{1-\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \left( \sigma^{\frac{1-\alpha}{2}}\langle0|_{E}U^{\dag
}\left( \left[ \left( \mathcal{N}\left( \sigma\right) ^{\frac{\alpha
-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \right] \otimes I_{E^{\prime}}\right) U|0\rangle_{E}\sigma
^{\frac{1-\alpha}{2}}\right) ^{\frac{1}{1-\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \left( AA^{\dag}\right) ^{\frac{1}{1-\alpha}}\right\}\end{aligned}$$ where$$\begin{aligned}
A & =\sigma^{\frac{1-\alpha}{2}}\langle0|_{E}U^{\dag}K_{\alpha}^{\dag},\\
K_{\alpha} & \equiv K_{\alpha}(\rho,\sigma,\mathcal{N})\equiv\mathcal{N}(\rho)^{\frac{1-\alpha}{2}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\otimes
I_{E^{\prime}}. \label{eq:k_alpha}$$ Then the above is equal to$$\text{Tr}\left\{ \left( A^{\dag}A\right) ^{\frac{1}{1-\alpha}}\right\}
=\text{Tr}\left\{ \left( K_{\alpha}U|0\rangle_{E}\sigma^{\frac{1-\alpha}{2}}\sigma^{\frac{1-\alpha}{2}}\langle0|_{E}U^{\dag}K_{\alpha}^{\dag}\right)
^{\frac{1}{1-\alpha}}\right\} , \label{eq:connect}$$ because the eigenvalues of $AA^{\dag}$ and $A^{\dag}A$ are the same for any operator $A$. Using that$$\begin{aligned}
U|0\rangle_{E}\sigma^{\frac{1-\alpha}{2}}\sigma^{\frac{1-\alpha}{2}}\langle0|_{E}U^{\dag} & =U\left( \sigma^{1-\alpha}\otimes|0\rangle
\langle0|_{E}\right) U^{\dag}\nonumber\\
& =U\left( \sigma\otimes|0\rangle\langle0|_{E}\right) ^{1-\alpha}U^{\dag
}\nonumber\\
& =\left[ U\left( \sigma\otimes|0\rangle\langle0|_{E}\right) U^{\dag
}\right] ^{1-\alpha},\end{aligned}$$ the right hand side of (\[eq:connect\]) is equal to$$\text{Tr}\left\{ \left( K_{\alpha}\left[ U\left( \sigma\otimes
|0\rangle\langle0|_{E}\right) U^{\dag}\right] ^{1-\alpha}K_{\alpha}^{\dag
}\right) ^{\frac{1}{1-\alpha}}\right\} . \label{eq:connect-up-2}$$ Let $\left\{ U_{E^{\prime}}^{i}\right\} $ be a set of Heisenberg-Weyl operators for the $E^{\prime}$ system and let $\pi_{E^{\prime}}$ denote the maximally mixed state on system $E^{\prime}$. Now we use Lemma \[lem:concave-norm\] to establish the inequality below:$$\begin{aligned}
& \text{Tr}\left\{ \left( K_{\alpha}\left[ U\left( \sigma\otimes
|0\rangle\langle0|_{E}\right) U^{\dag}\right] ^{1-\alpha}K_{\alpha}^{\dag
}\right) ^{\frac{1}{1-\alpha}}\right\} \nonumber\\
& =\frac{1}{d_{E^{\prime}}^{2}}\sum_{i}\text{Tr}\left\{ \left( K_{\alpha
}\left[ U_{E^{\prime}}^{i}U\left( \sigma\otimes|0\rangle\langle
0|_{E}\right) U^{\dag}\left( U_{E^{\prime}}^{i}\right) ^{\dag}\right]
^{1-\alpha}K_{\alpha}^{\dag}\right) ^{\frac{1}{1-\alpha}}\right\} \\
& \leq\text{Tr}\left\{ \left( K_{\alpha}\left[ \frac{1}{d_{E^{\prime}}^{2}}\sum_{i}U_{E^{\prime}}^{i}U\left( \sigma\otimes|0\rangle\langle
0|_{E}\right) U^{\dag}\left( U_{E^{\prime}}^{i}\right) ^{\dag}\right]
^{1-\alpha}K_{\alpha}^{\dag}\right) ^{\frac{1}{1-\alpha}}\right\} \\
& =\text{Tr}\left\{ \left( K_{\alpha}\left[ \text{Tr}_{E^{\prime}}\left\{
U\left( \sigma\otimes|0\rangle\langle0|_{E}\right) U^{\dag}\right\}
\otimes\pi_{E^{\prime}}\right] ^{1-\alpha}K_{\alpha}^{\dag}\right)
^{\frac{1}{1-\alpha}}\right\} .\end{aligned}$$ Continuing, the last line above is equal to$$\begin{aligned}
& \text{Tr}\left\{ \left( K_{\alpha}\left[ \mathcal{N}(\sigma)\otimes
\pi_{E^{\prime}}\right] ^{1-\alpha}K_{\alpha}^{\dag}\right) ^{\frac
{1}{1-\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \left( K_{\alpha}\left[ \mathcal{N}(\sigma)^{1-\alpha
}\otimes\pi_{E^{\prime}}^{1-\alpha}\right] K_{\alpha}^{\dag}\right)
^{\frac{1}{1-\alpha}}\right\} \label{eq:identities-1}\\
& =\text{Tr}\left\{ \left( \mathcal{N}(\rho)^{\frac{1-\alpha}{2}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\sigma)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{\frac{1-\alpha}{2}}\otimes\pi_{E^{\prime}}^{1-\alpha}\right) ^{\frac{1}{1-\alpha}}\right\}
\\
& =\text{Tr}\left\{ \left( \mathcal{N}(\rho)^{\frac{1-\alpha}{2}}[\mathcal{N}(\sigma)]^{0}\mathcal{N}(\rho)^{\frac{1-\alpha}{2}}\otimes
\pi_{E^{\prime}}^{1-\alpha}\right) ^{\frac{1}{1-\alpha}}\right\} \\
& \leq\text{Tr}\left\{ \left( \mathcal{N}(\rho)^{1-\alpha}\otimes
\pi_{E^{\prime}}^{1-\alpha}\right) ^{\frac{1}{1-\alpha}}\right\} \\
& =1, \label{eq:identities-2}$$ where the inequality follows because$$\mathcal{N}(\rho)^{\frac{1-\alpha}{2}}\left[ \mathcal{N}(\sigma)\right]
^{0}\mathcal{N}(\rho)^{\frac{1-\alpha}{2}}\leq\mathcal{N}(\rho)^{1-\alpha},$$ and$$\operatorname{Tr}\left\{ f(A)\right\} \leq\operatorname{Tr}\left\{
f(B)\right\} \label{eq:trace-monotone-1}$$ when $A\leq B$ and $f(x)\equiv x^{1/\left( 1-\alpha\right) }$ is a monotone non-decreasing function on $[0,\infty)$. The other inequality in (\[eq:other-trace-ineq-petz\]) follows by recognizing that$$\frac{1-\alpha}{\alpha}=1-\gamma, \label{eq:gamma-subs}$$ with $\gamma\equiv\left( 2\alpha-1\right) /\alpha$, and $\gamma\in\left(
0,1\right) $ when $\alpha\in(1/2,1)$. Thus, we can apply (\[eq:trace-ineq-rel-ent-diff\]) to conclude (\[eq:other-trace-ineq-petz\]).
To prove the last two statements, we exploit Choi’s inequality [@choi1974 Corollary 2.3] (see also [@B07 Theorem 2.3.6]), which states that$$\mathcal{M}(A)^{-1}\leq\mathcal{M}(A^{-1}),$$ for $\mathcal{M}$ a strictly positive map and $A\in\mathcal{B}(\mathcal{H})_{++}$. Consider that the adjoint $\mathcal{N}^{\dag}$ of a channel $\mathcal{N}$ is strictly positive because it is unital (recall that a map $\mathcal{M}$ is strictly positive if and only if $\mathcal{M}(I)>0$, which is the case if $\mathcal{M}$ is unital). So let $\alpha\in(1,2)$ and consider that$$\begin{gathered}
\operatorname{Tr}\left\{ \left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right)
\sigma^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}\right\} \\
=\operatorname{Tr}\left\{ \left[ \sigma^{-\frac{1-\beta}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\mathcal{N}(\rho)^{-\left( 1-\beta\right) }\mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\right) \sigma^{-\frac{1-\beta}{2}}\right] ^{-\frac{1}{1-\beta}}\right\} ,\end{gathered}$$ for $\beta=2-\alpha\in\left( 0,1\right) $. Then$$\begin{aligned}
& \left[ \sigma^{-\frac{1-\beta}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\mathcal{N}(\rho)^{-\left( 1-\beta\right)
}\mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\right) \sigma^{-\frac{1-\beta}{2}}\right] ^{-1}\nonumber\\
& =\sigma^{\frac{1-\beta}{2}}\left[ \mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\mathcal{N}(\rho)^{-\left( 1-\beta\right)
}\mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\right) \right] ^{-1}\sigma
^{\frac{1-\beta}{2}}\\
& \leq\sigma^{\frac{1-\beta}{2}}\mathcal{N}^{\dag}\left( \left[
\mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\mathcal{N}(\rho)^{-\left(
1-\beta\right) }\mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\right]
^{-1}\right) \sigma^{\frac{1-\beta}{2}}\\
& =\sigma^{\frac{1-\beta}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\beta-1}{2}}\mathcal{N}(\rho)^{1-\beta}\mathcal{N}(\sigma)^{\frac{\beta-1}{2}}\right) \sigma^{\frac{1-\beta}{2}},\end{aligned}$$ where the equalities follow from the assumptions that $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, $\mathcal{N}(\rho)\in\mathcal{S}(\mathcal{K})_{++}$, and $\mathcal{N}(\sigma
)\in\mathcal{B}(\mathcal{K})_{++}$, and the inequality is an application of Choi’s inequality. By applying and the result above, we find that$$\begin{gathered}
\operatorname{Tr}\left\{ \left[ \left[ \sigma^{-\frac{1-\beta}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{-\frac{\beta-1}{2}}\mathcal{N}(\rho)^{-\left( 1-\beta\right) }\mathcal{N}(\sigma)^{-\frac
{\beta-1}{2}}\right) \sigma^{-\frac{1-\beta}{2}}\right] ^{-1}\right]
^{\frac{1}{1-\beta}}\right\} \\
\leq\operatorname{Tr}\left\{ \left[ \sigma^{\frac{1-\beta}{2}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\beta-1}{2}}\mathcal{N}(\rho)^{1-\beta}\mathcal{N}(\sigma)^{\frac{\beta-1}{2}}\right)
\sigma^{\frac{1-\beta}{2}}\right] ^{\frac{1}{1-\beta}}\right\} \leq1,\end{gathered}$$ where the last inequality is a consequence of what we have previously shown. The other inequality (\[eq:other-trace-ineq-petz\]) for $\alpha\in\left(
1,\infty\right) $ follows from (\[eq:trace-ineq-rel-ent-diff\]) for $\alpha\in\left( 1,2\right) $ by making the same identification as in (\[eq:gamma-subs\]): note here that $\gamma\in(1,2)$ if $\alpha\in
(1,\infty)$.
As a corollary, we establish the following trace inequality, which was left as an open question in [@R02]:
For $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and a strict CPTP map $\mathcal{N}$ acting on $\mathcal{S}(\mathcal{H})$, the following inequality holds:$$\operatorname{Tr}\left\{ \exp\left\{ \log\sigma+\mathcal{N}^{\dag}\left(
\log\mathcal{N}( \rho) -\log\mathcal{N}( \sigma) \right) \right\} \right\}
\leq1.$$
This follows by taking the limit $\alpha\nearrow1$ in Lemma \[lem:trace-ineq-gen\] and using [@SBW14 Lemma 24]:$$\begin{gathered}
\exp\left\{ \log\sigma+\mathcal{N}^{\dag}\left( \log\mathcal{N}(\rho
)-\log\mathcal{N}(\sigma)\right) \right\} \\
=\lim_{\alpha\nearrow1}\left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag
}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \sigma^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}.\end{gathered}$$ The inequality in Lemma \[lem:trace-ineq-gen\] is preserved in the limit.
We can also solve an open question from [@Z14b] for some values of $\alpha$, simply by applying Lemma \[lem:trace-ineq-gen\] for the choices $\rho=\rho_{ABC}$, $\sigma=\rho_{AC}\otimes I_{B}$, and $\mathcal{N}=\operatorname{Tr}_{A}$:
\[cor:trace-ineq\]Let $\rho_{ABC}\in\mathcal{S}\left( \mathcal{H}_{ABC}\right) $. If $\alpha\in\left( 0,1\right) $, then$$\operatorname{Tr}\left\{ \left( \rho_{AC}^{\frac{1-\alpha}{2}}\rho
_{C}^{\frac{\alpha-1}{2}}\rho_{BC}^{1-\alpha}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{AC}^{\frac{1-\alpha}{2}}\right) ^{\frac{1}{1-\alpha}}\right\} \leq1.
\label{eq:1st-trace}$$ If $\alpha\in\left( 1/2,1\right) $, then$$\operatorname{Tr}\left\{ \left( \rho_{AC}^{\frac{1-\alpha}{2\alpha}}\rho
_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{BC}^{\frac{1-\alpha}{\alpha}}\rho
_{C}^{\frac{\alpha-1}{2\alpha}}\rho_{AC}^{\frac{1-\alpha}{2\alpha}}\right)
^{\frac{\alpha}{1-\alpha}}\right\} \leq1. \label{eq:2nd-trace}$$ If $\rho_{ABC}\in\mathcal{S}\left( \mathcal{H}_{ABC}\right) _{++}$, then holds for $\alpha\in\left( 1,2\right) $ and holds for $\alpha\in\left( 1,\infty\right) $.
Let $\rho_{ABC}\in\mathcal{S}\left( \mathcal{H}_{ABC}\right) _{++}$. The well-known inequality [@LR73]$$\operatorname{Tr}\left\{ \exp\left\{ \log\rho_{AC}+\log\rho_{BC}-\log
\rho_{C}\right\} \right\} \leq1$$ follows from Corollary \[cor:trace-ineq\] by taking the limit $\alpha
\rightarrow1$ and using the generalized Lie-Trotter product formula [@S85]:$$\exp\left\{ \log\rho_{AC}+\log\rho_{BC}-\log\rho_{C}\right\} =\lim
_{\alpha\rightarrow1}\left( \rho_{AC}^{\frac{1-\alpha}{2}}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{BC}^{1-\alpha}\rho_{C}^{\frac{\alpha-1}{2}}\rho_{AC}^{\frac{1-\alpha}{2}}\right) ^{\frac{1}{1-\alpha}}.$$
The following proposition establishes some important properties of the $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})$ and $\widetilde{\Delta}_{\alpha
}(\rho,\sigma,\mathcal{N})$ quantities, which were left as open questions in [@SBW14]:
\[lem:delta-positive\]Let $\rho\in\mathcal{S}(\mathcal{H})$, $\sigma
\in\mathcal{B}(\mathcal{H})_{+}$, and let $\mathcal{N}$ be a CPTP map. For all $\alpha\in\left( 0,1\right) $$$\Delta_{\alpha}(\rho,\sigma,\mathcal{N})\geq0, \label{delta-non-neg}$$ with equality occurring if and only if$$\rho=\left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \sigma^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}. \label{eq:Delta=0}$$ For all $\alpha\in\left( 1/2,1\right) $$$\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})\geq0,
\label{tilde-delta-non-neg}$$ with equality occurring if and only if$$\rho=\left[ \sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{\frac
{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\right)
\sigma^{\frac{1-\alpha}{2\alpha}}\right] ^{\frac{\alpha}{1-\alpha}}.
\label{eq:tilde-Delta=0}$$ Let $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and let $\mathcal{N}:\mathcal{B}(\mathcal{H})\mapsto\mathcal{B}(\mathcal{K})$ be a CPTP map such that $\mathcal{N}(\rho)\in\mathcal{S}(\mathcal{K})_{++}$, $\mathcal{N}(\sigma)\in
\mathcal{B}(\mathcal{K})_{++}$. For these choices, the inequality in and the equality condition in hold if $\alpha\in(1,2)$, and the inequality in and equality condition in hold if $\alpha\in(1,\infty)$.
Note that the quantities $\Delta_{\alpha}$ and $\widetilde{\Delta}_{\alpha}$ defined in (\[eq:Renyi-rel-ent-diff\]) and (\[eq:Renyi-rel-ent-diff-sand\]), respectively, can be expressed in terms of the $\alpha$-Rényi relative entropy and the $\alpha$-sandwiched Rényi relative entropy as follows [@SBW14]:$$\begin{aligned}
\Delta_{\alpha}(\rho,\sigma,\mathcal{N}) & =D_{\alpha}\left( \rho
\middle\Vert\left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left( \rho\right)
^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \sigma
^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}\right)
,\label{eq:ren-rel-ent-diff-1}\\
\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N}) & =\widetilde
{D}_{\alpha}\left( \rho\middle\Vert\left[ \sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N(}\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N(}\sigma)^{\frac
{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\right]
^{\frac{\alpha}{1-\alpha}}\right) . \label{eq:ren-rel-ent-diff-2}$$ The non-negativity and equality conditions of $\Delta_{\alpha}(\rho
,\sigma,\mathcal{N})$ for $\alpha\in(0,1)\cup(1,2)$, and that of $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})$ for $\alpha
\in(1/2,1)\cup(1,\infty)$, then follow by applying Lemmas \[lem:min-max-0\] and \[lem:trace-ineq-gen\].
Appendix \[app:alt-proofs\] gives alternate proofs of the inequality in for $\alpha\in(1,2)$ and the inequality in for $\alpha\in(1,\infty)$. These were the original proofs appearing in a preprint version of this paper and might be of independent interest. Appendix \[app:alt-proofs\] gives some other equality conditions.
Sufficiency of quantum channels and Rényi generalizations of relative entropy differences
=========================================================================================
\[thm:channel-suff-result\]Let $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and let $\mathcal{N}$ be a strict CPTP map. Then $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=0$ and $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for all $\alpha
\in\left( 0,1\right) \cup\left( 1,\infty\right) $ if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$. Furthermore, $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$ if $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for some $\alpha\in\left( 0,1\right) \cup\left( 1,2\right) $ or if $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})$ for some $\alpha
\in\left( 1/2,1\right) \cup\left( 1,\infty\right) $.
We begin by proving that $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for all $\alpha\in\left( 0,1\right) \cup\left( 1,\infty\right) $ if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$. So suppose that $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$. Then according to Theorem$~$\[thm:mos-petz-decomp\], this implies that the decompositions in (\[eq:sufficiency-states-decompose\])-(\[eq:sufficiency-channel-decompose\]) hold. To simplify things, we exploit the direct sum structure in (\[eq:sufficiency-states-decompose\])-(\[eq:sufficiency-channel-decompose\]) and first evaluate the contribution arising from the $j^{\text{th}}$ block, for any $j$. Then for a given block, evaluating the formula$$\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \sigma^{\frac{1-\alpha}{2}}$$ gives the following (where we suppress the index $j$, for simplicity)$$\begin{gathered}
\left( p\rho_{L}\otimes\tau_{R}\right) ^{\alpha}\left( q\sigma_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2}}\times\nonumber\\
\left[ \mathcal{U}^{\dag}\otimes\left( \mathcal{N}^{R}\right) ^{\dag
}\right] \left( \left( q\mathcal{U}(\sigma_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{\frac{\alpha-1}{2}}\left( p\mathcal{U}(\rho_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{1-\alpha}\left( q\mathcal{U}(\sigma_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{\frac{\alpha-1}{2}}\right) \times\\
\left( q\sigma_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2}}$$$$\begin{gathered}
=p\left( \rho_{L}\otimes\tau_{R}\right) ^{\alpha}\left( \sigma_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2}}\times\\
\left[ \mathcal{U}^{\dag}\otimes\left( \mathcal{N}^{R}\right) ^{\dag
}\right] \left( \left( \mathcal{U}(\sigma_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{\frac{\alpha-1}{2}}\left( \mathcal{U}(\rho_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{1-\alpha}\left( \mathcal{U}(\sigma_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{\frac{\alpha-1}{2}}\right) \times\\
\left( \sigma_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2}}$$$$\begin{gathered}
=p\left( (\rho_{L})^{\alpha}(\sigma_{L})^{\frac{1-\alpha}{2}}\otimes\tau
_{R}^{\left( \alpha+1\right) /2}\right) \times\\
\left[ \mathcal{U}^{\dag}\otimes\left( \mathcal{N}^{R}\right) ^{\dag
}\right] \left( \left( \mathcal{U}(\sigma_{L})^{\frac{\alpha-1}{2}}\mathcal{U}(\rho_{L})^{1-\alpha}\mathcal{U}(\sigma_{L})^{\frac{\alpha-1}{2}}\otimes\mathcal{N}^{R}(\tau_{R})^{\frac{\alpha-1}{2}}\mathcal{N}^{R}(\tau
_{R})^{1-\alpha}\mathcal{N}^{R}(\tau_{R})^{\frac{\alpha-1}{2}}\right)
\right) \times\\
\left( \sigma_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2}}.\end{gathered}$$ Continuing, the last line is equal to$$\begin{aligned}
& p\left( (\rho_{L})^{\alpha}(\sigma_{L})^{\frac{1-\alpha}{2}}\otimes
\tau_{R}^{\left( \alpha+1\right) /2}\right) \left[ \mathcal{U}^{\dag
}\otimes\left( \mathcal{N}^{R}\right) ^{\dag}\right] \left( \mathcal{U}\left[ (\sigma_{L})^{\frac{\alpha-1}{2}}(\rho_{L})^{1-\alpha}(\sigma
_{L})^{\frac{\alpha-1}{2}}\right] \otimes I_{R}\right) \left( \sigma
_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2}}\nonumber\\
& =p\left( (\rho_{L})^{\alpha}(\sigma_{L})^{\frac{1-\alpha}{2}}\otimes
\tau_{R}^{\left( \alpha+1\right) /2}\right) \left( \left[ (\sigma
_{L})^{\frac{\alpha-1}{2}}(\rho_{L})^{1-\alpha}(\sigma_{L})^{\frac{\alpha
-1}{2}}\right] \otimes I_{R}\right) \left( \sigma_{L}\otimes\tau
_{R}\right) ^{\frac{1-\alpha}{2}}\nonumber\\
& =p\left( (\rho_{L})^{\alpha}(\sigma_{L})^{\frac{1-\alpha}{2}}(\sigma
_{L})^{\frac{\alpha-1}{2}}(\rho_{L})^{1-\alpha}(\sigma_{L})^{\frac{\alpha
-1}{2}}(\sigma_{L})^{\frac{1-\alpha}{2}}\otimes\tau_{R}^{\left(
\alpha+1\right) /2}\tau_{R}^{\frac{1-\alpha}{2}}\right) \nonumber\\
& =p\left( (\rho_{L})^{\alpha}(\rho_{L})^{1-\alpha}\otimes\tau_{R}\right)
\nonumber\\
& =p\rho_{L}\otimes\tau_{R}.\end{aligned}$$ Taking the trace gives $p$ and thus we find that each block has trace equal to $p\left( j\right) $, so that$$\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=\frac{1}{\alpha-1}\log\sum_{j}p(j)=0.$$ So this proves that $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for all $\alpha\in\left( 0,1\right) \cup\left( 1,\infty\right) $ if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$.
We now prove that $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for all $\alpha\in\left( 0,1\right) \cup\left( 1,\infty\right) $ if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$. As above, the decompositions in (\[eq:sufficiency-states-decompose\])-(\[eq:sufficiency-channel-decompose\]) hold and we exploit this direct sum structure again. As in the previous proof, it suffices to evaluate the contribution arising from each block $j$, for any $j$, in the direct-sum decomposition. Evaluating the formula$$\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\right) \sigma^{\frac
{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}},$$ for the $j$th block, (where we once again suppress the index $j$) gives$$\begin{gathered}
\left( p\rho_{L}\otimes\tau_{R}\right) ^{\frac{1}{2}}\left( q\sigma
_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2\alpha}}\times\\
\left[ \mathcal{U}^{\dag}\otimes\left( \mathcal{N}^{R}\right) ^{\dag
}\right] \left( \left( q\mathcal{U}(\sigma_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{\frac{\alpha-1}{2\alpha}}\left( p\mathcal{U}(\rho
_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right) ^{\frac{1-\alpha}{\alpha}}\left( q\mathcal{U}(\sigma_{L})\otimes\mathcal{N}^{R}(\tau_{R})\right)
^{\frac{\alpha-1}{2\alpha}}\right) \times\\
\left( q\sigma_{L}\otimes\tau_{R}\right) ^{\frac{1-\alpha}{2\alpha}}\left(
p\rho_{L}\otimes\tau_{R}\right) ^{\frac{1}{2}}. \label{eq:jth-block-starting}$$ Then proceeding very similarly as before, we conclude that (\[eq:jth-block-starting\]) is equal to$$p^{\frac{1}{\alpha}}\rho_{L}^{\frac{1}{\alpha}}\otimes\tau_{R}^{\frac
{1}{\alpha}}.$$ We evaluate evaluate the norm $\left\Vert \cdot\right\Vert _{\alpha}^{\alpha}$ of the $j$th block to be$$\left\Vert p^{\frac{1}{\alpha}}\rho_{L}^{\frac{1}{\alpha}}\otimes\tau
_{R}^{\frac{1}{\alpha}}\right\Vert _{\alpha}^{\alpha}=p.$$ As a consequence, we find that$$\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})=\frac{1}{\alpha-1}\log\left( \sum_{j}p(j)\right) =0,$$ where we used that $\left\Vert A\oplus B\right\Vert _{\alpha}^{\alpha
}=\left\Vert A\right\Vert _{\alpha}^{\alpha}+\left\Vert B\right\Vert _{\alpha
}^{\alpha}$.
Now suppose that $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for some $\alpha\in\left( 0,1\right) \cup(1,2)$. From Proposition \[lem:delta-positive\], we have that$$\rho=\left[ \sigma^{\frac{1-\alpha}{2}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha
}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right) \sigma^{\frac{1-\alpha}{2}}\right] ^{\frac{1}{1-\alpha}}, \label{eq:step-ren-rel-ent-1}$$ which is equivalent to$$\sigma^{\frac{\alpha-1}{2}}\rho^{1-\alpha}\sigma^{\frac{\alpha-1}{2}}=\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right)
.$$ Multiply both sides by $\sigma$ and take the trace to get$$\begin{aligned}
\text{Tr}\left\{ \sigma\sigma^{\frac{\alpha-1}{2}}\rho^{1-\alpha}\sigma^{\frac{\alpha-1}{2}}\right\} & =\text{Tr}\left\{ \sigma
\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right)
\right\} \nonumber\\
\Leftrightarrow\text{Tr}\left\{ \rho^{1-\alpha}\sigma^{\alpha}\right\} &
=\text{Tr}\left\{ \mathcal{N}(\sigma)\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right\}
\nonumber\\
\Leftrightarrow\text{Tr}\left\{ \rho^{1-\alpha}\sigma^{\alpha}\right\} &
=\text{Tr}\left\{ \mathcal{N}(\rho)^{1-\alpha}\mathcal{N}(\sigma)^{\alpha
}\right\} \nonumber\\
\Leftrightarrow D_{\alpha}(\sigma\Vert\rho) & =D_{\alpha}(\mathcal{N}(\sigma)\Vert\mathcal{N}(\rho)). \label{eq:step-ren-rel-ent-last}$$ The last line is an equality of $f$-divergences (see Remark \[rem:renyi-f-div\]), which by Lemma \[lem-equiv\] allows us to conclude that $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$.
Suppose that $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for some $\alpha\in\left( 1/2,1\right) \cup(1,\infty)$. From Proposition \[lem:delta-positive\], we have that$$\rho=\left[ \sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{\frac
{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\right)
\sigma^{\frac{1-\alpha}{2\alpha}}\right] ^{\frac{\alpha}{1-\alpha}},$$ which can be rewritten as$$\rho=\left[ \sigma^{\frac{1-\gamma}{2}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\gamma-1}{2}}\mathcal{N}(\rho)^{1-\gamma
}\mathcal{N}(\sigma)^{\frac{\gamma-1}{2}}\right) \sigma^{\frac{1-\gamma}{2}}\right] ^{\frac{1}{1-\gamma}},$$ with $\gamma$ defined in (\[eq:gamma-subs\]), so that $\gamma\in
(0,1)\cup(1,2)$. By the development in -, we can conclude that $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$.
The following statement is a direct corollary of the correspondence in - and Theorem \[thm:channel-suff-result\]:
\[thm:Renyi-cmi-markov\]Let $\rho_{ABC}\in\mathcal{S}\left(
\mathcal{H}_{ABC}\right) _{++}$. Then $I_{\alpha}(A;B|C)_{\rho}=0$ and $\widetilde{I}_{\alpha}(A;B|C)_{\rho}=0$ for all $\alpha\in(0,1)\cup
(1,\infty)$ if $\rho_{ABC}$ is a short quantum Markov chain $A-C-B$. Furthermore, $\rho_{ABC}$ is a short quantum Markov chain $A-C-B\mathcal{\ }$if $I_{\alpha}(A;B|C)_{\rho}=0$ for some $\alpha\in\left( 0,1\right)
\cup\left( 1,2\right) $ or if $\widetilde{I}_{\alpha}(A;B|C)_{\rho}=0$ for some $\alpha\in\left( 1/2,1\right) \cup\left( 1,\infty\right) $.
Quantum Markov chains, sufficiency of quantum channels, and min- and max-information measures {#sec:min-max}
=============================================================================================
Min- and max- generalizations of a relative entropy difference
--------------------------------------------------------------
We consider the following generalizations of a relative entropy difference, motivated by the developments in [@BSW14; @SBW14].
\[def:CMI-quantities\]Let $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and let $\mathcal{N}$ be a strict CPTP map. Then, $$\Delta_{\min}(\rho,\sigma,\mathcal{N})\equiv D_{\min}(\rho\Vert\mathcal{R}_{\sigma,\mathcal{N}}(\mathcal{N(}\rho))),$$ where $\mathcal{R}_{\sigma,\mathcal{N}}$ is the Petz recovery channel defined in and $$\Delta_{\max}(\rho,\sigma,\mathcal{N})\equiv D_{\max}(\rho\Vert\mathcal{R}_{\sigma,\mathcal{N}}(\mathcal{N(}\rho))).$$
\[thm:renyi-rel-ent-min-max-equality\]For $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and a strict CPTP map $\mathcal{N}$, $$\Delta_{\min}(\rho,\sigma,\mathcal{N})\geq0,\ \ \ \ \quad\Delta_{\max}(\rho,\sigma,\mathcal{N})\geq0,$$ with equality holding if and only if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$.
The non-negativity conditions follow from the fact that $\mathcal{R}_{\sigma,\mathcal{N}}(\mathcal{N(}\rho))$ is a density operator (since $\mathcal{R}_{\sigma,\mathcal{N}}$ is a CPTP map) and Lemma \[lem:min-max-0\]. The equality conditions also follow from Lemma \[lem:min-max-0\] and the fact that $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$ if and only if $\mathcal{R}_{\sigma,\mathcal{N}}(\mathcal{N(}\rho))=\rho$.
Min- and max- generalizations of conditional mutual information
---------------------------------------------------------------
\[[@BSW14]\]For a tripartite state $\rho_{ABC}\in\mathcal{S}\left(
\mathcal{H}_{ABC}\right) _{++}$ the max-conditional mutual information is defined as follows:$$\begin{aligned}
I_{\max}(A;B|C)_{\rho} & \equiv D_{\max}\left( \rho_{ABC}\middle\Vert
\rho_{AC}^{\frac{1}{2}}\rho_{C}^{-\frac{1}{2}}\rho_{BC}\rho_{C}^{-\frac{1}{2}}\rho_{AC}^{\frac{1}{2}}\right) \label{imax1}\\
& =\inf\left\{ \lambda:\,\rho_{ABC}\leq\exp(\lambda)\rho_{AC}^{\frac{1}{2}}\rho_{C}^{-\frac{1}{2}}\rho_{BC}\rho_{C}^{-\frac{1}{2}}\rho_{AC}^{\frac{1}{2}}\right\} \label{imax2}\\
& =2\log\left\Vert \rho_{ABC}^{\frac{1}{2}}\rho_{AC}^{-\frac{1}{2}}\rho
_{C}^{\frac{1}{2}}\rho_{BC}^{-\frac{1}{2}}\right\Vert _{\infty},\label{imax3}$$ and the min-conditional mutual information is defined as follows:$$\begin{aligned}
I_{\min}(A;B|C)_{\rho} & \equiv D_{\min}\left( \rho_{ABC}\middle\Vert
\rho_{AC}^{\frac{1}{2}}\rho_{C}^{-\frac{1}{2}}\rho_{BC}\rho_{C}^{-\frac{1}{2}}\rho_{AC}^{\frac{1}{2}}\right) \label{imin1}\\
& =-\log F\left( \rho_{ABC},\rho_{AC}^{\frac{1}{2}}\rho_{C}^{-\frac{1}{2}}\rho_{BC}\rho_{C}^{-\frac{1}{2}}\rho_{AC}^{\frac{1}{2}}\right)
\label{imin2}\\
& =-2\log\left\Vert \rho_{ABC}^{\frac{1}{2}}\rho_{AC}^{\frac{1}{2}}\rho
_{C}^{-\frac{1}{2}}\rho_{BC}^{\frac{1}{2}}\right\Vert _{1}.\label{imin3}$$
As observed in [@BSW14], we have that$$I_{\max}(A;B|C)_{\rho},I_{\min}(A;B|C)_{\rho}\geq0, \label{eq:I_min_max_geq0}$$ due to Lemma \[lem:min-max-0\] and the fact that Tr$\{\rho_{AC}^{\frac{1}{2}}\rho_{C}^{-\frac{1}{2}}\rho_{BC}\rho_{C}^{-\frac{1}{2}}\rho_{AC}^{\frac
{1}{2}}\}=1$. These quantities are special cases of those from Definition \[def:CMI-quantities\], by using (\[eq:rel-ent-cmi-1\]). As such, the following is a corollary of Theorem \[thm:renyi-rel-ent-min-max-equality\]:
\[thm-imax-imin\] Let $\rho_{ABC}\in\mathcal{S}\left( \mathcal{H}_{ABC}\right) _{++}$. Then each of the following identities hold if and only if $\rho_{ABC}$ is a short quantum Markov chain $A-C-B$: $$\begin{aligned}
I_{\max}(A;B|C) & =0,\label{imax-a}\\
I_{\min}(A;B|C) & =0. \label{imin-b}$$
Conclusion and open questions
=============================
We have shown that the $\alpha$-Rényi quantities $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})$ and $\widetilde{\Delta}_{\alpha}(\rho
,\sigma,\mathcal{N})$ from [@SBW14] are non-negative and equal to zero if and only if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$. As a consequence, we find that the $\alpha$-Rényi conditional mutual informations $I_{\alpha}(A;B|C)_{\rho}$ and $\widetilde{I}_{\alpha
}(A;B|C)_{\rho}$ from [@BSW14] are equal to zero if and only if $\rho_{ABC}$ is a short quantum Markov chain $A-C-B$. Moreover, we have solved some open questions from [@R02; @Z14b].
There are some interesting open questions to consider going forward from here. We would like to know the equality conditions for monotonicity of the sandwiched Rényi relative entropies, i.e., for which triples $(\rho
,\sigma,\mathcal{N})$ is it true that$$\widetilde{D}_{\alpha}(\rho\Vert\sigma)=\widetilde{D}_{\alpha}\left(
\mathcal{N}(\rho)\Vert\mathcal{N}(\sigma)\right) \ ?$$ Apparently we cannot solve this question using the methods of [@HMPB11] because we cannot represent the sandwiched Rényi relative entropy as an $f$-divergence. Presumably, this equality occurs if and only if $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$, but it remains to be proved. Next, is there a characterization of states for which $I_{\alpha}(A;B|C)_{\rho}$ and $\widetilde{I}_{\alpha}(A;B|C)_{\rho}$ are nearly equal to zero? Also, is there a characterization of triples $(\rho,\sigma,\mathcal{N})$ for which $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})$ and $\widetilde{\Delta}_{\alpha
}(\rho,\sigma,\mathcal{N})$ are nearly equal to zero? Presumably the former has to do with $\rho_{ABC}$ being close to the Petz recovered $\rho_{BC}$ and the latter has to do with $\rho$ being close to the Petz recovered $\rho$, as recent developments in [@FR14; @BLW14] might suggest.
**Note**: After the posting of a preprint of this work to the arXiv, there have been improvements of the results detailed here for values of $\alpha<1$ [@W15; @DW15; @JRSWW15], which address some of the open questions mentioned above for these values of $\alpha$. In particular, the results of [@W15] provide stronger lower bounds for $\widetilde{I}_{\alpha
}(A;B|C)_{\rho}$ and $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})$ when $\alpha\in(1/2,1)$. This was further improved in [@JRSWW15]. The results of [@DW15] provide stronger lower bounds for $I_{\alpha
}(A;B|C)_{\rho}$ and $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})$ when $\alpha\in(0,1)$.
**Acknowledgements.** We acknowledge helpful discussions with Mario Berta, Kaushik Seshadreesan, and Marco Tomamichel and thank Perla Sousi for a helpful discussion on Markov chains. We are especially grateful to Milan Mosonyi for his very careful reading of our paper, pointing out a problem with our former justification of Lemma \[lem:trace-ineq-gen\] for $\alpha
\in\left( 1,2\right) $, and for communicating many other observations about our paper that go beyond those stated here. We are indebted to an anonymous referee for suggesting to use Choi’s inequality to prove some of the statements in Lemma \[lem:trace-ineq-gen\]. We acknowledge support from the Peter Whittle Fund, which helped to enable this research. MMW is grateful to the Statistical Laboratory in the Center for Mathematical Sciences at the University of Cambridge for hosting him for a research visit during January 2015. MMW acknowledges support from startup funds from the Department of Physics and Astronomy at LSU, the NSF under Award No. CCF-1350397, and the DARPA Quiness Program through US Army Research Office award W31P4Q-12-1-0019.
Alternative Proof of Non-Negativity of $\Delta_{\alpha}$ and $\widetilde{\Delta}_{\alpha}$ quantities {#app:alt-proofs}
=====================================================================================================
Here we provide an alternative proof of the non-negativity of the $\Delta_{\alpha}$ and $\widetilde{\Delta}_{\alpha}$ quantities, which appeared in the original preprint version of our work:
Let $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and let $\mathcal{N}$ be a strict CPTP map. Then holds for $\alpha\in\left( 1,2\right) $ and equality occurs if and only if$$\left[ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left(
\sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\right)
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}=\mathcal{N}(\rho). \label{eq:Delta=0-other-a}$$ Furthermore, for the same choices of $\rho$, $\sigma$, and $\mathcal{N}$, holds for $\alpha\in\left( 1,\infty\right) $.
We first prove the non-negativity of $\Delta_{\alpha}(\rho,\sigma
,\mathcal{N})$ for $\alpha\in(1,2)$, $\rho\in\mathcal{S}(\mathcal{H})_{++}$, $\sigma\in\mathcal{B}(\mathcal{H})_{++}$, and $\mathcal{N}$ a strict CPTP map. Using the definition (\[eq:Renyi-rel-ent-diff\]) of $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})$, cyclicity of trace, and the definition of the adjoint map, we can express $\Delta_{\alpha}(\rho
,\sigma,\mathcal{N})$ as$$\begin{aligned}
\Delta_{\alpha}(\rho,\sigma,\mathcal{N}) & =\frac{1}{\alpha-1}\log
\text{Tr}\left\{ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left(
\sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\right)
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}(\rho)^{1-\alpha}\right\}
\label{eq:Delta-rewrite}\\
& =D_{\alpha}\left( \left[ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma
^{\frac{1-\alpha}{2}}\right) \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right]
^{\frac{1}{\alpha}}\middle\Vert\mathcal{N}(\rho)\right)
\label{eq:Delta-rewrite-1}$$ where$$D_{\alpha}(P\Vert Q)\equiv\frac{1}{\alpha-1}\log\text{Tr}\left\{ P^{\alpha
}Q^{1-\alpha}\right\}$$ is the Rényi relative entropy between two positive definite operators $P$ and $Q$. It is known that $D_{\alpha}(P\Vert Q)\geq0$ if $P$ and $Q$ are such that$$\text{Tr}\left\{ P\right\} \geq\text{Tr}\left\{ Q\right\} =1.$$ This is because $D_{\alpha}(P\Vert Q)$ is monotone under quantum channels for $\alpha\in\left( 1,2\right) $, and one such quantum channel is the trace operation:$$\begin{aligned}
D_{\alpha}(P\Vert Q) & \geq D_{\alpha}\left( \text{Tr}\left\{ P\right\}
\Vert\text{Tr}\left\{ Q\right\} \right) \\
& =\frac{1}{\alpha-1}\log\left[ \left[ \text{Tr}\left\{ P\right\}
\right] ^{\alpha}\left[ \text{Tr}\left\{ Q\right\} \right] ^{1-\alpha
}\right] \\
& \geq0. \label{eq:renyi-non-neg}$$ If $D_{\alpha}(P\Vert Q)=0$ for some $\alpha\in\left( 1,2\right) $ and $P$ and $Q$ such that $\operatorname{Tr}\left\{ P\right\} \geq\operatorname{Tr}\left\{ Q\right\} $, then it is known that $P=Q$ (one can deduce this, e.g., from the above and [@HMPB11 Theorem 5.1]). Hence, to prove the non-negativity of $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})$ for $\alpha
\in\left( 1,2\right) $, it suffices to prove that$$\text{Tr}\left\{ \left[ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma
^{\frac{1-\alpha}{2}}\right) \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right]
^{\frac{1}{\alpha}}\right\} \geq\text{Tr}\left\{ \mathcal{N}(\rho)\right\}
=1.$$ Theorem 1.1 of [@Hiai20131568] establishes that the map$$\left( P,Q\right) \mapsto\text{Tr}\left\{ \left[ Q^{\frac{\alpha-1}{2}}PQ^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\}$$ is jointly concave for positive definite $P$ and $Q$ when $\alpha\in\left(
1,2\right) $. A straightforward argument allows to conclude its joint concavity for $\alpha\in\left( 1,2\right) $ and positive semidefinite $P$ and $Q$. Indeed, let $\varepsilon>0$, $\left\{ P_{x}\right\} $ and $\left\{
Q_{x}\right\} $ be sets of positive semidefinite operators, let $p_{X}(x)$ be a probability distribution, and let $\overline{P}\equiv\sum_{x}p_{X}(x)P_{x}$ and $\overline{Q}\equiv\sum_{x}p_{X}(x)Q_{x}$ . Consider that$$\begin{aligned}
& \sum_{x}p_{X}(x)\text{Tr}\left\{ \left[ Q_{x}^{\frac{\alpha-1}{2}}P_{x}Q_{x}^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\}
\nonumber\\
& \leq\sum_{x}p_{X}(x)\text{Tr}\left\{ \left[ Q_{x}^{\frac{\alpha-1}{2}}\left( P_{x}+\varepsilon I\right) Q_{x}^{\frac{\alpha-1}{2}}\right]
^{\frac{1}{\alpha}}\right\} \nonumber\\
& =\sum_{x}p_{X}(x)\text{Tr}\left\{ \left[ \left( P_{x}+\varepsilon
I\right) ^{\frac{1}{2}}Q_{x}^{\alpha-1}\left( P_{x}+\varepsilon I\right)
^{\frac{1}{2}}\right] ^{\frac{1}{\alpha}}\right\} \nonumber\\
& \leq\sum_{x}p_{X}(x)\text{Tr}\left\{ \left[ \left( P_{x}+\varepsilon
I\right) ^{\frac{1}{2}}\left( Q_{x}+\varepsilon I\right) ^{\alpha-1}\left(
P_{x}+\varepsilon I\right) ^{\frac{1}{2}}\right] ^{\frac{1}{\alpha}}\right\}
\nonumber\\
& =\sum_{x}p_{X}(x)\text{Tr}\left\{ \left[ \left( Q_{x}+\varepsilon
I\right) ^{\frac{\alpha-1}{2}}\left( P_{x}+\varepsilon I\right) \left(
Q_{x}+\varepsilon I\right) ^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\} \nonumber\\
& \leq\text{Tr}\left\{ \left[ \left( \overline{Q}+\varepsilon I\right)
^{\frac{\alpha-1}{2}}\left( \overline{P}+\varepsilon I\right) \left(
\overline{Q}+\varepsilon I\right) ^{\frac{\alpha-1}{2}}\right] ^{\frac
{1}{\alpha}}\right\} .\end{aligned}$$ The first inequality follows because $P_{x}\leq P_{x}+\varepsilon I$ and because $\operatorname{Tr}\left\{ f(A)\right\} \leq\operatorname{Tr}\left\{
f(B)\right\} $ for $A\leq B$ and $f(x)=x^{1/\alpha}$ a monotone non-decreasing function on $[0,\infty)$. The next inequality follows because $x^{\alpha-1}$ is operator monotone for $\alpha\in\left( 1,2\right) $ and for the same reason as above. The final inequality is a consequence of Theorem 1.1 of [@Hiai20131568]. By taking the limit $\varepsilon\searrow
0$, we can conclude that$$\sum_{x}p_{X}(x)\text{Tr}\left\{ \left[ Q_{x}^{\frac{\alpha-1}{2}}P_{x}Q_{x}^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\} \leq
\text{Tr}\left\{ \left[ \overline{Q}^{(\alpha-1)/2}\overline{P}\,\overline{Q}^{(\alpha-1)/2}\right] ^{\frac{1}{\alpha}}\right\} .
\label{eq:Hiai-concave}$$ Now consider the following chain of equalities:$$\begin{aligned}
1 & =\text{Tr}\left\{ \left[ \sigma^{\frac{\alpha-1}{2}}\sigma
^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\sigma
^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \left[ \sigma^{\frac{\alpha-1}{2}}\sigma^{\frac
{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\sigma^{\frac{\alpha
-1}{2}}\otimes|0\rangle\langle0|_{E}\right] ^{\frac{1}{\alpha}}\right\}
\nonumber\\
& =\text{Tr}\left\{ \left[ \left( \sigma^{\frac{\alpha-1}{2}}\otimes|0\rangle\langle0|_{E}\right) \left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\otimes|0\rangle\langle0|_{E}\right)
\left( \sigma^{\frac{\alpha-1}{2}}\otimes|0\rangle\langle0|_{E}\right)
\right] ^{\frac{1}{\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \left[ U\left( \sigma^{\frac{\alpha-1}{2}}\otimes|0\rangle\langle0|_{E}\right) U^{\dag}U\left( \sigma^{\frac{1-\alpha
}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\otimes|0\rangle\langle
0|_{E}\right) U^{\dag}U\left( \sigma^{\frac{\alpha-1}{2}}\otimes
|0\rangle\langle0|_{E}\right) U^{\dag}\right] ^{\frac{1}{\alpha}}\right\}
\nonumber\\
& =\text{Tr}\left\{ \left[ \left[ U\left( \sigma\otimes|0\rangle
\langle0|_{E}\right) U^{\dag}\right] ^{\frac{\alpha-1}{2}}U\left(
\sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\otimes|0\rangle\langle0|_{E}\right) U^{\dag}\left[ U\left( \sigma
\otimes|0\rangle\langle0|_{E}\right) U^{\dag}\right] ^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\} \nonumber\\
& =\frac{1}{d_{E^{\prime}}^{2}}\sum_{i}\text{Tr}\left\{ \left[ \left[
K^{i}\right] ^{\frac{\alpha-1}{2}}U_{E^{\prime}}^{i}U\left( \sigma
^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\otimes
|0\rangle\langle0|_{E}\right) U^{\dag}U_{E^{\prime}}^{i\dag}\left[
K^{i}\right] ^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\}
\label{eq:16}$$ where$$K^{i}\equiv U_{E^{\prime}}^{i}U\left( \sigma\otimes|0\rangle\langle
0|_{E}\right) U^{\dag}U_{E^{\prime}}^{i\dag},$$ with $\left\{ U_{E^{\prime}}^{i}\right\} $ a set of Heisenberg-Weyl operators. Then by the concavity result (\[eq:Hiai-concave\]) above for positive semidefinite $P$ and $Q$, it follows that the right-hand side of (\[eq:16\]) is no larger than$$\begin{gathered}
\text{Tr}\left\{ \left[ \left[ \mathcal{N}(\sigma)\otimes\pi_{E^{\prime}}\right] ^{\frac{\alpha-1}{2}}\left[ \mathcal{N}\left( \sigma
^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\right)
\otimes\pi_{E^{\prime}}\right] \left[ \mathcal{N}(\sigma)\otimes
\pi_{E^{\prime}}\right] ^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}}\right\} \\
=\text{Tr}\left\{ \left[ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma
^{\frac{1-\alpha}{2}}\right) \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right]
^{\frac{1}{\alpha}}\right\} .\end{gathered}$$ Thus we obtain the inequality$$\text{Tr}\left\{ \left[ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma
^{\frac{1-\alpha}{2}}\right) \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right]
^{\frac{1}{\alpha}}\right\} \geq1. \label{eq:helper-ineq}$$ This completes the proof of (\[delta-non-neg\]) for $\alpha\in\left(
1,2\right) $. The equality condition in (\[eq:Delta=0-other-a\]) follows from the representation in (\[eq:Delta-rewrite-1\]) and the equality condition stated after (\[eq:renyi-non-neg\]).
Next we prove that $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})\geq0$ for $\alpha\in\left( 1,\infty\right) $. We start with the definition (\[eq:Renyi-rel-ent-diff-sand\]), which we repeat here for convenience:$$\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})\equiv\frac{\alpha}{\alpha-1}\log\left\Vert \rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N(}\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N(}\sigma)^{\frac
{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac
{1}{2}}\right\Vert _{\alpha}. \label{eq-sandwich-rel-diff}$$ From [@MDSFT13 Lemma 12], it follows that the right-hand side of (\[eq-sandwich-rel-diff\]) can be written as $$\begin{gathered}
\left\Vert \rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N(}\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N(}\sigma)^{\frac{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}}\right\Vert
_{\alpha}\\
=\sup_{\tau\geq0,\text{Tr}\left\{ \tau\right\} \leq1}\text{Tr}\left\{
\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left(
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N(}\rho)^{\frac
{1-\alpha}{\alpha}}\mathcal{N(}\sigma)^{\frac{\alpha-1}{2\alpha}}\right)
\sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}}\tau^{\frac{\alpha
-1}{\alpha}}\right\} .\end{gathered}$$ Now define$$\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N};\tau)\equiv\frac{1}{\alpha-1}\log\text{Tr}\left\{ \rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha
}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha
-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}}\tau^{\frac{\alpha-1}{\alpha}}\right\} .$$ Let us focus on the trace functional in the above equation:$$\begin{aligned}
& \text{Tr}\left\{ \rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac
{\alpha-1}{2\alpha}}\right) \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac
{1}{2}}\tau^{\frac{\alpha-1}{\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}}\tau^{\frac{\alpha-1}{\alpha}}\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha
}{2\alpha}}\mathcal{N}^{\dag}\left( \mathcal{N}(\sigma)^{\frac{\alpha
-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\right) \right\} \nonumber\\
& =\text{Tr}\left\{ \mathcal{N}\left( \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}}\tau^{\frac{\alpha-1}{\alpha}}\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha}{2\alpha}}\right) \mathcal{N}(\sigma)^{\frac{\alpha
-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\right\} \nonumber\\
& =\text{Tr}\left\{ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}\left( \sigma^{\frac{1-\alpha}{2\alpha}}\rho^{\frac{1}{2}}\tau^{\frac{\alpha-1}{\alpha}}\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha
}{2\alpha}}\right) \mathcal{N}(\sigma)^{\frac{\alpha-1}{2\alpha}}\mathcal{N}(\rho)^{\frac{1-\alpha}{\alpha}}\right\} .\end{aligned}$$ By making the substitution $\alpha^{\prime}\equiv\left( 2\alpha-1\right)
/\alpha$, so that $\left( 1-\alpha\right) /\alpha=1-\alpha^{\prime}$ and thus $\alpha^{\prime}\in\left( 1,2\right) $ when $\alpha\in\left(
1,\infty\right) $, we see that the last line above is equal to$$\text{Tr}\left\{ \mathcal{N}(\sigma)^{\frac{\alpha^{\prime}-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha^{\prime}}{2}}\rho^{\frac{1}{2}}\tau^{\left( \alpha^{\prime}-1\right) }\rho^{\frac{1}{2}}\sigma
^{\frac{1-\alpha^{\prime}}{2}}\right) \mathcal{N}(\sigma)^{\frac
{\alpha^{\prime}-1}{2}}\mathcal{N}(\rho)^{\left( 1-\alpha^{\prime}\right)
}\right\} .$$ Observe that this is similar to (\[eq:Delta-rewrite\]). Hence we can write $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N};\tau)$ as$$\begin{gathered}
\frac{1}{\alpha^{\prime}-1}\log\text{Tr}\left\{ \mathcal{N}(\sigma
)^{\frac{\alpha^{\prime}-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha
^{\prime}}{2}}\rho^{\frac{1}{2}}\tau^{\left( \alpha^{\prime}-1\right) }\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha^{\prime}}{2}}\right) \mathcal{N}(\sigma)^{\frac{\alpha^{\prime}-1}{2}}\mathcal{N}(\rho)^{\left(
1-\alpha^{\prime}\right) }\right\} \\
=D_{\alpha^{\prime}}\left( \left[ \mathcal{N}(\sigma)^{\frac{\alpha^{\prime
}-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha^{\prime}}{2}}\rho^{\frac
{1}{2}}\tau^{\left( \alpha^{\prime}-1\right) }\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha^{\prime}}{2}}\right) \mathcal{N}(\sigma)^{\frac
{\alpha^{\prime}-1}{2}}\right] ^{\frac{1}{\alpha^{\prime}}}\middle\Vert
\mathcal{N}(\rho)\right) .\end{gathered}$$ This implies that$$\begin{aligned}
& \widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})\nonumber\\
& =\sup_{\tau\geq0,\text{Tr}\left\{ \tau\right\} \leq1}D_{\alpha^{\prime}}\left( \left[ \mathcal{N}(\sigma)^{\frac{\alpha^{\prime}-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha^{\prime}}{2}}\rho^{\frac{1}{2}}\tau^{\alpha^{\prime}-1}\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha^{\prime}}{2}}\right) \mathcal{N}(\sigma)^{\frac{\alpha^{\prime}-1}{2}}\right] ^{\frac
{1}{\alpha^{\prime}}}\middle\Vert\mathcal{N}(\rho)\right) \\
& \geq D_{\alpha^{\prime}}\left( \left[ \mathcal{N}(\sigma)^{\frac
{\alpha^{\prime}-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha^{\prime}}{2}}\rho^{\frac{1}{2}}\rho^{\left( \alpha^{\prime}-1\right) }\rho^{\frac{1}{2}}\sigma^{\frac{1-\alpha^{\prime}}{2}}\right) \mathcal{N}(\sigma
)^{\frac{\alpha^{\prime}-1}{2}}\right] ^{\frac{1}{\alpha^{\prime}}}\middle\Vert\mathcal{N}(\rho)\right) \\
& =D_{\alpha^{\prime}}\left( \left[ \mathcal{N}(\sigma)^{\frac
{\alpha^{\prime}-1}{2}}\mathcal{N}\left( \sigma^{\frac{1-\alpha^{\prime}}{2}}\rho^{\alpha^{\prime}}\sigma^{\frac{1-\alpha^{\prime}}{2}}\right)
\mathcal{N}(\sigma)^{\frac{\alpha^{\prime}-1}{2}}\right] ^{\frac{1}{\alpha^{\prime}}}\middle\Vert\mathcal{N}(\rho)\right) \\
& =\Delta_{\alpha^{\prime}}(\rho,\sigma,\mathcal{N})\\
& \geq0,\end{aligned}$$ where the first inequality follows by setting $\tau=\rho$ and the last inequality follows because we have already shown that $\Delta_{\alpha^{\prime
}}(\rho,\sigma,\mathcal{N})\geq0$ for $\alpha^{\prime}\in\left( 1,2\right)
$. The above inequality also demonstrates that$$\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})\geq\Delta_{\alpha
^{\prime}}(\rho,\sigma,\mathcal{N}). \label{eq:Delta-tilde-Delta}$$
We can use this result to give an alternative proof of the fact that $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$ if $\Delta_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for some $\alpha\in(1,2)$ or if $\widetilde
{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})=0$ for some $\alpha\in\left(
1,\infty\right) $:
If for some $\alpha\in(1,2)$ we have $\Delta_{\alpha}(\rho,\sigma
,\mathcal{N})=0$, then we know that$$\begin{aligned}
\left[ \mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\mathcal{N}\left(
\sigma^{\frac{1-\alpha}{2}}\rho^{\alpha}\sigma^{\frac{1-\alpha}{2}}\right)
\mathcal{N}(\sigma)^{\frac{\alpha-1}{2}}\right] ^{\frac{1}{\alpha}} &
=\mathcal{N}(\rho)\\
\Longleftrightarrow\mathcal{N}\left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha
}\sigma^{\frac{1-\alpha}{2}}\right) & =\mathcal{N}(\sigma)^{\frac{1-\alpha
}{2}}\mathcal{N}(\rho)^{\alpha}\mathcal{N}(\sigma)^{\frac{1-\alpha}{2}}$$ This implies that$$\text{Tr}\left\{ \mathcal{N}\left( \sigma^{\frac{1-\alpha}{2}}\rho^{\alpha
}\sigma^{\frac{1-\alpha}{2}}\right) \right\} =\text{Tr}\left\{
\mathcal{N}(\rho)^{\alpha}\mathcal{N}(\sigma)^{1-\alpha}\right\}$$ which in turn implies that$$\text{Tr}\left\{ \rho^{\alpha}\sigma^{1-\alpha}\right\} =\text{Tr}\left\{
\mathcal{N}(\rho)^{\alpha}\mathcal{N}(\sigma)^{1-\alpha}\right\} ,$$ because $\mathcal{N}$ is trace preserving. This is then an equality of $f$-divergences, which we can use to conclude the sufficiency property as in previous proofs.
If $\widetilde{\Delta}_{\alpha}(\rho,\sigma,\mathcal{N})=0$, for some $\alpha\in\left( 1,\infty\right) $, then from (\[eq:Delta-tilde-Delta\]) we have that $\Delta_{\alpha^{\prime}}(\rho,\sigma,\mathcal{N})=0$ for some $\alpha^{\prime}\in(1,2)$. Then we know from the above analysis that $\mathcal{N}$ is sufficient for $\rho$ and $\sigma$.
[^1]: Statistical Laboratory, Centre for Mathematical Sciences, University of Cambridge, Wilberforce Road, Cambridge CB3 0WB, UK
[^2]: Hearne Institute for Theoretical Physics, Department of Physics and Astronomy, Center for Computation and Technology, Louisiana State University, Baton Rouge, Louisiana 70803, USA
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Nelson R. F. Braga'
- ', M. A. Martin Contreras'
- and Saulo Diles
title: Holographic Picture of Heavy Vector Meson Melting
---
Introduction
==============
The suggestion [@Matsui:1986dk] (see [@Satz:2005hx] for a review) that $J/\psi $ supression in heavy ion collisions could be a signature for the formation of quark gluon plasma gave rise to a continuous interest in the thermal behavior of charmonium states. In particular, it is of great interest to know what are the temperature ranges at which the heavy vector mesons states melt. By melting one means the thermal dissociation in the medium that corresponds to the disappearance of the particle peak in the spectral function.
AdS/QCD models are very useful tools for studying spectral properties of hadronic states. Such models, inspired in the AdS/CFT correspondence [@Maldacena:1997re; @Gubser:1998bc; @Witten:1998qj], assume the existence of an approximate duality between a field theory living in an anti-de Sitter background deformed by the introduction of a dimension-full parameter and a gauge theory where the parameter plays the role of an energy scale. One of the earliest formulations, the hard wall AdS/QCD model, appeared in refs. [@Polchinski:2001tt; @BoschiFilho:2002ta; @BoschiFilho:2002vd] and consists in placing a hard geometrical cutoff in anti-de Sitter (AdS) space. In particular, the hard wall model was used in [@BoschiFilho:2002ta; @BoschiFilho:2002vd] as a tool for calculating masses of glueballs. Another AdS/QCD model, the soft wall, where the square of the mass grow linearly with the radial excitation number was introduced in ref. [@Karch:2006pv]. In this case, the background involves AdS space and a scalar field that acts effectively as a smooth infrared cutoff. A recent recent review of AdS/QCD with a wide list of references can be found in [@Brodsky:2014yha].
AdS/QCD models provide also a tool for calculating another important property of hadrons: the decay constant. The decay of mesons is represented as a transition from the initial state to the hadronic vacuum. For a meson at radial excitation level $n$ with mass $m_n$ the decay constant $ f_n $ is defined as: $ \langle 0 \vert \, J_\mu (0) \, \vert n \rangle = \epsilon_\mu f_n m_n $, where $ J_\mu $ is the gauge current and $ \epsilon_\mu $ the polarization. Expressing the two point correlator of gauge currents as a sum over transition matrix elements, one finds a holographic expression for decay constants [@Karch:2006pv; @Grigoryan:2007my].
A problem of the original formulations of the hard wall and soft wall models is that the experimental results available for charmonium and bottomonium vector states show that higher excited radial states have smaller decay constants. In other words, $f_n$ decrease with $n$. In contrast, the results obtained for decay constants of vector mesons in the soft wall are degenerate: all the decay constants of the radial excitations of a vector meson are equal. For the hard wall model the decay constants of radial excitations increase with the excitation level. A fit of the decay constants of charmonium states in soft wall appeared in ref. [@Grigoryan:2010pj], but introducing three extra parameters in the model. In ref. [@Grigoryan:2010pj] four experimental data, the masses and decay constants of $J /\psi $ and $\psi^\prime$, are used to fix four parameters introduced in the model, thus the formulation lacks of predictivity as a model for heavy vector mesons.
An alternative version of the soft wall model, consistent with the observed behavior of decay constants, was recently proposed in ref. [@Braga:2015jca]. In contrast to the original formulation, in this new framework the decay constants are obtained from two point correlators of gauge theory operators calculated at a finite value $ z = z_0$ of the radial coordinate of AdS space. This way an extra energy parameter $ 1/z_0 $, associated with an ultraviolet (UV) energy scale is introduced in the model. The masses and decay constants of charmonium and bottomonium $S$ wave states are calculated in ref. [@Braga:2015jca] using the quantity $ 1/z_0 $ as a flavour independent parameter and taking the usual infrared (IR) soft wall parameter $k$ to depend on the flavor, since it is associated with the quark mass. A total of eight masses and eight decay constants are determined using three parameters. The rms error is of $30 \% $ that is reasonable, given the simplicity of the model and the fact that two different properties of two different flavors are adjusted with just 3 parameters.
The purpose of the present article is to extend the model of ref. [@Braga:2015jca] to finite temperature in order to investigate the thermal spectra of $S$ wave states of charmonium and bottomonium. We will show that the spectral functions present the expected behavior: at low temperatures, sharp peaks for the lower level excitations, and, as the temperature increases, the peaks spread and decrease in height. The evolution of the spectral function with increasing temperature shows clearly the process of transition from well defined peaks to the disappearance of the states in the medium, for the states $ 1S$ , $2S$ and $3S$. The melting occurs at lower temperatures for the higher excitations.
It is worth mentioning that in refs. [@Hong:2003jm; @Kim:2007rt; @Fujita:2009wc; @Noronha:2009da; @Fujita:2009ca; @Grigoryan:2010pj; @Branz:2010ub; @Gutsche:2012ez; @Afonin:2013npa; @Hashimoto:2014jua], heavy vector mesons have been discussed in the context of AdS/QCD models. However, the holographic picture for the melting of $ 1S$ , $2S$ and $3S$ states of bottomonium and charmonium that we will show here, was not presented before in the literature.
The article is organized as follows: in section 2 we briefly review the model for heavy vector mesons at zero temperature presented recently in ref. [@Braga:2015jca]. Then is section 3 we build up a finite temperature version for this model and show how to calculate the corresponding thermal spectral function. In section 4 we show the results obtained by numerically solving the equations of motion. We analyze the melting of the states of charmonium and bottomonium as the temperature increase and estimate the temperature ranges where the thermal dissociation occurs. We leave for section 5 some final comments and remarks and present in the appendix more details of the melting of charmonium states. Appendix A shows more details of the temperature dependence of the thermal spectral functions and appendix B presents an analysis of the high frequency behavior.
Heavy Vector mesons in the vacuum
===================================
The holographic model proposed in ref. [@Braga:2015jca] contains two dimensionful parameters. One coming from a soft wall background and the other from a position in AdS space where the gauge theory correlators are calculated. The model leads to decay constants for heavy vector mesons decreasing with the radial excitation level, in agreement with the results obtained from experimental data.
One considers a vector field $V_m = (V_\mu,V_z)\,$ ($\mu = 0,1,2,3$) playing the role of the supergravity dual of the gauge theory current $ J^\mu = \bar{q}\gamma^\mu q \,$. The field lives in a five dimensional soft wall background governed by the action $$I \,=\, \int d^4x dz \, \sqrt{-g} \,\, e^{- \Phi (z) } \, \left\{ - \frac{1}{4 g_5^2} F_{mn} F^{mn}
\, \right\} \,\,,
\label{vectorfieldactionzeroptemp}$$ where $F_{mn} = \partial_mV_n - \partial_n V_m$ and $\Phi = k^2z^2 $ is the soft wall background, with the parameter $k$ playng the role of an IR, or mass, energy scale. The space is a Poincaré AdS chart: $$ds^2 \,\,= \,\, \frac{R^2}{z^2} \, \Big( - dt^2 + dz^2 + d\vec{x}\cdot d\vec{x} \Big) \,.
\label{metric1}$$
The second input parameter of the model, that is not present in the usual formulation of the soft wall model, is introduced by calculating the correlators at a finite position $ z = z_0$ instead of taking the boundary to be at $z=0$. The parameter $1/z_0$ is interpreted as an UV energy scale. A similar approach appeared in ref. [@Evans:2006ea] but for light vector mesons.
One considers the action of eq. (\[vectorfieldactionzeroptemp\]) to be defined in the region $ z_0 \le z < \infty $, then the on shell action takes the form $$I_{on \, shell }\,=\, - \frac{1}{2 {\tilde g}_5^2} \, \int d^4x \,\,\Big[ \, \frac{e^{- k^2 z^ 2 }}{z} \eta^{\mu \nu} V_\mu \partial_z V_ \mu \,\Big]
{\Big \vert}_{_{ \! z \to z_0 }}
\,,
\label{onshellactionzerotemp}$$ where $ {\tilde g}_5^2 = g_5^2 /R $ is the relevant dimensionless coupling of the vector field and $\eta^{\mu \nu} $ is the Minkowski metric.
The gauge $V_z = 0 $ is used, so that the boundary values of the other remaining components of the vector field: $ V^0_{\mu}(x) =
\lim_{z\to z_0} V_\mu (x,z)$ are the sources of the correlation functions of the boundary current operator $ J^\mu (x) \,\, (= \bar{q}\gamma^\mu q (x) \,)\,$. That means: $$\langle 0 \vert \, J_\mu (x) J_\nu (y) \, \vert 0 \rangle \, =\, \frac{\delta}{\delta V^{0\mu}(x)} \frac{\delta}{\delta V^{0\nu}(y)}
\exp \left( - I_{on shell} \right)\,.$$
Working in momentum space in the coordinates $x^ \mu$, or equivalently taking a plane wave solution, the field $ V_\mu (p,z) $ can be decomposed for convenience into a source factor times a $z$ dependent factor $$V_\mu (p,z) \,=\, v (p,z) V^0_\mu ( p ) \,,
\label{Bulktoboundary}$$ where $ v (p,z) $ is usually called bulk to boundary propagator and satisfies the equation of motion: $$\partial_z \Big( \frac{ e^{-k^ 2 z^ 2 }} { z} \partial_z v (p,z) \Big) + \frac{p^ 2 }{z} e^{-k^ 2 z^ 2 } v (p,z) \,=\, 0 \,.
\label{BulktoboundaryEOM}$$ In order that the factor $ V^0_\mu ( p )$, defined in the decomposition of eq. (\[Bulktoboundary\]), works as the source of the correlators of gauge theory currents, calculated at $z = z_0 $, one must impose the boundary condition: $$v (p, z=z_0) = 1\,.
\label{boundarycondition}$$ The solution of eq. (\[BulktoboundaryEOM\]) is a Tricomi function $U (-p^ 2/ 4k^ 2 , 0, k^2 z^ 2) $. The boundary condition can be trivially satisfied following ref. [@Afonin:2011ff; @Afonin:2012xq] and writing: $$v (p,z ) \, = \, \frac{ U (-p^ 2/ 4k^ 2 , 0, k^2 z^ 2 ) }{U (-p^ 2/ 4k^ 2 , 0, k^2 z_0^ 2 )}\,.
\label{bulktoboundary2}$$ The decay constants appear in the two point function: $$\Pi (p^2) = \sum_{n=1}^\infty \, \frac{f_n^ 2}{(- p^ 2) - m_n^ 2 + i \epsilon} \,.
\label{2point}$$ On the other hand, the two point function is related to the current current correlator $$\left( p^2 \eta_{\mu\nu} - p_\mu p_\nu \right) \, \Pi ( p^2 )
\, =\, \int d^4x \,\, e^{-ip\cdot x} \langle 0 \vert \, J_\mu (x) J_\nu (0) \, \vert 0 \rangle \, ,
\label{correlatorand2pointfunction}$$ that can be obtained holographically by differentiating the on shell action by the boundary values of the fields, with the result: $$\Pi ( p^2 ) \, =\, \frac{1}{{\tilde g}_5^ 2 \, (-p^ 2) } \left[ \frac{ e^{ -k^2 z^2 } \, v (p,z) \partial_z v (p,z) }{ \, z }
\right]_{_{ \! z \to z_0 }} \, .
\label{hol2point}$$
The expression (\[hol2point\]) has simple poles, although it does not have the exact simple pole structure of eq. (\[2point\]). But one can associate the coefficients of the approximate expansion near the poles with the decay constant $f_n$ in analogy with the exact expansion shown in eq. (\[2point\]). This way one finds the masses from the localization of the poles of the two point function and the decay constants from the corresponding coefficient. That means, if $\chi_n$ are the roots of the Tricomi function: $$U ( \chi_n \, , 0, k^2 z_0^ 2 ) = 0 \,,
\label{roots}$$ then the holographic vector meson masses are: $$m^2_n \, = \, 4 k^ 2 \, \chi_n \,.
\label{new masses}$$ The decay constants are calculated numerically from the fit to the approximate form of the simple pole of eq. (\[2point\]). That means: $$f_n^ 2 \,=\,
\lim_{p^2 \to - m^2_n} \Big( (-p^ 2) - m_n^ 2 \Big) \,
\Pi (p^2) \
\,.
\label{numerical decay}$$
The coupling ${\tilde g}_5 = g_5 /\sqrt{R} $ of the vector field in the AdS bulk is obtained by comparison with QCD (see refs. [@Karch:2006pv; @Grigoryan:2007my]) , wich gives: ${\tilde g}_5 \, =\, 2 \pi $.
The parameter $ k$ is flavor dependent, representing the mass of the heavy quarks. The energy scale $1/z_0$ is taken as having the same value for charmonium and bottomonium, representing a flavor independent factor associated with just color interaction. The parameters used in ref. [@Braga:2015jca] are: $$k_c = 1.2 GeV ; \, k_b = 3.4 GeV ; \, 1/z_0 = 12.5 GeV,
\label{parameters}$$ where $ k_c$ and $k_b$ are the values of the constants $k$ used for charmonium and bottomonium, respectively. Using these 3 parameters and the relations (\[new masses\]) and (\[numerical decay\]) the masses and decay constants of the states $1S, 2S, 3S, 4S$ of charmonium and bottomonium were estimated with an rms error of $30 \% $.
In the next section we extend this model to finite temperature and then, considering the same choice of parameters of eq. (\[parameters\]) we analyze the behavior of charmonium and bottomonium $S$ wave states in a thermal plasma.
Heavy Vector mesons at finite temperature
===========================================
Now we extend the zero temperature model of ref. [@Braga:2015jca] to finite temperature. It is important to mention that hadronic spectra at finite temperature have been studied in the context of AdS/QCD soft wall model before, for example, in refs. [@Fujita:2009wc; @Colangelo:2009ra; @Miranda:2009uw; @Fujita:2009ca; @Mamani:2013ssa]. In particular [@Mamani:2013ssa] describes light vector mesons in the soft wall model. However, a complete analysis of the thermal spectral function for vector states of bottomonium and charmonium like the one performed in this article is not present in the literature.
Dual space and Hawking Page transition
--------------------------------------
Gauge string duality at finite temperature was discussed originally in refs. [@Witten:1998qj; @Witten:1998zw]. Considering Euclidean signature and a compactified time coordinate, the geometry dual to a gauge theory at finite temperature is one of the two solutions of Einstein equations with constant negative curvature. One of these solutions is the AdS black hole space that in Euclidean signature reads $$ds^2 \,\,= \,\, \frac{R^2}{z^2} \, \Big( f(z) dt^2 + \frac{dz^2}{f(z) } + d\vec{x}\cdot d\vec{x} \Big) \,,
\label{EuclideanMetric}$$ where the Schwarschild factor is $ f (z) = 1 - z^ 4/ z_h^ 4 $ and $z_h$ is the horizon position. The other solution is the thermal AdS space, that is just AdS space corresponding to $ f(z) = 1 $, with a compactified time.
Following the work by Hawking and Page [@Hawking:1982dh], one uses the semiclassical argument that there is “competition” between the two solutions and the one with smaller Einstein Hilbert action will be stable at a given temperature. For the conformal gauge theory case (in a non compact space) the black hole is the stable solution for all temperatures[@Witten:1998zw]. So, the dual geometry is the black hole. For a non conformal gauge theory, as in the soft wall model case, the dual geometry has two different phases, as discussed in refs. [@Herzog:2006ra; @BallonBayona:2007vp]. For temperatures above a critical value $T_c$ the black hole is stable while for temperatures below $T_c$ the thermal AdS is stable. The so called Hawking Page transition between spaces was interpreted in [@Witten:1998zw] as a transition in the dual gauge theory from a deconfined ($T > T_c$) to a confined phase ($T < T_c$).
In order to compare the action integrals of the black hole AdS and the thermal AdS we must take into account the fact that the periodicity of the time coordinate is related to the temperature. In our model the gauge theory is at $ z= z_0$ where the transverse part of the metric of the black hole is: $$ds^ 2 = \frac{R^ 2}{z_0^ 2} \Big( f(z_0) dt^2 + d\vec{x}\cdot d\vec{x} \Big) \,.$$ The mapping of the supergravity theory to a gauge theory in flat space must be performed with the rule that the gauge theory time has to be $\tau = t \sqrt{ f(z_0) } $. Since the period is the inverse of the gauge theory temperature: $ \tau \sim \tau + 1/T$ and the period of the black hole coordinate $t$ must be $\pi z_h$ to avoid a conical singularity at the horizon, one finds:
$$T = \frac{1}{ \pi z_h \sqrt{ f (z_0) }} = \frac{ 1}{ \pi z_h \sqrt{ 1 - \frac{z_0^4 }{z_h^ 4 } \,}}\,.
\label{temp}$$
The action densities for the black hole AdS and thermal AdS in the soft wall model are calculated in ref. [@Herzog:2006ra]. The results of this article can be adapted to the model considered here, where there is an UV cutoff, by replacing the minimum value of the coordinate $z$ that in ref. [@Herzog:2006ra] is just an UV regulator $ z= \epsilon $ by the (inverse of the) UV energy scale: $z = z_0$. Using also the relation between the horizon position and the temperature in eq. (\[temp\]) one gets:
$$\begin{aligned}
V_{th\, AdS} &=& \frac{4 R^ 3 }{g_5^2} \,\frac{1}{T} \, \int_{z_0}^\infty dz \frac{ e^{-k^ 2 z^ 2}}{z^ 5}
\label{densities1} \\
V_{BH \, AdS} &=& \frac{4 R^ 3 }{g_5^2} \, \frac{1}{T \sqrt{f(z_0) }} \int_{z_0}^{z_h} dz \frac{ e^{-k^ 2 z^ 2}}{z^ 5} \,.
\label{densities2}
\end{aligned}$$
The critical temperature, where the two actions densities are equal, depends on the infrared parameter $k$ of the soft wall background. This parameter is flavor dependent. The formation of the plasma occurs when the lightest hadrons dissociate. So, we consider the confinement/deconfinement transition to be determined by the soft wall background that fits the masses of the $\rho$ vector mesons. In the present model $\rho$ vector mesons can be described taking as in ref. [@Braga:2015jca] $ 1/z_0 = 12.5 $ GeV and reproducing the calculation of the mass reviewed in section 2. One finds, using the parameter $ k $ = 0.388 GeV as in [@Herzog:2006ra], that the model with UV cut off leads to a mass of 777.6 MeV for the $1 S $ state.
The corresponding critical temperature is $ T_c = 191 $ MeV, the same result of ref. [@Herzog:2006ra]. In figure [**1** ]{} we show the difference $ \Delta V = V_{BH \, AdS} \,-\, V_{th\, AdS}\, $ between the action densities of eqs. (\[densities1\]) and (\[densities2\]) as a function of the temperature. The critical temperature corresponds to the point where the curve crosses the temperature axis.
Vector meson solutions in the black hole
-----------------------------------------
As in the zero temperature case, we take a vector field $V_m = (V_\mu,V_z)\,$ ($\mu = 0,1,2,3$) described by an action integral with the general form of eq. (\[vectorfieldactionzeroptemp\]) and soft wall background $\Phi = k^2z^2 $ . But for describing the thermal spectra one considers the geometry as the Minkowski version of the black hole metric (\[EuclideanMetric\]): $$ds^2 \,\,= \,\, \frac{R^2}{z^2} \, \Big( - f(z) dt^2 + \frac{dz^2}{f(z) } + d\vec{x}\cdot d\vec{x} \Big) \,.
\label{MinkoviskyMetric}$$ where again: $ f (z) = 1 - z^ 4/ z_h^ 4 $ and the gauge theory temperature is related to the horizon position by eq. (\[temp\]). It is important to note that this black hole geometry will be stable only for temperatures $T > T_c$. We will calculate the thermal spectral functions using this black hole metric for all temperatures with the interpretation that for $T < T_c$ it represents a super-cooled (unstable) phase.
As in the zero temperature case, we choose the gauge $V_z = 0 $ and assume $ V^0_{\mu}(x) =
\lim_{z\to z_0} V_\mu (x,z) $ to be the sources of the correlation functions of $ J^\mu (x) \,$. Now, with the radial AdS coordinate defined in the region: $ z_0 \le z \le z_h $, the on shell action takes the form: $$I_{on \, shell }\,=\, - \frac{1}{2 { g}_5^2} \, \int d^4x \,\,\,\Big[ e^{- k^2 z^ 2 } \sqrt{-g} \, g^{zz} g^{\mu \nu} V_\mu \partial_z V_\nu \Big]
{\Big \vert}_{_{ \! z \to z_0 }}^{_{ \! z \to z_h }}
\,.
\label{onshellaction2}$$ The imaginary part of the on shell action should generate holographically the thermal spectral function. However, it was pointed out in ref. [@Son:2002sd] that for an action like (\[onshellaction2\]) the imaginary part is $z $ independent. So the contributions from the two integration limits cancel each other out. This problem can be solved following again [@Son:2002sd] and using the additional prescription that only the boundary $z= z_0$ is considered. In other words, one takes eq. (\[onshellaction2\]) with only the lower integration limit. For an interesting discussion on the interpretation of the prescription for calculating the retarded Green’s function, see [@Iqbal:2008by].
The procedure to find the retarded Green’s function involves fourier transforming the fields and decomposing the momentum space fields as it was done in the finite temperature case in eq. (\[Bulktoboundary\]): $ V_\mu (q,z) \,=\, v (q,z) V^0_\mu ( q ) \,$. The on shell action takes the form $$I_{on \, shell }\,=\, \, \int d^4q \,\,\,\Big[ {V^ 0_\mu}^\ast ( q ) {\cal F}^{\mu\nu} (z,p) V^0_\nu ( q )
\Big]_{_{ \! z \to z_0 }}
\,,
\label{onshellaction3}$$ where $${\cal F^{\mu \nu} } (z,q) \,=\, \frac{1}{2 g_5^2} e^{- k^2 z^ 2 } \sqrt{-g} \, g^{zz} g^{\mu\nu}
v^\ast ( q,z) \partial_z v (q,z) ]\,.$$
The corresponding retarded Green’s function is: $$G_R^{\mu\nu} (q) \,=\, {\cal F}^{\mu\nu} (z=z_0 , q) \,,$$ and the spectral function is the imaginary part of the retarded Green’s function: $$\rho^{\mu\nu} (q) \, =\, - {\cal I}m \{ G_R^{\mu\nu} (q) \} \,.$$ The bulk to boundary propagators $ v (q,z) $ are solutions of the equations of motion. These equations have different forms for the temporal $V_0$ and spatial $V_i$ components of the vector field. For the case of a plane wave solution with momentum $q^\mu = (\omega , \vec q ) $ they are: $$\begin{aligned}
\partial_z \Big( \frac{ e^{-k^ 2 z^ 2 } } { z} \partial_z V_0 (q,z) \Big) &-&
\frac{ e^{-k^ 2 z^ 2 } }{ z f(z) } \Big( \frac{\omega^ 2 }{ f(z) } -\vert {\vec q}\vert^2 \Big)
V_0 (q,z) \,=\, 0 \cr \cr
\partial_z \Big( \frac{ e^{-k^ 2 z^ 2 } f(z) } { z} \partial_z V_i (q,z) \Big) &+&
\frac{ e^{-k^ 2 z^ 2 } }{ z } \Big( \frac{\omega^ 2 }{ f(z) } - \vert {\vec q}\vert^2 \Big)
V_i (q,z) \,=\, 0 \,.
\label{EOMfiniteT}\end{aligned}$$
It is convenient [@Grigoryan:2010pj] to choose the momentum $q^\mu = (\omega , \vec 0 ) $ where the transversality of the current $ q^\mu J_\mu = 0 $ translates into the vanishing of the temporal component $J_0$. Then we just need to solve the equation for the spatial component: $ V_i (\omega ,z) \,=\, v (\omega ,z) V^0_i ( \omega ) $. In this case $ v (\omega ,z) $ satisfies the equation:
$$\partial_z \Big( \frac{ e^{-k^ 2 z^ 2 } f(z) } { z} \partial_z v (\omega ,z) \Big) +
\frac{ e^{-k^ 2 z^ 2 } }{ z } \frac{\omega^ 2 }{ f(z) }
v (\omega ,z) \,=\, 0 \,.
\label{eqmotionbtb}$$
The bulk to boundary propagator has to satisfy two boundary conditions. One is $$v (\omega, z=z_0) = 1\,,
\label{conditionbtb}$$ that was present in the zero temperature case and implies that the field components work as the sources of the correlation functions at $z= z_0$. The other is the condition that the solution behaves as an incoming wave in the near horizon limit $ z \to z_h$. The absence of outgoing solutions represents the absorption by the black hole horizon. In order to implement this condition one can use the Regge-Wheeler tortoise coordinate that makes it explicit the decomposition of the solutions of the equations of motion in incoming plus outgoing solutions. One introduces the coordinate $ r_\ast $ such that $ \partial_{r_\ast} = - f(z) \partial_z $ that implies $$r_\ast = \frac{1}{2} z_h \Big[ -\tan^{-1} \left( \frac{z}{z_h} \right) +
\frac{1}{2} \ln \left( \frac{z_h - z}{z_h + z} \right) \Big]\,,$$ in the interval $ z \le z_h $ where $z$ is defined.
Performing a Bogoliubov transformation $ v (\omega , z) = e^{B/2} \psi (\omega , z) $ with $ e^B = z e^{k^ 2 z^ 2} $ one finds that the equation of motion (\[EOMfiniteT\]) takes the form
$$\partial^2_{r_\ast} \psi + \omega^2 \psi = U \psi \,,$$
where the potential $$U (z) \,=\, \left( 1 - \frac{z^ 4}{z_h^4} \right) \left[ \left( k^4 z^2 + \frac{3}{4z^ 2} \right) \left( 1 - \frac{z^ 4}{z_h^4} \right) + 2 z^2 \frac{(1 + 2 k^2 z^2 )}{z_h^4} \right]$$ vanishes at the horizon. Thus, the function $\psi$ has the asymptotic near horizon solutions $\psi_{in/out} = e^{ \mp i \omega r_\ast} $ representing incoming and outgoing waves respectively.
Expanding the incoming wave solution near the horizon as $$\psi_{in} = e^{ - i \omega r_\ast} \left[ 1 + a_1 (z - z_h) + a_2 (z - z_h )^2 + ... \right] \,,$$ and inserting in the equation of motion one finds the relevant coefficient: $$a_1 = \frac{ 1 + 2k^2 z_h^2 }{ z_h ( i \omega z_h - 2 ) } \,.
\label{a1}$$ In order to implement the incoming wave condition we write the bulk to boundary propagator as $$v (\omega, z) \,=\,e^{ - i \omega r_\ast} \, F(\omega,z) \,,
\label{vdecomp}$$ so that the function $F$ takes the form: $$F(\omega,z) = \sqrt{z} e^{\frac{k^2z^2}{2}} \left[ 1 + a_1 (z - z_h) + a_2 (z - z_h )^2 + ... \right]
\label{BCF}$$ and the derivative of $F $ at the horizon is obtained from this expansion and the expression for $a_1$ in eq. (\[a1\]).
Finally, the spectral function for spatial components $\rho^{ii} $ with the choice of momentum $q^\mu = (\omega , \vec 0 ) $ and written in terms of $F $ takes the form (omitting the indices $ i i $) $$\rho (w) \,=\, \frac{w } {2 {\tilde g}_5^2} \, \frac{ e^{- k^2 z_h^ 2 } } { z_h} \,
\vert F( \omega , z_h) \vert^ 2 \,,
\label{spectralf}$$ where we defined the dimensionless coupling $ {\tilde g}_5^2 = g_5^2 /R $, as in the zero temperature case. This is the object that will describe the thermal behavior of the heavy vector mesons. In the next section we present the results of the numerical calculations of $\rho$.
Spectral functions for charmonium and bottomonium S-wave states
================================================================
We solved numerically equation (\[eqmotionbtb\]) for the bulk to boundary propagator $v (\omega, z) $, written in terms of the function $ F$ as in eq. (\[vdecomp\]), with the boundary conditions described in the previous section. The parameters used are the zero temperature ones, from ref. [@Braga:2015jca], namely a flavor independent UV cutoff $ 1/z_0 = 12.5 $ GeV and flavor dependent soft wall parameters with values $ k_c = 1.2 $ GeV for charmonium and $ k_b = 3.4 $ GeV for bottomonium S-wave states.
The spectral function (\[spectralf\]) was calculated for different temperatures. An important non trivial fact emerged from the analysis of the large frequency asymptotic behavior. It is well know that when one calculates the spectral function from correlators at the conformal boundary $z \to 0$, the spectral function in the limit $\omega \to \infty$ grows up as $\rho \sim \omega^2 $. This results comes from conformal invariance and dimensional analyisis (see for example ref. [@Myers:2007we]).
In the present case we do not calculate the correlators at the $z\to 0$ conformal limit. There is an extra dimensionfull quantity, the position $z_0$, that appears in the calculation of the spectral function. So, the argument of simple dimensional analysis does not hold in the same way here. The numerical results obtained show a behavior that is different from the conformal case. For large frequencies the spectral function grows linearly with the frequency: $ \rho \sim \omega $. We present in appendix [**B**]{} an analysis of this behavior. We show there that if in the present model one takes the limit of $z_0 \to 0 $ one finds spectral functions growing with $\omega^2$, as expected in the conformal case. But for the finite value of $ z_0$ explored here they grow with $\omega$ for large $\omega$. So, we analyzed the behavior of the relevant (normalized) quantity: $$\frac{ \rho(\omega ) }{\omega } \,.$$
We show in figure [**2**]{} the spectral functions for the bottomonium vector states at four illustrative temperatures. In these plots one can clearly see the following situation:
- $\bullet$ at $T = 200$ MeV three peaks corresponding to $ 1S$, $2S$ and $3S$ states;
- $\bullet$ at $T = 260$ MeV two peaks corresponding to the melting of the $3S$ state;
- $\bullet$ one peak at $T = 340 $ MeV where only the $1S$ states survives and
- $\bullet$ at $T = 660 $ MeV the complete melting of the states.
We present in the appendix [**A**]{} a more detailed picture of the melting process by showing more plots that illustrate the temperature evolution of the spectral function. From this analysis one can infer that the states $1S, 2S $ and $3S$ melt at different temperatures, as expected. In particular, the $1S$ states survives at temperatures much larger than the critical temperature. The complete disappearance of the $1 S$ states happens at $ T \sim 600 $ MeV, corresponding to $ T/T_c \sim 3.2 $. For the $ 2S$ state there will be no trace of the peak for temperatures above $ T \sim 360 $ Mev, corresponding to $ T/T_c \sim 1.9 $ while for the $3 S$ states the total melting happens at $ T \sim 220 $ MeV, that means $T/T_c \sim 1.2 $.
Then figure [**3** ]{} shows the spectral functions for the charmonium vector states at four different temperatures that illustrate the melting process. More details for the thermal evolution of charmonium states are shown in appendix [**A**]{} . One can clearly see the change from the case with three well defined peaks corresponding to the states $1S, 2S$ and $3S$ to the case where there is no well defined quasi particle state. An important difference with respect to the bottomonium case is that the melting process occurs at temperatures below $T_c$. At the critical temperature there is only a very small peak of the state $1S$, so one can interpret this situation as meaning that the charmonium states $ 2S$ and $3S$ do not survive in the deconfined plasma phase, while there could be some trace of the $1 S$ state up to temperatures of $ 1.2 T_c$.
The present results for bottomonium states are consistent with the ones obtained using lattice QCD in [@Suzuki:2012ze]. This article predicts a lower bound for the melting temperature of the 1$S$ state of $ 2.3 T_c$. They are also consistent with the lattice results of [@Aarts:2011sm] where the temperature range between $0.4 T_c$ and $2.1 T_c$ was analyzed and the $1 S$ state survives for higher temperatures whereas the higher excitations melt around $ 1.4 T_c$. It is interesting to mention that experiments show that in Au + Au collisions with center of mass energy of 200 GeV the bottomonium states $ 2 S $ and $ 3 S$ are completely suppressed [@Adare:2014hje].
Using a potential model, ref. [@Mocsy:2007jz] finds that the excited states of charmonium melt below $ T_c$ while the $1S$ state melts at $ 1.2 T_c$, that is consistent with our results, taking into account the error that will be discussed in the next section.
The results obtained here are also consistent with the anlysis of the thermal behavior of quarkonium states using QCD sum rules developed in refs. [@Dominguez:2009mk; @Dominguez:2010ve; @Dominguez:2013fca] regarding the survival of quarkonium states above the critical temperature.
Conclusions
=============
It is shown in this paper that a consistent picture for the thermal behavior of $S-$wave states of bottomonium and charmonium emerges from a finite temperature version of the model for heavy vector mesons masses and decay constants proposed in ref. [@Braga:2015jca]. The spectral functions obtained numerically for bottomonium and charmonium states exihibit clear peaks for the states $1S, 2S $ and $3S$ at low temperatures. As the temperature increases, the peaks spread and disapear, with the expected result that highly excited states melt in the thermal medium (plasma) at lower temperatures.
One point that must be remarked is that the model of reference [@Braga:2015jca] presents a rms error of $ 30 \% $ when one fits the decay constants and masses of the four initial $S$ wave states of charmonium and bottomonium. So, one should not consider our numerical results for the melting temperatures of the states with a precision larger than that. We mean, our (rougth) estimate for the error in the melting temperatures is of the order of $ 30 \%$.
Even with this error, one can infer that the model predicts a very distinct behavior for bottomonium and charmonium states. This could be an iteresting tool to investigate not only the formation of quark gluon plasma but also the temperature os the thermal medium. The strong supression of charmonium states with a low supression of bottomonium states would indicate temperatures not much larger then the critical one. On the other hand, an eventual observation of supression of bottomonium $S$ wave states could indicate the formation of plamas at higher temperatures.
One question that could be asked is if one could find more acurate etimates for the melting temperatures using holography. With more accurate results one could be more confident in analysing the temperture of the plasma from the relative supression of the different states. An alternative model for calculating masses of heavy vector mesons was recently proposed in ref. [@Braga:2015lck] . In this reference the masses of charmonium and bottomonium states are estimated with an rms error of $2.0 \%$. It would be nice to formulate a finite temperature version of this model also, in order to compare the thermal behavior with the one found here. There is however an obstruction to this task. The incoming wave condition that has to be used for the field that describes a vector meson at finite temperature is apparently inconsitent with the zero temperature limit of the incoming wave condition at the black hole horizon. More precisely, at any finite temperature, the incoming wave condition implies that the derivative of the bulk to boundary propagator is infinite at the horizon. In the limit of zero temperature this would mean that the derivative should be infinite at $ z \to \infty$. In contrast, in the model of ref. [@Braga:2015lck] there is the boundary condition that the derivative of the bulk to boundary propagator is zero at $z\to \infty$. We leave for a future work the non trivial task of finding a consistent finite temperature for this model.
[**Acknowledgments:**]{} We thank Jorge Noronha for suggesting that we study the finite temperature version of ref. [@Braga:2015jca] and for very important comments about a draft of the manuscript. N.B. and S.D. are partially supported by CNPq and M.A.M. is supported by Vicerrectoria de Investigaciones de La Universidad de los Andes.
[**Appendix** ]{}
Temperature dependence of the spectral functions
=================================================
In order to present a more detailed view of the bottomonium melting process, we show in figure [**4**]{} the thermal spectral function for nine different representative temperatures. At 200 MeV, we have three defined vector states $\Upsilon (1S)$, $\Upsilon^\prime (2S)$ and $\Upsilon^{\prime\prime} (3S)$. At 220 MeV, one can see that the 3S state disappears. So, the 3S melting temperature in this model is between 200 MeV and 220 MeV. Then the 2S peak disappears near 300 MeV (left middle panel). Finally, near 580 MeV one observes the 1S melting. Lattice calculations [@Adare:2014hje] show that the $\Upsilon$ (1S) melting temperature lies inside an interval 350 MeV–612 MeV, while for the $\Upsilon$ (2S) and $\Upsilon$ (3S) the melting temperatures are in a 200 MeV– 300 MeV region. Our results are consistent with these calculations.
Figure [**5** ]{} shows the behavior of charmonium spectral function. The panels correspond to temperatures varying in steps of 20 MeV. Starting at the upper left panel, at the temperature of 70 MeV there are three peaks corresponding to $J/\Psi$, $\Psi'$ and $\Psi''$. At higher temperatures one observes the melting starting by the heavier states. At $T =$ 90 MeV, the 3S state melts. Then at temperatures about 110 MeV the 2S melts. Then at $ T = 250 $ MeV the 1$S$ peak has virtually disappeared.
It is important to take into account the fact that for temperatures below $T_c$ the black hole phase is unstable due to Hawking Page transition, as explained in section [**3**]{}. So, the transitions described in the plots of lower temperatures could be absent if the plasma phase is not formed and the medium is confined. So, the thermal spectrum is more reliable for $T > T_c = 191$ MeV.
High energy behaviour of the spectral functions
================================================
At high frequencies, the spectral functions studied in this article show a non trivial behavior. The holographic model presented in section [**3** ]{} and extended to finite temperature is section [**4**]{}, with two point correlation functions calculated at a finite position $z = z_0=1/(12.5 GeV)$ of AdS space, leads to spetral functions $\rho(\omega) \propto \omega$ in the limit of large $\omega$. This result contrasts with the situation when gauge theories correlators are calculated at $ z=0$ and conformal symmetry is manifest implying: $\rho(\omega)\,\propto \, \omega^2$.
In order to display the effect of the $z_0$ parameter in the assimptotic behaviour of spectral functions, we plot in logarithm scale in separate panels $\rho(\omega)\,,~\rho(\omega)/\omega\,$ and $ ~\rho(\omega)/\omega^2$ for frequencies up to $10^4 GeV$ using two different choices of $z_0$. Since we are interested only in the role played by the parameter $z_0$, we fix the temperature and the dilaton constant $k$ in all plots to the values: $T=400 MeV,~ k=3.4 GeV$.
In figure [**6**]{} we choose the parameter: $z_0=1/(12.5 GeV)$ that was used in the present article. One clearly see in the second panel that $\rho/\omega$ reaches a constant value for $\omega \gtrsim 50 GeV$. As a check, the first panel shows the increase of $\rho $ and the third the decrease of $\rho/\omega^2$ for large $\omega$.
Then, as check, one can take the limit where the present model should recovers the usual soft wall case, namely, a very small $z_0$. We show in figure [**7**]{} the situation at $z_0 = 10^{-6}GeV^{-1}$. Consistently, one observes that in this case where $z_0$ approximately ceases to be a parameter of the model, the ultraviolet behavior of the spectral function changes to $\rho \propto \omega^2$.
[ABC]{}
T. Matsui and H. Satz, Phys. Lett. B [**178**]{}, 416 (1986). doi:10.1016/0370-2693(86)91404-8
H. Satz, J. Phys. G [**32**]{}, R25 (2006) doi:10.1088/0954-3899/32/3/R01 \[hep-ph/0512217\].
J. M. Maldacena, Int. J. Theor. Phys. [**38**]{}, 1113 (1999) \[Adv. Theor. Math. Phys. [**2**]{}, 231 (1998)\] doi:10.1023/A:1026654312961 \[hep-th/9711200\].
S. S. Gubser, I. R. Klebanov and A. M. Polyakov, Phys. Lett. B [**428**]{}, 105 (1998) doi:10.1016/S0370-2693(98)00377-3 \[hep-th/9802109\].
E. Witten, Adv. Theor. Math. Phys. [**2**]{} (1998) 253 \[hep-th/9802150\].
J. Polchinski and M. J. Strassler, Phys. Rev. Lett. [**88**]{}, 031601 (2002) \[arXiv:hep-th/0109174\]. H. Boschi-Filho and N. R. F. Braga, Eur. Phys. J. C [**32**]{}, 529 (2004) \[arXiv:hep-th/0209080\]. H. Boschi-Filho and N. R. F. Braga, JHEP [**0305**]{}, 009 (2003) \[arXiv:hep-th/0212207\].
A. Karch, E. Katz, D. T. Son and M. A. Stephanov, Phys. Rev. D [**74**]{}, 015005 (2006) doi:10.1103/PhysRevD.74.015005 \[hep-ph/0602229\].
S. J. Brodsky, G. F. de Teramond, H. G. Dosch and J. Erlich, Phys. Rept. [**584**]{}, 1 (2015) doi:10.1016/j.physrep.2015.05.001 \[arXiv:1407.8131 \[hep-ph\]\].
H. R. Grigoryan and A. V. Radyushkin, Phys. Rev. D [**76**]{}, 095007 (2007) doi:10.1103/PhysRevD.76.095007 \[arXiv:0706.1543 \[hep-ph\]\].
H. R. Grigoryan, P. M. Hohler and M. A. Stephanov, Phys. Rev. D [**82**]{}, 026005 (2010) doi:10.1103/PhysRevD.82.026005 \[arXiv:1003.1138 \[hep-ph\]\].
N. R. F. Braga, M. A. M. Contreras and S. Diles, arXiv:1507.04708 \[hep-th\].
S. Hong, S. Yoon and M. J. Strassler, JHEP [**0404**]{}, 046 (2004) \[hep-th/0312071\].
Y. Kim, J. P. Lee and S. H. Lee, Phys. Rev. D [**75**]{}, 114008 (2007) \[hep-ph/0703172 \[HEP-PH\]\].
M. Fujita, K. Fukushima, T. Misumi and M. Murata, Phys. Rev. D [**80**]{}, 035001 (2009) doi:10.1103/PhysRevD.80.035001 \[arXiv:0903.2316 \[hep-ph\]\].
J. Noronha and A. Dumitru, Phys. Rev. Lett. [**103**]{}, 152304 (2009) doi:10.1103/PhysRevLett.103.152304 \[arXiv:0907.3062 \[hep-ph\]\].
M. Fujita, T. Kikuchi, K. Fukushima, T. Misumi and M. Murata, Phys. Rev. D [**81**]{}, 065024 (2010) \[arXiv:0911.2298 \[hep-ph\]\].
T. Branz, T. Gutsche, V. E. Lyubovitskij, I. Schmidt and A. Vega, Phys. Rev. D [**82**]{}, 074022 (2010) \[arXiv:1008.0268 \[hep-ph\]\].
T. Gutsche, V. E. Lyubovitskij, I. Schmidt and A. Vega, Phys. Rev. D [**87**]{}, no. 5, 056001 (2013) \[arXiv:1212.5196 \[hep-ph\]\].
S. S. Afonin and I. V. Pusenkov, Phys. Lett. B [**726**]{}, 283 (2013) \[arXiv:1306.3948 \[hep-ph\]\].
K. Hashimoto, N. Ogawa and Y. Yamaguchi, JHEP [**1506**]{}, 040 (2015) \[arXiv:1412.5590 \[hep-th\]\].
N. Evans and A. Tedder, Phys. Lett. B [**642**]{}, 546 (2006) doi:10.1016/j.physletb.2006.10.019 \[hep-ph/0609112\].
S. S. Afonin, Phys. Rev. C [**83**]{}, 048202 (2011) doi:10.1103/PhysRevC.83.048202 \[arXiv:1102.0156 \[hep-ph\]\].
S. S. Afonin, Int. J. Mod. Phys. A [**27**]{}, 1250171 (2012) doi:10.1142/S0217751X12501710 \[arXiv:1207.2644 \[hep-ph\]\].
P. Colangelo, F. Giannuzzi and S. Nicotri, Phys. Rev. D [**80**]{}, 094019 (2009) doi:10.1103/PhysRevD.80.094019 \[arXiv:0909.1534 \[hep-ph\]\].
A. S. Miranda, C. A. Ballon Bayona, H. Boschi-Filho and N. R. F. Braga, JHEP [**0911**]{}, 119 (2009) doi:10.1088/1126-6708/2009/11/119 \[arXiv:0909.1790 \[hep-th\]\].
L. A. H. Mamani, A. S. Miranda, H. Boschi-Filho and N. R. F. Braga, JHEP [**1403**]{}, 058 (2014) doi:10.1007/JHEP03(2014)058 \[arXiv:1312.3815 \[hep-th\]\].
E. Witten, Adv. Theor. Math. Phys. [**2**]{}, 505 (1998) \[hep-th/9803131\].
S. W. Hawking and D. N. Page, Commun. Math. Phys. [**87**]{}, 577 (1983). doi:10.1007/BF01208266
C. P. Herzog, Phys. Rev. Lett. [**98**]{}, 091601 (2007) doi:10.1103/PhysRevLett.98.091601 \[hep-th/0608151\].
C. A. Ballon Bayona, H. Boschi-Filho, N. R. F. Braga and L. A. Pando Zayas, Phys. Rev. D [**77**]{}, 046002 (2008) doi:10.1103/PhysRevD.77.046002 \[arXiv:0705.1529 \[hep-th\]\].
D. T. Son and A. O. Starinets, JHEP [**0209**]{}, 042 (2002) doi:10.1088/1126-6708/2002/09/042 \[hep-th/0205051\].
N. Iqbal and H. Liu, Phys. Rev. D [**79**]{}, 025023 (2009) doi:10.1103/PhysRevD.79.025023 \[arXiv:0809.3808 \[hep-th\]\].
R. C. Myers, A. O. Starinets and R. M. Thomson, JHEP [**0711**]{}, 091 (2007) doi:10.1088/1126-6708/2007/11/091 \[arXiv:0706.0162 \[hep-th\]\].
K. Suzuki, P. Gubler, K. Morita and M. Oka, Nucl. Phys. A [**897**]{}, 28 (2013) doi:10.1016/j.nuclphysa.2012.09.011 \[arXiv:1204.1173 \[hep-ph\]\].
G. Aarts, C. Allton, S. Kim, M. P. Lombardo, M. B. Oktay, S. M. Ryan, D. K. Sinclair and J. I. Skullerud, JHEP [**1111**]{}, 103 (2011) doi:10.1007/JHEP11(2011)103 \[arXiv:1109.4496 \[hep-lat\]\].
A. Adare [*et al.*]{} \[PHENIX Collaboration\], Phys. Rev. C [**91**]{}, no. 2, 024913 (2015) doi:10.1103/PhysRevC.91.024913 \[arXiv:1404.2246 \[nucl-ex\]\].
A. Mocsy and P. Petreczky, Phys. Rev. Lett. [**99**]{}, 211602 (2007) doi:10.1103/PhysRevLett.99.211602 \[arXiv:0706.2183 \[hep-ph\]\].
C. A. Dominguez, M. Loewe, J. C. Rojas and Y. Zhang, Phys. Rev. D [**81**]{}, 014007 (2010) doi:10.1103/PhysRevD.81.014007 \[arXiv:0908.2709 \[hep-ph\]\].
C. A. Dominguez, M. Loewe, J. C. Rojas and Y. Zhang, Nucl. Phys. Proc. Suppl. [**207-208**]{}, 273 (2010) doi:10.1016/j.nuclphysbps.2010.10.070 \[arXiv:1009.1169 \[hep-ph\]\].
C. A. Dominguez, M. Loewe and Y. Zhang, Phys. Rev. D [**88**]{}, 054015 (2013) doi:10.1103/PhysRevD.88.054015 \[arXiv:1307.5766 \[hep-ph\]\].
N. R. F. Braga, M. A. M. Contreras and S. Diles, arXiv:1511.06373 \[hep-th\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The thermal response of nonequilibrium systems requires the knowledge of concepts that go beyond entropy production. This is showed for systems obeying overdamped Langevin dynamics, either in steady states or going through a relaxation process. Namely, we derive the linear response to perturbations of the noise intensity, mapping it onto the quadratic response to a constant small force. The latter, displaying divergent terms, is explicitly regularized with a novel path-integral method. The nonequilibrium equivalents of heat capacity and thermal expansion coefficient are two applications of this approach, as we show with numerical examples.'
address:
- '$^1$Max Planck Institute for Mathematics in the Sciences, Inselstr. 22, 04103 Leipzig, Germany'
- '$^2$ Dipartimento di Fisica ed Astronomia, Università di Padova, Via Marzolo 8, I-35131 Padova, Italy'
- '$^3$ INFN, Sezione di Padova, Via Marzolo 8, I-35131 Padova, Italy'
author:
- 'G Falasco$^{1}$ and M Baiesi$^{2,3}$'
bibliography:
- 'bib\_noneq\_150723.bib'
title: Nonequilibrium temperature response for stochastic overdamped systems
---
Introduction
============
The determination of response functions is arguably one of the most topical issues in statistical physics. Even though its history dates back to the works of Einstein, Nyquist and Onsager [@ein05; @nyq28; @ons31; @ons31b], it was Kubo [@kub57; @tod92] who subsumed the later developments [@cal51; @gre52; @gre54] under a general theory. For a system slightly driven off equilibrium, the Kubo formula gives the linear response of an observable in terms of the equilibrium time-correlation between the observable itself and the entropy produced by the perturbation. The first systematic application of Kubo’s theory —along with kinetic theories based on generalized Boltzmann equations— underscored the endeavor to calculate the transport coefficients of moderately dense gasses [@ern69]. These efforts culminated in the discovery of the algebraic decay in time of the correlation functions entering Kubo formulas [@ald70; @dor70; @zwa70], which prevents the existence of transport coefficients in low dimensions.
Later, the possibility to perform progressively more efficient computer simulations and thus to compute response functions numerically, led to the extension of the original theory to thermostatted systems arbitrarily perturbed from an initial equilibrium state [@eva90]. Remarkably, it was established that the (nonlinear) response to an external driving is largely insensitive to the choice of the thermostatting mechanisms [@eva93b], represented by the artificial forces required to maintain nonequilibrium steady-state conditions [@ron07].
In contrast to such major achievements, the related theory for the response upon perturbation of nonequilibrium states has progressed far more slowly. Apart from the obvious obstacle represented by the lack of knowledge of nonequilibrium phase-space distributions, further difficulties are met when dealing rigorously with deterministic dynamical systems, owing to the fractal nature of their invariant distribution [@rue09; @mar08; @col12; @col14]. Nonequilibrium response theories have rather flourished for stochastic dynamics [@han78; @fal90; @cug94; @rue98; @nak08; @che08; @spe06; @spe09; @sei10; @pro09; @ver11; @lip05; @lip07; @bai09; @bai09b], which is applicable to a wide variety of complex systems in physics as well as in related sciences. However, most of these results are usually restricted to [*mechanical*]{} perturbations and do not consider [*thermal*]{} perturbations. Thus, they do not allow one to compute quantities such as nonequilibrium heat capacities and thermal expansions coefficients, which would arise as the (integrated) linear response to step variations of the temperature, i.e., of the noise intensity in the stochastic dynamical equations. Besides some previous formal results [@che09b; @sei10], only recently there appeared formulas for the thermal response of driven stochastic systems, which are given in terms of correlations between state observables calculated in the unperturbed state. Apparently, the mathematical difficulties entailed by handling noise variations require either to introduce an explicit time-discretization to avoid divergences in the response [@bai14; @yol15] or to rely on a rescaling of the stochastic dynamics in order to derive regular results [@fal15b].
The present work is devoted to show that neither of these expedients is actually necessary. A well-defined thermal response formula can be derived by standard path integral techniques, in close analogy to the case of deterministic perturbations. After introducing the model equations in section \[sec:model\], we define in section \[sec:path\] the linear response to a temperature perturbation of a generic observable of the system. In section \[sec:map\] after a brief explanation of the formal differences from the ordinary response to a deterministic forcing, we tackle the problem first showing that the thermal response is equivalent to a portion of the quadratic (i.e. second-order) response to a constant force. Such expression, which displays divergent terms, is then explicitly regularized in section \[sec:reg\] and is showed to be equivalent to a Kubo formula in equilibrium. In section \[sec:num\] we illustrate two applications of these results: the energy susceptibility of a driven quenched particle (that is the non-equilibrium specific heat for zero driving) and the thermal expansion coefficient of an anharmonic lattice subjected to large heat flows. Moreover, in the simplest tractable case of a freely diffusing particle we connect our formulas to the Einstein relation. A summary and an outlook are finally given in the conclusions.
Overdamped Langevin dynamics {#sec:model}
============================
The overdamped diffusive system we consider consists of $N$ degrees of freedom, denoted ${\boldsymbol{x}} = \{x_1,\dots x_N \}$. For instance, $x_j$ may be a component of a particle position vector in $d$-dimensions, so that $N= n d$ if the system is composed by $n$ particles. The dynamics is given by the overdamped Langevin equation $$\begin{aligned}
\label{ods}
\dot x_j(t) = \mu_j F_j({\boldsymbol{x}}(t)) + \sqrt{2 \mu_j T_j} \xi_j(t),\end{aligned}$$ where each Gaussian white noise $\xi_j$ is uncorrelated from the others, $$\begin{aligned}
\label{noise}
{{\left< \xi_j(t)\xi_{j'}(t') \right>}} = \delta(t-t') \delta_{jj'} .\end{aligned}$$ The $j$-th bath temperature $T_j$ and mobility $\mu_j$ (which is the inverse of a damping constant) determine the strength of the noise term, while the drift depends on $\mu_j$ and on the mechanical force $F_j({\boldsymbol{x}}(t))$. Such structure respects local detailed balance and thus assumes that the baths are noninteracting with each other and always in equilibrium, regardless of the nonequilibrium conditions experienced by the system. Temperatures and mobilities in our formalism do not depend on the coordinates, hence there is no ambiguity in the interpretation of the stochastic equation. Throughout this paper we will always consider the Stratonovich convention, that is the midpoint rule is employed to discretise in time [@sek10]. Hence none of the integrals will be of the Ito type and the rules of standard calculus can be applied.
In this context of temperature response, even if equations have a noise prefactor that does not depend on the system’s state ${\boldsymbol{x}}$, it turns out that the choice of using Stratonovich path-weights rather than Ito ones is not trivial. As discussed previously [@yol15], by differentiating with respect to temperature one proves a response formula that depends on the choice of the path-weight. One can check that the formulas in this paper are indeed different from those found adopting the Ito convention [@bai14]. Ultimately, the path-weight, and thus the corresponding discretization of , have to be chosen consistently with the physical process that is meant to model. The Ito convention, for example, is by construction suitable for numerical data generated by integration of with the Euler scheme [@bai14]. On the other hand, the Wong-Zakai theorem [@oksendal] ensures the Stratonovich convention to be adequate to experimental data, for which the white noise is an idealized limit of the short correlation times of the microscopic degrees of freedom.
The $F_i$’s are generic nonconservative forces that may bring the system arbitrarily far from equilibrium. In the resulting statistical averages, denoted ${{\left< \ldots \right>}}$, there is an understood dependence on the initial density of states $\rho_0({\boldsymbol{x}}_0)$, with ${\boldsymbol{x}}_0= {\boldsymbol{x}}(0)$. This may coincide or not with the steady state density. Finally, we introduce the backward generator of the Markovian dynamics , written as a sum of “one-coordinate" operators ${{\mathbb L}}_j$, $$\begin{aligned}
\label{L}
{{\mathbb L}}= \sum_{j=1}^N {{\mathbb L}}_j
\quad \textrm{with} \quad
{{\mathbb L}}_j = \mu_j F_j({\boldsymbol{x}}) \partial_{j} + \mu_jT_j \partial^2_{j},\end{aligned}$$ where we set $\partial_{x_j} \equiv \partial_j$ to avoid clutter. It gives the average time derivative of a state observable ${{\mathcal O}}(t)$ as $\frac{{d}}{{d}t} {{\left< {{\mathcal O}}(t) \right>}}= {{\left< {{\mathbb L}}{{\mathcal O}}(t) \right>}}$. Hereafter for any state observable we use the shorthand notation ${{\mathcal O}}({\boldsymbol{x}}(t),t) \equiv {{\mathcal O}}(t)$ to indicate the implicit (and possibly explicit) dependence on the time $t$.
Linear response in path integral formalism {#sec:path}
==========================================
We imagine to perturb the system varying the noise amplitude through a time dependent parameter ${\theta(t) \ll 1}$ switched on at time $t=0$, namely $$\begin{aligned}
\label{Theta}
T_i\, \to\, \Theta_i(t)\equiv T_i + \epsilon_i \theta(t),\end{aligned}$$ where $\epsilon_i$ is a constant determining the $i$-th amplitude of the perturbation. This renders for a perturbed degree of freedom into the form $$\begin{aligned}
\label{ods_h}
\dot x_i(t) = \mu_i F_i({\boldsymbol{x}}(t)) + \sqrt{2 \mu_i \Theta_i(t)} \xi_i(t).\end{aligned}$$ Without loss of generality we assume the mobility to be independent of temperature. The extension to the case where $\mu_i=\mu_i(\Theta_i)$ does not involve particular difficulties, since the linear response would be just the sum of the temperature response here described plus a standard response to a deterministic perturbation [@bai09; @bai09b], which arises linearising the term $\mu_i F_i$.
The aim is to calculate the linear response of a generic observable ${{\mathcal O}}(t)$ to the just introduced temperature change, defined by $$\begin{aligned}
\label{resp}
R_{{{\mathcal O}},\theta}(t,t')
\equiv \frac{\delta{{\left< {{\mathcal O}}(t) \right>}}_{\theta}}{\delta \theta(t')}\Bigg|_{\theta=0}
= \frac{\delta}{\delta \theta(t')} \int {{\mathcal D}}{\boldsymbol{x}}_{\theta} {d}{\boldsymbol{x}}_0 {{\mathcal O}}(t) P_{\theta}[{\boldsymbol{x}}] \rho_0({\boldsymbol{x}}_0)\Bigg|_{\theta=0} .\end{aligned}$$
Here ${{\left< \dots \right>}}_{\theta}$ denotes an average performed in the perturbed dynamics starting from the state $\rho_0({\boldsymbol{x}}_0)$, which is unaltered by the perturbation. The associated path weight, proportional to the probability of a trajectory $[{\boldsymbol{x}}] \equiv \{{\boldsymbol{x}}(s): 0 \leqslant s \leqslant t\}$ solution of , is expressed as [@zinn02] $$\begin{aligned}
\label{path}
P_{\theta}[{\boldsymbol{x}}]= \exp {{\mathcal A}}_{\theta}[{\boldsymbol{x}}],\end{aligned}$$ with the action functional $$\begin{aligned}
\label{action}
{{\mathcal A}}_{\theta}[{\boldsymbol{x}}] =
-\sum_{j=1}^{N} \bigg\{\int_0^t {d}s \frac{(\dot x_j(s) - \mu_j F_j(s))^2}{4\mu_j \Theta_j(s)}
+ \frac {\mu_j} 2 \int_0^t {d}s \partial_j F_j(s) \bigg\}.\end{aligned}$$ The last term in appears as the functional Jacobian in deriving the the path-weight for $[{\boldsymbol{x}}]$ from the Gaussian path-weight associated to the noise $\xi_i$, and depends on the convention used to discretise (e.g. it would be absent with the Ito convention). In the following we will also make use of the unperturbed action ${{\mathcal A}}\equiv {{\mathcal A}}_{\theta}|_{\theta=0}$, which amounts to replace $\Theta_j$ with $T_j$ in .
Deep physical insights come from separating any action of the form into time-antisymmetric (${{\mathcal S}}$) and time-symmetric (${{\mathcal K}}, {{\mathcal K}}_0$) components: $$\begin{aligned}
\label{Asmart}
{{\mathcal A}}[{\boldsymbol{x}}] &= \frac 1 2 {{\mathcal S}}[{\boldsymbol{x}}] - {{\mathcal K}}[{\boldsymbol{x}}] - {{\mathcal K}}_0[{\boldsymbol{x}}]\end{aligned}$$ with $$\begin{aligned}
\label{term-act-S}
{{\mathcal S}}[{\boldsymbol{x}}] & \equiv \sum_{j=1}^N \frac 1 {T_j}\int_0^t {d}s F_j(s) \dot x_j(s)\,,\\
\label{term-act-K}
{{\mathcal K}}[{\boldsymbol{x}}] & \equiv \sum_{j=1}^N \int_0^t {d}s
\frac{\mu_j}{4 T_j} \left[ F_j^2(s)+ 2 T_j \partial_j F_j(s) \right],\\
\label{term-act-K'}
{{\mathcal K}}_0[{\boldsymbol{x}}] & \equiv \sum_{j=1}^N \int_0^t {d}s \frac{\dot x_j^2(s)}{4\mu_j T_j} \,.\end{aligned}$$ The integrated entropy flux ${{\mathcal S}}[{\boldsymbol{x}}]$ is the antisymmetric part of the action ${{\mathcal A}}$ under the time-reversal transformation $x_j(s) \to x_j(t-s)$. It is defined consistently with thermodynamics as the sum of the individual heat fluxes into the reservoirs, each weighted by the respective bath temperature [@sek10]. The time-symmetric terms have been studied in connection with the notion of dynamical activity, formerly introduced in the context of jump systems [@lec05; @mer05; @gar09], where it counts the number of jumps and provides important informations, e.g., on the state of glassy systems. Both ${{\mathcal K}}[{\boldsymbol{x}}]$ and ${{\mathcal K}}_0[{\boldsymbol{x}}]$ in fact may quantify an amount of activity in the diffusive system we are considering [@ful13]. Being ${{\mathcal K}}_0[{\boldsymbol{x}}]$ related to the mean square displacement of the $N$ degrees of freedom, it offers a direct estimate of the trajectory frenzy. Nevertheless, this kinetic-like term should be understood as part of the functional measure [@zinn02 Sec. 2.2], as it selects from all possible trajectories the Brownian paths that make ${{\mathcal K}}_0$ finite in the limit $ds \to 0$ (i.e. those that satisfy $d x_j^2 \sim ds$). The functionals ${{\mathcal S}}$ and ${{\mathcal K}}$ are then the statistical weights of such selected trajectories. Therefore, in the following we will reserve the name [*dynamical activity*]{} for ${{\mathcal K}}$, which was shown to be a good measure of the system activity [@ful13]. Written as $$\begin{aligned}
{{\mathcal K}}[{\boldsymbol{x}}] &\equiv \int_0^t {d}s V_{\rm eff}({\boldsymbol{x}}(s)),\end{aligned}$$ it may be seen as a time-integral of a state variable $V_{\rm eff}({\boldsymbol{x}})$ that, for systems with interactions deriving from an energy potential $U({\boldsymbol{x}})$ and with a global bath temperature $T$, would read $$V_{\rm eff}({\boldsymbol{x}}) = \frac{1}{4 T} \sum_j \mu_j \left[ (\partial_jU({\boldsymbol{x}}))^2- 2 T \partial^2_j U({\boldsymbol{x}}) \right].$$ Such quantity was called effective potential [@aut09; @pit11] and is proportional to the escape rate from a configuration ${\boldsymbol{x}}$, as the probability to remain in ${\boldsymbol{x}}$ for a short time $\Delta t$ is $\sim \exp( -V_{\rm eff} \Delta t)$. For our nonequilibrium systems we generalise such concept by writing $V_{\rm eff}(s) = \sum_{j=1}^N \lambda_j(s) $, with $$\begin{aligned}
\lambda_j(s)\equiv \frac{\mu_j}{4 T_j} \left[ F_j^2(s)+ 2 T_j \partial_j F_j(s) \right].\end{aligned}$$ The escape rate of the degree of freedom $x_j$, denoted $\lambda_j$, follows from evaluating the action at fixed ${\boldsymbol{x}}$ along a very short trajectory of duration $\Delta t \ll 1$, that is, $\lim_{\Delta t\to 0}\text{Prob}({\boldsymbol{x}},s+ \Delta t| {\boldsymbol{x}}, s)/\Delta t =\exp(-\sum_{j=1}^N \lambda_j(s))$.
In the following sections we will sometimes also use the name [*frenesy*]{} for describing correlation functions in the response formulas involving time-symmetrical features. This alternate naming originated in the response-theory framework [@bai13] and usually refers to quantities akin to ${{\mathcal K}}$ —more specifically, to its excess generated by a perturbing force—, namely to quantities assessing the system [*impatience*]{} for changing its state (rather than direct measures of the trajectory zigzags). Hopefully the double terminology is guiding the reader through the connections with the recent literature.
Response to heating as response to a force {#sec:map}
==========================================
We are now in the position to develop the thermal linear response theory, but we immediately find an obstacle. Since the path weight is normalised to one, $\int {{\mathcal D}}{\boldsymbol{x}}_{\theta} P_{\theta}[{\boldsymbol{x}}]=1$, the functional measure ${{\mathcal D}}{\boldsymbol{x}}_{\theta}$ in contains the noise temperatures $\Theta_j$ (see e.g. [@zinn02; @lau07]), and therefore depends itself on the external parameter $\theta$. This is a major difference with respect to an external perturbation of the deterministic forces, which leads to the formal difficulties reported in [@bai14], namely the introduction of an explicit time-mesh to avoid singularities in the results. To overcome this problem we first seek a more manageable expression for the path average. That is obtained through an Hubbard-Stratonovich transformation [@negele88] of the action that, introducing an auxiliary variable ${\boldsymbol{y}}$, linearises the quadratic term in and removes the $\theta$ dependence from the functional measure of the path weight (see e.g. [@lau07]). By doing so, it is easy to bring in the form (see \[sec:HS\]) $$\begin{aligned}
&\frac{\delta{{\left< {{\mathcal O}}(t) \right>}}_{\theta}}{\delta \theta(t')}\Bigg|_{\theta=0}
= \sum_{i}\frac{\epsilon_i}{\mu_i}R^{(2)}_{{{\mathcal O}},f_i}(t,t',t'). \label{R2}\end{aligned}$$ Here $R_{{{\mathcal O}},f_i}^{(2)}$ is the second-order response function to a constant force perturbation $f_i$ of the $i$-th degree of freedom [@lip08], namely $$\begin{aligned}
R^{(2)}_{{{\mathcal O}},f_i}(t,t',t'') \equiv \frac{\delta^2 {{\left< {{\mathcal O}}(t) \right>}}_{{\boldsymbol{f}}}}{\delta f_i(t')\delta f_i(t'')}\Bigg|_{{\boldsymbol{f}}=0},\end{aligned}$$ where ${{\left< \dots \right>}}_{{\boldsymbol{f}}}$ now denotes the average with respect to the perturbed dynamics $$\begin{aligned}
\label{const}
\dot x_i = \mu_i (F_i({\boldsymbol{x}})+f_i )+ \sqrt{2 \mu_i T_i} \xi_i.\end{aligned}$$ Formal calculation of response functions to external forces poses no technical difficulty [@cug94; @lip05; @bai09]. After integrating out the auxiliary variable ${\boldsymbol{y}}$, it is straightforward to find for $$\begin{aligned}
R^{(2)}_{{{\mathcal O}},f_i}(t,t',t')
&= \frac{1}{2 T_i} \frac{\delta}{\delta f_i(t')} {{\left< (\dot x_i(t')- \mu_i F_i(t') -\mu_i f_i(t')) {{\mathcal O}}(t) \right>}}_{{\boldsymbol{f}}} \bigg|_{{\boldsymbol{f}}=0} \nonumber \\
&= \frac{1}{4 T_i^2} \Big[ {{\left< (\dot x_i(t') - \mu_i F_i(t'))^2{{\mathcal O}}(t) \right>}} - 2\mu_i T_i \delta(0) {{\left< {{\mathcal O}}(t) \right>}} \Big]. \label{delta}\end{aligned}$$ Summing up, a standard Hubbard-Stratonovich transformation has allowed us to write the linear response of an observable ${{\mathcal O}}$ to a temperature change as the second-order response to a state-independent force, thus arriving at the intermediate result $$\begin{aligned}
\label{response}
R_{{{\mathcal O}},\theta}(t,t')= \sum_i \frac{\epsilon_i}{4 \mu_i T_i^2} \Big[ &
\Big<{{\mathcal O}}(t) \big(\dot x_i^2(t') -2\mu_i \dot x_i(t') F_i(t')
+ \mu_i^2 F_i^2(t')\big) \Big> - 2 \mu_i T_i \delta(0){{\left< {{\mathcal O}}(t) \right>}} \Big]. \end{aligned}$$ As anticipated, this result is slightly different from that of a previous approach [@bai14] where the Ito convention was adopted for the path-integrals.
Regularization of the response {#sec:reg}
==============================
In the divergence caused by the Dirac delta formally compensates the divergence in the squared velocity. This can be heuristically understood recalling that , despite being formally expressed in continuous time notation, can be interpreted in terms of discrete, albeit small, time intervals $\Delta t$ [@zinn02; @gro98]. Therefore one has $\dot x_i^2\sim 1/\Delta t$, being the dynamics diffusive at short times, and clearly $\delta(0) \sim 1/\Delta t$. However, it would be convenient to recast as an explicit result devoid of singular terms. In the following we perform such operation, first for a single degree of freedom ($N=1$), and then extending the result to arbitrary $N$.
One degree of freedom {#sec:one-dof}
---------------------
With one degree of freedom the parameter $\epsilon_i$ is superfluous and is thus set to $1$. We first focus on the kinetic-like term by starting with the rewriting (valid for $t > t'$) $$\begin{aligned}
\label{diff}
{{\left< \dot x^2(t') {{\mathcal O}}(t) \right>}}=\frac{1}{2}\frac{{d}^2 \phantom{c}}{{d}t'^2}{{\left< x^2(t') {{\mathcal O}}(t) \right>}} -{{\left< \ddot x(t') x(t') {{\mathcal O}}(t) \right>}},\end{aligned}$$ and by seeking a replacement for the correlation function ${{\left< \ddot x(t') x(t') {{\mathcal O}}(t) \right>}}$. This can be achieved recalling that the integral of a total derivative involving the path weight is null. Therefore, we may exploit the identity $$\begin{aligned}
\label{identity}
0 = \int {{\mathcal D}}x \frac{\delta}{\delta x(t')} {{\mathcal B}}[x] P[x] ={{\left< \frac{\delta {{\mathcal B}}}{\delta x(t')} \right>}} + {{\left< {{\mathcal B}}\frac{\delta {{\mathcal A}}}{\delta x(t')} \right>}}\,,\end{aligned}$$ where ${{\mathcal B}}$ is any functional of $\{x(s):0 \leqslant s \leqslant t \}$, and ${{\mathcal A}}[x]$ is the unperturbed action $$\label{S}
{{\mathcal A}}[x]=-\frac{1}{4\mu T}\int_0^t {d}s (\dot x(s) - \mu F(s))^2 - \frac \mu 2 \int_0^t {d}s \partial_x F(s)\,,$$ corresponding to calculated at $\theta=0$, with $N=1$. First, we evaluate the second term in making use of the expression for the functional variation of the action derived in \[sec:path\_rel\], see . The entropy variation is shown to vanish, while the variation of ${{\mathcal K}}[{\boldsymbol{x}}]$ expressed in terms of the backward generator ${{\mathbb L}}$ gives $$\begin{aligned}
{{\left< {{\mathcal B}}\frac{\delta {{\mathcal A}}}{\delta x(t')} \right>}}=
{{\left< {{\mathcal B}}\frac{\delta {{\mathcal K}}}{\delta x(t')} \right>}}=
\frac{1}{2\mu T} {{\left< {{\mathcal B}}\Big[ \ddot x(t')- \mu {{\mathbb L}}F(t')\Big] \right>}}\label{deS}\,.\end{aligned}$$ Hereafter we restrict to the case in which $F$ does not depend explicitly on time, but only via $x$. In order to extract from the sought substitute for ${{\left< \ddot x(t') x(t') {{\mathcal O}}(t) \right>}}$, we choose ${{\mathcal B}}= {{\mathcal O}}(t) x(t')$ and the first term in becomes $$\begin{aligned}
\label{diffA}
{{\left< \frac{\delta {{\mathcal B}}}{\delta x(t')} \right>}}= {{\left< \frac{\delta {{\mathcal O}}(t)}{ \delta x(t')} x(t') \right>}} + {{\left< {{\mathcal O}}(t) \right>}}\delta(0).\end{aligned}$$ If ${{\mathcal O}}$ is a state observable, i.e., it depends only on the trajectory endpoint, the first term on the right hand side of drops for all $t' \neq t$, since it reads $\frac{\delta {{\mathcal O}}(t)}{ \delta x(t')}= \partial_x {{\mathcal O}}(t) \delta(t-t')$. Putting all the pieces together we get the compact expression $$\begin{aligned}
{{\left< \ddot x(t') x(t') {{\mathcal O}}(t) \right>}}& = \mu {{\left< {{\mathcal O}}(t) x(t') {{\mathbb L}}F(t') \right>}} -2\mu T \delta(0),\end{aligned}$$ which, plugged in the response formula , gives finally $$\begin{aligned}
\label{R1}
R_{{{\mathcal O}},\theta}(t,t') = \frac{1}{4 T^2}\bigg[
& \frac{1}{2 \mu}\frac{{d}^2 \phantom{c}}{{d}t'^2}{{\left< {{\mathcal O}}(t) x^2(t') \right>}}+ {{\left< {{\mathcal O}}(t) \mu F^2(t') \right>}} \nonumber\\
& -{{\left< {{\mathcal O}}(t) x(t'){{\mathbb L}}F(t') \right>}}
-2{{\left< {{\mathcal O}}(t) \dot x(t') F(t') \right>}}
\bigg].\end{aligned}$$ for $t' < t$. This is a regularised version of valid for $N=1$ and any state observable ${{\mathcal O}}$. We have traded the kinetic-like term and the Dirac delta in with a second-order time derivative and a correlation involving the backward generator. The second-order time derivative, even tough unusual for a linear response formula (but not for a second-order response function [@lip08]), is indeed necessary to obtain the correct result, as it can be easily verified in the analytically solvable case of a particle in free diffusion (see section \[sec:free\]).
If one is interested in the response of path-dependent observables (namely, ${{\mathcal O}}$ is a functional of the trajectory up to time $t$), the first summand in is non-zero and hence has to be supplemented by the term $- 2 \mu T {{\left< \frac{\delta {{\mathcal O}}(t)}{ \delta x(t')} x(t') \right>}} $. As an example we may consider the heat exchanged with the thermal bath in a time $t$, ${{\mathcal Q}}[x]\equiv\int_0^t {d}s F(s) \dot x(s)$. It turns out that the response formula requires no additional term in this case, since $$\begin{aligned}
\frac{\delta {{\mathcal Q}}(t)}{ \delta x(t')}&=\partial_x F(t') \dot x(t')+ \int_0^t {d}s \dot \delta(s-t') F(s) \nonumber\\
&=\partial_x F(t') \dot x(t')- \partial_x F(t') \dot x(t')=0\,. \label{Q1}\end{aligned}$$
Many degrees of freedom {#sec:many-dof}
-----------------------
The procedure is easily extended to a system composed of $N>1$ degrees of freedom. Equations , and are still valid replacing $x$ with $x_i$, and taking the action (corresponding to calculated at $\theta=0$) $$\begin{aligned}
\label{action_un}
{{\mathcal A}}[{\boldsymbol{x}}] = &
-\sum_{j=1}^{N} \bigg\{ \frac{1}{4\mu_j T_j}\int_0^t {d}s (\dot x_j(s) - \mu_j F_j(s))^2
+ \frac {\mu_j} 2 \int_0^t {d}s \partial_j F_j(s) \bigg\},\end{aligned}$$ where we reverted to the notation accommodating the particle labels. Equation is then generalised to (see \[sec:path\_rel\]) $$\begin{aligned}
{{\left< {{\mathcal B}}\frac{\delta {{\mathcal A}}}{\delta x_i(t')} \right>}} &= \frac{1}{2 \mu_i T_i} {{\left< {{\mathcal B}}\ddot x_i(t') \right>}}
- {{\left< {{\mathcal B}}\frac{\delta {{\mathcal K}}}{\delta x_i(t')} \right>}}
+ \frac 1 2 {{\left< {{\mathcal B}}\frac{\delta {{\mathcal S}}}{\delta x_i(t')} \right>}}.
\label{noifU}\end{aligned}$$ In the following we focus on systems with two-body potential interactions, deferring the more general result (valid for arbitrary $d$, generic driving and interactions) to \[sec:path\_rel\]. Yet, the results reported here are general enough to describe the thermal response of a broad class of non-equilibrium systems, such as heat conducting lattices in contact with different heat baths \[Eq. \], and aging systems \[Eq. \]. Under the above assumption, the variation of ${{\mathcal K}}[{\boldsymbol{x}}]$ in is given by $$\begin{aligned}
\label{varK}
\frac{\delta {{\mathcal K}}}{\delta x_i(t')} = {{\mathbb L}}^{(T_i)}F_i(t'),\end{aligned}$$ where we identified the operator $$\begin{aligned}
\label{genLT}
{{\mathbb L}}^{(T_i)} & \equiv \sum_{j=1}^N \frac{T_i}{T_j} {{\mathbb L}}_j
\, = \sum_{j=1}^N \left( \frac{T_i}{T_j} \mu_j \partial_j + T_i \partial_j^2 \right)\end{aligned}$$ which acts on the observables as if all temperatures were equal to $T_i$ and all forces $F_j$ were rescaled by $T_i/T_j$. Interesting, this rescaling is found by rewriting the Langevin dynamics in terms of a new time variable, the *thermal time* $\tau_j \equiv t \frac{T_j}{T_i}$, by which reads $$\begin{aligned}
\label{odstau}
\frac{{d}x_j}{{d}\tau_j} = \mu_j \frac{T_i}{T_j} F_j + \sqrt{2 \mu_j T_i} \xi_j.\end{aligned}$$ While ${{\mathbb L}}$ is the generator of the stochastic dynamics in the kinematic time $t$, in view of , the operator ${{\mathbb L}}^{(T_i)}$ acts as the generator of the corresponding dynamics in thermal time coordinates. This permits to rationalize the variation of the dynamical activity as the tendency to change $F_i$ measured with respect to the thermal time.
Coming back to the regularization of we operate as before. We choose ${{\mathcal B}}= {{\mathcal O}}(t) x_i(t')$ and obtain, by means of , and $$\begin{aligned}
{{\left< {{\mathcal O}}(t)\ddot x_i(t') x_i(t') \right>}}= & \mu_i {{\left< {{\mathcal O}}(t) x_i(t') {{\mathbb L}}^{(T)} F_i(t') \right>}}
+ \mu_i T_i {{\left< {{\mathcal O}}(t) x_i(t') \frac{\delta {{\mathcal S}}}{\delta x_i(t')} \right>}} -2\mu_i T_i \delta(0),\end{aligned}$$ where a state observable ${{\mathcal O}}$ was considered. Finally, using the explicit form of the entropy variation , we find for the response function ($t' < t$) $$\begin{aligned}
\label{many_dof}
R_{{{\mathcal O}},\theta}(t,t') =
\sum_i \frac{\epsilon_i}{4 T_i^2} \bigg[ & \frac 1 {2 \mu_i} \frac{{d}^2}{{d}t'^2}{{\left< {{\mathcal O}}(t)x_i^2(t') \right>}}
- {{\left< {{\mathcal O}}(t) x_i(t') {{\mathbb L}}^{(T_i)} F_i(t') \right>}} \nonumber\\
&+ {{\left< {{\mathcal O}}(t) \mu_i F_i^2(t') \right>}}
- 2 {{\left< {{\mathcal O}}(t) \mu_i \dot x_i(t') F_i(t') \right>}}
\nonumber\\ &
+ \sum_{j=1}^N {{\left< {{\mathcal O}}(t) [x_i \dot x_j \partial_j F_i](t') \right>}}\left(\frac{T_i}{T_j}-1\right)\bigg].\end{aligned}$$ This equation simplifies if the system is *isothermal* before the perturbation is applied, i.e., the heat reservoirs are all at the same temperature $T_j= T \, \forall j$. In this case $\frac{\delta {{\mathcal S}}}{\delta x_i}$ vanishes and boils down to $$\begin{aligned}
{{\left< {{\mathcal B}}\frac{\delta {{\mathcal A}}}{\delta x_i(t')} \right>}} =& \frac{1}{2 \mu_i T_i} {{\left< {{\mathcal B}}\Big[ \ddot x_i(t')- \mu_i {{\mathbb L}}F_i(t') \Big] \right>}},\end{aligned}$$ once we recognise $ {{\mathbb L}}^{(T_i)}|_{T_j=T} = \sum_{j=1}^N {{\mathbb L}}_j $ as the total generator of the dynamics in the complete state space. Consequently, for isothermal systems the response formula takes the simpler form ($t' < t$) $$\begin{aligned}
R_{{{\mathcal O}},\theta}(t,t') = \frac{1}{4 T^2}\sum_i \epsilon_i \bigg[
\frac 1 {2 \mu_i} \frac{{d}^2}{{d}t'^2}{{\left< {{\mathcal O}}(t) x_i^2(t') \right>}}
- {{\left< {{\mathcal O}}(t) x_i(t') {{\mathbb L}}F_i(t') \right>}} & \nonumber\\
+ {{\left< {{\mathcal O}}(t) \mu_i F_i^2(t') \right>}}
-2 {{\left< {{\mathcal O}}(t) \dot x_i(t') F_i(t') \right>}} &
\bigg] ,\label{R3}\end{aligned}$$ which is a straightforward generalization of to a many-body system.
As noted above, if ${{\mathcal O}}$ is a path-dependent observable one needs to include in the response formula the additional term $$\begin{aligned}
- 2 \mu_i T_i {{\left< \frac{\delta {{\mathcal O}}(t)}{ \delta x_i(t')} x_i(t') \right>}},\end{aligned}$$ coming from the first summand of . For the example of the total heat flux into the reservoirs, ${{{\mathcal Q}}[{\boldsymbol{x}}]\equiv\sum_{j=1}^N \int_0^t {d}s F_j(s) \dot x_j(s)}$, the supplementary term contains $$\begin{aligned}
\frac{\delta {{\mathcal Q}}(t)}{ \delta x_i(t')}&=\sum_{j=1}^N \big(\partial_i F_j(t') - \partial_j F_i(t')\big)\dot x_j(t')\,,\end{aligned}$$ and thus vanishes when the interactions derive from a two-body potential.
Susceptibility
--------------
Upon integration of we get an equation for the susceptibility of the system, $$\begin{aligned}
\label{chi}
\chi_{{{\mathcal O}},\theta}(t) \equiv & \int_0^t {d}t' R_{{{\mathcal O}},\theta}(t,t') = S_1 + S_2 + K_1 + K_2\end{aligned}$$ with
$$\begin{aligned}
\label{chi_S1}
S_1 &= - {{\left< {{\mathcal O}}(t)\,
\sum_i \frac{\epsilon_i}{2 T_i^2}\int_0^t d t' \dot x_i(t') F_i(t') \right>}}
\\
\label{chi_S2}
S_2 &=
{{\left< {{\mathcal O}}(t) \,
\sum_i \frac{\epsilon_i}{4 T_i^2}
\sum_{j=1}^N \left(\frac{T_i}{T_j}-1 \right) \int_0^t d t' [x_i \dot x_j \partial_j F_i](t') \right>}}
\\
\label{chi_K1}
K_1 &= {{\left< {{\mathcal O}}(t)\,
\sum_i \frac{\epsilon_i}{4 T_i^2} \int_0^t d t' \left[
\mu_i F_i^2(t') + x_i(t') {{\mathbb L}}^{(T_i)} F_i(t') \right] \right>}}
\\
\label{chi_K2}
K_2 & = \left.\frac{{d}}{{d}t'}{{\left< {{\mathcal O}}(t) \,
\sum_i \frac{\epsilon_i}{8 \mu_i T_i^2} x_i^2(t') \right>}}\right|_{t'=0}^{t'=t}\end{aligned}$$
where we recall that integrals are in the Stratonovich sense and ${{\mathbb L}}^{(T_i)}$ was introduced in . The term $S_1$ is the standard correlation between observable and entropy production, appearing with a $1/2$ prefactor with respect to the equilibrium version (see next section), in which it would be the only correlation relevant for determining the linear response. The term $S_2$ is a novel correlation between observable and a time-antisymmetric quantity, proportional to the functional variation of the bath entropy $\frac{\delta {{\mathcal S}}[{\boldsymbol{x}}]}{\delta x}$, which may be non-zero only if $T_j \ne T_i$ for some $j$. The remaining correlations, the [*frenetic*]{} terms [@bai13] $K_1$ and $K_2$, collect correlations between the observable and time-symmetric dynamical features. As in previous studies of force perturbations, both $S$’s and $K$’s contain, respectively, the entropy and [*frenesy*]{} [@bai13] in excess due to the perturbation.
In order to correctly evaluate the time derivative of the correlation in $K_2$, when dealing with data it is important to avoid taking discrete-time derivatives with $t'>t$ because cusps are not unusual in correlation functions for $t'\to t$. To compute numerically $\frac{d}{d t'}{{\left< x_i^2(t') {{\mathcal O}}(t) \right>}}|_{t'=t}$, in the examples of the following section we have estimated the slope of data for ${{\left< x_i^2(t') {{\mathcal O}}(t) \right>}}$ with $t' \lesssim t$.
Only if averages are evaluated in a steady state, $K_2$ can be modified as $$\begin{aligned}
\label{chi_K2ss}
K_2^{s} & = {{\left< {{\mathbb L}}{{\mathcal O}}(t) \sum_i \frac{\epsilon_i}{8 \mu_i T_i^2}[x_i^2(0)-x_i^2(t)] \right>}} \end{aligned}$$
because $$\begin{aligned}
\frac{d}{d t'}{{\left< {{\mathcal O}}(t) x_i^2(t') \right>}}\bigg|_{t'=0}^{t'=t}
= -\frac{d}{d t}{{\left< {{\mathcal O}}(t) x_i^2(t') \right>}}\bigg|_{t'=0}^{t'=t}
= -{{\left< {{\mathbb L}}{{\mathcal O}}(t) x_i^2(t') \right>}}\bigg|_{t'=0}^{t'=t}\end{aligned}$$ A steady state susceptibility $\chi^s_{{{\mathcal O}},\theta}(t) = S_1 + S_2 + K_1 + K_2^s$ is associated with $K_2^s$.
A steady state formula and its reduction to the Kubo formula at equilibrium {#subsec:Kubo}
---------------------------------------------------------------------------
Every thermal response formulation should reduce to the standard Kubo formula when the system is under complete equilibrium conditions at temperature $T$. These conditions are met if conservative forces $F_i=- \partial_i U$ (with $U({\boldsymbol{x}})$ the system’s energy) are present, if $T_i=T \,\, \forall i$ and the perturbation is applied to a thermalised system, namely $\rho_0({\boldsymbol{x}})$ is the canonical distribution at temperature $T$. In equilibrium, the Kubo formula expresses the response function as $$\begin{aligned}
\label{FDT}
R^{\rm Kubo}_{{{\mathcal O}}, \theta}(t-t') =& \frac{1}{T^2} \frac{d}{d t'}{{\left< {{\mathcal O}}(t) U(t') \right>}},\end{aligned}$$ and the corresponding susceptibility is $$\begin{aligned}
\label{FDT_chi}
\chi^{\rm Kubo}_{{{\mathcal O}}, \theta}(t) =& \frac{1}{T^2} {{\left< {{\mathcal O}}(t) [U(t)-U(0)] \right>}},
\nonumber\\
=& \frac{1}{T^2} {{\left< {{\mathcal O}}(t) {{\mathcal Q}}(t) \right>}},\end{aligned}$$ where ${{\mathcal Q}}(t) = U(t)-U(0)$ is the heat transferred to the system in the time interval $[0,t]$. This formula shows that the temperature response in equilibrium is totally determined by the correlation between observable and the entropy ${{\mathcal Q}}(t)/T$ [*paid*]{} by the reservoir to change the system energy.
When a global perturbation is applied to an isothermal steady state regime, say with $\epsilon_i=1 \,\, \forall i$, eq. may be recast in an alternative form, that correctly reduces to the Kubo formula in equilibrium, as we show in the following. In the derivation we stay in a generic steady state condition until the very end, so that in turn we obtain another quite general formula for the response function, eq. below, in which the genuine nonequilibrium contribution is well distinguished from the Kubo correlation. A possible practical issue of such elegant separation is that it can be computed explicitly only if one knows the microscopic probability density of states.
We start noticing that the last term in is in equilibrium half of the expected result: $$\begin{aligned}
-\frac{1}{2 T^2} \sum_i {{\left< {{\mathcal O}}(t) \dot x_i(t')\partial_i U (t') \right>}} =\frac{1}{2 T^2} \frac{d}{d t'} {{\left< {{\mathcal O}}(t) U(t') \right>}}. \label{half}\end{aligned}$$ The remaining frenetic terms yield an analogous contribution at equilibrium. To show that, we first use that the system is in a stationary state. This implies that correlations are functions of the time difference only, hence $\frac{d}{d t'}$ can be exchanged with $-\frac{d}{d t}$. Moreover, the backward generator can be expressed in terms of the generator of the time-reversed dynamics, ${{\mathbb L}}^*$, through the relation ${{\mathbb L}}={{\mathbb L}}^*+ 2 \sum_{j=1}^N v_j \partial_j$, where ${\boldsymbol{v}}_j \equiv J^s_j/\rho^s $ is the state velocity, that is the probability current $J^s_j$ associated to $x_j$, over the steady state density of the system $\rho^s$ [@che09; @bai13]. We will ultimately exploit the time-reversal invariance of equilibrium states, which formally manifests in the equality ${{\mathbb L}}={{\mathbb L}}^*$, as the probability currents $v_j$ are by definition absent at equilibrium.
The time derivatives in can be manipulated as $$\begin{aligned}
\frac{1}{2 \mu_i}\frac{d^2}{d t'^2}{{\left< x_i^2(t') {{\mathcal O}}(t) \right>}}&= - \frac{1}{2 \mu_i}\frac{d}{d t'} \frac{d}{d t}{{\left< x_i^2(t') {{\mathcal O}}(t) \right>}} \nonumber\\
&= - \frac{1}{2 \mu_i}\frac{d}{d t'} {{\left< x_i^2(t') {{\mathbb L}}{{\mathcal O}}(t) \right>}} \nonumber \\
&= - \frac{1}{2 \mu_i}\frac{d}{d t'} {{\left< x_i^2(t') \left({{\mathbb L}}^*+ 2\sum_j v_j \partial_j \right){{\mathcal O}}(t) \right>}} \nonumber\\
&= - \frac{1}{2 \mu_i}\frac{d}{d t'} \left[ {{\left< \left({{\mathbb L}}x_i^2(t') \right) {{\mathcal O}}(t) \right>}} + 2 \sum_j{{\left< x_i^2(t') v_j \partial_j {{\mathcal O}}(t) \right>}} \right] \nonumber \\
& = - \frac{d}{d t'} \left[{{\left< x_i(t')F_i(t') {{\mathcal O}}(t) \right>}} + 2\sum_j {{\left< x_i^2(t') v_j \partial_j {{\mathcal O}}(t) \right>}} \right]
\label{first_quarter}\end{aligned}$$ Together with stationarity, we used that ${{\mathbb L}}^*$ is the adjoint of ${{\mathbb L}}$, and the equality ${{\mathbb L}}x_i^2= 2\mu_i F_i x_i + \text{const}$ in the last passage. We then turn to the second and third summand in , starting with the rewriting $\mu_i F_i^2= F_i{{\mathbb L}}x_i$: $$\begin{aligned}
{{\left< {{\mathcal O}}(t) \left( F_i(t') {{\mathbb L}}x_i(t') - x_i(t') {{\mathbb L}}F_i(t') \right) \right>}}&= {{\left< {{\mathcal O}}(t) \left( F_i(t') [{{\mathbb L}}, x_i](t') - x_i(t') [ {{\mathbb L}},F_i] (t') \right) \right>}} \nonumber \\
&= - {{\left< {{\mathcal O}}(t) \left( F(t') \dot x(t') - x(t') \dot F (t') \right) \right>}} . \label{second_quarter}\end{aligned}$$ Here we introduced the commutator acting as, e.g., $[x_i,{{\mathbb L}}] \equiv x_i{{\mathbb L}}- {{\mathbb L}}x_i$, and exploit the fact that in the operator formalism time derivatives within average values are given by $\dot {{\mathcal O}}= [{{\mathcal O}}, {{\mathbb L}}]$, for any state observable ${{\mathcal O}}$ (see \[sec:comm\]). Putting together equations , and we obtain an expression of the thermal response valid under stationary isothermal conditions, $$\begin{aligned}
\label{R_with_v}
R_{{{\mathcal O}}, \theta}(t,t') =-
\frac{1}{T^2} \sum_i \bigg[ {{\left< O(t) \dot x_i(t') F_i(t') \right>}}
+ \frac{1}{ \mu_i} \frac{d}{d t'}{{\left< \sum_j v_j(t) \partial_j O(t) \, x^2_i(t') \right>}} \bigg].\end{aligned}$$ Finally, at equilibrium the Kubo formula is retrieved by setting $v_j=0 \,\,\forall j$ and using the rewriting for potential forces. Equation is a thermal response counterpart of previous results for the steady-state force response based on the notion of state velocity [@spe06; @che08].
Examples {#sec:num}
========
Specific heat for a quenched toy system
---------------------------------------
In this first example we want to highlight that this framework is valid not only for steady states but also for transient regimes. There is to recall an understood dependence of the statistical averages ${{\left< \ldots \right>}}$ on the initial density of states $\rho_0$.
Let us consider a paradigmatic model of nonequilibrium overdamped systems, namely a single particle in a periodic potential $U(x) = \cos x$ and subject to an additional constant force $f$, for simplicity with mobility $\mu=1$. Thus $F(x) = \sin x + f$, in the evolution equation of the unperturbed system. The backward operator acts on the force as ${{\mathbb L}}F(x) = \sin x \cos x - T \sin x$.
To generate a transient condition we choose to thermalise the particle at $T_0\ne T$ and to switch to $T$ only at $t=0$, when the perturbation is also applied. In this way, even for $f=0$ one cannot apply the Kubo formula for equilibrium systems, as the initial state in not in equilibrium at temperature $T$. Due to the periodic potential, as an arbitrary procedure for obtaining a well defined $\rho_0(x)$, we shift to the interval $[0,2 \pi]$ any $x$ obtained from a long simulation run. However, averages such as ${{\left< x^2(t') {{\mathcal O}}(t) \right>}}$ need to be computed with $x$ interpreted as a non-periodic coordinate. We adopted a Heun scheme [@sek10] to integrate the stochastic equation, because it yields trajectories that are consistent with the Stratonovich path-weights used in our theory.
![Temperature susceptibility of the energy $U(x) = \cos x$ of a single particle, computed with the formula $(\chi)$ and by actually perturbing the system $(\chi^h)$. Also the single terms of the formula are shown. The system is out of equilibrium because of a quench at time $t=0$ from an initial $T_0=5$ to $T=0.3$. Consistently, the response is not given by twice the correlation $S_1$ between entropy produced and observable. In (a) there is no additional constant force ($f=0$), while $f=0.7$ in (b) generates a nonequilibrium steady state previous to the quench. Averages are over $4 \times 10^7$ trajectories, integrated with finite time step $dt=2.5 \times 10^{-3}$. []{data-label="fig:1p1"}](fig_long_1a.pdf "fig:"){width="48.00000%"} ![Temperature susceptibility of the energy $U(x) = \cos x$ of a single particle, computed with the formula $(\chi)$ and by actually perturbing the system $(\chi^h)$. Also the single terms of the formula are shown. The system is out of equilibrium because of a quench at time $t=0$ from an initial $T_0=5$ to $T=0.3$. Consistently, the response is not given by twice the correlation $S_1$ between entropy produced and observable. In (a) there is no additional constant force ($f=0$), while $f=0.7$ in (b) generates a nonequilibrium steady state previous to the quench. Averages are over $4 \times 10^7$ trajectories, integrated with finite time step $dt=2.5 \times 10^{-3}$. []{data-label="fig:1p1"}](fig_long_1b.pdf "fig:"){width="48.00000%"}
In figure \[fig:1p1\] we show examples of susceptibilities of the internal energy (${{\mathcal O}}= U$) to a change of $T$ for $T_0=5$ and $T=0.3$, both for $f=0$ and $f=0.7$. We compare the susceptibility $\chi_{U,\theta}(t)$ from with that computed directly as $$\chi_{U,\theta}^h(t) = \frac{{{\left< U(t) \right>}}_{\theta=h} - {{\left< U(t) \right>}}_{\theta=0}}{h}$$ with $h = T/100$ active from $t=0$ on. We note that, for $f=0$, the force $F$ is potential and thus the heat exchanged with the bath reduces to an energy difference, ${{\mathcal Q}}= -\int_0^t d t' \partial_x U(t') \dot x(t')= U(0)-U(t)$. Therefore, the susceptibility of the energy gives in the long-time limit the specific heat $C$ of the system: $$\begin{aligned}
C\equiv -\lim_{t\to \infty} \int_0^t d t' \frac{\delta {{\left< {{\mathcal Q}}(t) \right>}}_{\theta}}{\delta \theta(t')}
\Bigg|_{\theta=0}=
\lim_{t\to \infty} \chi_{U,\theta}(t) \end{aligned}$$ If a Kubo formula were valid, twice the entropic term would yield the response. One can note that this is not the case, rather all terms in the response formula are relevant for determining the correct form of the susceptibility. In these examples, in particular, the term is especially important. Being the derivative of a correlation function, it is however the noisiest one. One could resort to some high-frequency filtering for better results. In the example of the following subsection we will show that is a good alternative to in case one is dealing with steady states.
Thermal expansion in a temperature gradient
-------------------------------------------
In equilibrium at a given temperature $T$, the correlation function between the heat absorbed by a system and its length may be used to predict the thermal expansion response. In this example we show how this picture breaks down out of equilibrium, where, as exposed in the previous sections, one needs to know also correlations between length and time-symmetric observables, given by and or , as well as the new entropic form due to temperature unbalances. This example specialises to steady state conditions but, with respect to the previous examples, it includes the more general setup of multiple heat baths, in which one can exploit the general formulation with perturbation amplitudes $\epsilon_i$.
Let us consider the $N$ degrees of freedom arranged in a one-dimensional chain. The system has an energy $$U({\boldsymbol{x}}) = \frac {x_1^2}2 + \sum_{i=1}^{N-1} u(x_{i+1}- x_i),
\qquad\textrm{with}\quad u(r) = \frac{(r-1)^4}4 + r -\frac 1 4$$ which determines the forces, $F_i({\boldsymbol{x}})= -\partial_i U({\boldsymbol{x}})$, and again mobilities $\mu_i$ are set equal to $1$ for simplicity. The $x_1^2/2$ term is a pinning potential on the first site, and $x_i$’s represent the displacements from the average positions. The length of the system in excess with respect to the length at zero temperature, $X\equiv x_N-x_1$, increases on average for increasing $T_i$’s due to the asymmetric two-body potential $u(r)$ (see the inset of figure \[fig:chain\](b)). As a paradigm of nonequilibrium conditions, the system is driven by a set of temperatures varying linearly from $T_1$ to $T_N>T_1$.
![Temperature steady-state susceptibility of the length $X$ of the overdamped chain ($N=11$), computed with the formula $(\chi^s)$ and by actually perturbing the system $(\chi^h)$. Also here the single terms of the formula are displayed. In these examples, $T_i$ varies linearly from $T_1=1$ to $T_N=2$. In (a) the response is to a global temperature rise, while in (b) it is to an increase of the gradient $T_N-T_1$ preserving the average bath temperature (the inset shows the interaction potential). Averages are over $10^7$ trajectories, integrated with finite time step $dt=10^{-3}$. []{data-label="fig:chain"}](fig_long_2a.pdf "fig:"){width="48.00000%"} ![Temperature steady-state susceptibility of the length $X$ of the overdamped chain ($N=11$), computed with the formula $(\chi^s)$ and by actually perturbing the system $(\chi^h)$. Also here the single terms of the formula are displayed. In these examples, $T_i$ varies linearly from $T_1=1$ to $T_N=2$. In (a) the response is to a global temperature rise, while in (b) it is to an increase of the gradient $T_N-T_1$ preserving the average bath temperature (the inset shows the interaction potential). Averages are over $10^7$ trajectories, integrated with finite time step $dt=10^{-3}$. []{data-label="fig:chain"}](fig_long_2b.pdf "fig:"){width="48.00000%"}
We study the response of the length $X$ to temperature variations, in the form of (a) a global constant increase of the temperatures given by a constant $\epsilon_i=1$, and (b) an increment of the gradient $T_N-T_1$, chosen so that the average temperature is unaltered by varying $\epsilon_i$ linearly from $\epsilon_1=-1$ to $\epsilon_N=1$. For both cases, in figure \[fig:chain\] we see that the susceptibility $\chi^s_{X,\theta}$ computed with the steady state term agrees fairly well with the direct estimate of the response, $$\chi_{X,\theta}^h(t) = \frac{{{\left< X(t) \right>}}_{\theta=h} - {{\left< X(t) \right>}}_{\theta=0}}{h},$$ obtained with a constant $h=0.005$ turned on at $t=0$. From figure \[fig:chain\] one also sees that the entropic and frenetic terms have opposite trends, between each other and with switched roles in the two cases, complementing each other to sum up to the correct response level. In figure \[fig:chain\](b) we also show the response $\chi_{X,\theta}$ obtained by an evaluation of the time-derivative in (the local variation in time of the correlation function is obtained through a linear fit of data relative to four nearby time steps). It results more noisy than the estimate via $\chi^s_{X,\theta}$.
Free diffusion of one degree of freedom {#sec:free}
---------------------------------------
Let us consider the equations of motion for free diffusion of a single degree of freedom, $\dot x(t)=\hat \xi(t)$ with $\hat \xi=\sqrt{2 \mu T} \xi$. The noise prefactor $\sqrt{2 \mu T}$ comes from assuming the bath to be in equilibrium. In this way the mean square displacement of a free particle in a time $t$ is simply ${{\left< x^2(t) \right>}} = 2 \mu T t \equiv 2 D t$, the response of the mean velocity to a small force is the free-particle mobility $\mu$, and the Einstein relation $\mu = D/ T$ between diffusion constant $D$ and mobility is found. One can note that the susceptibility of the observable ${{\mathcal O}}(t) = x^2(t)$ to a change of $T$ is expected to be $2 \mu t$, hence the corresponding response function is $2\mu$. We show how our formalism reduces to this result.
For free diffusion all terms in drop but the one involving the second derivative. In this case, the response function can be calculated directly from its definition and one can thus prove analytically that both sides of are equal to the same quantity. As we argued above, the response of the mean square displacement to the perturbation $T\to\Theta(t) = T+\theta(t)$ is $$\begin{aligned}
\frac{\delta {{\left< x^2(t) \right>}}_{ h}}{\delta \theta(t')} &=\frac{\delta }{\delta \theta(t')} \Big<x_0^2 +2 x_0 \int_0^t {d}s \hat \xi(s) + \int_0^t {d}s \int_0^t {d}u \hat \xi(s) \hat \xi(u)\Big>_{ \theta}
\nonumber\\
& = 2\mu \frac{\delta }{\delta \theta(t')} \int_0^t {d}s \int_0^t {d}u \delta(s-u) \Theta(s) \nonumber\\
& = 2 \mu.\end{aligned}$$ where we used that the initial condition is independent of the perturbation and noise, thus only the noise autocorrelation contributes. On the other hand, the response formula becomes $$\begin{aligned}
R_{x^2 T}(t,t') & = \frac{1}{8 \mu T^2} \frac{{d}^2}{{d}t'^2}{{\left< x^2(t') x^2(t) \right>}}\\
&= \frac{1}{8 \mu T^2}\frac{{d}^2}{{d}t'^2}\left( {{\left< x^2(t') \right>}}{{\left< x^2(t) \right>}} +2 {{\left< x(t') x(t) \right>}}^2\right),\nonumber\end{aligned}$$ making use of Wick’s theorem to split the 4-point correlation into products of 2-point correlations. The latter read $$\begin{aligned}
{{\left< x(t')x(t) \right>}} &= {{\left< x_0^2 \right>}} + \int_0^{t'} {d}s \int_0^t {d}u {{\left< \xi(s)\xi(u) \right>}}= {{\left< x_0^2 \right>}} + 2 \mu T \min(t',t),\end{aligned}$$ leading to a result in agreement with the previous calculation: $$\begin{aligned}
R_{x^2 T}(t,t') &= \frac{1}{8 \mu T^2} \frac{{d}^2}{{d}t'^2}\Big[
\left( {{\left< x_0^2 \right>}} + 2 \mu Tt' \right) \left( {{\left< x_0^2 \right>}} + 2 \mu Tt \right)+ 2 \left( {{\left< x_0^2 \right>}} + 2 \mu Tt' \right)^2 \Big] \nonumber \\
&= \frac{1}{8 \mu T^2} \frac{{d}^2}{{d}t'^2}\Big[ 3{{\left< x_0^2 \right>}}^2 + 2\mu T {{\left< x_0^2 \right>}}( t' + t ) + (2\mu T)^2 t t' + 8 \mu T t' {{\left< x_0^2 \right>}} + 2 (2 \mu T)^2 {t'}^2 \Big] \nonumber\\
&= 2 \mu.\end{aligned}$$ As expected, interchanging $\frac{{d}}{{d}t'^2}$ with $\frac{{d}}{{d}t^2}$ would give an incorrect result as the system is not in a steady state. It is also trivial to verify , namely that the this result coincides with the second order response to a state-independent force, giving rise to the dynamics $ \dot x(t)=\mu f(t)+\hat \xi(t)$. Indeed, using again the conditions of independency of the initial condition, one finds $$\begin{aligned}
\frac 1 \mu \frac{\delta^2 {{\left< x^2(t) \right>}}_{ f}}{\delta f^2(t')}
& = \frac 1 \mu \frac{\delta^2 }{\delta f^2(t')} {{\left< \left(\int_0^t {d}s \big(\mu f(s)- \hat \xi(s)\big) \right)^2 \right>}}\nonumber\\
& = \mu \frac{\delta^2 }{\delta f^2(t')} \int_0^t {d}s \int_0^t {d}u f(s) f(u)
\nonumber\\&
= 2 \mu.\end{aligned}$$
Conclusions
===========
For overdamped stochastic systems far from equilibrium we have obtained the linear response function of generic state observables to a change in the temperature of the Langevin heat baths. Improving a previous result [@bai14], we need not express the response in terms of a finite time mesh, being all the divergencies appearing in the continuous limit removed, and being all terms in the susceptibility standard integrals or derivatives. This was achieved by deriving a sort of Dyson-Schwinger equation [@zinn02], i.e., a relation between unperturbed correlation functions involving an arbitrary observable. This method complements and expands our recent results [@fal15b] obtained via a different approach, in which the additional noise stemming from the perturbation was turned into mechanical forces by means of a space rescaling.
As in many previous examples, in order to describe a nonequilibrium systems, one needs to know more than just the entropy production. The additional information concerns the knowledge of dynamical quantities that are even under the reversal of the arrow of time (squares of forces, etc.). Among them we have recognized the change of the time-integral of the effective potential (i.e., the total escape rate integrated along trajectories) upon variation of the perturbed degree of freedom, $\frac{\delta {{\mathcal K}}}{ \delta x_i}$. This quantity emerges from the regularization procedure we set up, along with the change of the total bath entropy flow $\frac{\delta {{\mathcal S}}}{ \delta x_i}$, which complements, perhaps surprisingly, the usual entropy production entering Kubo formula.
For the common scenario of isothermal systems in a steady state, we have also shown how to convert the results in a formula that separates the Kubo term from a nonequilibrium additional correlation that includes the state velocity, see . Such version is complementary to the others in the sense that it requires the knowledge of the density of states rather than that of dynamical details.
Future developments of this framework should include multiplicative noise, i.e. those cases where the temperature experienced by the particle depends on their positions.
Derivation of the second order response function {#sec:HS}
================================================
The derivation of starts with a Hubbard-Stratonovich transformation of the path weight, which is a functional generalization of the integral identity $\int {d}y e^{-Dy^2-{\rm i} z y}= e^{-\frac{z^2}{4D}} \sqrt{\frac \pi D}$ (below the $\sqrt{\pi/D}$ is adsorbed in the path measure ${{\mathcal D}}y$) valid for real $y$ and $D>0$. When applied to and it renders the response in the form through the following manipulations: $$\begin{aligned}
\frac{\delta{{\left< {{\mathcal O}}(t) \right>}}_{\theta}}{\delta \theta(t')}\Bigg|_{\theta=0} =
& \frac{\delta}{\delta \theta(t')} \int {{\mathcal D}}{\boldsymbol{x}} {d}{\boldsymbol{x}}_0 {{\mathcal D}}{\boldsymbol{y}} \rho_0({\boldsymbol{x}}_0) {{\mathcal O}}(t) \times \nonumber\\
& \quad \prod_{j=1}^N \exp \bigg\{ - \int_{0}^{t} {d}s \bigg[
\mu_j \Theta_j y_j^2 - \mathrm{i} y_j (\dot x_j - \mu_j F_j)
+ \frac{1}{2}\mu_j \partial_j F_j \bigg]\bigg\} \Bigg|_{\theta=0}
\nonumber\\
=
& \int {{\mathcal D}}{\boldsymbol{x}} {d}{\boldsymbol{x}}_0 {{\mathcal D}}{\boldsymbol{y}} \rho_0({\boldsymbol{x}}_0) {{\mathcal O}}(t) \bigg[-\sum_i\epsilon_i \mu_i y_i^2(t')\bigg]
\times
\nonumber\\
& \quad \prod_{j=1}^N\exp\bigg\{
- \int_{0}^{t} {d}s \bigg[
\mu_j T_j y_j^2 - \mathrm{i} y_j (\dot x_j - \mu_j F_j) +\frac{1}{2}\mu_j \partial_jF_j\bigg]\bigg\} \label{y^2}\\
=
& \sum_i\frac{\epsilon_i}{\mu_i}\frac{\delta^2}{\delta f_i ^2(t')}
\int {{\mathcal D}}{\boldsymbol{x}} {d}{\boldsymbol{x}}_0 {{\mathcal D}}{\boldsymbol{y}} \rho_0({\boldsymbol{x}}_0) {{\mathcal O}}(t)
\times
\label{forceh}\\
& \quad \prod_{j=1}^N\exp\bigg\{ - \int_{0}^{t} {d}s \bigg[
\mu_j T_j y_j^2 - \mathrm{i}y_j (\dot x_j - \mu_j F_j-\mu_j f_j)
+ \frac{1}{2}\mu_j \partial_j F_j \bigg]\bigg\}\Bigg|_{{\boldsymbol{f}}=0} \nonumber\\
=
&
\sum_i\frac{\epsilon_i}{\mu_i}R^{(2)}_{{{\mathcal O}},f_i}(t,t',t'),
\nonumber\end{aligned}$$ where we rewrote introducing the derivatives of a state-independent force $f_i$, and recognised in the Martin-Siggia-Rose path-weight [@mar73] associated to the perturbed dynamics .
Variation of the action functional {#sec:path_rel}
==================================
Here we detail the calculation of the functional variation of the path-weight action ${{\mathcal A}}[{\boldsymbol{x}}]$ that was used in Section \[sec:reg\]. For the sake of clarity we distinguish the single-particle from the many-particle case.
One degree of freedom {#one-degree-of-freedom}
---------------------
For $N=1$, the action is given by and its variation is $$\begin{aligned}
\label{deA}
\frac{\delta {{\mathcal A}}}{\delta x(t')} &= \frac 1 2 \frac{\delta {{\mathcal S}}}{\delta x(t')}
- \frac{\delta {{\mathcal K}}}{\delta x(t')} + \frac{\ddot x (t')}{2 \mu T}\,.\end{aligned}$$ The variation of the bath entropy is identically zero, unless $F$ is an explicit function of time $F(t')=F(x(t'),t')$: $$\begin{aligned}
\frac{\delta {{\mathcal S}}}{ \delta x(t')}&=\frac{1}{T}\left(\partial_x F(t') \dot x(t')+ \int_0^t {d}s \dot \delta(s-t') F(s)\right) \nonumber\\
&=\frac 1 T \left(\partial_x F(t') \dot x(t')-\partial_{t'} F(t')- \partial_x F(t') \dot x(t') \right)=- \frac 1 T \partial_{t'} F(t')\,. \nonumber\end{aligned}$$ Since the dynamical activity is independent of $\dot x$, its variation is simply the derivative of the escape rate from $x(t')$: $$\begin{aligned}
\label{deK}
\frac{\delta {{\mathcal K}}}{ \delta x(t')}&=
\partial_x \lambda(t')=\frac 1 {2 T} \left(\mu F(t')\partial_x F(t') - \mu T \partial_x^2 F(t') \right).\end{aligned}$$ Therefore, introducing in the backward generator ${{\mathbb L}}$, becomes $$\begin{aligned}
\frac{\delta {{\mathcal A}}}{\delta x(t')}
&= \frac{1}{2\mu T} \Big[ \ddot x(t')- \mu {{\mathbb L}}F(t') - \mu \partial_{t'}F(t')\Big].\label{derivative} \end{aligned}$$ As a side note, plugging this result into with ${{\mathcal B}}=1$ one obtains (if $F$ deepens on $x$ only) $$\begin{aligned}
{{\left< \ddot x \right>}}= \mu{{\left< {{\mathbb L}}F \right>}},\end{aligned}$$ i.e., the mean trajectory satisfies Newton’s equation with an effective force $\mu {{\mathbb L}}F$. In the weak-noise limit $T\ll 1$, such trajectory becomes the most probable one, being the minimiser of the action. This expression could be obtained directly by applying the backward generator ${{\mathbb L}}$ to the Langevin equation , and using that $\xi$ does not depend on $x$.
Many degrees of freedom {#many-degrees-of-freedom}
-----------------------
For $N>1$, thanks to the independency of the different thermal noises, the action is simply the sum of “single-coordinate” actions: ${{\mathcal A}}[{\boldsymbol{x}}]= \sum_{j=1}^N {{\mathcal A}}^{(j)}[{\boldsymbol{x}}]$ with ${{\mathcal A}}^{(j)}$ following the structure . Nevertheless, its variation is not just equal to but in general it will contain additional terms owing to the interactions between different degrees of freedom. One indeed finds modified expressions for the variation of the total entropy flux into the (unperturbed) reservoirs, $$\begin{aligned}
\frac{\delta {{\mathcal S}}[{\boldsymbol{x}}] }{\delta x_i(t')} &= \sum_{j=1}^N \dot x_j(t')\left(\frac{\partial_i F_j(t')}{T_j} -\frac{\partial_j F_i(t')}{T_i} \right) - \frac{1}{T_i}\partial_{t'}F_i(t')\, ,\end{aligned}$$ and for the variation of the total dynamical activity $$\begin{aligned}
\label{deK_b}
\frac{\delta {{\mathcal K}}}{ \delta x(t')}&= \sum_{j=1}^N\partial_i \lambda_j(t')=
\sum_{j=1}^N\frac{1}{ 2 T_j } \left( \mu_j F_j(t')\partial_i F_j(t') + \mu_j T_j \partial_i\partial_j F_j(t') \right),\end{aligned}$$ which in general cannot be cast in terms of the total backward generator ${{\mathbb L}}$. The variation of the action is thus given by $$\begin{aligned}
\frac{\delta {{\mathcal A}}}{\delta x_i(t')}
& = \frac 1 2 \frac{\delta {{\mathcal S}}}{\delta x_i(t')}
- \frac{\delta {{\mathcal K}}}{\delta x_i(t')} + \frac{\ddot x_i (t')}{2 \mu_i T_i} \nonumber\\
& = \frac{1}{2 \mu_i T_i} \bigg[ \ddot x_i(t')- \mu_i T_i \sum_{j=1}^N\partial_i \lambda_j(t')- \mu_i \partial_{t'}F_i(t') \nonumber \\
& \qquad \qquad + \mu_i T_i \sum_{j=1}^N \dot x_j(t')\left(\frac{\partial_i F_j(t')}{T_j} -\frac{\partial_j F_i(t')}{T_i} \right)
\bigg] . \label{deStot} \end{aligned}$$ Equation is completely general, and thus, when combined with , provides a regularised expression for the thermal response of overdamped systems under any nonequilibrium conditions: $$\begin{aligned}
\label{general}
R_{{{\mathcal O}},\theta}(t,t') = \sum_i \frac{\epsilon_i}{4 T_i^2} \Bigg[
& \frac{1}{\mu_i}
\frac{{d}^2}{{d}t'^2}{{\left< {{\mathcal O}}(t)x_i^2(t') \right>}} - {{\left< {{\mathcal O}}(t) x_i(t')\partial_{t'}F_i(t') \right>}}\nonumber\\
& + {{\left< {{\mathcal O}}(t) \left( \mu_i F_i^2(t')-2 \dot x_i(t') F_i(t') -x_i(t') T_i
\sum_{j=1}^N \partial_i \lambda_j(t') \right) \right>}} \nonumber\\
&+ T_i \sum_{j=1}^N {{\left< {{\mathcal O}}(t) x_i(t') \dot x_j(t')
\left(\frac{\partial_i F_j(t')}{T_j} -\frac{\partial_j F_i(t')}{T_i} \right) \right>}}\Bigg].\end{aligned}$$
Nevertheless, the cross-terms $\partial_i F_j$ with $i \neq j$ appearing in simplify considerably if we assume that the degrees of freedom interact with each others via a two-body potential ${{\mathcal U}}(\{x_i-x_j\})$. Hence we can exploit the relation $$\begin{aligned}
\label{simply}
\partial_i F_j= - \partial_i \partial_j {{\mathcal U}}= - \partial_j \partial_i {{\mathcal U}}=\partial_j F_i,\end{aligned}$$ which is nothing but the action-reaction principle. Equation then becomes $$\begin{aligned}
\frac{\delta {{\mathcal A}}}{\delta x_i(t')} &= \frac{1}{2 \mu_i T_i}\bigg[ \ddot x_i(t')- \mu_i{{\mathbb L}}^{(T_i)} F_i(t')- \mu_i \partial_{t'}F_i(t') \bigg] + \sum_{j=1}^N \dot x_j(t') \partial_j F_i(t')\left(\frac{1}{2T_j} -\frac{1}{2T_i} \right) \,. \label{deStot_sim}\end{aligned}$$ We remark that for systems in $d=1$ does not impose any limitation on the driving, that is, one-body non-conservative forces can be present as well, they simply do not enter in , which concerns only the interactions between different particles. Instead, in $d>1$, different indexes $i$ and $j$ in may refer to the coordinates of the same particle, thus cannot be simplified to in the presence of generic non-conservative forces.
It is worth noting that when the equality $\partial_j F_i=\partial_i F_j$ holds, the choice ${{\mathcal B}}=1$ in the identity yields the effective Newton’s equation for the mean trajectory $$\begin{aligned}
{{\left< \ddot x_i \right>}}=\mu_i {{\left< {{\mathbb L}}^{(T)} F_i \right>}}-\mu_i T_i {{\left< \frac{\delta {{\mathcal S}}}{\delta x_i} \right>}}.\end{aligned}$$ On the other hand, direct application of the operator ${{\mathbb L}}$ to the Langevin equation gives ${{\left< \ddot x_i \right>}}=\mu_i {{\left< {{\mathbb L}}F_i \right>}}$. By comparison, one concludes that there exists a natural splitting of the effective force, namely $$\begin{aligned}
{{\left< {{\mathbb L}}F_i \right>}}=\mu_i {{\left< {{\mathbb L}}^{(T)} F_i \right>}}-\mu_i T_i {{\left< \frac{\delta {{\mathcal S}}}{\delta x_i} \right>}},\end{aligned}$$ where the first component originates from variations of the force $F_i$ in thermal time, while the second is a gradient-like force in which the entropy flux into the bath acts a free-energy.
Time derivative in operator formalism {#sec:comm}
=====================================
Consider the state observables ${{\mathcal O}}_\alpha$, that are arbitrary functions of ${\boldsymbol{x}}$. In the operator formalism, their (steady-state) evolution over a time-span $t-t'$ is given by the action of the operator $e^{{{\mathbb L}}(t-t')}$. Therefore, the typical correlation functions we are interested in are expressed by (with $t>t'$) $$\begin{aligned}
\label{eg.corr}
{{\left< {{\mathcal O}}_3(t) {{\mathcal O}}_2(t') {{\mathcal O}}_1(t') \right>}}= \int d {\boldsymbol{x}}_0 \rho_0({\boldsymbol{x}}_0) e^{{{\mathbb L}}t'} {{\mathcal O}}_1 {{\mathcal O}}_2 e^{{{\mathbb L}}(t-t')} {{\mathcal O}}_3,\end{aligned}$$ where the dependence of ${{\mathcal O}}_\alpha$ on ${\boldsymbol{x}}_0$ is omitted for brevity [@bai13]. In analogy to the Heisenberg picture in quantum mechanics, one may include the dependency on time in the observables by the definition ${{\mathcal O}}_\alpha(t')\equiv e^{{{\mathbb L}}t'} {{\mathcal O}}_\alpha e^{-{{\mathbb L}}t'} $. Hence, a time derivative applied to one of the operators in gives, e.g., $$\begin{aligned}
\label{eg.corr2}
{{\left< {{\mathcal O}}_3(t) \dot {{\mathcal O}}_2(t') {{\mathcal O}}_1(t') \right>}}&= {{\left< {{\mathcal O}}_3(t)({{\mathbb L}}e^{{{\mathbb L}}t'} {{\mathcal O}}_2 e^{-{{\mathbb L}}t'}- e^{{{\mathbb L}}t'} {{\mathcal O}}_2 e^{-{{\mathbb L}}t'} {{\mathbb L}}){{\mathcal O}}_1(t') \right>}}\nonumber \\
&={{\left< {{\mathcal O}}_3(t)({{\mathbb L}}{{\mathcal O}}_2(t') - {{\mathcal O}}_2(t') {{\mathbb L}}){{\mathcal O}}_1(t') \right>}}\nonumber\\
&={{\left< {{\mathcal O}}_3(t)[{{\mathbb L}},{{\mathcal O}}_2(t')] {{\mathcal O}}_1(t') \right>}}.\end{aligned}$$
References {#references .unnumbered}
==========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Since the identification of these stars by Morgan et al. in 1943, various definitions have been proposed for the stars of the Lambda Bootis group. We present here the various definitions which have been given to these objects in order to induce a general discussion on this topic.'
author:
- 'R. Faraggiana'
- 'M. Gerbaldi'
title: Definitions of the Lambda Boo stars
---
Introduction {#intr}
============
The criteria to detect this class of peculiar A-type stars rely upon the choices made by various authors in the last 50 years. Therefore several definitions of lambda Boo stars are found in the literature.
Both photometric and spectroscopic criteria have been used. Some definitions so far proposed, concern only stars of spectral type near A0 while others include A and F stars; no restrictions appear about the luminosity class and therefore the evolutionary stage of the lambda Boo stars.
The common character of these stars, according to the various definitions, is the weakness of the metallic lines; however requirements such as high $v\sin i$ and deficiency of specific elements are introduced by some, but not all authors; the same remark concerns their kinematic properties.
We shall present below the criteria used by various authors in order to understand the differences between the different lists of such stars published up to now, and to open a discussion for the future.
The discovery of the Lambda Boo Stars
=====================================
[**... a definition based on Spectral Classification criteria ...**]{}
Morgan et al. (1943) gave in fact implicitly the first definition of this group describing the peculiarities of Lambda Boo itself.
“The spectral type of Lambda Boo is near A0, as far as can it be determined. The spectral lines, while not unusually broad, are very weak, so that the only features easily visible are a weak K line and the Balmer series of hydrogen”. However these authors did not define a group of stars; this was done later on from spectroscopy by Burbidge and Burbidge (1956). Occasionally stars similar to Lambda Boo were discovered later by Slettebak. (Slettebak 1952 and 1954).
[**... and the first abundance analysis ...**]{}
Burbidge and Burbidge (1956) have analyzed Lambda Boo and 29 Cyg and they found a metal deficiency for the elements Mg, Ca, Fe, Sc, Ti and Sr. Later on, Baschek and Searle suggested that the oxygen abundance should be normal (quoted in the Annual Report of the Director, Mount Wilson and Palomar Observatories, 1962-63 page 12). This was shown by Kodaira (1967) with infra-red spectra.
Very soon the photometry was used to detect Lambda Boo stars (Parenago 1958).
In 1965 Sargent (1965) showed that these stars can be distinguished from other weak lined stars such as horizontal branch stars by the fact that their space velocities are those of Population I stars and that they have moderately large rotational velocities.
In 1968, Settebak et al. introduced the first spectroscopic definition of the lambda Boo class : “These objects are defined spectroscopically as A-type stars (as classified from Ca[ii]{} K line to Balmer line ratio) with weakened metallic lines. They may be distinguished from other stars with the same characteristics (such as horizontal branch stars) by the fact that all show moderately large rotational velocities and small space velocities”
From an abundance analysis of 5 so-defined Lambda Boo stars, Baschek and Searle (1969) found that only 3 of them (Lambda Boo, 29 Cygn and $\pi^1$ Ori) form a group from the composition point of view; these authors suggested that these stars constitute a type of peculiar A stars.
We recall that at that time, a list of 7 lambda Boo stars was available (see for example Sargent, 1965), detected by spectroscopy or photometry. Occasionally a star with low abundances was discovered and often related to the lambda Bootis stars (see for example ADS 3910B in Sargent, 1966).
So from the beginning some confusion exists in the literature about which objects should be considered as “lambda Boo stars”. It is clear that insufficient credit has been given to the true definition of this class found in Baschek & Searle (1969); “...we define the Lambda Boo stars as stars whose composition resembles that of Lambda Boo itself...”.
The eighties
============
The lambda Boo stars were forgotten until the 1980s when both photometric and spectroscopic researches underwent a revival of interest starting with the paper by Hauck and Slettebak (1983) where the spectroscopic definition was expanded to the A-F stars. Since then, different candidates have been selected by different criteria, either photometric (for example: Hauck (1986) or spectroscopic (for example: Abt 1984a).
At that time it became clear that the peculiarities of the Lambda Boo spectrum in the UV were easily detectable (Cucchiaro et al. 1980 ) even at the low TD1 resolution. Baschek et al. (1984) pointed out that characteristic features of the lambda Boo stars can be easily seen on low resolution IUE spectra. Faraggiana et al. (1990) extended this research and defined the UV criteria useful to detect the lambda Boo stars in the range 120-200 nm.
An extensive spectroscopic classification
=========================================
An extensive spectroscopic classification has been made by Gray (1988) in the classical wavelength range. Gray (1991) described in details the spectroscopic peculiarities detectable in the photographic domain at moderate resolution, providing a precise working definition of the lambda Boo stars : “Spectra of these stars are characterized by a weak Mg[ii]{} 4481 line, a K line type of A0 or slightly later and hydrogen-line type between A0 and F0. For their hydrogen line type, the metallic-line spectrum is weak”. Moreover their space velocities are those of Population I stars, and their rotational velocities are moderately high. The shape of the hydrogen lines profiles, peculiar in some lambda Boo stars, is introduced as a further criterium to separate two classes of these objects.
In 1990, Renson et al. collected all the stars that in the literature have been called lambda Boo or candidate lambda Boo at least once, as well as objects called “weak-lines” stars that may be lambda Boo candidates. This catalogue contains 101 stars.
The confusing situation is well illustrated by the two lists of lambda Boo candidates extracted from the same sample of stars (the Bright Star Catalogue) and based on similar classification by Abt (1984a), Abt and Morrell (1995) and by Gray (1988). Few stars are in common between these authors and some stars classified as Lambda Boo by Abt are considered normal by Gray and Garrison (1987).
On the basis of the selection by Gray (1988 and 1991) and UV criteria by Faraggiana et al. (1990), a list of stars fulfilling the visible and/or UV properties of lambda Boo itself has been established by Gerbaldi and Faraggiana (1993). We consider all these stars to be reliable Lambda Boo candidates on the basis of the fact that all the stars classified as Lambda Boo by Gray and observed by IUE belong to the same class according to UV criteria, and vice versa not one star among those rejected on the basis of UV criteria appears as lambda Boo in Gray’s list.
The nineties
============
The abundance determination by Venn and Lambert (1990), their interpretation of the abundance pattern of the lambda Boo stars and the IR excess detected by IRAS around some lambda Boo stars (Sadakane and Nishida 1986) were the starting points for observations in new directions :
- detection of lambda Boo stars in young clusters by Gray and Corbally (1993) and by Levato et al.(1994).
- detection of gas or dust shells around the Lambda Boo stars (Holweger & Rentzsch-Holm 1995; Grady et al. 1996; King & Patten 1992; Holweger & Stürenburg 1991; Bohlender & Walker 1994; King 1994; Hauck et al. 1995; Hauck et al. 1997)
- new abundance analysis (Stürenburg 1993)
- discussion on the position in the HR diagramme (Gerbaldi et al. 1993; Iliev & Barzova 1995)
- observations of Lambda Boo candidates with asteroseismic techniques (Paunzen et al. 1997) and detection of pulsating objects among them.
Charbonneau (1993) and Turcotte and Charbonneau (1993) computed diffusion effects in the atmosphere of a lambda Boo stars giving time scale for the duration of this phenomenon in the context of accreted material on the surface.
A new hypothesis on the origin of the Lambda Boo stars in the framework of the evolution of a close binary system has been recently developed by Andrievsky (1997).
The revival of “theoretical” interest for these stars prompted search for new candidates and it lead Gray (1997) to re-iterate and expand upon what he feels to be the best optical spectroscopic definition for the class of lambda Boo stars.
At the same time a new definition of the lambda Boo stars is given by Pauzen et al. (1997): “ Pop I hydrogen burning metal poor (except of C, N, O and S) A-type stars”; a list of 45 stars is given in their catalogue.
Only 8 stars among the 26 proposed by Abt (1984a) are present in this catalogue. We notice also that Abt and Morrell (1995) classified 46 lambda Boo stars, but only 9 are in common with the catalogue by Paunzen et al. and for 9 other stars the classification by Paunzen et al. as lambda Boo stars is not shared by Abt and Morrell.
Moreover, in the framework of spectral analysis of binary stars some authors have classifed the nearly unseen component as a lambda Boo star (Griffin et al. 1992; Hack et al. 1997).
[**What will be the group of Lambda Boo stars in the next millennium ?**]{}
Abt, H.A.: 1984a, in: the [*MK Process and Stellar Classification*]{}, ed. R.F. Garrison, David Dunlap Obs. Toronto, p. 340 Abt, H.A.: 1984b, [*Astrophys. J*]{}, [**285**]{}, 247 Abt, H.A. Morrell, N.: 1995, [*Astrophys. J. Suppl.*]{}, [**99**]{}, 135 Andrievsky, S.M.: 1997, [*Astron. Astrophys.*]{}, [**321**]{}, 838 Baschek, B. and Searle, L.: 1969, [*Astrophys. J*]{}, [**155**]{}, 537 Baschek B., Heck, A., Jaschek, C., Jaschek, M., Kopper, J., Scholtz, M., Wehrse, R.: 1984, [*Astron. Astrophys.*]{}, [**131**]{}, 378 Burbidge, R.M. and Burbidge, G.R.: 1956, [*Astrophys. J*]{}, [**124**]{}, 116 Bohlender, DA., Walker G.A.H.: 1994, [*Mon. Not. R. Astron. Soc.*]{}, [**266**]{}, 891 Charbonneau, P.: 1993, [*Astrophys. J*]{}, [**405**]{}, 720 Cucchiaro, A., Jaschek, M., Jaschek, C., Macau-Hercot, D.: 1980, [*Astron. Astrophys. Suppl.*]{}, [**40**]{}, 207 Faraggiana, R., Gerbaldi, M., Boehm, C.,: 1990, [*Astron. Astrophys.*]{}, [**235**]{}, 311 Gerbaldi, M., Faraggiana, R.: 1993, [*ASP Conf. Series*]{}, [**44**]{}, 368 Gerbaldi, M., Zorec, J., Castelli, F., Faraggiana, R.: 1993, [*ASP Conf. Serie*]{}, [**44**]{}, 413 Grady, C.A., McCollom, B., Ramley, L.A., England, M.N., Groebner, A., Schlegel, M.: 1996, [*Astrophys. J.*]{}, [**464**]{}, L183 Gray, R.O.: 1988, [*Astron.J.*]{}, [**95**]{}, 220 Gray, R.O.: 1991, in [*Precision Photometry*]{}, eds. A.G. Davis Philip, A.R. Upgren, K. A. Janes L. Davis Press, page 309 Gray, R.O.: 1997, [*Third Colloquium on Faint Blue Stars*]{}, to be published Gray, R.O. and Garrison, R.F. : 1987, [*Astrophys. J. Suppl.*]{}, [**65**]{}, 581). Gray, R.O. and Corbally, C.J.: 1993, [*Astron. J.*]{}, [**106**]{}, 632 Griffin, R.E.M., Schroder, K.P., Mish, A., Griffin, R.F.: 1992, [*Astron. Astrophys.*]{}, [**254**]{}, 289 Hack, M., Polosukhina, N.S., Malanushenko, V.P., Castelli, F.: 1997, [*Astron. Astrophys.*]{}, [**319**]{}, 637 Hauck, B.: 1986, [*Astron. Astrophys.*]{}, [**154**]{}, 349 Hauck, B., Ballereau, D., Chauville, J.: 1995, [*Astron. Astrophys. Sup*]{}, [**109**]{}, 505 Hauck, B., Ballereau , D., Chauville, J.: 1997, to be published, [*Astrophys. Astrophys. Sup.*]{} Hauck, B. and Slettebak, A.,: 1983, [*Astron. Astrophys.*]{}, [**127**]{}, 231 Holweger, H. Stürenburg, S.: 1991, [*Astron. Astrophys.*]{}, [**252**]{}, 255 Holweger, H. and Rentzsch-Holm I.: 1995, [*Astron. Astrophys.*]{}, [**303**]{}, 819 Iliev, I.Kh. ,Barzova, I.S.,: 1995, [*Astron. Astrophys.*]{}, [**302**]{}, 735 King, J.R.: 1994, [*Mon. Not. R. Astron. Soc.*]{}, [**269**]{}, 209 King, J.R., Patten, B.M.: 1992, [*Mon. Not. R. Astron. Soc.*]{}, [**256**]{}, 571 Kodaira, K.: 1967, [*Publ. Astr. Soc. Japan*]{}, [**19**]{}, 556 Levato, H., Malaroda, S., Grosso, M., Morrell, N.I.: 1994, [*ASP Conf. Serie*]{}, [**60**]{}, 93 Morgan, W.W., Keenan, P.C., Kellman, E.: 1943, [*An Atlas of Stellar Spectra*]{}, University of Chicago press Parenago, P.P.: 1958, [*Sov. Astron.*]{}, [**2**]{}, 151 Paunzen, E., Kuschnig, R., Handler, G., Gelbmann, M., Weiss, W.W. : 1997, [*Astron. Astrophys.*]{}, [**124**]{}, 23 Paunzen, E., Weiss, W.W., Heiter, U., North, P. : 1997, [*Astron. Astrophys. Sup.*]{}, [**123**]{}, 93 Renson , P., Faraggiana, R., Boehm, C.: 1990, [*Bull. Inf. CDS*]{}, [**38**]{}, 137 Sadakane K. and Nishida M. 1986, [*Pub. Astron. Soc. Pac.*]{}, [**98**]{}, 685 Sargent, W.L.W.: 1965 in [*The magnetic and related stars*]{}, AAS-NASA Symposium, Greenbelt, ed. R.C. Cameron, page 329 Sargent, W.L.W.: 1965, [*Astrophys. J*]{}, [**142**]{}, 787 Sargent, W.L.W.: 1966, [*Astrophys. J*]{}, [**144**]{}, 1128 Slettebak A.: 1952, [*Astrophys. J*]{}, [**115**]{}, 575 Slettebak A. 1954, [*Astrophys. J*]{}, [**119**]{}, 146 Slettebak, A., Wright, R.R. and Graham, J.A.: 1968, [*Astron. J.*]{}, [**73**]{}, 152 Stürenburg, S.: 1993, [*Astron. Astrophys.*]{}, [**277**]{}, 139 Turcotte, S., Charbonneau, P.: 1993, [*Astrophys. J*]{}, [**413**]{}, 376 Venn, K.A., Lambert, D.L.: 1990, [*Astrophys. J*]{}, [**363**]{}, 234
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Denis Osipov [^1]'
title: 'Adeles on $n$-dimensional schemes and categories $C_n$.'
---
Introduction
============
In this note we want to introduce by induction some class of infinite-dimensional vector spaces and morphisms between them. These spaces depend on integer $n$, we call such spaces as $C_n$-spaces.
$C_0$-spaces are finite dimensional spaces. A first basic example of $C_1$-space is the field of Laurent series $k((t))$. If ${{\cal O}}(n)
= t^n k[[t]]$, then we have a filtration $$\ldots {{\cal O}}(m) \subset {{\cal O}}(m-1) \subset {{\cal O}}(m-2) \ldots$$ and every factor space ${{\cal O}}(m - k)/ {{\cal O}}(m) $ is a finite dimensional vector space over $k$.
We can consider the morphisms between two such spaces as continuous linear maps. We remark that these morphisms can be described only in terms of filtration ${{\cal O}}_n$, without considering topology on $k((t))$. From this point of view, the space of adeles on an algebraic curve has a structure of $C_1$-space, which is filtered by partially ordered set of coherent sheaves on the curve.
We construct an iterated version of $C_1$-spaces, which we call a $C_n$-space. But in our construction we do not consider the structure of completion. So, we consider the filtered vector spaces. For example, the discrete valuation field is also a $C_1$-space, the space of rational adeles from [@S] is also a $C_1$-space.
The constructions of similar categories were introduced also in [@B1], [@Ka], [@K]. Our construction of $C_n$ is very close to the iterated functor $\mathop{\lim}\limits_{\longleftrightarrow}$, which was introduced by A.A.Beilinson in appendix to [@B1]. The main difference is that we consider noncompleted version of $\mathop{\lim}\limits_{\longleftrightarrow}$, i.e., filtered spaces, but with morphisms which come from $\mathop{\lim}\limits_{\longleftrightarrow}$. From this point of view the categories $C_n$ are rather closed to the dir-inv modules which were considered by A. Yekutieli in [@Y] for $n=1$.
The main result of this note is the theorem \[th1\], where we prove that the space of Parshin-Beilinson adeles on an $n$-dimensional Noetherian scheme $V$ over $k$ is a $C_n$-space, which is filtered by partially ordered set of coherent sheaves on the scheme. We calculate also the endomorphism algebra of n-dimensional local field.
Adeles on algebraic surfaces were introduced by A.N.Parshin in [@P2]. A.A. Beilinson generalized it to arbitrary Noetherian schemes in [@B]. Adeles on higher-dimensional schemes were applied to a lot of problems of algebraic geometry, see [@PF].
A.N. Parshin has pointed out to me that the categories $C_n$ can be useful for constructing of harmonic analysis on higher-dimensional schemes and higher-dimensional adeles and local fields, see [@P1].
In this note we consider the categories $C_n$ and the schemes over a field $k$. But all the constructions, for example, can be moved to arithmetical schemes, where $C_0$ are finite abelian groups, $C_1$ are filtered abelian groups with finite abelian group factors and so on.
Categories $C_n$
================
Constructions
-------------
### Objects in $C_n$
\[def1\] We say that $(I, F, V)$ is a filtered $k$-vector space, if
1. $V$ is a vector space over the field $k$,
2. $I$ is a partially ordered set, such that for any $i,j \in I$ there are $k,l \in I$ with $k \le i \le l$ and $k \le j \le l$,
3. $F$ is a function from $I$ to the set of $k$-vector subspaces of $V$ such that if $i \le j $ are any from $I$, then $F(i) \subset F(j)$,
4. $ \bigcap\limits_{i \in I} F(i) = 0$ and $
\bigcup\limits_{i \in I} F(i) = V $.
We say that a filtered vector space $(I_1, F_1, V)$ dominates another filtered vector space $(I_2, F_2, V)$ when there is a preserving order function $\phi: I_2 \to I_1$ such that
1. for any $i \in I_2$ we have $F_1(\phi (i)) = F_2 (i)$
2. for any $j \in I_1$ there are $i_1, i_2 \in I_2$ such that $ \phi(i_1) \le j \le \phi(i_2)$.
Now we define by induction the category of $C_n$-spaces and morphisms between them.
1. The category $C_0$ is the category of finite-dimensional vector spaces over $k$ with morphisms coming from $k$-linear maps between vector spaces.
2. The triple from $C_0$ $$0 {\longrightarrow}V_0 {\longrightarrow}V_1 {\longrightarrow}V_2 {\longrightarrow}0$$ is admissible when it is an exact triple of vector spaces
Now we define the objects of the category $C_n$ by induction. We suppose that we have already defined the objects of the category $C_{n-1}$ and the notion of admissible triple in $C_{n-1}$.
\[def4\]
1. Objects of the category $C_n$, i.e. $Ob(C_n)$, are filtered $k$-vector spaces $(I, F, V)$ with the following additional structures
1. for any $i \le j \in I$ on the $k$-vector space $F(j) / F(i)$ it is given a structure $E_{i,j} \in Ob(C_{n-1})$,
2. for any $i \le j \le k \in I$ $$0 {\longrightarrow}E_{i,j} {\longrightarrow}E_{i,k} {\longrightarrow}E_{j,k} {\longrightarrow}0$$ is an admissible triple from $C_{n-1}$.
2. Let $E_1 = (I_1, F_1, V_1)$, $E_2 = (I_2, F_2, V_2)$ and $E_3=(I_3, F_3, V_3)$ be from $Ob(C_n)$. Then we say that $$0 {\longrightarrow}E_1 {\longrightarrow}E_2 {\longrightarrow}E_3 {\longrightarrow}0$$ is an admissible triple from $C_n$ when the following conditions are satisfied
1. $$0 {\longrightarrow}V_1 {\longrightarrow}V_2 {\longrightarrow}V_3 {\longrightarrow}0$$ is an exact triple of $k$-vector spaces
2. \[itaa\] the filtration $(I_1, F_1, V_1)$ dominates the filtration $(I_2, F'_1, V_1)$, where $F'_1 (i) = F_2(i) \cap V_1$ for any $i \in I_2$,
3. \[itbb\] the filtration $(I_3, F_3, V_3)$ dominates the filtration $(I_2, F'_3, V_3)$, where $F'_3(i) = F_2(i) / F_2(i) \cap V_1$,
4. for any $i \le j \in I_2$ $$\label{trojkaa1}
0 {\longrightarrow}\frac{F'_1(j)}{F'_1(i)} {\longrightarrow}\frac{F_2(j)}{F_2(i)} {\longrightarrow}\frac{F'_3(j)}{F'_3(i)} {\longrightarrow}0$$ is an admissible triple from $C_{n-1}$. (By definition of $Ob(C_n)$, on every vector space from triple (\[trojkaa1\]) it is given the structure of $Ob(C_{n-1})$).
### Morphisms in $C_n$
By induction, we define now the morphisms in the category $C_n$. We suppose that we have already defined the morphisms in $C_{n-1}$.
\[d1\] Let $E_1 = (I_1, F_1, V_1)$ and $E_2 = (I_2, F_2, V_2)$ be from $Ob(C_n)$. Then $Mor_{C_n}(E_1, E_2)$ consists of elements $A \in Hom_k (V_1, V_2)$ such that the following conditions hold.
1. \[i1\] for any $i \in I_1$ there is an $j \in I_2$ such that $A (F_1(i)) \subset F_2(j)$,
2. \[i2\] for any $j \in I_2$ there is an $i \in I_1$ such that $A (F_1(i)) \subset F_2(j)$,
3. \[i3\] for any $i_1 \le i_2 \in I_1$ and $j_1 \le j_2 \in I_2$ such that $A (F_1(i_1)) \subset F_2(j_1)$ and $A (F_1(i_2)) \subset F_2(j_2)$ we have that the induced $k$-linear map $$\bar{A} : \frac{F_1(i_2)}{F_1(i_1)} {\longrightarrow}\frac{F_2(j_2)}{F_2(j_1)}$$ is an element from $$Mor_{C_{n-1}}(\frac{F_1(i_2)}{F_1(i_1)}, \frac{F_2(j_2)}{F_2(j_1)})$$
Now we want to prove that the compositions of so defined morphisms in $C_n$ will be again a morphism. We need the following definition.
Let $E_1, E_2$ be from $Ob(C_n)$.
1. A $k$-linear map $C: E_1 \to E_2$ is an admissible $C_n$-monomorphism when it is the part of an admissible triple from $C_n$ $$0 {\longrightarrow}E_1 \stackrel{C}{{\longrightarrow}} E_2 {\longrightarrow}E_3 {\longrightarrow}0$$
2. A $k$-linear map $D: E_1 \to E_2$ is an admissible $C_n$-epimorphism when it is the part of an admissible triple $$0 {\longrightarrow}E_3 {\longrightarrow}E_1 \stackrel{D}{{\longrightarrow}} E_2 {\longrightarrow}0$$
We have the following proposition.
Let $E_1= (I_1, F_1, V_1)$, $E_2= (I_2, F_2, V_2)$, $E_1'$, $E_2'$ be from $Ob(C_n)$ and $A$ is a $k$-linear map form $Hom_k(V_1, V_2)$.
1. \[item1\] If $B: E_3 \to E_1$ is an admissible $C_n$-epimorphism then $A \in Mor_{C_n}(E_1, E_2)$ if and only if $ A \circ B \in Mor_{C_n}(E_3, E_2)$.
2. \[item2\] If $B: E_2 \to E_3$ is an admissible $C_n$-monomorphism, then $A \in Mor_{C_n}(E_1, E_2)$ if and only if $ B \circ A \in Mor_{C_n}(E_1, E_3)$.
3. If the filtered vector space $E_1$ dominates the filtered vector space $E'_1$ and the filtered vector space $E_2$ dominates the filtered vector space $E'_2 $, then $A \in Mor_{C_n}(E_1, E_2)$ if and only if $A \in Mor_{C_n}(E'_1, E'_2)$.
4. $Mor_{C_n}(E_1, E_2)$ is a $k$-linear subspace of $Hom_k(V_1, V_2)$.
5. If $E_3$ is an object of $C_n$, then $$Mor_{C_n}(E_2, E_3) \circ Mor_{C_n}(E_1, E_2) \subset Mor_{C_n}(E_1, E_3)$$
[[**Proof **]{}]{}. The first two statements follow by induction on $n$.
The third statement follows from the first and the second statement.
The other statements follow by induction on $n$ using the previous statements.
We give the proof of the fifth statement. Let $A \in Mor_{C_n}(E_1, E_2)$ and $B \in Mor_{C_n}(E_2, E_3)$. We have to prove that $B \circ A \in Mor_{C_n}(E_1, E_3)$. We have to check for $B \circ A$ the conditions \[i1\], \[i2\], \[i3\] of definition \[d1\]. Let $E_3 = (I_3, F_3, V_3)$.
For any $i_1 \in I_1$ there is $i'_2 \in I_2$ such that $A (F_1(i_1)) \subset F_2(i'_2)$. For $i'_2 \in I_2$ there is $i'_3 \in I_3$ such that $B (F_2(i'_2)) \subset F_3 (i'_3)$. Therefore $B \circ A (F_1 (i_1)) \subset F_3 (i'_3)$.
Analogously for any $j_3 \in I_3$ we find $j'_2 \in I_2$ such that $B (F_2(j'_2)) \subset F(j_3)$. For $j'_2 \in I_2$ we find $j'_1 \in I_1$ such that $A (F_1(j'_1)) \subset F_2 (j'_2)$. Then $B \circ A (F_1(j'_1)) \subset F_3(j_3)$.
Now let $i_1 \ge j_1 \in I_1$ and $i_3 \ge j_3 \in I_3$ such that $$B \circ A (F_1(j_1)) \subset F_3(j_3) \qquad \mbox{and} \qquad B \circ A (F_1(i_1)) \subset F_3(i_3) \mbox{.}$$
Now we fix any $i^{''}_3 \in I_3$ such that $i^{''}_3 \ge i'_3$ and $i^{''}_3 \ge i_3$. We fix any $j^{''}_1 \in I_1$ such that $j^{''}_1 \le j_1$ and $j^{''}_1 \le j'_1$. Then, by items \[item1\] and \[item2\] of this proposition, the map induced by $B \circ A$ belongs to $Mor_{C_{n-1}}(\frac{F_1(i_1)}{F_1(j_1)},
\frac{F_3(i_3)}{F_3(j_3)})$ if and only if the map induced by $B
\circ A$ belongs to $Mor_{C_{n-1}}(\frac{F_1(i_1)}{F_1(j^{''}_1)},
\frac{F_3(i^{''}_3)}{F_3(j_3)})$. But the last induced map is the composition of the induced map by $A$ in $Mor_{C_{n-1}}(\frac{F_1(i_1)}{F_1(j^{''}_1)},
\frac{F_2(i'_2)}{F_2(j'_2)})$ and of the induced map by $B$ in $Mor_{C_{n-1}}(\frac{F_2(i_2)}{F_2(j'_2)},
\frac{F_3(i^{''}_3)}{F_3(j_3)})$. By induction, the composition of morphisms from $C_{n-1}$ is a morphism from $C_{n-1}$. The proposition is proved.
Examples
--------
### Linearly locally compact spaces. {#lcs}
A topological $k$-vector space $V$ (over a discrete field $k$) is linearly compact (see [@L ch.2 ,§6], [@Bo ch.III, §2, ex.15-21]) when the following conditions hold
1. $V$ is complete and Hausdorff,
2. $V$ has a base of neighbourhoods of $0$ consisting of vector subspaces,
3. each open subspace of $V$ has finite codimension.
A topological $k$-vector space $W$ (over a discrete field $k$) is linearly locally compact (see [@L]) when it has a basis of neighborhoods of $0$ formed by linearly compact open subspaces.
Any topological linearly locally compact space is a $C_1$-space, where filtration is given by linearly compact open subspaces.
Any field of discrete valuation is $C_1$-space. And the completion functor gives us the linearly locally compact vector space. Moreover, if $E = (I, F, V)$ is a $C_1$-space, we take $$\Phi_1(E) = \mathop{{\mathop {\rm lim}}_{\rightarrow}}_{i \in I} \mathop{{\mathop {\rm lim}}_{\leftarrow}}_{j \le i} F(i) / F(j) \mbox{.}$$ The space $\Phi_1 (E)$ is a linearly locally compact space. And all the linearly locally compact spaces can be obtained in a such way.
Moreover, we define by induction on $n$ the functor of completion $\Phi_n$ from $C_n$-spaces to $C_n$-spaces. We put $$\Phi_n (E) =
\mathop{{\mathop {\rm lim}}_{\rightarrow}}_{i \in I} \mathop{{\mathop {\rm lim}}_{\leftarrow}}_{j \le i} \Phi_{n-1} ( F(i) / F(j))$$ where $E= (I, F, V) $ is a $C_n$-space. From properties of $C_n$ we obtain that this functor $\Phi_n$ is well-defined.
### Contragredient space
We will define by induction on $n$ a contravariant functor of contragredient space $D_n$ from $C_n$-spaces to $C_n$-spaces. Let $E
=(I, F, V)$ be a $C_n$-space.
If $n=0$, then $D_0 (V) = V^*$, where $V^*$ is the dual space to $V$.
If $n \ge 1$, then we put $$D_n (V) = \mathop{{\mathop {\rm lim}}_{\rightarrow}}_{j \in I}
\mathop{{\mathop {\rm lim}}_{\leftarrow}}_{i \ge j} D_{n-1} ( F(i) / F(j))
\mbox{.}$$
Then $D_n(E) = (I^0, F^0, D_n(V) )$ is a $C_n$-space, where $I^0$ is a partially ordered set, which has the same set as $I$, but with the inverse order then $I$, and $$F^0 (j) =
\mathop{{\mathop {\rm lim}}_{\leftarrow}}_{i \le j \in I^0} D_{n-1} ( F(i) / F(j))
\mbox{.}$$
It is easy to see by induction on $n$ that $D_n(V) \subset V^*$, the functor $D_n$ maps admissible triples to admissible triples, and $D_n(D_n (E)) = \Phi_n(E)$.
### Adelic space.
The next example of $C_n$-space is the space of adeles on a Noetherian $n$-dimensional scheme $V$ (see [@B], [@H], [@Osi]).
Let $P(V)$ be the set of points of the scheme $V$. Consider $\eta, \nu \in P(V)$. Define $\eta \ge \nu$ if $\nu \in \bar{\{ \eta\}}$. $\ge$ is a half ordering on $P(X)$. Let $S(V)$ be the simplicial set induced by $(P(X), \ge)$, i.e. $$S(V)_m = \{ (\nu_0, \ldots, \nu_m) \mid \nu_i \in P(V); \nu_i \ge \nu_{i+1} \}$$ is the set of $m$-simplices of $S(V)$ with the usual boundary and degeneracy maps.
In [@B], [@H] for any $K \subset S(V)_m$ it was constructed the space ${{\mbox{\dbl A}}}(K, {{\cal F}})$ for any quasicoherent sheaf ${{\cal F}}$ on $V$ such that $${{\mbox{\dbl A}}}(K, {{\cal F}}) \subset
\prod_{\delta \in K} {{\mbox{\dbl A}}}(\delta, {{\cal F}})$$
\[th1\] Let $V$ be a Noetherian $n$-dimensional scheme over $k$, $K
\subset S(V)_n$, and ${{\cal F}}$ be a coherent sheaf on $V$. Then the adelic space ${{\mbox{\dbl A}}}(K, {{\cal F}})$ has a structure of $C_n$-space.
[[**Proof **]{}]{}. We denote $$K_0 = \{\eta \in S(V)_0 \mid (\eta > \eta_1 \ldots > \eta_n ) \in K \quad
\mbox{for some} \quad \eta_i \in P(V) \} \mbox{.}$$
For $\eta \in K_0$ we denote $${}_{\eta} K = \{ (\eta_1 > \ldots > \eta_{n}) \in S(V)_{n-1}
\mid (\eta > \eta_1 \ldots > \eta_n) \in K \} \mbox{.}$$
We have $${{\mbox{\dbl A}}}(K,{{\cal F}}) = \prod_{\eta \in K_0} {{\mbox{\dbl A}}}({}_{\eta} K ,{{\cal F}}_{\eta}) \mbox{.}$$
If $(I_1, F_1, V_1)$ and $(I_2, F_2, V_2)$ are $C_n$-spaces, then $(I_1 \times I_2, F_1 \times F_2, V_1 \times V_2)$ is a $C_n$-space as well. Moreover, any finite product of $C_n$-spaces is a $C_n$-space in the same way. The set $K_0$ is finite, therefore it is enough to define a $C_n$-structure on ${{\mbox{\dbl A}}}({}_{\eta} K ,{{\cal F}}_{\eta})$ for every $\eta \in K_0$.
For $\eta \in K_0$ we define a partially ordered set $$I_{\eta}({{\cal F}}) = \{ {{\cal G}}\subset {{\cal F}}_{\eta} \mid \quad {{\cal G}}\quad \mbox{is a coherent sheaf on }
\bar{\eta}, \quad {{\cal G}}_{\eta} = {{\cal F}}_{\eta} \} \mbox{,}$$ which is ordered by inclusions of sheafs.
The functor ${{\mbox{\dbl A}}}({}_{\eta} K, \;)$ is an exact functor. Therefore for any ${{\cal G}}\in I_{\eta}({{\cal F}})$ we have an embedding $${{\mbox{\dbl A}}}({}_{\eta} K,{{\cal G}}) {\longrightarrow}{{\mbox{\dbl A}}}({}_{\eta} K,{{\cal F}}_{\eta}) \mbox{.}$$
If ${{\cal G}}_1 \subset {{\cal G}}_2$ are from $I_{\eta} ({{\cal F}})$, then from the exactness of the functor ${{\mbox{\dbl A}}}({}_{\eta} K, \;)$ we have $${{\mbox{\dbl A}}}({}_{\eta} K, {{\cal G}}_2) /
{{\mbox{\dbl A}}}({}_{\eta} K, {{\cal G}}_1) =
{{\mbox{\dbl A}}}({}_{\eta} K, {{\cal G}}_2 / {{\cal G}}_1)
\mbox{.}$$
We have $({{\cal G}}_2 / {{\cal G}}_1)_{\eta} = 0$. Therefore ${{\cal G}}_2 / {{\cal G}}_1 $ is a coherent sheaf on some subscheme $Y$ of dimension $n-1$. Therefore ${{\mbox{\dbl A}}}({}_{\eta} K, {{\cal G}}_2 / {{\cal G}}_1) =
{{\mbox{\dbl A}}}( S(Y) \cap {}_{\eta} K , {{\cal G}}_2 / {{\cal G}}_1 )
$. We apply induction on dimension of scheme to the $n-1$-dimensional scheme $Y$. Therefore we have that ${{\mbox{\dbl A}}}( S(Y) \cap {}_{\eta} K, {{\cal G}}_2 / {{\cal G}}_1 )$ is a $C_{n-1}$-space.
We check now that this $C_n$-structure is well defined. It is enough to prove only conditions \[itaa\] and \[itbb\] of definition \[def4\]. They follow from the following three statements.
The first statement claims that every exact triple of quasicoherent sheaves on $V$ is a direct limit of exact triples of coherent sheaves on $V$, see [@H lemma 1.2.2].
The second statement claims that on any irreducible Noetherian scheme $X$ for any exact triple of coherent sheaves $$0 {\longrightarrow}{{\cal F}}_1 {\longrightarrow}{{\cal F}}_2 {\longrightarrow}{{\cal F}}_3 {\longrightarrow}0$$ and any coherent subsheaves ${{\cal G}}_1 \subset {{\cal F}}_1$ and ${{\cal G}}_3 \subset {{\cal F}}_3$ such that $({{\cal F}}_1 / {{\cal G}}_1)_{\eta} = 0$ and $({{\cal F}}_3 / {{\cal G}}_3 )_{\eta}= 0$, where $\eta $ is the general point of $V$, there exists a coherent subsheaf ${{\cal G}}_2 \subset {{\cal F}}_2$ such that $ ({{\cal G}}_2 \cap {{\cal F}}_1) \subset {{\cal G}}_1 $, $ ({{\cal G}}_2 / ({{\cal G}}_2 \cap {{\cal F}}_1)) \subset {{\cal G}}_3 $ and $({{\cal F}}_2 / {{\cal G}}_2)_{\eta} = 0$. It is enough to construct ${{\cal F}}_2$ locally, where this sheaf exists by the Artin-Rees lemma.
The third statement claims that for any two quasicoherent subsheaves ${{\cal F}}_1$, ${{\cal F}}_2$ of a quasicoherent sheaf ${{\cal F}}_3$ on $V$ we have $${{\mbox{\dbl A}}}(K, {{\cal F}}_1 \cap {{\cal F}}_2) = {{\mbox{\dbl A}}}(K, {{\cal F}}_1) \cap {{\mbox{\dbl A}}}(K, {{\cal F}}_2) \mbox{,}$$ $${{\mbox{\dbl A}}}(K, {{\cal F}}_1 + {{\cal F}}_2) = {{\mbox{\dbl A}}}(K, {{\cal F}}_1) + {{\mbox{\dbl A}}}(K, {{\cal F}}_2) \mbox{,}$$ where the intersection and sum is taken inside of ${{\mbox{\dbl A}}}(K, {{\cal F}}_3)$. It follows from exactness of the functor ${{\mbox{\dbl A}}}(K, \,)$ and the following commutative diagram: $$\begin{array}{ccc}
\frac{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_1 + {{\cal F}}_2)}{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_1)} & = &
\frac{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_2)}{{{\mbox{\sdbl A}}}(K, \, {{\cal F}}_1 \cap {{\cal F}}_2)} \\
\uparrow & & \downarrow \\
\frac{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_1) + {{\mbox{\sdbl A}}}(K, \: {{\cal F}}_2)}{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_1)} & =
& \frac{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_2)}{{{\mbox{\sdbl A}}}(K, \: {{\cal F}}_1) \cap {{\mbox{\sdbl A}}}(K, \: {{\cal F}}_2)}
\quad \mbox{,}
\end{array}$$ which gives that the vertical arrows are isomorphisms.
The theorem is proved.
[*For the smooth surface $V$ a $C_2$-structure on ${{\mbox{\dbl A}}}(K,{{\cal F}})$ can be defined by filtration of Cartier divisors. (We use that the Cartier divisors coincide with the Weil divisors for the smooth varieties).* ]{}
[*The structure of $C_n$ space can be defined on introduced in [@H §5.2] spaces of rational adeles $a(K, {{\cal F}})$ as well. Then the functor of completion $\Phi_n$ (see section \[lcs\]) applied to the space of rational adeles gives the space ${{\mbox{\dbl A}}}(K,
{{\cal F}})$.*]{}
The endomorphism algebra of $n$-dimensional local field
-------------------------------------------------------
We suppose now that $\delta = (\eta_0 > \ldots > \eta_n) \in S(V)_n$ on the $n$-dimensional scheme $V$ has the following property: $\eta_n$ is a smooth point on every scheme $\bar{\eta_i}$.
Then $ ({{\cal O}}_V)_{\delta} =
{{\mbox{\dbl A}}}(\delta, {{\cal O}}_V)=
k'((t_n)) \ldots ((t_1))$ is an $n$-dimensional local field, where $k' = k(\eta_n)$. We demand that the local parameters $t_i \in \widehat{({{\cal O}}_V)}_{\eta_n}$ for any $i$.
Then we define the filtration on $k' ((t_n)) \ldots ((t_1))$ by $ E_l = t_1^{l} k'((t_n)) \ldots ((t_2)) [[t_1]] $ for $l \in {{\mbox{\dbl Z}}}$. On each factor $E_{l_1} / E_{l_2}$ of this filtration we define the new filtration given in $E_{l_1} / E_{l_2}$ by images of $ t_2^{m} t_1^{l_1} k'((t_n)) \ldots ((t_{3})) [[t_2, t_1]]$ for $m \in {{\mbox{\dbl Z}}}$ and so on.
We obtained the structure of $C_n$-space on $({{\cal O}}_V)_{\delta}$. And the structure of $C_n$-space constructed on $({{\cal O}}_V)_{\delta}$ in theorem \[th1\] dominates the constructed now structure of $C_n$-space.
Now let $K = k((t_n)) \ldots ((t_1))$. We define the $k$-algebra $$End_K = Mor_{C_n} (K, K)$$ Let $\bar{K} = k((t_{n})) \ldots ((t_2))$. Then $K = \bar{K}((t_1))$. For any element $A \in {\mathop {\rm End}}_K$ we consider the matrix $\{(A_{ij})_{i,j \in {{\mbox{\sdbl Z}}}} \mid A_{ij} \in End_{k}(\bar{K}) \}$ given by $$A (x t^i_n) = \sum_j A_{ij} ( x) t^j_n \quad \mbox{with} \quad x
\in \bar{K} \mbox{.}$$
An endomorphism $A \in End_k(K)$ belongs to $End_K$ if and only if the following conditions are satisfied.
1. There is a nondecreasing function $a: {{\mbox{\dbl Z}}}\to {{\mbox{\dbl Z}}}$, $a(i) \to \infty$ when $i \to \infty$ such that for $j < a(i)$ all elements $A_{ij} = 0$.
2. Any element $A_{ij}$ belongs to $End_{\bar{K}}$.
[[**Proof **]{}]{}follows directly from the definition of morphisms between $C_n$-spaces and the definition of $C_n$-structure on $K$. The proposition is proved.
We consider on $K$ the topology of $n$-dimensional local field (see [@P]). Let $End^{c}_k(K)$ be the algebra of continuous $k$-linear endomorphisms of $K$.
1. \[iit1\] $End_K \subset End^{c}_k(K) $
2. \[iit2\] If $n =1$, then $End_K = End^{c}_k(K)$.
3. \[iit3\] If $n > 1$, then $End^{c}_k(K)$ bigger then $End_K$.
[[**Proof **]{}]{}. We denote by ${{\cal O}}_K = \bar{K}[[t_1]]$. The base of neighbourhoods of $0$ in $\bar{K}((t_1))$ consists of the following $k$-vector subspaces: $$\sum_i U_i t_1^i + {{\cal O}}_K t_1^m \mbox{,}$$ where $U_i$ are open $k$-vector subspaces from $\bar{K}$ and $m$ is an integer.
Then statement \[iit2\] is the same as condition \[i2\] of definition \[d1\]. Statement \[iit1\] follows by induction on $n$ from the definition \[d1\], for $n=2$ see [@O lemma 2]. Now we give an example for statement \[iit3\].
Let $n=2$. Then we consider a $k$-linear map $\phi: K \to K $ defined as following. For $U = k[[t_2]] ((t_1)) + k((t_2))[[t_1]])$ we put $\phi (U) = 0$ and the induced map $\phi : K/U \to K $ we put on monomials $t_2^l t_1^m$ by the rule: if $m=-1 $ then $\phi(t_2^l t_1^m) = t_2^l t_1^{m+l}$, for other monomials we put $\phi = 0$. Then $\phi$ is continuous, but $\phi$ is not from $End_K$. The proposition is proved.
\[pro\]
1. For any $m$ we have an embedding $End_K^{\oplus m} \hookrightarrow End_K$.
2. For any $m$ we have an embedding $ gl(m, K) \hookrightarrow End_K$.
[[**Proof **]{}]{}We have $K = \bar{K}((t_n))$. Let $e_i$, $1 \le i \le m$ be the standard basis of $K^{\oplus m}$, i.e. $K^m = \oplus K e_m$. We consider a $\bar{K}$-isomorphism $\phi$ of $\bar{K}$-vector spaces $$\phi : K^{\oplus m} \to K \quad \quad \phi(t_n^j e_i) = t_n^{mj +i -1} \qquad 1 \le i \le m \mbox{.}$$ This isomorphism induces an isomorphism $$Mor_{C_n} (K^{\oplus m}, K^{\oplus m}) \hookrightarrow End_K \mbox{.}$$ Now the proposition follows from the natural embeddings: $$End_K^{\oplus m} \hookrightarrow Mor_{C_n} (K^{\oplus m}, K^{\oplus m}) \qquad
\mbox{and} \qquad gl(m, K) \hookrightarrow Mor_{C_n} (K^{\oplus m}, K^{\oplus m}) \mbox{.}$$
[*From proposition \[pro\] we have embeddings of $$K^* \hookrightarrow End_K$$ and toroidal Lie algebras $$gl (m, k[t_1, \ldots, t_n, t_1^{-1}, \ldots, t_n^{-1}])
\hookrightarrow End_K \mbox{.}$$* ]{}
Acknowledgments {#acknowledgments .unnumbered}
===============
I am very grateful to A.N.Parshin for interesting and valuable comments and the stimulation the author to write this note.
I am also grateful to H.Kurke, A.Zheglov and I.Zhukov for interesting comments. This note was prepared during my stay at the Humboldt University of Berlin. I am grateful to the Humboldt University of Berlin for the hospitality.
[99]{}
Beilinson A.A., [Residues and Adeles]{}, Funct. Anal. Pril., 14 (1980), no. 1, 44-45; English transl. in Func. Anal. Appl., 14, no. 1 (1980), 34-35.
Beilinson A.A., How to glue perverse sheaves, in K-theory, Arithmetic, Geometry, Lecture Notes in Mathematics 1289.
Bourbaki N., [Algebre commutative]{}, Hermann, 1961.
Fimmel T., Parshin A.N., [An introduction to the higher adelic theory]{}, preprint 1999.
Huber A., [ On the Parshin-Beilinson Adeles for Schemes]{}, Abh. Math. Sem. Univ. Hamburg, 61 (1991), 249-273.
Kapranov M.M., [Semiinfinite symmetric powers]{}, e-print math.QA/0107089.
Kato K., [Existence theorem for higher local class field theory]{}, Invitation to higher local fields (Münster, 1999), 165–195 (electronic), Geom. Topol. Monogr., 3, Geom. Topol. Publ., Coventry, 2000, see also math.AG/0012150.
Lefschetz S., [Algebraic topology]{}, AMS Colloquium Publications 27, Amer. Math. Soc. New York, (1942).
Osipov D.V., [Central extensions and reciprocity laws on algebraic surfaces]{}, (Russian) Mat. Sb. 196:10, 111-136; translation in Sb. Math. 196:10, 1503-1527, see also math.NT/0501155.
Osipov D.V., [$n$-dimensional local fields and adeles on $n$-dimensional schemes]{}, e-print math.AG/0508205.
Parshin A. N., [Local class field theory]{}, Proc. Steklov Inst. Math. 165 (1985), 157-185.
Parshin A. N., [Higher-dimensional local fields and L-functions]{}, Invitation to higher local fields (Münster, 1999), 199–213 (electronic), Geom. Topol. Monogr., 3, Geom. Topol. Publ., Coventry, 2000, see also math.AG/0012151.
Parshin A.N., [On the arithmetic of two-dimensional schemes I, Repartitions and residues]{}, Math. USSR Izv. Vol.10 (1976) No.4, 695-729 (Engl.)
Serre J.-P., [Groupes algebriques et corps de classes]{}, Hermann, 1959.
Yekutieli A., [Continuous and twisted $L_{\infty}$ morphisms]{}, math.QA/0502137, to appear in Journal of Pure and Applied Algebra.
Steklov Mathematical Institute,\
Gubkina str. 8,\
119991, Moscow, Russia\
e-mail ${d}_{-} [email protected]$
[^1]: supported by DFG-Schwerpunkt ”Globale Methoden in der Komplexen Geometrie”, by RFBR grant no. 05-01-00455 and grant no. 04-01-00702, by grant of Leading Scientific Schools no. 9969.2006.1, by INTAS grant 05-100000-8118, and by grant of Russian Science Support Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a coercive Hamiltonian which is quasi-convex with respect to the gradient variable and periodic with respect to time and space at least “far away from the origin”, we consider the solution of the Cauchy problem of the corresponding Hamilton-Jacobi equation posed on the real line. Compact perturbations of coercive periodic quasi-convex Hamiltonians enter into this framework for example. We prove that the rescaled solution converges towards the solution of the expected effective Hamilton-Jacobi equation, but whose “flux” at the origin is “limited” in a sense made precise by the authors in [@im]. In other words, the homogenization of such a Hamilton-Jacobi equation yields to supplement the expected homogenized Hamilton-Jacobi equation with a junction condition at the single discontinuous point of the effective Hamiltonian. We also illustrate possible applications of such a result by deriving, for a traffic flow problem, the effective flux limiter generated by the presence of a finite number of traffic lights on an ideal road. We also provide meaningful qualitative properties of the effective limiter.'
author:
- 'G. Galise[^1], C. Imbert[^2], R. Monneau[^3]'
bibliography:
- 'gim.bib'
nocite: '[@*]'
title: |
A junction condition by specified homogenization\
and application to traffic lights
---
#### AMS Classification:
35F21, 49L25, 35B27
#### Keywords:
Hamilton-Jacobi equations, quasi-convex Hamiltonians, homogenization,\
junction condition, flux-limited solution, viscosity solution.
Introduction
============
Setting of the general problem
------------------------------
This article is concerned with the study of the limit of the solution $u^\varepsilon(t,x)$ of the following equation $$\label{eq:hj-eps}
u^\eps_t +H\left(\frac{t}\eps,\frac{x}\eps,u^\eps_x\right) = 0 \quad \text{ for }
(t,x)\in(0,T)\times \R$$ submitted to the initial condition $$\label{eq:ic}
u^\eps(0,x)=u_0(x) \quad \text{for } x\in \R$$ for a Hamiltonian $H$ satisfying the following assumptions:
- (Continuity) $H\colon \R^3 \to \R$ is continuous.
- (Time periodicity) For all $k\in \Z$ and $(t,x,p)\in \R^3$, $$H(t+k,x,p)=H(t,x,p).$$
- (Uniform modulus of continuity in time) There exists a modulus of continuity $\omega$ such that for all $t,s,x,p \in \R$, $$H(t,x,p)-H(s,x,p)\le \omega(\left|t-s\right|\left(1+\max
\left(H(s,x,p),0\right)\right)).$$
- (Uniform coercivity) $$\lim_{|q|\to +\infty} H(t,x,q)=+\infty$$ uniformly with respect to $(t,x)$.
- (Quasi-convexity of $H$ for large $x$’s) There exists some $\rho_0>0$ such that for all $x\in \R\setminus
(-\rho_0,\rho_0)$, there exists a continuous map $t\mapsto p^0(t,x)$ such that $$\left\{\begin{array}{l}
H(t,x,\cdot) \quad \text{is non-increasing in}\quad (-\infty,p^0(t,x)),\\
H(t,x,\cdot) \quad \text{is non-decreasing in}\quad (p^0(t,x),+\infty).
\end{array}\right.$$
- (Left and right Hamiltonians) There exist two Hamiltonians $H_\alpha(t,x,p)$, $\alpha=L,R$, such that $$\left\{\begin{array}{l}
H(t,x+k,p) -H_L(t,x,p)\to 0 \quad \text{as}\quad \Z\ni k\to -\infty \\
H(t,x+k,p) -H_R(t,x,p)\to 0 \quad \text{as}\quad \Z\ni k\to +\infty
\end{array}\right.$$ uniformly with respect to $(t,x,p)\in [0,1]^2\times \R$, and for all $k,j\in \Z$, $(t,x,p)\in \R^3$ and $\alpha \in \{L,R\}$, $$H_\alpha(t+k,x+j,p)=H_\alpha(t,x,p).$$
We have to impose some condition in order to ensure that effective Hamiltonians $\bar H_\alpha$ are quasi-convex; indeed, we will see that the effective equation should be solved with *flux-limited solutions* recently introduced by the the second and third authors [@im]; such a theory relies on the quasi-convexity of the Hamiltonians.
- (Quasi-convexity of the left and right Hamiltonians) For each $\alpha=L,R$, $H_\alpha$ does not depend on time and there exists $p_\alpha^0$ (independent on $(t,x)$) such that $$\left\{\begin{array}{l}
H_\alpha(x,\cdot) \quad \text{is non-increasing on}\quad (-\infty,p^0_\alpha),\\
H_\alpha(x,\cdot) \quad \text{is non-decreasing on}\quad (p^0_\alpha,+\infty).
\end{array}\right.$$
- (Convexity of the left and right Hamiltonians) For each $\alpha=L,R$, and for all $(t,x)\in \R\times \R$, the map $p\mapsto H_\alpha(t,x,p)$ is convex.
A simple example of such a Hamiltonian is $$H(t,x,p)= \left|p\right| - f(t,x)$$ with a continuous function $f$ satisfying $f(t+1,x)=f(t,x)$ and $f(t,x)\to 0$ as $\left|x\right|\to +\infty$ uniformly with respect to $t\in \R$.
Main results
------------
Our main result is concerned with the limit of the solution $u^\eps$ of -. It makes part of the huge literature dealing with homogenization of Hamilton-Jacobi equation, starting with the pioneering work of Lions, Papanicolaou and Varadhan [@lpv]. In particular, we need to use the perturbed test function introduced by Evans [@evans]. As pointed out to us by the referee, there are few papers dealing with Hamiltonians that depend on time; it implies in particular that so-called correctors also depend on time. The reader is in particular referred to [@bs; @br] for the large time behaviour and to [@fim09a; @fim09b; @fim12] for homogenization results. This limit satisfies an effective Hamilton-Jacobi equation posed on the real line whose Hamiltonian is discontinuous. More precisely, the effective Hamiltonian equals the one which is expected (see **(A5)**) in $(-\infty;0)$ and $(0;+\infty)$; in particular, it is discontinuous in the space variable (piecewise constant in fact). In order to get a unique solution, a flux limiter should be identified [@im].
### Homogenized Hamiltonians and effective flux limiter {#homogenized-hamiltonians-and-effective-flux-limiter .unnumbered}
The homogenized left and right Hamiltonians are classically determined by the study of some “cell problems”.
\[prop:quasi-conv\] Assume **(A0)-(A5)**, and either **(B-i)** or **(B-ii)**. Then for every $p\in \R$, and $\alpha=L,R$, there exists a unique $\lambda\in\R$ such that there exists a bounded solution $v^\alpha$ of $$\label{eq:cell-alpha}
\left\{\begin{array}{l}
v^\alpha_t + H_\alpha(t,x,p+v^\alpha_x)=\lambda \quad \text{in}\quad \R\times \R,\\
v^\alpha \text{ is $\Z^2$-periodic}.
\end{array}\right.$$ If $\bar H_\alpha(p)$ denotes such a $\lambda$, then the map $p\mapsto
\bar H_\alpha(p)$ is continuous, coercive and quasi-convex.
We recall that a function $\bar H_\alpha$ is quasi-convex if the sets $\{ \bar H_\alpha \le \lambda \}$ are convex for all $\lambda
\in \R$. If $\bar H_\alpha$ is also coercive, then $\bar p^0_\alpha$ denotes in proofs some $p \in \operatorname{argmin}\bar H_\alpha$.
The effective flux limiter $\bar A$ is the smallest $\lambda \in \R$ for which there exists a solution $w$ of the following global-in-time Hamilton-Jacobi equation $$\label{eq:cell}
\left\{\begin{array}{ll}
w_t + H(t,x,w_x)=\lambda, \quad (t,x)\in \R\times \R,\\
w \text{ is $1$-periodic w.r.t. } t.
\end{array}\right.$$
\[thm:bar-A\] Assume **(A0)-(A5)** and either **(B-i)** or **(B-ii)**. The set $$E= \{ \lambda \in \R: \exists \text{$w$ sub-solution of
\eqref{eq:cell}} \}$$ is not empty and bounded from below. Moreover, if $\bar A$ denotes the infimum of $E$, then $$\label{eq::g1}
\bar A\ge A_0:=\max_{\alpha=L,R}\left(\min \bar H_\alpha\right).$$
We will see below (Theorem \[thm:corrector\]) that the infimum is in fact a minimum: there exists a global corrector which, in particular, can be rescaled properly.
We can now define the effective junction condition.
\[defi:F-bar\] The *effective junction function* $F_{\bar A}$ is defined by $$F_{\bar A}(p_L,p_R):=\max
(\bar A,\bar H_L^+(p_L), \bar H_R^-(p_R))$$ where $$\bar H_\alpha^-(p)=\left\{\begin{array}{ll}
\bar H_\alpha (p) &\quad \text{if}\quad p< \bar p_\alpha^0,\\
\bar H_\alpha(\bar p_\alpha^0)&\quad \text{if}\quad p\ge \bar p_\alpha^0
\end{array}\right.
\quad \text{and}\quad
\bar H_\alpha^+(p)=\left\{\begin{array}{ll}
\bar H_\alpha(\bar p_\alpha^0) &\quad \text{if}\quad p\le \bar p_\alpha^0,\\
\bar H_\alpha (p) &\quad \text{if}\quad p> \bar p_\alpha^0
\end{array}\right.$$ where $\bar p_\alpha^0 \in \operatorname{argmin}\bar H_\alpha$.
### The convergence result {#the-convergence-result .unnumbered}
Our main result is the following theorem.
\[thm:conv\] Assume **(A0)-(A5)** and either **(B-i)** or **(B-ii)**. Assume that the initial datum $u_0$ is Lipschitz continuous and for $\eps>0$, let $u^\eps$ be the solution of -. Then $u^\eps$ converges locally uniformly to the unique flux-limited solution $u^0$ of $$\label{eq:hj-homog}
\left\{\begin{array}{ll}
u^0_t + \bar H_L(u^0_x)=0, & t>0,x<0,\\
u^0_t + \bar H_R(u^0_x)=0, & t>0,x>0,\\
u^0_t + F_{\bar A}(u^0_x(t,0^-),u^0_x(t,0^+))=0, & t>0,x=0
\end{array}\right.$$ submitted to the initial condition .
We recall that the notion of flux-limited solution for (\[eq:hj-homog\]) is introduced in [@im].
This theorem asserts in particular that the slopes of the limit solution at the origin are characterized by the effective flux limiter $\bar A$. Its proof relies on the construction of a global “corrector”, *i.e.* a solution of , which is close to an appropriate $V$-shaped function after rescaling. This latter condition is necessary so that the slopes at infinity of the corrector fit the expected slopes of the solution of the limit problem at the origin. Here is a precise statement.
\[thm:corrector-simple\] Assume **(A0)-(A5)** and either **(B-i)** or **(B-ii)**. There exists a solution $w$ of with $\lambda = \bar A$ such that, the function $$w^\eps(t,x)=\eps w(\eps^{-1}t,\eps^{-1}x)$$ converges locally uniformly (along a subsequence $\eps_n\to
0$) towards a function $W=W(x)$ which satisfies $W(0)=0$ and $$\label{eq:W-estim-2}
\hat p_R x 1_{\left\{x>0\right\}} + \hat p_L x
1_{\left\{x<0\right\}}\ge W(x)\ge \bar p_R x
1_{\left\{x>0\right\}} + \bar p_L x 1_{\left\{x<0\right\}}$$ where $$\begin{aligned}
\label{def:pR}
\begin{cases} \bar p_R = \min E_R \\ \hat p_R = \max E_R \end{cases}
& \quad \text{with}\quad E_R:= \left\{p\in \R,\quad \bar H_R^+(p)=\bar
H_R(p)=\bar A\right\} \\
\label{def:pL}
\begin{cases}\bar p_L = \max E_L \\\hat p_L = \min E_L \end{cases}
& \quad \text{with}\quad E_L:= \left\{p\in \R,\quad \bar
H_L^-(p)=\bar H_L(p)=\bar A\right\}.\end{aligned}$$
The construction of this global corrector is the reason why homogenization is referred to as being “specified”. See also Section \[subsec:rel\] about related results. As a matter of fact, we will prove a stronger result, see Theorem \[thm:corrector\].
### Extension: application to traffic lights {#extension-application-to-traffic-lights .unnumbered}
The techniques developed to prove the Theorem \[thm:conv\] allow us to deal with a different situation inspired from traffic flow problems. As explained in [@imz], such problems are related to the study of some Hamilton-Jacobi equations. The problem that we address in Theorem \[thm:conv-time\] below is motivated by its meaningful application to traffic lights. We aim at figuring out how the fraffic flow on an ideal (infinite, straight) road is modified by the presence of a finite number of traffic lights.
We can consider a Hamilton-Jacobi equation whose Hamiltonian does not depend on $(t,x)$ for $x$ outside a (small) interval of the form $N_\eps = (b_1 \eps, b_N \eps)$ and is piecewise constant with respect to $x$ in $(b_1 \eps,b_N \eps)$. At space discontinuities, junction conditions are imposed with $\eps$-time periodic flux limiters. The limit solution satifies the equation after the “neighbourhood” $N_\eps$ disappeared. We will see that the equation keeps memory of what happened there through a flux limiter at the origin $x=0$.
Let us be more precise now. For $N\ge 1$, (a finite number of) junction points $-\infty=b_0<b_1<b_2<\dots <b_N<b_{N+1}=+\infty$ and (a finite number of) times $0=\tau_0<\tau_1<\dots
<\tau_K<1=\tau_{K+1}$, $K \in \N$ are given. For $N\ge 1$ and $\alpha \in \{0,\dots,N\}$, $\ell_\alpha$ denotes $b_{\alpha+1}-b_\alpha$. Note that $\ell_\alpha=+\infty$ for $\alpha=0,N$.
We then consider the solution $u^\eps$ of where the Hamiltonian $H$ satifies the following conditions.
- The Hamiltonian is given by $$H (t,x,p) =
\begin{cases}\bar H_{\alpha}(p) & \text{ if } b_{\alpha}
<x < b_{\alpha+1} \\
\max (\bar H_{\alpha-1}^+(p^-),\bar
H_{\alpha}^-(p^+), a_\alpha(t) ) & \text{ if } x=b_\alpha, \alpha
\neq 0.
\end{cases}$$
- The Hamiltonians $\bar{H}_\alpha$, for $\alpha=0,\dots, N$, are continuous, coercive and quasi-convex.
- The flux limiters $a_\alpha$, for $\alpha=1,\dots,N$ and $i=0,\dots,K$, satisfy $$a_\alpha(s+1)=a_\alpha(s) \quad \text{with}\quad a_\alpha(s)=
A_\alpha^i \quad \text{for all}\quad s\in
\left[\tau_i,\tau_{i+1}\right)$$ with $(A_\alpha^i)_{\alpha=1,\dots,N}^{i=0,\dots,K}$ satisfying $A_\alpha^i \ge \max_{\beta=\alpha-1,\alpha}\left(\min \bar
H_\beta\right).$
The Hamiltonians outside $N_\eps$ are denoted by $\bar H_\alpha$ instead of $H_\alpha$ in order to emphasize that they do not depend on time and space.
In view of the litterature in traffic modeling, the Hamiltonians could be assumed to be convex. But we prefer to stick to the quasi-convex framework since it seems to us that it is the natural one (in view of [@im]).
The equation is supplemented with the following initial condition $$\label{eq:ic-bis}
u^\eps (0,x) = U^\eps_0 (x) \quad \text{ for } x \in \R$$ with $$\label{hyp:uo}
U^\eps_0 \text{ is equi-Lipschitz continuous and } U^\eps_0 \to u_0
\text{ locally uniformly}.$$ Then the following convergence result holds true.
\[thm:conv-time\] Assume **(C1)-(C3)** and . Let $u^\eps$ be the solution of - for all $\eps >0$. Then:
1. [(Homogenization)]{} There exists some $\bar A\in \R$ such that $u^\eps$ converges locally uniformly as $\eps$ tends to zero towards the unique viscosity solution $u^0$ of - with $$\bar H_L:=\bar H_0,\quad \bar H_R:= \bar H_N.$$
2. [(Qualitative properties of $\bar A$)]{} For $\alpha=1,\dots,N$, $\langle a_\alpha \rangle$ denotes $\int_0^1 a_\alpha(s)\ ds$. The effective limiter $\bar A$ satisfies the following properties.
- For all $\alpha$, $\bar A$ is non-increasing w.r.t. $\ell_\alpha$.
- For $N=1$, $$\label{eq:N1}
\bar A = \langle a_1 \rangle.$$
- For $N \ge 1$, $$\label{eq:Nge1}
\bar A \ge \max_{\alpha=1,\dots,N}\quad \langle a_\alpha \rangle .$$
- For $N \ge 2$, there exists a critical distance $d_0 \ge 0$ such that $$\label{eq:critical}
\bar A = \max_{\alpha=1,\dots,N}\quad \langle a_\alpha \rangle \quad
\quad \mbox{if}\quad \quad \min_\alpha \ell_\alpha \ge d_0;$$ this distance $d_0$ only depends on $\displaystyle
\max_{\alpha=1,\dots,N} \|a_\alpha\|_\infty$, $\displaystyle
\max_{\alpha=1,\dots,N} \langle a_\alpha\rangle$ and the $\bar
H_\alpha$’s.
- We have $$\label{eq:limit}
\bar A \to \langle \bar a \rangle \quad \mbox{as}\quad (\ell_1,\dots,\ell_{N-1}) \to (0,\dots,0)$$ where $\bar a(\tau)=\max_{\alpha=1,\dots,N} a_\alpha(\tau)$.
\[rem:meaning\] Since the function $a(t)$ is piecewise constant, the way $u^\eps$ satisfies has to be made precise. An $L^1$ theory in time (following for instance the approach of [@b1; @b2]) could probably be developed for such a problem, but we will use here a different, elementary approach. The Cauchy problem is understood as the solution of successive Cauchy problems. This is the reason why we will first prove a global Lipschitz bound on the solution so that there indeed exists such a solution.
\[rem::g1\] Note that the result of Theorem \[thm:bar-A\] still holds for equation under Assumptions **(C1)-(C3)**, with the set $E$ defined for sub-solutions which are moreover assumed to be globally Lipschitz (without fixed bound on the Lipschitz constant). The reader can check that the proof is unchanged.
It is somewhat easy to get when the Hamiltonians $\bar
H_\alpha$ are convex by using the optimal control interpretation of the problem. In the more general case of quasi-convex Hamiltonians, the result still holds true but the proof is more involved.
\[rem::g81\] We may have $\bar A > \max_{\alpha=1,\dots,N}\langle a_\alpha
\rangle$. It is possible to deduce it from in the case $N=2$ by using the traffic light interpretation of the problem. If we have two traffic lights very close to each other (let us say that the distance in between is at most the place for only one car), and if the common period of the traffic lights are exactly in opposite phases (with for instance one minute for the green phase, and one minute for the red phase), then the effect of the two traffic lights together, gives a very low flux which is much lower than the effect of a single traffic light alone (*i.e.* here at most one car every two minutes will go through the two traffic lights).
Traffic flow interpretation of Theorem \[thm:conv-time\]
--------------------------------------------------------
We mentioned above that there are some connections between our problem and traffic flows.
Inequality has a natural traffic interpretation, saying that the average limitation on the traffic flow created by several traffic lights on a single road is higher or equal to the one created by the traffic light which creates the highest limitation. Moreover this average limitation is smaller if the distances between traffic lights are bigger, as says the monotonicity of $\bar A$ with respect to the distances $\ell_\alpha$.
Property (\[eq:critical\]) says that the minimal limitation is reached if the distances between the traffic lights are bigger than a critical distance $d_0$. The proof of this result is quite involved and is reflected in the fact that the bounds that we have on $d_0$ are not continuous on the data ($\displaystyle \max_{\alpha=1,\dots,N}
\|a_\alpha\|_\infty$, $\displaystyle \max_{\alpha=1,\dots,N}\langle
a_\alpha \rangle$ and the $\bar H_\alpha$’s).
Finally property (\[eq:limit\]) is very natural from the point of view of traffic, since it corresponds to the case where all the traffic lights would be at the same position.
Related results {#subsec:rel}
---------------
Achdou and Tchou [@at] studied a singular perturbation problem which has the same flavor as the one we are looking at in the present paper. More precisely, they consider the simplest network (a so-called junction) embedded in a star-shaped domain. They prove that the value function of an infinite horizon control problem converges, as the star-shaped domain “shrinks” to the junction, to the value function of a control problem posed on the junction. We borrow from them the idea of studying the cell problem on truncated domains with state constraints. We provide a different approach, which is also in some sense more general because it can be applied to problems outside the framework of optimal control theory. Our approach relies in an essential way on the general theory developed in [@im].
The general theme of Lions’s 2013-2014 lectures at Collège de France [@lions] is “Elliptic or parabolic equations and specified homogenization”. As far as first order Hamilton-Jacobi equations are concerned, the term “specified homogenization” refers to the problem of constructing correctors to cell problems associated with Hamiltonians that are typically the sum of a periodic one $H$ and a compactly supported function $f$ depending only on $x$, say. Lions exhibits sufficient conditions on $f$ such that the effective Hamilton-Jacobi equation is not perturbed. In terms of flux limiters [@im], it corresponds to look for sufficient conditions such that the effective flux limiter $\bar A$ given by Theorem \[thm:bar-A\] is (less than or) equal to $A_0=
\min H$.
Barles, Briani and Chasseigne [@bbc Theorem 6.1] considered the case $$H(x,p ) = \varphi \left(\frac{x}\eps\right) H_R (p) +
\left(1-\varphi\left(\frac{x}\eps\right)\right) H_L (p)$$ for some continuous increasing function $\varphi: \R \to \R$ such that $$\lim_{s\to -\infty} \varphi (s)=0 \quad \text{ and } \quad \lim_{s\to
+\infty} \varphi (s) =1.$$ They prove that $u^\eps$ converges towards a value function denoted by $U^-$, that they characterize as the solution to a particular optimal control problem. It is proved in [@im] that $U^-$ is the solution of with $\bar
H_\alpha = H_\alpha$ and $\bar A$ replaced with $A_I^+= \max(A_0,A^*)$ with $$A_0 = \max (\min H_R,\min H_L) \quad \text{ and } \quad A^* =
\max_{q \in [\min(p_R^0,p_L^0),\max(p_R^0,p_L^0)]} (\min (H_R(q),H_L(q))).$$
In [@gh], Giga and Hamamuki develop a theory which allows in particular to prove existence and uniqueness for the following Hamilton-Jacobi equation (changing $u$ in $-u$) in $\R^d$, $$\begin{cases} \partial_t u + |\nabla u | = 0 & \text{ for } x \neq 0 \\
\partial_t u + |\nabla u | + c = 0 & \text{ at } x =0.\end{cases}$$ The solutions of [@gh] are constructed as limits of the following equation $$\partial_t u^\eps + |\nabla u^\eps | + c (1-|x|/\eps)^+ = 0.$$ In the monodimensional case ($d=1$), Theorem \[thm:conv\] implies that $u^\eps$ converges towards $$\begin{cases} \partial_t u + |\nabla u | = 0 & \text{ for } x \neq 0 \\
\partial_t u + \max(A,|\nabla u |) = 0 & \text{ at } x =0
\end{cases}$$ for some $A \in \R$. In view of Theorem \[thm:bar-A\], it is not difficult to prove that $A= \max
(0,c)$. The Hamiltonian $\max(c,|\nabla u |)$ is identified in [@gh] and is referred to as the *relaxed* one.
It is known that homogenization of Hamilton-Jacobi equations is closely related to the study of the large time behaviour of solutions. In [@hamamuki], the large time behaviour of Hamilton-Jacobi equations with discontinuous source terms is discussed in two cases: for compactly supported ones and periodic ones. Remark that in our setting, we can adress both and even the sum of a periodic source term and of a compactly supported one. It would be interesting to adress such a problem in the case of traffic lights. In [@jy], the authors study the large time behaviour of the solutions of a Hamilton-Jacobi equations with an $x$-periodic Hamiltonian and what can be interpreted as a flux-limiter depending periodically in time.
Further extensions {#ss1}
------------------
It is also possible to adress the time homogenization problem of Theorem \[thm:conv-time\] with any finite number of junctions (with limiter functions $a_\alpha(t)$ piecewise constants – or continuous – and $1$-periodic), either separated with distance of order $O(1)$ or with distance of order $O(\varepsilon)$, or mixing both, and even on a complicated network. See also [@jy] for other connexions between Hamilton-Jacobi equations and traffic light problems and [@ags] for green waves modelling.
Note that the method presented in this paper can be readily applied (without modifying proofs) to the study of homogeneization on a finite number of branches and not only two branches; the theory developed in [@im] should also be used for the limit problem.
Similar questions in higher dimensions with point defects of other co-dimensions will be addressed in future works.
Organization of the article
---------------------------
Section \[sec:conv\] is devoted to the proof of the convergence result (Theorem \[thm:conv\]). Section \[sec:homog\] is devoted to the construction of correctors far from the junction point (Proposition \[prop:quasi-conv\]) while the junction case, *i.e.* the proof of Theorem \[thm:corrector\], is addressed in Section \[sec:trunc\]. We recall that Theorem \[thm:corrector-simple\] is a straightforward corollary of this stronger result. The proof of Theorem \[thm:corrector\] makes use of a comparison principle which is expected but not completely standard. This is the reason why a proof is sketched in Appendix, together with two other ones that are rather standard but included for the reader’s convenience.
#### Notation.
A ball centered at $x$ of radius $r$ is denoted by $B_r(x)$. If $\{u^\eps\}_\eps$ is locally bounded, the upper and lower relaxed limits are defined as $$\begin{cases}
\displaystyle \limsup_\eps{}^* u^\eps (X) = \limsup_{Y \to X, \eps \to 0} u^\eps
(Y), \\
\displaystyle \liminf_\eps{}_* u^\eps (X) = \liminf_{Y \to X, \eps \to 0} u^\eps
(Y).
\end{cases}$$ In our proofs, constants may change from line to line.
Proof of convergence {#sec:conv}
====================
This section is devoted to the proof of Theorem \[thm:conv\]. We first construct barriers.
\[Barriers\] There exists a nonnegative constant $C$ such that for any $\varepsilon>0$ $$\label{barriers1}
\left|u^\varepsilon(t,x)-u_0(x)\right|\leq Ct\quad\text{for}\quad(t,x)\in(0,T)\times\R\,.$$
Let $L_0$ be the Lipschitz constant of the initial datum $u_0$. Taking $$C=\sup_{\stackrel{(t,x)\in\R\times\R}{|p|\leq
L_0}}\left|H(t,x,p)\right|<+\infty$$ owing to **(A0)** and **(A5)**, the functions $u^\pm(t,x)=u_0(x)\pm Ct$ are a super- and a sub-solution of (\[eq:hj-eps\])-(\[eq:ic\]) respectively and (\[barriers1\]) follows via comparison principle.
We can now prove the convergence theorem.
We classically consider the upper and lower relaxed semi-limits $$\begin{cases}
\overline u =\displaystyle \limsup_\eps{}^* u^\varepsilon,\\
\underline u =\displaystyle \liminf_\eps{}_* u^\varepsilon\,.
\end{cases}$$ Notice that these functions are well defined because of Lemma \[Barriers\]. In order to prove convergence of $u^\varepsilon$ towards $u^0$, it is sufficient to prove that $\overline u$ and $\underline u$ are a sub- and a super-solution of (\[eq:hj-homog\])-(\[eq:ic\]) respectively. The initial condition immediately follows from (\[barriers1\]). We focus our attention on the sub-solution case since the super-solution one can be handled similarly.
We first check that $$\label{eq:wk}
\overline u (t,0) = \limsup_{(s,y) \to (t,0), y >0} \overline u(s,y) =
\limsup_{(s,y) \to (t,0), y <0} \overline u(s,y).$$ This is a consequence of the stability of such a “weak continuity” condition, see [@im]. Indeed, it is shown in [@im] that classical viscosity solution can be viewed as flux-limited one; in particular, $u^\eps$ solves $$u^\eps_t + H^- \left(\frac{t}\eps,\frac{0}\eps,u^\eps_x (t,0^+) \right) \vee
H^+ \left(\frac{t}\eps,\frac{0}\eps,u^\eps_x (t,0^-)) \right) = 0 \quad \text{ for } t >0.$$ Since these $\eps$-Hamiltonians are uniformly coercive and $u^\eps$ is continuous, we conclude that holds true.
Let $\varphi$ be a test function such that $$\label{eq::g2}
(\overline u-\varphi)(t,x)<(\overline u-\varphi)(\overline
t,\overline x)=0 \quad\forall(t,x)\in B_{\overline r}(\overline
t,\overline x)\setminus\left\{(\overline t,\overline x)\right\}.$$ We argue by contradiction by assuming that $$\label{conv2}
\varphi_t(\overline t,\overline x)+\bar H\left(\bar
x,\varphi_x(\overline t,\overline x)\right)=\theta>0,$$ where $$\bar H\left(\bar x,\varphi_x(\overline t,\overline x)\right) := \left\{
\begin{array}{ll}
\bar H_R(\varphi_x(\overline t,\overline x)) &\quad \mbox{if}\quad \overline x >0,\\
\bar H_L(\varphi_x(\overline t,\overline x)) &\quad \mbox{if}\quad \overline x <0,\\
F_{\bar A}(\varphi_x(\overline t,0^-),\varphi_x(\overline t,0^+)) &\quad \mbox{if}\quad \overline x =0.
\end{array}\right.$$ We only treat the case where $\overline x=0$ since the case $\overline
x \neq 0$ is somewhat classical. This latter case is detailed in Appendix for the reader’s convenience. Using [@im Proposition 2.8], we may suppose that $$\label{eq::g3}
\varphi(t,x)=\phi(t)+\bar {p}_L x1_{\left\{x<0\right\}}+\bar {p}_R x 1_{\left\{x>0\right\}}$$ where $\phi$ is a $C^1$ function defined in $(0,+\infty)$. In this case, Eq. becomes $$\label{conv7}
\phi'(\bar t)+F_{\bar A}\left(\bar{p}_L,\bar{p}_R\right)=\phi'(\bar t)+\bar A=\theta>0.$$ Let us consider a solution $w$ of the equation $$\label{conv8}
w_t+H(t,x,w_x)=\bar A$$ provided by Theorem \[thm:corrector-simple\], which is in particular $1$-periodic with respect to time. We recall that the function $W$ is the limit of $w^\eps = \eps w (\cdot/\eps)$ as $\eps \to 0$. We claim that, if $\varepsilon>0$ is small enough, the perturbed test function $\varphi^\varepsilon(t,x)=\phi(t)+w^\varepsilon(t,x)$ [@evans] is a viscosity super-solution of $$\varphi^\varepsilon_t+H\left(\frac t\varepsilon, \frac
x\varepsilon,\varphi^\varepsilon_x\right)
=\frac\theta2\quad\text{in}\quad B_r(\overline t,0)$$ for some sufficiently small $r>0$. In order to justify this fact, let $\psi(t,x)$ be a test function touching $\varphi^\varepsilon$ from below at $(t_1,x_1)\in
B_r(\overline t,0)$. In this way $$w\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon}\right)
=\frac1\varepsilon\left(\psi(t_1,x_1)-\phi(t_1)\right)$$ and $$w\left(s,y\right) \ge \frac1\varepsilon\left(\psi(\varepsilon
s,\varepsilon y) -\phi(\varepsilon s)\right)$$ for $(s,y)$ in a neighborhood of $\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon}\right)$. Hence from - $$\begin{split}
\psi_t(t_1,x_1)+H\left(\frac {t_1}\varepsilon, \frac
{x_1}\varepsilon,\psi_x(t_1,x_1)\right)& \ge \bar A+\phi'(t_1)\\
&\geq\bar A+\phi'(\overline t)-\frac\theta2\geq\frac\theta2
\end{split}$$ provided $r$ is small enough. Hence, the claim is proved.
Combining (\[eq:W-estim-2\]) from Theorem \[thm:corrector-simple\] with (\[eq::g2\]) and (\[eq::g3\]), we can fix $\kappa_r>0$ and $\eps>0$ small enough so that $$u^\varepsilon+\kappa_r \le \varphi^\eps \quad\text{on}\quad\partial B_r(\overline t,0).$$ By comparison principle the previous inequality holds in $B_r(\overline t,0)$. Passing to the limit as $\eps \to 0$ and $(t,x)
\to (\bar t,\bar x)$, we get the following contradiction $$\overline u(\overline t,0)+\kappa_r\le\varphi(\overline
t,0)=\overline u(\overline t,0).$$ The proof of convergence is now complete.
\[rem::1bis\] For the super-solution property, $\varphi$ in (\[eq::g3\]) should be replaced with $$\varphi(t,x)=\phi(t)+ \hat{p}_L x 1_{\left\{x<0\right\}} + \hat{p}_R x 1_{\left\{x>0\right\}}.$$
Homogenized Hamiltonians {#sec:homog}
========================
In order to prove Proposition \[prop:quasi-conv\], we first prove the following lemma. Even if the proof is standard, we give it in full details since we will adapt it when constructing global correctors for the junction.
\[lem:exis-base\] There exists $\lambda \in \R$ and a bounded (discontinuous) viscosity solution of .
\[rem:time-indep\] If $H_\alpha$ does not depend on $t$, then it is possible to construct a corrector which does not depend on time either. We leave details to the reader.
For any $\delta >0$, it is possible to construct a (possibly discontinuous) viscosity solution $v^\delta$ of $$\begin{cases}
\delta v^\delta +v^\delta_t + H_\alpha (t,x,p+v_x^\delta) = 0 \quad
\text{ in } \R \times \R, \\
v^\delta \text{ is $\Z^2$-periodic}.
\end{cases}$$ First, the comparison principle implies $$\label{eq:naive}
|\delta v^\delta | \le C_\alpha$$ where $$C_\alpha = \sup_{(t,x)\in \left[0,1\right]^2} |H_\alpha (t,x,p)|.$$ Second, the function $$m^\delta (x) = \sup_{t \in \R} (v^\delta)^* (t,x)$$ is a sub-solution of $$H_\alpha (t(x),x,p+m^\delta_x ) \le C_\alpha$$ (for some function $t(x)$). Assumptions **(A3)** and **(A5)** imply in particular that there exists $C>0$ independent of $\delta$ such that $$|m^\delta_x | \le C$$ and $$v^\delta_t \le C.$$ In particular, the comparison principle implies that for all $t \in
\R$ and $x \in \R$ and $h \ge 0$, $$v^\delta (t+h,x) \le v^\delta (t,x) + Ch.$$ Combining this inequality with the time-periodicity of $v^\delta$ yields $$|v^\delta (t,x) - m^\delta (x) | \le C;$$ in particular, $$\label{eq::g4}
|v^\delta (t,x) - v^\delta (0,0)| \le C.$$ Hence, the half relaxed limits $$\bar v = \limsup_{\delta\to 0}{}^*(v^\delta - v^\delta (0,0)) \quad \text{
and } \quad \underline v = \liminf_{\delta\to 0}{}_*(v^\delta - v^\delta
(0,0))$$ are finite. Moreover, implies that $\delta v^\delta (0,0) \to - \lambda$ (at least along a subsequence). Hence, discontinuous stability of viscosity solutions implies that $\bar v$ is a $\Z^2$-periodic sub-solution of and $\underline v$ is a $\Z^2$-periodic super-solution of the same equation. Perron’s method then allows us to construct a corrector between $\bar v$ and $\underline v + C$ with $C = \sup (\bar v -
\underline v)$. The proof of the lemma is now complete.
The following lemma is completely standard; the proof is given in Appendix for the reader’s convenience.
\[lem:unique\] The real number $\lambda$ given by Lemma \[lem:exis-base\] is unique. If $\bar H_\alpha(p)$ denotes such a real number, the function $\bar H_\alpha$ is continuous.
\[lem:coercive\] The continuous function $\bar H_\alpha$ is coercive, $$\lim_{|p|\to +\infty} \bar H_\alpha (p)= +\infty.$$
In view of the uniform coercivity in $p$ of $H_\alpha$ with respect to $(t,x)$ (see **(A3)**), for any $R>0$ there exists a positive constant $C_R$ such that $$\label{coercivity1}
|p|\geq C_R\quad\Rightarrow\quad \forall (t,x)\in\R\times\R, \quad H_\alpha(t,x,p)\geq R.$$ Let $v^\alpha$ be the discontinuous corrector given by Lemma \[lem:exis-base\] and $(\bar t,\bar x)$ be point of supremum of its upper semi-continuous envelope $(v^\alpha)^*$. Then we have $$H_\alpha(\bar t,\bar x,p)\le \bar H_\alpha(p)$$ which implies $$\label{coercivity2}
\bar H_\alpha(p)\geq R\quad\text{for}\quad |p|\geq C_R.$$ The proof of the lemma is now complete.
We first prove the quasi-convexity of $\bar H_\alpha$ under assumption **(B-ii)**. We prove in fact more: the effective Hamiltonian is convex in this case.
\[lem:conv\] Assume **(A0)-(A5)** and **(B-ii)**. Then the function $\bar H_\alpha$ is convex.
For $p,\,q\in\R$, let $v_p$, $v_q$ be solutions of (\[eq:cell-alpha\]) with $\lambda=\bar H_\alpha(p)$ and $\bar
H_\alpha(q)$ respectively. We also set $$u_p(t,x)=v_p(t,x) + px - t\bar H_\alpha(p)$$ and define similarly $u_q$.
#### Step 1: $u_p$ and $u_q$ are locally Lipschitz continuous.
In this case, we have almost everywhere: $$\left\{\begin{array}{l}
(u_p)_t + H_\alpha(t,x,(u_p)_x) = 0,\\
(u_q)_t + H_\alpha(t,x,(u_q)_x) = 0.\\
\end{array}\right.$$ For $\mu\in \left[0,1\right]$, let $$\bar u = \mu u_p + (1-\mu)u_q.$$ By convexity, we get almost everywhere $$\label{eq::g5}
\bar u_t + H_\alpha(t,x,\bar u_x) \le 0.$$ We claim that the convexity of $H_\alpha$ (in the gradient variable) implies that $\bar u$ is a viscosity sub-solution. To see it, we use an argument of [@bcd Proposition 5.1]. For $P=(t,x)$, we define a mollifier $\rho_\delta(P)=\delta^{-2}\rho(\delta^{-1}P)$ and set $$\bar u_\delta=\bar u\star \rho_\delta$$ Then by convexity, we get with $Q=(s,y)$: $$(\bar u_\delta)_t + H_\alpha(P,(\bar u_\delta)_x) \le \int dQ\
\left\{H_\alpha(P,\bar u_x(Q))-H_\alpha(Q,\bar u_x(Q)\right\}\rho_\delta(P-Q).$$
The fact that $\bar u_x$ is locally bounded and the fact that $H_\alpha$ is continuous imply that the right hand side goes to zero as $\delta \to
0$. We deduce (by stability of viscosity sub-solutions) that (\[eq::g5\]) holds true in the viscosity sense. Then the comparison principle implies that $$\label{eq::g10}
\mu \bar H_\alpha(p)+ (1-\mu) \bar H_\alpha(q) \ge \bar H_\alpha(\mu p + (1-\mu)q).$$
#### Step 2: $u_p$ and $u_q$ are continuous.
We proceed in two (sub)steps.
<span style="font-variant:small-caps;">Step 2.1: the case of a single function $u$.</span> We first want to show that if $u=u_p$ is continuous and satisfies (\[eq::g5\]) almost everywhere, then $u$ is a viscosity sub-solution. To this end, we will use the structural assumptions satisfied by the Hamiltonian. The ones that were useful to prove the comparison principle will be also useful to prove the result we want. Indeed, we will revisit the proof of the comparison principle. We also use the fact that $$\label{eq::g6}
u(t,x)-px + t\bar H_\alpha(p) \quad \mbox{is bounded}.$$
For $\nu>0$, we set $$u^\nu(t,x)=\sup_{s\in \R} \left(u(s,x)-\frac{(t-s)^2}{2\nu}\right)=u(s_\nu,x)-\frac{(t-s_\nu)^2}{2\nu}.$$ As usual, we get from (\[eq::g6\]) that $$\label{eq::g8}
\left|t-s_\nu\right| \le C\sqrt{\nu} \quad \mbox{with}\quad C=C(p,T)$$ for $t \in (-T,T)$. In particular $ s_\nu \to t$ locally uniformly. If a test function $\varphi$ touches $u^\nu$ from above at some point $(t,x)$, then we have $\displaystyle \varphi_t(t,x)=-\frac{t-s_\nu}{\nu}$ and $$\begin{aligned}
\varphi_t(t,x)+ H_\alpha(t,x, \varphi_x(t,x)) &
\le H_\alpha(t,x, \varphi_x(t,x))-H_\alpha(s_\nu,x, \varphi_x(t,x))\nonumber \\
& \le \omega(\left|t-s_\nu\right|(1+\max (0, H_\alpha(s_\nu,x,
\varphi_x(t,x))))) \nonumber \\
& \le \omega\left(\frac{(t-s_\nu)^2}{\nu}+ \left|t-s_\nu\right|\right)
\label{eq::g7}\end{aligned}$$ where we have used [**(A2)**]{} in the third line. The right hand side goes to zero as $\nu$ goes to zero since $$\frac{(t-s_\nu)^2}{\nu} \to 0 \quad \mbox{locally uniformly w.r.t. $(t,x)$}$$ (recall $u$ is continuous). Indeed, this can be checked for $(t,x)$ replaced by $(t_\nu,x_\nu)$ because for any sequence $(t_\nu,s_\nu,x_\nu)\to (t,t,x)$, we have $$u(t_\nu,x_\nu) \le u^\nu(t_\nu,x_\nu)=u(s_\nu,x_\nu) - \frac{(t_\nu-s_\nu)^2}{2\nu}$$ where the continuity of $u$ implies the result. For a given $\nu>0$, we see that (\[eq::g8\]) and (\[eq::g7\]) imply that $$\left|\varphi_t \right|, \left|\varphi_x\right| \le C_{\nu,p}.$$ This implies in particular that $u^\nu$ is Lipschitz continuous, and then $$u^\nu_t + H(t,x,u^\nu_x)\le o_\nu(1) \quad \mbox{a.e.}$$ where $o_\nu(1)$ is locally uniform with respect to $(t,x)$.
[Step 2.2: application.]{} Applying Step 2.1, we get for $z=p,q$ $$(u^\nu_z)_t + H(t,x,(u^\nu_z)_x)\le o_\nu(1) \quad \mbox{a.e.}$$ where $o_\nu(1)$ is locally uniform with respect to $(t,x)$. Step 1 implies that $$\bar u^\nu:= \mu u^\nu_p + (1-\mu)u^\nu_q$$ is a viscosity sub-solution of $$(\bar u^\nu)_t + H_\alpha(t,x,(\bar u^\nu)_x)\le o_\nu(1)$$ where $o_\nu(1)$ is locally uniform with respect to $(t,x)$. In the limit $\nu\to 0$, we recover (by stability of sub-solutions) that $\bar u$ is a viscosity sub-solution, *i.e.* satisfies (\[eq::g5\]) in the viscosity sense. This gives then the same conclusion as in Step 1.
#### Step 3: the general case.
To cover the general case, we simply replace $u_p$ by $\tilde{u}_p$ which is the solution to the Cauchy problem $$\left\{\begin{array}{l}
(\tilde{u}_p)_t + H_\alpha(t,x,(\tilde{u}_p)_x)=0,\quad \mbox{for}\quad (t,x)\in (0,+\infty)\times \R\\
\tilde{u}_p(0,x)=px,
\end{array}\right.$$ Then $\tilde{u}_p$ is continuous and satisfies $\left|\tilde{u}_p-u_p\right|\le C$. Proceeding similarly with $\tilde{u}_q$ and using Step 2, we deduce the desired inequality (\[eq::g10\]). The proof is now complete.
We finally prove the quasi-convexity of $\bar H_\alpha$ under assumption **(B-i)**.
\[lem:quasi-conv\] Assume **(A0)-(A5)** and **(B-i)**. Then the function $\bar H_\alpha$ is quasi-convex.
We reduce quasi-convexity to convexity by composing with an increasing function $\gamma$; notice that such a reduction was already used in optimization and in partial differential equations, see for instance [@lions-convexe; @ka].
We first assume that $H_\alpha$ satisfies $$\label{eq::g11}
\left\{\begin{array}{l}
H_\alpha\in C^2,\\
D^2_{pp}H_\alpha(x,p^0_\alpha)>0,\\
D_p H_\alpha(x,p)<0 \quad \mbox{for}\quad p\in (-\infty,p^0_\alpha),\\
D_p H_\alpha(x,p)>0 \quad \mbox{for}\quad p\in (p^0_\alpha, +\infty),\\
H_\alpha(x,p)\to +\infty \quad \mbox{as}\quad |p|\to +\infty \quad
\mbox{uniformly w.r.t. } x\in \R.
\end{array}\right.$$ For a function $\gamma$ such that $$\gamma \quad \text{is convex},\quad \gamma \in C^2(\R) \quad \text{and}\quad \gamma' \ge \delta_0>0$$ we have $$D^2_{pp} (\gamma \circ H_\alpha) >0$$ if and only if $$\label{eq:cns}
(\ln \gamma')'(\lambda) > -\frac{D^2_{pp}H_\alpha(x,p)}{(D_p H_\alpha(x,p))^2}\quad \mbox{for}\quad
p=\pi_\alpha^\pm(x,\lambda)\quad \text{and}\quad \lambda\ge H_\alpha(x,p)$$ where $\pi_\alpha^\pm(x,\lambda)$ is the only real number $r$ such that $\pm r \ge 0$ and $H_\alpha (x,r) = \lambda$. Because $D^2_{pp}H_\alpha(x,p^0_\alpha)>0$, we see that the right hand side is negative for $\lambda$ close enough to $H_\alpha(x,p^0_\alpha)$ and it is indeed possible to construct such a function $\gamma$.
In view of Remark \[rem:time-indep\], we can construct a solution of $\delta v^\delta + \gamma\circ H_\alpha(x,p+v^\delta_x)=0$ with $-\delta v^\delta \to \overline{\gamma\circ H_\alpha}(p)$ as $\delta\to 0$, and a solution of $$\gamma\circ H_\alpha(x,p+v_x)=\overline{\gamma\circ H_\alpha}(p)$$ This shows that $$\bar H_\alpha = \gamma^{-1} \circ \overline{\gamma \circ H_\alpha}.$$ Thanks to Lemmas \[lem:coercive\] and \[lem:conv\], we know that $\overline{\gamma \circ H_\alpha}$ is coercive and convex. Hence $\bar
H_\alpha$ is quasi-convex.
If now $H_\alpha$ does not satisfies (\[eq::g11\]), then for all $\varepsilon>0$, there exists $H_\alpha^\varepsilon \in C^2$ such that $$\begin{cases}
(D^2_{pp}H_\alpha^\varepsilon)(x,p^0_\alpha)>0 \\
D_p H_\alpha^\eps(x,p)<0 \quad \mbox{for}\quad p\in (-\infty,p^0_\alpha),\\
D_p H_\alpha^\eps(x,p)>0 \quad \mbox{for}\quad p\in (p^0_\alpha, +\infty),\\
|H_\alpha^\varepsilon - H_\alpha| < \varepsilon.
\end{cases}$$ Then we can argue as in the proof of continuity of $\bar H_\alpha$ and deduce that $$\bar H_\alpha (p ) = \lim_{\varepsilon \to 0} \bar H_\alpha^\varepsilon
(p).$$ Moreover, the previous case implies that $\bar H_\alpha^\varepsilon$ is quasi-convex. Hence, so is $\bar H_\alpha$. The proof of the lemma is now complete.
Combine Lemmas \[lem:exis-base\], \[lem:unique\], \[lem:coercive\], \[lem:conv\] and \[lem:quasi-conv\].
Truncated cell problems {#sec:trunc}
=======================
We consider the following problem: find $\lambda_\rho\in \R$ and $w$ such that $$\label{eq:cell-trunc}
\left\{\begin{array}{ll}
w_t + H(t,x,w_x)=\lambda_\rho, & (t,x)\in \R\times (-\rho,\rho),\\
w_t + H^-(t,x,w_x)=\lambda_\rho, & (t,x)\in \R\times \left\{-\rho\right\},\\
w_t + H^+(t,x,w_x)=\lambda_\rho, & (t,x)\in \R\times
\left\{\rho\right\}, \medskip\\
w \text{ is $1$-periodic w.r.t. } t.
\end{array}\right.$$ Even if our approach is different, we borrow here an idea from [@at] by truncating the domain and by considering correctors in $[-\rho,\rho]$ with $\rho \to +\infty$.
A comparison principle
----------------------
\[prop:comp\] Let $\rho_2>\rho_1>\rho_0$ and $\lambda\in \R$ and $v$ be a super-solution of the following boundary value problem $$\label{eq:super-comp}
\left\{\begin{array}{ll}
v_t + H(t,x,v_x) \ge \lambda & \text{ for } (t,x)\in \R\times (\rho_1,\rho_2),\\
v_t + H^+(t,x,v_x) \ge \lambda & \text{ for }(t,x)\in \R\times \left\{\rho_2\right\},\\
v(t,x) \ge U_0(t) & \text{ for }(t,x)\in \R\times
\left\{\rho_1\right\}, \medskip\\
v \text{ is $1$-periodic w.r.t. } t
\end{array}\right.$$ where $U_0$ is continuous and for $\eps_0>0$ and $u$ be a sub-solution of the following one $$\label{eq:sub-comp}
\left\{\begin{array}{ll}
u_t + H(t,x,u_x)\le \lambda -\eps_0 & \text{ for } (t,x)\in \R\times (\rho_1,\rho_2),\\
u_t + H^+(t,x,u_x)\le \lambda -\eps_0 &\text{ for } (t,x)\in \R\times \left\{\rho_2\right\},\\
u(t,x) \le U_0(t) & \text{ for } (t,x)\in \R\times
\left\{\rho_1\right\},\medskip\\
u \text{ is $1$-periodic w.r.t. } t.
\end{array}\right.$$ Then $u \le v$ in $\R \times [\rho_1,\rho_2]$.
A similar result holds true if the Dirichlet condition is imposed at $x= \rho_2$ and junction conditions $$\begin{aligned}
v_t + H^-(t,x,v_x) &\ge \lambda \quad \quad \quad \text{ at } x = \rho_1 \\
u_t + H^-(t,x,u_x)&\le \lambda -\eps_0 \quad \; \text{ at } x = \rho_1\end{aligned}$$ are imposed at $x = \rho_1$.
The proof of Proposition \[prop:comp\] is very similar to (in fact simpler than) the proof of the comparison principle for Hamilton-Jacobi equations on networks contained in [@im]. The main difference lies in the fact that in our case, $u$ and $v$ are global in time and the space domain is bounded. A sketch of the proof is provided in Appendix shedding some light on the main differences. Here the parameter $\varepsilon_0>0$ in (\[eq:sub-comp\]) is used in place of the standard correction term $-\eta/(T-t)$ for a Cauchy problem.
Correctors on truncated domains
-------------------------------
\[prop:cor-trunc\] There exists a unique $\lambda_\rho \in \R$ such that there exists a solution $w^\rho=w$ of . Moreover, there exists a constant $C>0$ independent of $\rho\in (\rho_0,+\infty)$ and a function $m^\rho
\colon [-\rho,\rho] \to \R$ such that $$\label{eq:trunc-cor}
\left\{\begin{array}{ll}
\left|\lambda_\rho\right| \le C,\\
\left|m^\rho(x)-m^\rho(y)\right|\le C \left|x-y\right| & \text{ for } x,y\in [-\rho,\rho],\\
\left|w^\rho(t,x)-m^\rho(x)\right|\le C & \text{ for } (t,x)\in \R\times [-\rho,\rho].
\end{array}\right.$$
In order to construct a corrector on the truncated domain, we proceed classically by considering $$\label{eq:cell-trunc-approx}
\left\{\begin{array}{ll}
\delta w^\delta + w^\delta_t + H(t,x,w^\delta_x)= 0, & (t,x)\in \R\times (-\rho,\rho)\,,\\
\delta w^\delta +w^\delta_t + H^-(t,x,w^\delta_x)= 0, & (t,x)\in \R\times \left\{-\rho\right\},\\
\delta w^\delta +w^\delta_t + H^+(t,x,w^\delta_x)=0, & (t,x)\in
\R\times \left\{\rho\right\}, \medskip\\
w^\delta \text{ is $1$-periodic w.r.t. } t.
\end{array}\right.$$ A discontinuous viscosity solution of is constructed by Perron’s method (in the class of $1$-periodic functions with respect to time) since $\pm \delta^{-1} C$ are trivial super-/sub-solutions if $C$ is chosen as follows $$C = \sup_{t \in \R,\ x \in \R} |H(t,x,0)|.$$ In particular, the solution $w^\delta$ satisfies by construction $$\label{eq:lli}
|w^\delta | \le \frac{C}\delta.$$ We next consider $$m^\delta (x) = \sup_{t \in \R} (w^\delta)^* (t,x).$$ We remark that the supremum is reached since $w^\delta$ is periodic with respect to time; we also remark that $m^\delta$ is a viscosity sub-solution of $$H(t(x),x,m^\delta_x)\le C, \quad x \in (-\rho,\rho)$$ (for some function $t(x)$). In view of **(A3)**, we conclude that $m^\delta$ is globally Lipschitz continuous and $$\label{eq:grad-md}
|m^\delta_x | \le C$$ for some constant $C$ which still only depends on $H$. Assumption **(A3)** also implies that, $$w^\delta_t \le C$$ (with $C$ only depending on $H$). In particular, the comparison principle implies that for all $t \in \R$, $x \in (-\rho,\rho)$ and $h
\ge 0$, $$w^\delta (t+h,x) \le w^\delta (t,x) + Ch.$$ Combining this information with the periodicity of $w^\delta$ with respect to $t$, we conclude that for $t \in \R$ and $x \in (-\rho,\rho)$, $$|w^\delta(t,x) - m^\delta (x)| \le C.$$ In particular, $$|w^\delta (t,x) - w^\delta (0,0)| \le C.$$ We then consider $$\overline w = \limsup_\delta{}^* (w^\delta - w^\delta (0,0)) \quad
\text{ and } \quad \underline w = \liminf_\delta{}_* (w^\delta -
w^\delta (0,0)).$$ We next remark that and imply that there exists $\delta_n \to 0$ such that $$\begin{aligned}
m^{\delta_n} - m^{\delta_n} (0) \to m^\rho \quad \text{ as } n \to
+\infty \\
\delta_n w^{\delta_n} (0,0) \to - \lambda_\rho \quad \text{ as } n \to +\infty\end{aligned}$$ (the first convergence being locally uniform). In particular, $\lambda$, $\overline w$, $\underline w$ and $m^\rho$ satisfies $$\begin{aligned}
| \lambda_\rho | \le C \\
|\overline w - m^\rho | \le C \\
|\underline w -m^\rho | \le C\\
|m^\rho_x | \le C.\end{aligned}$$ Discontinuous stability of viscosity solutions of Hamilton-Jacobi equations imply that $\overline w-2C$ and $\underline w$ are respectively a sub-solution and a super-solution of and $$\overline w - 2 C \le \underline w .$$ Perron’s method is used once again in order to construct a solution $w^\rho$ of which is $1$-periodic with respect to time. In view of the previous estimates, $\lambda_\rho$, $m^\rho$ and $w^\rho$ satisfy . Proving the uniqueness of $\lambda_\rho$ is classical so we skip it. The proof of the proposition is now complete.
\[prop:bar-A\] The map $\rho\mapsto \lambda_\rho$ is non-decreasing and bounded in $(0,+\infty)$. In particular, $$\bar A = \lim_{\rho \to +\infty} \lambda_\rho$$ exists and $\bar A \ge \lambda_\rho$ for all $\rho >0$.
For $\rho'>\rho>0$, we see that the restriction of $w^{\rho'}$ to $\left[-\rho,\rho\right]$ is a sub-solution, as a consequence of [@im Proposition 2.19]. The boundedness of the map follows from Proposition \[prop:cor-trunc\]. The proof is thus complete.
We next prove that we can control $w^\rho$ from below under appropriate assumptions on $\bar A$.
\[prop:slopes-trunc\] Assume first that $\bar A > \min \bar H_R$. Then for all $\delta>0$, there exists $\rho_\delta>0$ and $C_\delta >0$ (independent on $\rho$) such that for $x \ge
\rho_\delta$ and $h \ge 0$, $$\label{eq:estim-slope}
w^\rho (t,x+h) -w^\rho (t,x) \ge (\bar p_R-\delta) h - C_\delta.$$ If now we assume that $\bar A > \min \bar H_L$, then for $x\le
-\rho_\delta$ and $h\ge 0$, $$\label{eq:estim-slope-bis}
w^\rho (t,x-h) -w^\rho (t,x) \ge (-\bar p_L-\delta) h - C_\delta$$ for some $\rho_\delta>0$ and $C_\delta >0$ as above.
We only prove since the proof of follows along the same lines. Let $\delta
>0$. In view of **(A5)**, we know that there exists $\rho_\delta$ such that $$\label{eq:error-H}
|H(t,x,p) - H_R (t,x,p)| \le \delta \quad \text{ for } \quad x \ge
\rho_\delta.$$ Assume that $\bar A > \min \bar H_R$. Then Proposition \[prop:quasi-conv\] implies that we can pick $p^\delta_R$ such that $$\bar H_R (p^\delta_R) =\bar H_R^+ (p^\delta_R) = \lambda_\rho - 2
\delta$$ for $\rho \ge \rho_0$ and $\delta \le \delta_0$, by choosing $\rho_0$ large enough and $\delta_0$ small enough.
We now fix $\rho \ge \rho_\delta$ and $x_0\in \left[\rho_\delta,\rho\right]$. In view of Proposition \[prop:quasi-conv\] applied to $p = p_R^\delta$, we know that there exists a corrector $v_R$ solving with $\alpha =R$. Since it is $\Z^2$-periodic, it is bounded and $w_R
= p_R^\delta x + v_R (t,x)$ solves $$(w_R)_t + H_R(t,x,(w_R)_x)= \lambda_\rho -2\delta, \quad (t,x)\in \R\times \R.$$ In particular, the restriction of $w_R$ to $[\rho_\delta,\rho]$ satisfies (see [@im Proposition 2.19]), $$\left\{\begin{array}{ll}
(w_R)_t + H_R(t,x,(w_R)_x)\le \lambda_\rho -2\delta & \text{ for } (t,x)\in \R\times (\rho_\delta,\rho),\\
(w_R)_t + H^+_R(t,x,(w_R)_x)\le \lambda_\rho -2\delta &\text{ for } (t,x)\in \R\times \left\{\rho\right\}.
\end{array}\right.$$ In view of , this implies $$\left\{\begin{array}{ll}
(w_R)_t + H(t,x,(w_R)_x)\le \lambda_\rho -\delta & \text{ for } (t,x)\in \R\times (\rho_\delta,\rho),\\
(w_R)_t + H^+(t,x,(w_R)_x)\le \lambda_\rho -\delta &\text{ for } (t,x)\in \R\times \left\{\rho\right\}.
\end{array}\right.$$ Now we remark that $v = w^\rho - w^\rho (0,x_0)$ and $u = w_R -w_R
(0,x_0) -2C - 2 \|v_R\|_\infty$ satisfies $$v(t,x_0) \ge - 2C \ge u (t,x_0)$$ where $C$ is given by . Thanks to the comparison principle from Proposition \[prop:comp\], we thus get for $x \in [x_0,\rho]$, $$w^\rho (t,x) -w^\rho (t,x_0) \ge p_R^\delta (x-x_0) - C_\delta$$ where $C_\delta$ is a large constant which does not depend on $\rho$. In particular, we get , reducing $\delta$ if necessary.
Construction of global correctors
---------------------------------
We now state and prove a result which implies Theorem \[thm:corrector-simple\] stated in the introduction.
\[thm:corrector\] Assume **(A0)-(A5)** and either **(B-i)** or **(B-ii)**.
1. [(General properties)]{} There exists a solution $w$ of with $\lambda = \bar A$ such that for all $(t,x) \in
\R^2$, $$\label{eq:w-x-osc}
\left|w(t,x)-m(x)\right|\le C$$ for some globally Lipschitz continuous function $m$, and $$\bar A\ge A_0.$$
2. [(Bound from below at infinity)]{} If $\bar A > \max_{\alpha=L,R}\left(\min \bar H_\alpha\right)$, then there exists $\delta_0>0$ such that for every $\delta\in
(0,\delta_0)$, there exists $\rho_\delta>\rho_0$ such that $w$ satisfies $$\label{eq:w-lb-slope}
\begin{cases}
w(t,x+h)-w(t,x)\ge (\bar p_R -\delta) h -C_\delta & \quad \text{ for } x\ge
\rho_\delta \quad \text{and}\quad h\ge 0 , \\
w(t,x-h)-w(t,x)\ge (-\bar p_L -\delta) h -C_\delta & \quad \text{ for } x\le
-\rho_\delta \quad \text{and}\quad h\ge 0.
\end{cases}$$ The first line of (\[eq:w-lb-slope\]) also holds if we have only $\bar A> \min
\bar H_R$, while the second line of (\[eq:w-lb-slope\]) also holds if we have only $\bar A> \min \bar H_L$.
3. [(Rescaling $w$)]{} For $\eps>0$, we set $$w^\eps(t,x)=\eps w(\eps^{-1}t,\eps^{-1}x).$$ Then (along a subsequence $\eps_n\to 0$), we have that $w^\eps$ converges locally uniformly towards a function $W=W(x)$ which satisfies $$\label{eq:W-estim}
\left\{\begin{array}{ll}
\left|W(x)-W(y)\right|\le C \left|x-y\right| &\quad \text{for all}\quad x,y\in \R,\\
\bar{H}_R(W_x)=\bar A \quad \text{and}\quad \hat{p}_R \ge W_x\ge \bar p_R &\quad \text{ for } x\in (0,+\infty),\\
\bar{H}_L(W_x)=\bar A \quad \text{and}\quad \hat{p}_L \le W_x\le \bar p_L &\quad \text{ for } x\in (-\infty,0).
\end{array}\right.$$ In particular, we have $W(0)=0$ and $$\label{eq:W-estim-2-bis}
\hat p_R x 1_{\left\{x>0\right\}} + \hat p_L x
1_{\left\{x<0\right\}}\ge W(x)\ge \bar p_R x
1_{\left\{x>0\right\}} + \bar p_L x 1_{\left\{x<0\right\}}.$$
We consider (up to some subsequence) $$\overline w = \limsup_{\rho\to +\infty}{}^* (w^\rho - w^\rho (0,0)),\quad \underline w = \liminf_{\rho\to +\infty}{}_* (w^\rho -
w^\rho (0,0))\quad \text{ and } \quad m=\lim_{\rho\to +\infty} (m^\rho-m^\rho(0)).$$ We derive from that $\underline w$ and $\overline
w$ are finite and $$m- C \le \underline w \le \overline w \le m+C.$$ Moreover, discontinuous stability of viscosity solutions imply that $\overline w-2C$ and $\underline w$ are respectively a sub-solution and a super-solution of with $\lambda = \bar A$ (recall Proposition \[prop:bar-A\]). Hence, a discontinuous viscosity solution $w$ of can be constructed by Perron’s method (in the class of functions that are $1$-periodic with respect to time).
Using again , $w$ and $m$ satisfy . We also get from Proposition \[prop:slopes-trunc\] (use and pass to the limit with $m$ instead of $w$ if necessary).
We now study $w^\eps(t,x)=\varepsilon
w(\varepsilon^{-1}t,\varepsilon^{-1}x)$. Remark that implies in particular that $$w^\eps (t,x) = \eps m(\eps^{-1} x) + O (\eps) .$$ In particular, we can find a sequence $\eps_n \to 0$ such that $$w^{\eps_n} (t,x) \to W(x) \quad \text{ locally uniformly as } n \to
+ \infty,$$ with $W(0)=0$. Arguing as in the proof of convergence away from the junction point (see the case $\bar x \neq 0$ in Appendix), we deduce that $W$ satisfies $$\begin{aligned}
\bar H_R (W_x ) = \bar A & \text{ for } x >0, \\
\bar H_L (W_x ) = \bar A & \text{ for } x<0.\end{aligned}$$ We also deduce from that for all $\delta >0$ and $x>0$, $$W_x \ge \bar p_R - \delta$$ in the case where $\bar A > \min \bar H_R$. Assume now that $\bar A =
\min \bar H_R$. This implies that $$\bar p_R \le W_x \le \hat p_R$$ and, in all cases, we thus get (\[eq:W-estim-2-bis\]) for $x>0$.
Similarly, we can prove for $x<0$ that $$\hat p_L\le W_x \le \bar p_L$$ and the proof of of is achieved. This implies (\[eq:W-estim-2-bis\]). The proof of Theorem \[thm:corrector\] is now complete.
Proof of Theorem \[thm:bar-A\]
------------------------------
Let $\bar A$ denote the limit of $A_\rho$ (see Proposition \[prop:bar-A\]). We want to prove that $\bar A = \inf E$ where we recall that $$E= \{ \lambda \in \R: \exists \text{$w$ sub-solution of
\eqref{eq:cell}} \}.$$ In view of , sub-solutions are assumed to be periodic with respect to time; we will see that they also automatically satisfy some growth conditions at infinity, see below.
We argue by contradiction by assuming that there exist $\lambda <
\bar A$ and a sub-solution $w_\lambda$ of . The function $$m_\lambda (x) = \sup_{t \in \R} (w_\lambda)^* (t,x)$$ satisfies $$H (t(x),x,(m_\lambda)_x ) \le C$$ (for some function $t(x)$). Assumption **(A3)** implies that $m_\lambda$ is globally Lipschitz continuous. Moreover, since $w_\lambda$ is $1$-periodic w.r.t. time and $(w_\lambda)_t
\le C$, then $$|w_\lambda(t,x) -m_\lambda(x) | \le C.$$ Hence $$w_\lambda^\eps (t,x) = \eps w_\lambda (\eps^{-1}t,\eps^{-1}x)$$ has a limit $W^\lambda$ which satisfies $$\bar H_R (W^\lambda_x) \le \lambda \quad \text{ for } x >0 .$$ In particular, for $x>0$, $$W^\lambda_x \le \hat p^\lambda_R := \max \{ p \in \R: \bar H_R (p) =
\lambda \} < \bar p_R$$ where $\bar p_R$ is defined in . Similarly, $$W^\lambda_x \ge \hat p^\lambda_L := \min \{ p \in \R: \bar H_L(p)=
\lambda \} > \bar p_L$$ with $\bar p_L$ defined in . Those two inequalities imply in particular that for all $\delta>0$, there exists $\tilde{C}_\delta$ such that $$\label{eq:growth}
w_\lambda (t,x) \le \begin{cases}
(\hat p^\lambda_R + \delta) x + \tilde{C}_\delta & \text{ for } x >0, \\
(\hat p^\lambda_L + \delta) x + \tilde{C}_\delta & \text{ for } x <0.
\end{cases}$$ In particular, $$w_\lambda < w \text{ for } |x| \ge R$$ if $\delta$ is small enough and $R$ is large enough. In particular, $$w_\lambda < w + C_R \text{ for } x \in \R.$$ Remark finally that $u (t,x) = w(t,x) + C_R - \bar A t$ is a solution and $u_\lambda (t,x)
= w_\lambda (t,x) - \lambda t$ is a sub-solution of with $\eps = 1$ and $u_\lambda (0,x) \le u (0,x)$. Hence the comparison principle implies that $$w_\lambda(t,x) - \lambda t \le w(t,x) - \bar A t + C_R.$$ Dividing by $t$ and letting $t$ go to $+\infty$, we get the following contradiction $$\bar A \le \lambda.$$ The proof is now complete.
Proof of Theorem \[thm:conv-time\] {#sec:extension}
==================================
This section is devoted to the proof of Theorem \[thm:conv-time\]. As pointed out in Remark \[rem:meaning\] above, the notion of solutions for has to be first made precise because the Hamiltonian is discontinuous with respect to time.
#### Notion of solutions for .
For $\varepsilon=1$, a function $u$ is a [ *solution*]{} of if it is globally Lipschitz continuous (in space and time) and if it solves successively the Cauchy problems on time intervals $[\tau_i+k,\tau_{i+1}+k)$ for $i=0,\dots, K$ and $k\in \N$.
Because of this definition (approach), we have to show that if the initial datum $u_0$ is globally Lipschitz continuous, then the solution to the successive Cauchy problems is also globally Lipschitz continuous (which of course insures its uniqueness from the classical comparison principle). See Lemma \[lem:glob-lip\] below.
In view of the proof of Theorem \[thm:conv\], the reader can check that it is enough to get a global Lipschitz bound on the solution $u^\eps$ and to construct a global corrector in this new framework. The proof of these two facts is postponed, see Lemmas \[lem:glob-lip\] and \[lem:cor-ok\] following this proof. Notice that half-relaxed limits are not necessary anymore and that the reasoning can be completed by considering locally converging subsequences of $\{u^\eps\}_\eps$. Notice also that the perturbed test function method of Evans [@evans] still works. As usual, if the viscosity sub-solution inequality is not satisfied at the limit, this implies that the perturbed test function is a super-solution except at times $\varepsilon\left(\Z +
\left\{\tau_0,\dots,\tau_K\right\}\right)$. Still a localized comparison principle in each slice of times for each Cauchy problem is sufficient to conclude.
\[lem:glob-lip\] The function $u^\eps$ is equi-Lipschitz continuous with respect to time and space.
Remark that it is enough to get the result for $\eps =1$ since $u(t,x) = \eps^{-1} u^\eps (\eps t,\eps x)$ satisfies the equation with $\eps =1$ and the initial condition $$u^\eps_0 (x) = \frac1\eps U^\eps_0 (\eps x)$$ is equi-Lipschitz continuous. For the sake of clarity, we drop the $\eps$ superscript in $u^\eps_0$ and simply write $u_0$.
We first derive bounds on the time interval $[\tau_0,\tau_1)=[0,\tau_1)$. In order to do so, we assume that the initial data satisfies $|(u_0)_x|\le L$. Then as usual, there is a constant $C>0$ such that $$u^\pm(t,x)=u_0(x)\pm C t$$ are super-/sub-solutions of - with $H$ given by **(C1)** with for instance $$\label{eq::g30}
C:=\max\left( \max_{\alpha=1,\dots,N}\|a_\alpha\|_\infty, \; \max_{\alpha=0,\dots,N}\left(\max_{|p|\le L} |\bar H_\alpha(p)|\right)\right).$$ Let $u$ be the standard (continuous) viscosity solution of (\[eq:hj-eps\]) on the time interval $(0,\tau_1)$ with initial data given by $u_0$ (recall that $\eps =1$). Then for any $h>0$ small enough, we have $-Ch\le u(h,x)-u(0,x)\le Ch$. The comparison principle implies for $t\in (0,\tau_1-h)$ $$-Ch\le u(t+h,x)-u(t,x)\le Ch$$ which shows the Lipschitz bound in time, on the time interval $[0,\tau_1)$: $$\label{eq::g31}
|u_t|\le C.$$ From the Hamilton-Jacobi equation, we now deduce the following Lipschitz bound in space on the time interval $(0,\tau_1)$: $$\label{eq::g32}
|\bar H_\alpha(u_x(t,\cdot))|_{L^\infty(b_{\alpha},b_{\alpha+1})}\quad \le C\quad \mbox{for}\quad \alpha=0,\dots,N.$$
We can now derive bounds on the time interval $[\tau_1,\tau_2)$ as follows. We deduce first that (\[eq::g32\]) still holds true at time $t=\tau_1$. Combined with our definition (\[eq::g30\]) of the constant $C$, we also deduce that $$v^\pm(t,x)=u(\tau_1,x) \pm C(t-\tau_1)$$ are sub/super-solutions of (\[eq:hj-homog\]) for $t\in
(\tau_1,\tau_2)$ where $H$ is given by **(C1)**. Reasoning as above, we get bounds (\[eq::g31\]) and (\[eq::g32\]) on the time interval $[\tau_1,\tau_2)$.
Such a reasoning can be used iteratively to get the Lipschitz bounds (\[eq::g31\]) and (\[eq::g32\]) for $t\in
[0,+\infty)$. The proof of the lemma is now complete.
\[lem:cor-ok\] The conclusion of Theorem \[thm:corrector\] still holds true in this new framework.
The proof proceeds in several steps.
#### Step 1. Construction of a time periodic corrector $w^\rho$ on $[-\rho,\rho]$.
We first construct a Lipschitz corrector on a truncated domain. In order to do so, we proceed in several steps.
[Step 1.1. First Cauchy problem on $(0,+\infty)$.]{} The method presented in the proof of Proposition \[prop:cor-trunc\], using a term $\delta w^\delta$ has the inconvenience that it would not clearly provide a Lipschitz solution. In order to stick to our notion of globally Lipschitz solutions, we simply solve the Cauchy problem for $\rho>\rho_0 :=\max_{\alpha=1,\dots,N}|b_\alpha|$: $$\label{eq::g33}
\left\{\begin{array}{ll}
w^\rho_t + H(t,x,w^\rho_x)=0 &\quad \mbox{on}\quad (0,+\infty)\times
(-\rho,\rho)\, ,\\
w^\rho_t + \bar H^-_N (w^\rho_x)=0 &\quad \mbox{on}\quad (0,+\infty)\times \left\{-\rho\right\},\\
w^\rho_t + \bar H^+_0 (w^\rho_x)=0 &\quad \mbox{on}\quad (0,+\infty)\times \left\{\rho\right\},\\
w^\rho(0,x)=0 &\quad \mbox{for}\quad x\in [-\rho,\rho]\, .
\end{array}\right.$$ As in the proof of the previous lemma, we get global Lipschitz bounds with a constant $C$ (independent on $\rho>0$ and independent on the distances $\ell_\alpha = b_{\alpha+1}-b_\alpha$): $$\label{eq::g34}
|w^\rho_t|,\quad |\bar
H_\alpha(w^\rho_x(t,\cdot))|_{L^\infty((b_{\alpha},b_{\alpha+1})\cap
(-\rho,\rho))} \quad \le C,\quad \mbox{for}\quad \alpha=0,\dots,N.$$ Arguing as in [@im0] for instance, we deduce that there exists a real number $\lambda_\rho$ with $$|\lambda_\rho|\le C$$ and a constant $C_0$ (that depends on $\rho$) such that we have $$\label{eq::g35}
|w^\rho(t,x)+\lambda_\rho t|\le C_0.$$ Details are given in Appendix for the reader’s convenience.
[Step 1.2. Getting global sub and super-solutions.]{} Let us now define the following function (up to some subsequence $k_n\to +\infty$): $$w^\rho_\infty(t,x)=\lim_{k_n\to +\infty} \left(w^\rho(t+k_n,x)+\lambda_\rho k_n\right)$$ which still satisfies (\[eq::g34\]) and (\[eq::g35\]). Then we also define the two functions $$\overline{w}^\rho_\infty(t,x)=\inf_{k\in \Z} \left(w^\rho_\infty(t+k,x)+k\lambda_\rho\right),\quad
\underline{w}^\rho_\infty(t,x)=\sup_{k\in \Z}
\left(w^\rho_\infty(t+k,x)+k\lambda_\rho\right).$$ They still satisfy and and are respectively a super- and a sub-solution of the problem in $\R\times [-\rho,\rho]$. They satisfy moreover that $\overline{w}^\rho_\infty(t,x) +\lambda_\rho t$ and $\underline{w}^\rho_\infty(t,x) +\lambda_\rho t$ are $1$-periodic in time, which implies the following bounds $$|\overline{w}^\rho_\infty(t,x) -\overline{w}^\rho_\infty(0,x)
+\lambda_\rho t|\le C,\quad |\underline{w}^\rho_\infty(t,x) -\underline{w}^\rho_\infty(0,x) +\lambda_\rho t|\le C.$$
<span style="font-variant:small-caps;">Step 1.3: A new Cauchy problem on $(0,+\infty)$ and construction of a time periodic solution.</span> We note that $\overline{w}^\rho_\infty + 2C_0 \ge
\underline{w}^\rho_\infty$, and we now solve the Cauchy problem with new initial data $\underline{w}^\rho_\infty(0,x)$ instead of the zero initial data and call $\tilde{w}^\rho$ the solution of this new Cauchy problem. From the comparison principle, we get $$\underline{w}^\rho_\infty \le \tilde{w}^\rho \le \overline{w}^\rho_\infty + 2C_0.$$ In particular, $$\tilde{w}^\rho(1,x)\ge \underline{w}^\rho_\infty(1,x) \ge \tilde{w}^\rho(0,x)-\lambda_\rho.$$ This implies, by comparison, that $$\label{eq::g36}
\tilde{w}^\rho(k+1,x) \ge \tilde{w}^\rho(k,x) - \lambda_\rho.$$ Moreover $\tilde{w}^\rho$ still satisfies (\[eq::g34\]) (indeed with the same constant because, by construction, this is also the case for $\underline{w}^\rho_\infty$). We now define (up to some subsequence $k_n\to +\infty$): $$\tilde{w}^\rho_\infty(t,x)=\lim_{k_n\to +\infty} \left(\tilde{w}^\rho(t+k_n,x)+\lambda_\rho k_n\right)$$ which, because of (\[eq::g36\]) and the fact that $\tilde{w}^\rho(t,x)+\lambda_\rho t$ is bounded, satisfies $$\tilde{w}^\rho_\infty(k+1,x)+\lambda=\tilde{w}^\rho_\infty(k,x)$$ and then $\tilde{w}^\rho_\infty(t,x)+\lambda_\rho t$ is $1$-periodic in time. Moreover $\tilde{w}^\rho_\infty$ is still a solution of the Cauchy problem and satisfies (\[eq::g34\]). We define $$w^\rho:=\tilde{w}^\rho_\infty + \lambda_\rho t$$ which satisfies (\[eq:trunc-cor\]) and then provides the analogue of the function given in Proposition \[prop:cor-trunc\].
#### Step 2. Contruction of $w$ on $\R$.
The result of Theorem \[thm:corrector\] still holds true for $$w=\lim_{\rho\to +\infty} \left(w^\rho-w^\rho(0,0)\right)$$ which is globally Lipschitz continuous in space and time and satisfies (\[eq::g34\]) with $\rho=+\infty$, and $$\bar A =\lim_{\rho\to +\infty} \lambda_\rho. \qedhere$$
We recall that $\bar H_L=\bar H_0$ and $\bar H_R=\bar H_1$ and set $a=a_1$ and (up to translation) $b_1=0$.
#### Step 1: The convex case: identification of $\bar A$.
\
<span style="font-variant:small-caps;">Step 1.1: A convex subcase.</span> We first work in the particular case where both $\bar H_\alpha$ for $\alpha=L,R$ are convex and given by the Legendre-Fenchel transform of convex Lagrangians $L_\alpha$ which satisfy for some compact interval $I_\alpha$: $$\label{eq::g45}
L_\alpha(p)= \left\{\begin{array}{ll}
\mbox{finite} &\quad \mbox{if} \quad q\in I_\alpha,\\
+\infty &\quad \mbox{if} \quad q\not\in I_\alpha.
\end{array}\right.$$ Then it is known (see for instance the section on optimal control in [@im]) that the solution of on the time interval $[0,\varepsilon\tau_1)$, is given by $$\label{eq::g40}
u^\varepsilon(t,x)=\inf_{y\in \R} \left(\inf_{X\in S_{0,y;t,x}}
\left\{u^\varepsilon(0,X(0)) + \int_0^t
L_\varepsilon(s,X(s),\dot{X}(s))\, ds \right\} \right)$$ with $$L_\varepsilon(s,x,p)=\left\{\begin{array}{ll}
\bar H_L^*(p) &\quad \mbox{if}\quad x<0,\\
\bar H_R^*(p) &\quad \mbox{if}\quad x>0,\\
\min(-a(\varepsilon^{-1}s), \displaystyle \min_{\alpha=L,R} L_\alpha(0)) &\quad \mbox{if}\quad x=0,\\
\end{array}\right.$$ and for $s<t$, the following set of trajectories: $$S_{s,y;t,x} = \left\{X\in \mbox{Lip}((s,t);\R), \quad X(s)=y,\quad X(t)=x\right\}.$$ Combining this formula with the other one on the time interval $[\varepsilon\tau_1,\varepsilon \tau_2)$, and iterating on all necessary intervals, we get that (\[eq::g40\]) is a representation formula of the solution $u^\varepsilon$ of for all $t>0$. We also know (see the section on optimal control in [@im]), that the optimal trajectories from $(0,y)$ to $(t_0,x_0)$ intersect the axis $x=0$ at most on a time interval $[t_1^\varepsilon, t_2^\varepsilon]$ with $0\le t_1^\varepsilon\le t_2^\varepsilon\le t_0$. If this interval is not empty, then we have $t^\varepsilon_i\to t^0_i$ for $i=1,2$ and we can easily pass to the limit in (\[eq::g40\]). In general, $u^\varepsilon$ converges to $u^0$ given by the formula $$u^0(t,x)=\inf_{y\in \R} \left(\inf_{X\in S_{0,y;t,x}}
\left\{u^0(0,X(0)) + \int_x^t L_0(s,X(s),\dot{X}(s)) \, ds \right\}\right)$$ with $$L_0(s,x,p)=\left\{\begin{array}{ll}
\bar H_L^*(p) &\quad \mbox{if}\quad x<0,\\
\bar H_R^*(p) &\quad \mbox{if}\quad x>0,\\
\min(-\langle a\rangle, \displaystyle \min_{\alpha=L,R} L_\alpha(0)) &\quad \mbox{if}\quad x=0,\\
\end{array}\right.$$ and from [@im] we see that $u^0$ is the unique solution of - with $\bar A = \langle a\rangle$.
<span style="font-variant:small-caps;">Step 1.2: The general convex case.</span> The general case of convex Hamiltonians is recovered, because for Lipschitz continuous initial data $u_0$, we know that the solution is globally Lipschitz continuous. Therefore, we can always modify the Hamiltonians $\bar H_\alpha$ outside some compact intervals such that the modified Hamiltonians satisfy (\[eq::g45\]).
#### Step 2: General quasi-convex Hamiltonians: identification of $\bar A$.
\
<span style="font-variant:small-caps;">Step 2.1: Sub-Solution inequality.</span> From Theorem 2.10 in [@im], we know that $w(t,0)$, as a function of time only, satisfies in the viscosity sense $$w_t(t,0) + a(t)\le \bar A \quad \mbox{for all}\quad
t\notin\bigcup_{i=1,\dots,K+1}\tau_i + \Z.$$ Using the $1$-periodicity in time of $w$, we see that the integration in time on one period implies: $$\label{eq::g66}
\langle a\rangle \le \bar A.$$
<span style="font-variant:small-caps;">Step 2.2: Super-solution inequality.</span> Recall that $\bar A\ge \ \langle a\rangle\ \ge A_0:=\displaystyle
\max_{\alpha=L,R}\min (\bar H_\alpha)$. If $\bar A= A_0$, then obviously, we get $\bar A= \langle a\rangle$. Hence, it remains to treat the case $\bar A>A_0$.
<span style="font-variant:small-caps;">Step 2.3: Construction of a super-solution for $x\not=0$.</span> Recall that $\bar p_R$ and $\bar p_L$ are defined in and and the minimum of $\bar H_\alpha$ is reached for $\bar p^0_\alpha$, $\alpha = R,L$. Since $\bar A >
A_0$, there exists some $\delta>0$ such that $$\label{eq::g60}
\bar p_L +2\delta < \bar p^0_L\quad \quad \mbox{and}\quad \bar p^0_R<\bar p_R-2\delta.$$ If $w$ denotes a global corrector given by Lemma \[lem:cor-ok\] (or Theorem \[thm:corrector\]), let us define $$\underline{w}_R(t,x)=\inf_{h\ge 0} \left(w(t,x+h)-\bar p^0_R h\right)\quad \mbox{for}\quad x\ge 0,$$ and similarly $$\underline{w}_L(t,x)=\inf_{h\ge 0} \left(w(t,x-h)+\bar p^0_L h\right)\quad \mbox{for}\quad x\le 0.$$ From (\[eq:w-lb-slope\]) with $\rho_\delta=0$, we deduce that we have for some $\bar h\ge 0$ $$w(t,x)\ge \underline{w}_R(t,x)=w(t,x+\bar h)-\bar p^0_R \bar h\ge w(t,x)+(\bar p_R-\delta-\bar p^0_R)\bar h - C_\delta.$$ From (\[eq::g60\]), this implies $$\label{eq::re1}
0\le \bar h\le C_\delta/\delta$$ and using the fact that $w$ is globally Lipschitz continuous, we deduce that for $\alpha=R$: $$\label{eq::g61}
w\ge \underline{w}_\alpha\ge w -C_1.$$ Moreover, by constrution (as an infimum of (globally Lipschitz continuous) super-solutions), $\underline{w}_R$ is a (globally Lipschitz continuous) super-solution of the problem in $\R\times (0,+\infty)$. We also have for $x=y+z$ with $z\ge 0$: $$\begin{aligned}
\underline{w}_R(t,x)-\underline{w}_R(t,y) &=w(t,x+\bar h)-\bar p^0_R \bar h - \underline{w}_R(t,y)\\
&\ge w(t,x+\bar h)-\bar p^0_R \bar h -\left(w(t,y+\bar h +z)-\bar p^0_R (\bar h + z)\right)\\
&\ge \bar p^0_R z = \bar p^0_R (x-y)\end{aligned}$$ which shows that $$\label{eq::g62}
(\underline{w}_R)_x\ge \bar p^0_R.$$ Similarly (and we can also use a symmetry argument to see it), we get that $\underline{w}_L$ is a (globally Lipschitz continuous) super-solution in $\R\times (-\infty,0)$, it satisfies (\[eq::g61\]) with $\alpha =L$ and $$\label{eq::g63}
(\underline{w}_L)_x\le \bar p^0_L.$$ We now define $$\label{eq::g70}
\underline{w}(t,x)=\left\{\begin{array}{ll}
\underline{w}_R(t,x) &\quad \mbox{if}\quad x>0,\\
\underline{w}_L(t,x) &\quad \mbox{if}\quad x<0,\\
\min(\underline{w}_L(t,0), \underline{w}_R(t,0))&\quad \mbox{if}\quad x=0
\end{array}\right.$$ which by constrution is lower semi-continuous and satisfies , and is a super-solution for $x\not=0$.
<span style="font-variant:small-caps;">Step 2.4: Checking the super-solution property at $x=0$.</span> Let $\varphi$ be a test function touching $\underline{w}$ from below at $(t_0,0)$ with $t_0\notin\bigcup_{i=1,\dots,K+1}
\tau_i +\Z$. We want to check that $$\label{eq::g64}
\varphi_t(t_0,0) + F_{a(t_0)}(\varphi_x(t_0,0^-),\varphi_x(t_0,0^+))\ge \bar A.$$ We may assume that $$\underline{w}(t_0,0)=\underline{w}_R(t_0,0)$$ since the case $\underline{w}(t_0,0)=\underline{w}_L(t_0,0)$ is completely similar. Let $\bar h \ge 0$ be such that $$\underline{w}_R(t_0,0) =w(t_0,0+\bar h)-\bar
p^0_R \bar h.$$
We distinguish two cases. Assume first that $\bar h>0$. Then we have for all $h\ge 0$ $$\varphi(t,0)\le w(t,0+h)-\bar p^0_R h$$ with equality for $(t,h)=(t_0,\bar h)$. This implies the viscosity inequality $$\varphi_t(t_0,0) + \bar H_R(\bar p^0_R) \geq \bar A$$ which implies (\[eq::g64\]), because $F_{a(t_0)}(\varphi_x(t_0,0^-),\varphi_x(t_0,0^+)) \geq a(t_0)\geq
A_0\geq \min \bar H_R = \bar H_R(\bar p^0_R)$.
Assume now that $\bar h=0$. Then we have $\varphi\le \underline{w}\le w$ with equality at $(t_0,0)$. This implies immediately (\[eq::g64\]).
<span style="font-variant:small-caps;">Step 2.5: Conclusion.</span> We deduce that $\underline{w}$ is a super-solution on $\R\times \R$. Now let us consider a $C^1$ function $\psi(t)$ such that $$\psi (t)\le \underline{w}(t,0)$$ with equality at $t=t_0$. Because of (\[eq::g62\]) and (\[eq::g63\]), we see that $$\varphi(t,x)=\psi(t) + \bar p^0_L x 1_{\left\{x<0\right\}} + \bar p^0_R x 1_{\left\{x>0\right\}}$$ satisfies $$\varphi\le \underline{w}$$ with equality at $(t_0,0)$. This implies (\[eq::g64\]), and at almost every point $t_0$ where the Lipschitz continuous function $\underline{w}(t,0)$ is differentiable, we have $$\underline{w}_t(t_0,0) + a(t_0) \ge \bar A.$$ Because $w$ is $1$-periodic in time, we get after an integration on one period, $$\label{eq::re2}
\langle a \rangle\ge \bar A.$$ Together with (\[eq::g66\]), we deduce that $\langle a \rangle =\bar
A$, which is the desired result, for $N=1$.
. We simply remark, using the sub-solution viscosity inequality at each junction condition, that for $\alpha=1,\dots,N$, $$\bar A \ge \langle a_\alpha \rangle$$ which is the desired result. This achieves the proof of and .
Let $N\ge 2$, and for $i=c,d$, let us assume some given $b_1^i<\dots < b_N^i$. and let us call $w^i$ a global corrector given by Lemma \[lem:cor-ok\] (or Theorem \[thm:corrector\]) with $\lambda=\bar A^i$ and $H=H^i$ with $i=c,d$ respectively.
We call $\ell_\alpha^i = b_{\alpha+1}^i -b_\alpha^i>0$ and assume that $$0<\ell_{\alpha_0}^d - \ell_{\alpha_0}^c=:\delta_{\alpha_0} \quad \mbox{for some}\quad \alpha_0 \in \left\{1,\dots,N-1\right\}$$ and $$\ell_{\alpha}^d = \ell_{\alpha}^c \quad \mbox{for all}\quad \alpha \in \left\{1,\dots,N-1\right\}\backslash \left\{\alpha_0\right\}.$$ Calling $\bar p^0_{\alpha_0}$ a point of global minimum of $\bar H_{\alpha_0}$, we define $$\tilde{w}^d(t,x)=\left\{\begin{array}{ll}
w^c(t,x-b_{\alpha_0}^d+b_{\alpha_0}^c) & \quad \mbox{if}\quad x\le b_{\alpha_0}^d + \ell_{\alpha_0}^c/2=:x_-,\\
w^c(t,x_--b_{\alpha_0}^d+b_{\alpha_0}^c) +\bar p^0_{\alpha_0}(x-x_-)& \quad \mbox{if}\quad x_- \le x \le x_+,\\
w^c(t,x-b_{\alpha_0+1}^d+b_{\alpha_0+1}^c) +\bar p^0_{\alpha_0}(x_+-x_-)& \quad \mbox{if}\quad x\ge b_{\alpha_0+1}^d-\ell_{\alpha_0}^c/2=:x_+.
\end{array}\right.$$ Recall that $w^i$ for $i=c,d$, are globally Lipschitz continuous in space and time. This shows that $\tilde{w}^d$ is also Lipschitz continuous in space and time by construction, because it is continuous at $x=x_-,x_+$. Moreover $\tilde{w}^d$ is $1$-periodic in time. We now want to check that $\tilde{w}^d$ is a sub-solution of the equation satisfied by $w^d$ with $\bar A^c$ on the right hand side instead of $\bar A^d$. We only have to check it for all times $\bar t\not\in
\left\{\tau_0,\dots, \tau_K\right\}$ and $\bar x\in
\left[x_-,x_+\right]$, *i.e.* we have to show that $$\label{eq::jr0}
\tilde{w}^d_t(\bar t,\bar x) + \bar H_{\alpha_0}(\tilde{w}^d_x(\bar t,\bar x))\le \bar A^c \quad
\mbox{for all}\quad \bar x\in [x_-,x_+].$$ Assume that $\varphi$ is a test function touching $\tilde{w}^d$ from above at such a point $(\bar t,\bar x)$ with $\bar x\in [x_-,x_+]$. Then this implies in particular that $\psi(t,x)=\varphi(t,x)-\bar
p_{\alpha_0}^0(x-x_-)$ touches $\tilde{w}^d(\cdot,x_-)=w^c(\cdot,x_0)$ from above at time $\bar t$ with $x_0=b_{\alpha_0}^c +
\ell_{\alpha_0}^c/2$. Recall that $w^c$ is solution of $$w^c_t + \bar H_{\alpha_0}(w^c_x)=\bar A^c \quad \mbox{on}\quad (b^c_{\alpha_0},b^c_{\alpha_0+1}).$$ From the characterization of sub-solutions (see Theorem 2.10 in [@im]), we then deduce that $$\psi_t(\bar t)+\bar H_{\alpha_0}(\bar p^0_{\alpha_0}) \le \bar A^c.$$ If $\bar x \in (x_-,x_+)$, then we have $\varphi_x(\bar t,\bar x)=\bar p^0_{\alpha^0}$. This means in particular $$\label{eq::jr1}
\varphi_t + \bar H_{\alpha_0}(\varphi_x) \le \bar A^c \quad \mbox{at}\quad (\bar t,\bar x) \quad \mbox{if}\quad \bar x \in (x_-,x_+).$$ Using now (\[eq::jr1\]), and still from Theorem 2.10 in [@im], we deduce that we have in the viscosity sense $$\label{eq::jr2}
\tilde{w}^d_t(\bar t,\bar x) + \max\left(\bar H^-_{\alpha_0}(\tilde{w}^d_x(\bar t,\bar x^+)),\bar H^+_{\alpha_0}(\tilde{w}^d_x(\bar t,\bar x^-))\right)\le \bar A^c \quad
\mbox{for}\quad \bar x = x_\pm.$$ Therefore (\[eq::jr1\]) and (\[eq::jr2\]) imply (\[eq::jr0\]).
Let us now call $H^d$ the Hamiltonian in assumption **(C1)** constructed with the points $\{ b_\alpha^d \}_{\alpha=1, \dots, N}$. Then we have $$\tilde{w}^d_t + H^d(t,x,\tilde{w}^d_x)\le \bar A^c \quad \mbox{for all}\quad t\not\in \left\{\tau_0,\dots \tau_K\right\}.$$ Note that the proof of Theorem \[thm:bar-A\] is unchanged for the present problem, and then Theorem \[thm:bar-A\] still holds true. This shows that $$\label{eq::jr3}
\bar A^d \le \bar A^c$$ which shows the expected monotonicity. The proof is now complete.
Note that, in the previous proof, it would also be possible to compare the sub-solution given by the restriction of $\tilde{w}^d$ on some interval $[-\rho,\rho]$ with $\rho>0$ large enough (see [@im Proposition 2.19]), with the approximation $w^{d,\rho}$ of $w^d$ on $[-\rho,\rho]$ with $\bar A^d\ge \bar A^d_\rho \to \bar
A^d$ as $\rho\to +\infty$. The comparison for large times would imply $\bar A^d_\rho \le \bar A^c$. As $\rho\to +\infty$, this would give the same conclusion (\[eq::jr3\]).
Let $w$ be a global corrector associated to $\bar A$.
Recall that $$\label{eq::jr10}
\bar A \ge \bar A_0:= \max_{\alpha=1,\dots,N} \langle a_\alpha\rangle
\ge A_0:=\max_{\alpha=1,\dots,N} A_0^\alpha
\quad \mbox{with}\quad A^\alpha_0=\max_{\beta=\alpha-1,\alpha} (\min \bar H_\beta).$$ Our goal is to prove that $\bar A=\bar A_0$ when all the distances $\ell_\alpha$ are large enough, *i.e.* (\[eq:critical\]). Let us assume that $$\bar A > \bar A_0.$$
<span style="font-variant:small-caps;">Step 1: Considering another corrector with the same $\langle \widehat a_\alpha \rangle = \bar A_0$.</span> Let $\mu_\alpha\ge 0$ such that $$\widehat a_\alpha = \mu_\alpha + a_\alpha \quad \mbox{with}\quad \langle \widehat a_\alpha\rangle = \bar A_0
\quad \mbox{for all}\quad \alpha=1,\dots,N.$$ Let us call $\widehat w$ the corresponding corrector with associated constant $\widehat A$. Then Theorem \[thm:bar-A\] (still valid here) implies that $$\widehat A\ge \bar A > \bar A_0.$$ We also split the set $\left\{1,\dots,N\right\}$ into two disjoint sets $$I_0=\left\{\alpha\in \left\{1,\dots,N\right\},\quad \bar A_0 = A^\alpha_0\right\}$$ and $$I_1=\left\{\alpha\in \left\{1,\dots,N\right\},\quad \bar A_0 > A^\alpha_0\right\}.$$ Note that by (\[eq::jr10\]), if $\alpha\in I_0$, then $\langle a_\alpha\rangle = A^\alpha_0$, and then by [**(C3)**]{}, we have $a_\alpha(t)=const=A^\alpha_0$ for all time $t\in \R$. For later use, we then claim that $\widehat w$ satisfies $$\label{eq::jr11}
\widehat w_t(t,x)+ \max(\bar H_\alpha^-(\widehat w_x(t,x^+)), \bar
H_{\alpha-1}^+(\widehat w_x(t,x^-)))= \hat A \quad \mbox{for all}\quad (t,x)\in \R\times \left\{b_\alpha\right\}$$ and not only for $t\in \R\backslash \left(\Z +
\left\{\tau_0,\dots,\tau_K\right\}\right)$. Let us show it for sub-solutions (the proof being similar for super-solutions). Let $\varphi$ be a test function touching $\widehat w$ from above at some point $(\bar t,\bar x)=(j+\tau_k,b_\alpha)$ for some $j\in\Z$, $k\in
\left\{0,\dots,K\right\}$. Assume also that the contact between $\varphi$ and $\widehat w$ only holds at that point $(\bar t,\bar x)$. The proof is a variant of a standard argument. For $\eta>0$, let us consider the test function $$\varphi_\eta(t,x)=\varphi(t,x)+\frac{\eta}{\bar t-t} \quad \mbox{for}\quad t\in (-\infty,\bar t).$$ Then for $r>0$ fixed, we have $$\inf_{(t,x)\in \overline{B_r(\bar t,\bar x)},\ t<\bar t} (\varphi_\eta-\widehat w)(t,x)=(\varphi_\eta-\widehat w)(t_\eta,x_\eta)$$ with $$\left\{\begin{array}{l}
P_\eta=(t_\eta,x_\eta)\to (\bar t,\bar x)=\bar P \quad \text{ as } \quad
\eta\to 0, \medskip \\
\varphi_t(\bar P) \le \limsup_{\eta \to 0}(\varphi_\eta)_t(P_\eta).
\end{array}\right.$$ This implies that $\widehat w$ is a relaxed viscosity sub-solution at $(\bar t,\bar x)$ in the sense of Definition 2.2 in [@im]. By [@im Proposition 2.6], we deduce that $\widehat w$ is also a standard (*i.e.* not relaxed) viscosity sub-solution at $(\bar
t,\bar x)$. Finally we get (\[eq::jr11\]).
<span style="font-variant:small-caps;">Step 2: Defining a space super-solution.</span> Let us define the function $$M(x)=\inf_{t\in \R} \widehat w(t,x).$$ Because $\widehat w$ is globally Lispschitz continuous, we deduce that $M$ is also globally Lipschitz continuous. Moreover we have the following viscosity super-solution inequality $$\bar H_{\alpha}(M_x(x))\ge \widehat A > \bar A_0\quad \mbox{for all}\quad x\in (b_{\alpha},b_{\alpha+1}),
\quad \mbox{for all}\quad \alpha=0,\dots,N.$$ Let us call for $\alpha=0,\dots,N$: $$\bar p_{\alpha,R} = \min E_{\alpha,R} \quad \mbox{with}\quad E_{\alpha,R}=\left\{p\in\R,\quad \bar H^+_{\alpha}(p)=\bar H_{\alpha}(p)=\bar A_0\right\},$$ $$\bar p_{\alpha,L} = \max E_{\alpha,L} \quad \mbox{with}\quad E_{\alpha,L}=\left\{p\in\R,\quad \bar H^-_{\alpha}(p)=\bar H_{\alpha}(p)=\bar A_0\right\}.$$ Let us now consider $\alpha =0,\dots, N$ and two points $x_-<x_+$ with $x_\pm \in (b_\alpha,b_{\alpha+1})$. Let us assume that there is a test function $\varphi^\pm$ touching $M$ from below at $x_\pm$. Then we have $$\bar H_\alpha(\varphi^\pm_x(x_\pm))\ge \widehat A>\bar A_0$$ with $$\varphi^\pm_x(x_\pm)\ge \bar p_{\alpha,R} \quad \mbox{or}\quad \varphi^\pm_x(x_\pm)\le \bar p_{\alpha,L}.$$ Moreover, if $\bar A_0> \min \bar H_\alpha$, then we have $$\bar p_{\alpha,L}<\bar p^0_\alpha < \bar p_{\alpha,R}$$ for any $\bar p^0_{\alpha}$ which is a point of global minimum of $\bar H_{\alpha}$.
<span style="font-variant:small-caps;">Step 3: A property of the space super-solution.</span> We now claim that the following case is impossible: $$p^-:=\varphi^-_x(x_-)< \varphi^+_x(x_+)=:p^+ \quad \mbox{and}\quad \inf_{[p^-,p^+]}\bar H_\alpha < \widehat A.$$ If it is the case, then let $\bar p\in (p^-,p^+)$ such that $\bar
H_\alpha(\bar p)<\widehat A$. Therefore the geometry of the graph of the function $M$ implies that $$\inf_{x\in [x_-,x_+]} (M(x)- x \bar p) = M(\bar x)- \bar x \bar
p\quad \mbox{for some}\quad \bar x\in (x_-,x_+)$$ and then we have the viscosity super-solution inequality at $\bar x$: $$\bar H_\alpha(\bar p)\ge \widehat A$$ which leads to a contradiction. Therefore (in all cases $\bar A_0
>\min \bar H_\alpha$ or $\bar A_0=\min \bar H_\alpha$), it is possible to check that there is a point $\bar x_\alpha\in
[b_{\alpha},b_{\alpha+1}]$ such that the Lipschitz continuous function $M$ satisfies in the viscosity sense $$\left\{\begin{array}{ll}
M_x\ge \bar p_{\alpha,R} &\quad \mbox{in}\quad (b_\alpha,\bar
x_\alpha),\\ - M_x \ge - \bar p_{\alpha,L} &\quad \mbox{in}\quad (\bar
x_\alpha,b_{\alpha+1}).
\end{array}\right.$$ Moreover from Theorem \[thm:corrector\] ii) (see Lemma \[lem:cor-ok\]), we deduce from $\widehat A>\max (\min
H_N, \min H_0)$ that $$\bar x_{N}=+\infty \quad \mbox{and}\quad \bar x_0=-\infty.$$ In particular, we deduce that there exists at least one $\alpha_0\in
\left\{1,\dots,N\right\}$ such that $$\label{eq::jr5} \bar x_{\alpha_0} - b_{\alpha_0}\ge
\ell_{\alpha_0}/2\quad \mbox{and}\quad b_{\alpha_0} - \bar
x_{\alpha_0-1}\ge \ell_{\alpha_0-1}/2.$$
<span style="font-variant:small-caps;">Step 4: The case $\alpha_0\in I_0$.</span> In this case, we see that there exists a time $\bar t$ such that the test function $$\varphi(t,x)=\left\{\begin{array}{ll}
\bar p_{\alpha_0,R}(x-b_{\alpha_0}) & \quad \mbox{for}\quad x\ge b_{\alpha_0},\\
\bar p_{\alpha_0-1,L}(x-b_{\alpha_0}) & \quad \mbox{for}\quad x\le b_{\alpha_0}
\end{array}\right.$$ is a test function touching (up to some additive constant) $\widehat w$ from below at $(\bar t,b_{\alpha_0})$. By (\[eq::jr11\]), this implies $$\bar A_0=\max(\bar H_{\alpha_0}(\bar p_{\alpha_0,R}),\bar H_{\alpha_0-1}(\bar p_{\alpha_0-1,L})) \ge \widehat A\ge \bar A.$$ Contradiction.
<span style="font-variant:small-caps;">Step 5: Consequences on $\widehat w$.</span> From the fact that $\widehat w$ is $1$-periodic in time and $C$-Lipschitz continuous in time (with a constant $C$ depending only on $\displaystyle \max_{\alpha=1,\dots,N}\|\widehat a_\alpha\|_\infty$ and the $\bar H_\alpha$’s, see (\[eq::g30\])), we deduce that we have $$\label{eq::jr6}
\left\{\begin{array}{ll}
\widehat w(t,x+h)-\widehat w(t,x)\ge \bar p_{\alpha,R} h -2C &\quad \mbox{for}\quad x,x+h \in (b_\alpha,\bar x_\alpha),\\
\widehat w(t,x-h)-\widehat w(t,x)\ge -\bar p_{\alpha,R} h -2C &\quad \mbox{for}\quad x,x+h \in (\bar x_\alpha,b_{\alpha+1}).
\end{array}\right.$$
<span style="font-variant:small-caps;">Step 6: The case $\alpha_0\in I_1$: definition of a space-time super-solution.</span> Proceeding similarly to Step 3 of the proof of , we define $$\underline{\widehat w}_{\alpha_0,R}(t,x)=\inf_{\frac{\ell_{\alpha_0}}{4}\ge h\ge 0} \left(\widehat w(t,x+h)-\bar p^0_{\alpha_0}h\right)\quad \mbox{for}\quad b_{\alpha_0}\le x \le b_{\alpha_0} +\frac{\ell_{\alpha_0}}{4}$$ and $$\underline{\widehat w}_{(\alpha_0-1),L}(t,x)=\inf_{\frac{\ell_{(\alpha_0-1)}}{4}\ge h\ge 0} \left(\widehat w(t,x-h)+\bar p^0_{\alpha_0-1}h\right)\quad \mbox{for}\quad b_{\alpha_0} - \frac{\ell_{\alpha_0-1}}{4} \le x\le b_{\alpha_0}.$$ From (\[eq::jr6\]), we deduce that we have for some $\bar h\in [0,\frac{\ell_{\alpha_0}}{4}]$ $$\widehat w(t,x)\ge \underline{\widehat w}_{\alpha_0,R}(t,x) = \widehat w(t,x+\bar h)-\bar p^0_{\alpha_0}\bar h
\ge \widehat w (t,x)+(\bar p_{\alpha_0,R}-\bar p^0_{\alpha_0}) \bar h -2C$$ which implies $$0\le \bar h\le \frac{2C}{\bar p_{\alpha_0,R}-\bar p^0_{\alpha_0}}.$$ As in Step 3 of the proof of , if $$\label{eq::jr7}
\frac{\ell_{\alpha_0}}{4} > \frac{2C}{\bar p_{\alpha_0,R}-\bar p^0_{\alpha_0}}$$ this implies that $\underline{\widehat w}_{\alpha_0,R}$ is a super-solution for $x\in (b_{\alpha_0},b_{\alpha_0}+\frac{\ell_{\alpha_0}}{4})$. Similarly, if $$\label{eq::jr8}
\frac{\ell_{\alpha_0-1}}{4} > \frac{2C}{\bar p^0_{\alpha_0-1}-\bar p_{\alpha_0-1,L}}$$ then $\underline{\widehat w}_{\alpha_0-1,L}$ is a super-solution for $x\in (b_{\alpha_0}-\frac{\ell_{\alpha_0-1}}{4},b_{\alpha_0})$. We now define $$\underline{\widehat w}(t,x)=\left\{\begin{array}{ll}
\underline{\widehat w}_{\alpha_0,R}(t,x) & \quad \mbox{if}\quad x\in (b_{\alpha_0},b_{\alpha_0}+\frac{\ell_{\alpha_0}}{4}),\\
\underline{\widehat w}_{\alpha_0-1,L}(t,x) & \quad \mbox{if}\quad x\in (b_{\alpha_0}-\frac{\ell_{\alpha_0-1}}{4},b_{\alpha_0}),\\
\min (\underline{\widehat w}_{\alpha_0-1,L}(t,b_{\alpha_0}), \ \underline{\widehat w}_{\alpha_0,R}(t,b_{\alpha_0})) & \quad \mbox{if}\quad x= b_{\alpha_0}.
\end{array}\right.$$ Then as in Steps 4 and 5 of the proof , we deduce that $\underline{\widehat w}$ is a super-solution up to the junction point $x=b_{\alpha_0}$ and that $$\bar A_0= \langle \widehat a_{\alpha_0}\rangle \ge \widehat A \ge \bar A.$$ Contradiction.
<span style="font-variant:small-caps;">Step 7: Conclusion.</span> If (\[eq::jr7\]) and (\[eq::jr8\]) hold true for any $\alpha_0\in I_1$, then we deduce that $\bar A \le \bar A_0$, which implies $\bar A = \bar A_0$. This ends the proof of in Theorem \[thm:conv-time\].
Let us consider $$\bar a(t)=\max_{\alpha=1,\dots,N} a_\alpha(t),$$ and $(w,\bar{\bar{A}})$ a solution (given by Theorem \[thm:corrector\] (see also Lemma \[lem:cor-ok\])) of $$\left\{\begin{array}{ll}
w_t + \bar H_0(w_x) = \bar{\bar{A}}& \quad \mbox{if}\quad x<0,\\
w_t + \bar H_N(w_x) = \bar{\bar{A}}& \quad \mbox{if}\quad x>0,\\
w_t(t,0)+\max (\bar a(t),\bar H_N^-(w_x(t,0^+)), \bar
H_0^+(w_x(t,0^-))) = \bar{\bar{A}} & \quad \mbox{if}\quad x=0, \medskip\\
w \quad \mbox{is $1$-periodic with respect to $t$}.
\end{array}\right.$$ From Theorem \[thm:conv-time\], we also know that $$\bar{\bar{A}} = \langle \bar a \rangle.$$ For $N\ge 2$, we set $\ell=(\ell_1,\dots,\ell_{N-1})\in (0,+\infty)^{N-1}$ and consider $b_0=-\infty<b_1<\dots<b_N<b_{N+1}=+\infty$ with $$\ell_\alpha=b_{\alpha+1}-b_\alpha \quad \mbox{for}\quad \alpha=1,\dots N-1.$$ We now call $(w^\ell,\bar A^\ell)$ a global corrector given by Theorem \[thm:corrector\] (see also Lemma \[lem:cor-ok\]). The remaining of the proof is divided into several steps.
<span style="font-variant:small-caps;">Step 1: Bound from above on $\bar A^\ell$.</span> We define $$\tilde{w}(t,x)=\left\{\begin{array}{ll}
w(t,x-b_{1}) & \quad \mbox{if}\quad x\le b_{1},\\
\\
\displaystyle w(t,0)+ \bar p^0_\alpha(x-b_\alpha) +
\sum_{\beta=1,\dots,\alpha-1} \bar p^0_\beta (b_{\beta+1}-b_\beta) &
\quad \mbox{if}\quad \left\{\begin{array}{l}
b_{\alpha} \le x\le b_{\alpha+1},\\
\alpha\in \left\{1,\dots,N-1\right\},\\
\end{array}\right.\\
\\
\displaystyle w(t,x-b_N) + \sum_{\beta=1,\dots,N-1} \bar p^0_\beta (b_{\beta+1}-b_\beta) & \quad \mbox{if}\quad x\ge b_N.
\end{array}\right.$$ Proceeding as in Step 1 of the proof of Theorem \[thm:conv-time\] ii), it is then easy to check that $\tilde{w}$ is a sub-solution of the equation satisfied by $w^\ell$ with $\bar{\bar{A}}$ on the right hand side instead of $\bar A^\ell$. Then Theorem \[thm:bar-A\] implies that $$\label{eq::jr15}
\bar A^\ell \le \bar{\bar{A}} = \langle \bar a \rangle.$$
<span style="font-variant:small-caps;">Step 2: Bound from below on $\bar A^\ell$.</span> From Theorem 2.10 in [@im], we deduce that we have in the viscosity sense (in time only) $$w^\ell_t(t,b_\alpha)+ a_\alpha(t)\le \bar A^\ell \quad \mbox{for
all}\quad t \notin \cup_{k=0}^K \{ \tau_k + \Z\}.$$ Let us call $$\underline{A}=\liminf_{\ell \to 0} \bar A^\ell.$$ We also know that $w^\ell$ is $1$-periodic and globally Lipschitz continuous with a constant which is independent on $\ell$. Therefore there exists a $1$-periodic and Lipschitz continuous function $g=g(t)$ such that $$w^\ell(t,b_\alpha) \to g(t) \quad \mbox{for all}\quad \alpha=1,\dots,N,\quad \mbox{as}\quad \ell\to 0.$$ The stability of viscosity solutions implies in the viscosity sense $$g'(t) + a_\alpha(t) \le \underline{A},\quad \mbox{for all}\quad
\alpha=1,\dots,N,\quad \mbox{for all} \quad t \notin \cup_{k=0}^K \{ \tau_k + \Z\}.$$ Because $g$ is Lipschitz continuous, this inequality also holds for almost every $t\in\R$. This implies $$g'(t) + \bar a(t)\le \underline{A} \quad \mbox{for a.e.}\quad t\in\R.$$ An integration on one period gives $$\label{eq::jr16}
\langle \bar a \rangle \le \underline{A}.$$
<span style="font-variant:small-caps;">Step 3: Conclusion.</span> Combining (\[eq::jr15\]) with (\[eq::jr16\]) finally yields that $\bar A^\ell \to \langle \bar
a \rangle$ as $\ell\to 0$. The proof of in Theorem \[thm:conv-time\] is now complete.
#### Acknowledgements.
The authors thank the referees for their valuable comments. The authors thank Y. Achdou, K. Han and N. Tchou for stimulating discussions. The authors thank N. Seguin for interesting discussions on green waves. C. I. thanks Giga for the interesting discussions they had together and for drawing his attention towards papers such as [@hamamuki]. R. M. thanks G. Costeseque for his comments on traffic lights modeling and his simulations which inspired certain complementary results. The second and third authors are partially supported by ANR-12-BS01-0008-01 HJnet project.
Proofs of some technical results
================================
The case $\bar x \neq 0$ in the proof of convergence
----------------------------------------------------
We only deal with the subcase $\bar x >0$ since the subcase $\bar x<0$ is treated in the same way. Reducing $\overline r$ if necessary, we may assume that $B_{\overline r}(\overline t,\overline x)$ is compactly embedded in the set $\left\{(t,x)\in(0,+\infty)\times(0,+\infty):\,x>0\right\}$: there exists a positive constant $c_{\overline r}$ such that $$\label{conv1}
(t,x)\in B_{\overline r}(\overline t,\overline x)\quad\Rightarrow \quad x>c_{\overline r}\,.$$ Let $p=\varphi_x(\overline t,\overline x)$ and let $v^R=v^R(t,x)$ be a solution of the cell problem $$\label{conv3}
v_t^R+H_R\left(t,x,p+v^R_x\right)=\bar H_R(p)\quad\text{in}\quad\R\times\R\,.$$ We claim that if $\varepsilon>0$ is small enough, the perturbed test function [@evans] $$\varphi^\varepsilon(t,x)=\varphi(t,x)+\varepsilon v^R\left(\frac t\varepsilon,\frac x\varepsilon\right)$$ satisfies, in the viscosity sense, the inequality $$\label{conv4}
\varphi^\varepsilon_t+H\left(\frac t\varepsilon,\frac x\varepsilon,\varphi_x^\varepsilon\right)
\geq\frac\theta2\quad\text{in}\quad B_r(\overline t,\overline x)$$ for sufficiently small $r>0$. To see this let $\psi$ be a test function touching $\varphi^\varepsilon$ from below at $(t_1,x_1)\in
B_r(\overline t,\overline x)\subseteq B_{\overline r}(\overline
t,\overline x)$. In this way the function $$\eta(s,y)=\frac{1}{\varepsilon}\left(\psi(\varepsilon s,\varepsilon y)
-\varphi(\varepsilon s, \varepsilon y)\right)$$ touches $v^R$ from below at $(s_1,y_1)=\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon}\right)$ and (\[conv3\]) yields $$\label{conv5}
\psi_t(t_1,x_1)-\varphi_t(t_1,x_1)+H_R\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},p
+\psi_x(t_1,x_1)-\varphi_x(t_1,x_1)\right)\geq \bar H_R(p).$$ Since (\[conv1\]) implies that $\frac x\varepsilon\to+\infty$, as $\varepsilon\to0$, uniformly with respect to $(t,x)\in B_{\overline
r}(\overline t,\overline x)$, we can find, owing to **(A5)**, an $\varepsilon_0>0$ independent of $\psi$ and $(t_1,x_1)$ such that the inequality $$\label{conv6}
H\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},\psi_x(t_1,x_1)\right)\geq
H_R\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},\psi_x(t_1,x_1)\right)
-\frac\theta4$$ holds true for $\varepsilon<\varepsilon_0$. Combining (\[conv2\])-(\[conv5\])-(\[conv6\]) and using the continuity of $\varphi_x$ and $\varphi_t$ we have $$\begin{split}
&\psi_t(t_1,x_1)+H\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},\psi_x(t_1,x_1)\right)\\
&\geq\psi_t(t_1,x_1)+H_R\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},p+\psi_x(t_1,x_1)-\varphi_x(t_1,x_1)\right)+\\
&\quad
H_R\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},\psi_x(t_1,x_1)\right)-H_R\left(\frac{t_1}{\varepsilon},\frac{x_1}{\varepsilon},\varphi_x(\bar
t, \bar x)+\psi_x(t_1,x_1)-\varphi_x(t_1,x_1)\right)-\frac\theta4\\
&\geq\frac\theta2
\end{split}$$ if $r$ is sufficiently close to 0. The claim (\[conv4\]) is proved.
Since $\varphi$ is strictly above $\overline u$, if $\varepsilon$ and $r$ are small enough $$u^\varepsilon+\kappa_r\leq\varphi^\varepsilon\quad\text{on}\quad\partial B_r(\overline t, \overline x)$$ for a suitable positive constant $\kappa_r$. By comparison principle we deduce $$u^\varepsilon+\kappa_r\leq\varphi^\varepsilon\quad\text{in}\quad B_r(\overline t, \overline x)$$ and passing to the limit as $\eps \to 0$ and $(t,x) \to (\bar t,\bar
x)$ on both sides of the previous inequality, we produce the contradiction $$\overline u(\overline t,\overline x)<\overline u(\overline t,\overline x)
+\kappa_r\leq \varphi(\overline t,\overline x)=\overline u(\overline
t,\overline x)\,. \qedhere$$
Proof of Lemma \[lem:unique\]
-----------------------------
We first adress uniqueness. Let us assume that we have two solutions $(v^i,\lambda^i)$ for $i=1,2$ of (\[eq:cell-alpha\]). Let $$u^i(t,x)=v^i(t,x)+px - \lambda^i t$$ Then $u^i$ solves $$u^i_t + H_\alpha(t,x,u^i_x)=0$$ with $$u^1(0,x)\le u^2(0,x) +C$$ The comparison principle implies $$u^1\le u^2 +C \quad \mbox{for all}\quad t>0$$ and then $\lambda^1\ge \lambda^2$. Similarly we get the reverse inequality and then $\lambda^1=\lambda^2$.
We now turn to the continuity of the map $p\mapsto \bar
H_\alpha(p)$. It follows from the stability of viscosity sub- and super-solutions, from the fact that the constant $C$ in is bounded for bounded $p$’s and from the comparison principle. This achieves the proof of the lemma.
Sketch of the proof of Proposition \[prop:comp\] {#sec:comp}
------------------------------------------------
Consider $$M_\nu = \sup_{x \in [\rho_1,\rho_2], s,t \in \R} \left\{
u(t,x) - v(s,x) - \frac{(t-s)^2}{2\nu} \right\}.$$ We want to prove that $$M = \lim_{\nu \to 0} M_\nu \le 0.$$ We argue by contradiction by assuming that $M>0$. The supremum defining $M_\nu$ is reached, let $s_\nu,t_\nu,x_\nu$ denote a maximizer. Choose $\nu$ small enough so that $M_\nu \ge \frac{M}2 >0$. We classically get, $$|t_\nu - s_\nu| \le C \sqrt{\nu}.$$
If there exists $\nu_n \to 0$ such that $x_{\nu_n} = \rho_1$ for all $n \in \N$, then $$\frac{M}2 \le M_{\nu_n} \le
U_0(t_{\nu_n}) -U_0(s_{\nu_n}) \le \omega_0 (t_{\nu_n} -s_{\nu_n})
\le \omega_0 (C \sqrt{\nu_n})$$ where $\omega_0$ denotes the modulus of continuity of $U_0$. The contradiction $M\le 0$ is obtained by letting $n$ go to $+\infty$.
Hence, we can assume that for $\nu$ small enough, $x_\nu > \rho_1$. Reasoning as in [@im Theorem 7.8], we can easily reduce to the case where $H(t_\nu,x_\nu,\cdot)$ reaches its minimum for $p=p_0=0$. We can also consider the vertex test function $G^\gamma$ associated with the single Hamiltonian $H$ (using notation of [@im], it corresponds to the case $N=1$) and the free parameter $\gamma$. If $x_\nu<\rho_2$, then $G^\gamma(x,y)$ reduces to the standard test function $\frac{(x-y)^2}{2}$.
We next consider $$M_{\nu,\eps} = \sup_{\stackrel{x,y \in [\rho_1,\rho_2] \cap \overline{B_r(x_\nu)}}{s,t \in \R}} \left\{
u(t,x) - v(s,y) - \frac{(t-s)^2}{2\nu} - \eps G^\gamma (\eps^{-1}x,\eps^{-1}y)
- \varphi^\nu (t,s,x) \right\}$$ where $r=r_\nu$ is chosen so that $\rho_1 \notin \overline{B_r (x_\nu)}$ and the localization function $$\varphi^\nu (t,s,x) = \frac12((t-t_\nu)^2 + (s-s_\nu)^2 +
(x-x_\nu)^2).$$ The supremum defining $M_{\nu,\eps}$ is reached and if $(t,s,x,y)$ denotes a maximizer, then $$(t,s,x,y) \to (t_\nu,s_\nu,x_\nu,x_\nu) \quad \text{ as }
(\eps,\gamma) \to 0 .$$ In particular, $x,y \in B_r (x_\nu)$ for $\eps$ and $\gamma$ small enough. The remaining of the proof is completely analogous (in fact much simpler).
Construction of $\lambda_\rho$ in the proof of Lemma \[lem:cor-ok\]
-------------------------------------------------------------------
In order to get $\lambda_\rho$, it is enough to apply the following lemma.
Let $u$ be the solution of a Hamilton-Jacobi equation of evolution-type submitted to the initial condition: $u (0,x) = 0$ and posed on a compact set $K$. Assume that
- the comparison principle holds true;
- $u$ is $L$-globally Lipschitz continuous in time and space;
- $u(k+\cdot,\cdot)+C$ is a solution for all $k \in \N$ and $C \in
\R$.
There then exists $\lambda \in \R$ such that $$|u(t,x) - \lambda t | \le C_0$$ and $$|\lambda | \le L$$ where $C_0 = L (2+3\rho)$ if $\rho$ denotes the diameter of $K$.
Define $$\lambda^+ (T) = \sup_{\tau \ge 0 } \frac{u (\tau+T,0) -
u(\tau,0)}{T} \quad \text{ and } \quad
\lambda^- (T) = \inf_{\tau \ge 0 } \frac{u (\tau+T,0) -
u(\tau,0)}{T}.$$ Remark that $T \mapsto \pm T \lambda^\pm (T)$ is sub-additive. Remark that the fact that $u$ is $L$-Lipschitz continuous with respect to time implies that $\lambda^\pm(T)$ are both finite: $$|\lambda^\pm(T)| \le L.$$ the ergodic theorem implies that $\lambda^\pm (T)$ converges towards $\lambda^\pm$ and $$\lambda^+ = \inf_{T>0} \lambda^+ (T) \quad \text{ and } \quad
\lambda^- = \sup_{T>0} \lambda^-(T).$$ If moreover $$\label{estim:lambda-pm}
| \lambda^+ (T) - \lambda^-(T) | \le \frac{C}T,$$ then the proof of the lemma is complete. Indeed, implies in particular that $\lambda^+=\lambda^-$ and $$- \frac{C}T \le \lambda^- (T) - \lambda \le \lambda^+ (T) - \lambda
\le \frac{C}T.$$ This implies that $$|u(t,0) - \lambda t | \le C.$$ Finally, we get $$|u (t,x) - \lambda t | \le C + L \rho.$$
It remains to prove . There exists $k\in \Z$ and $\beta \in [0,1)$ such that $\tau^+ = k + \tau^- +
\beta$. Moreover, $$u(\tau^+,x) \le u (\tau^-+\beta,x) + u(\tau^+,0) -
u(\tau^-+\beta,0) + 2 L \rho$$ where $\rho = \mathrm{diam} \; K$. Now remark that $u(\tau^-+\beta+t,x)+D$ is a solution in $[\tau^+,+\infty)$ for all constant $D$. Hence, we get by comparison that for all $t>0$ and $x
\in K$, $$u (\tau^++t,x) \le u(\tau^-+\beta +t,x) + u(\tau^+,0) -
u(\tau^-+\beta,0) + 2 L \rho .$$ In particular, $$\begin{aligned}
u (\tau^+ +T,0) -u(\tau^+,0) & \le u(\tau^-+ \beta + T,0) -
u (\tau^-+\beta, 0) + 2 L \rho \\
& \le u(\tau^-+ T,0) - u (\tau^-, 0) + 2L (1+\rho).\end{aligned}$$ Finally, we get (after letting $\eps \to 0$), $$\lambda^+(T) \le \lambda^- (T) + \frac{2L(1+\rho)}{T}.$$ Similarly, we can get $$\lambda^+(T) \ge \lambda^- (T) - \frac{2L(1+\rho)}{T}.$$ This implies with $C = 2 L (1+\rho)$. The proof of the lemma is now complete.
[^1]: Department of Mathematics, University of Salerno, Via Giovanni Paolo II, 132, 84084 Fisciano (SA), Italy
[^2]: CNRS, UMR 7580, Université Paris-Est Créteil, 61 avenue du Général de Gaulle, 94 010 Créteil cedex, France
[^3]: Université Paris-Est, CERMICS (ENPC), 6-8 Avenue Blaise Pascal, Cité Descartes, Champs-sur-Marne, F-77455 Marne-la-Vallée Cedex 2, France
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We introduce a model for a growing random graph based on simultaneous reproduction of the vertices. The model can be thought of as a generalisation of the reproducing graphs of Southwell and Cannings and Bonato et al to allow for a random element, and there are three parameters, $\alpha$, $\beta$ and $\gamma$, which are the probabilities of edges appearing between different types of vertices. We show that as the probabilities associated with the model vary there are a number of phase transitions, in particular concerning the degree sequence. If $(1+\alpha)(1+\gamma)<1$ then the degree distribution converges to a stationary distribution, which in most cases has an approximately power law tail with an index which depends on $\alpha$ and $\gamma$. If $(1+\alpha)(1+\gamma)>1$ then the degree of a typical vertex grows to infinity, and the proportion of vertices having any fixed degree $d$ tends to zero. We also give some results on the number of edges and on the spectral gap.\
AMS 2000 Subject Classification: Primary 05C82; secondary 60G99, 60J10\
Key words and phrases: reproducing graphs, random graphs, degree distribution, phase transition
author:
- Jonathan Jordan
bibliography:
- 'js.bib'
title: Randomised reproducing graphs
---
\[thm\][Lemma]{} \[thm\][Corollary]{} \[thm\][Proposition]{}
Introduction
============
In this paper we introduce a new model for a growing random graph based on simultaneous reproduction of the vertices in the graph, with edges being formed between the new vertices and each other and between the new vertices and the existing ones according to a random mechanism conditioned on the pattern of edges between the vertices in the previously existing graph. The model is a generalisation of the models introduced by Southwell and Cannings [@rc1; @rc3; @gamenets] and the Iterated Local Transitivity (ILT) model of [@ILT], introducing stochasticity, which causes the regular structure found in the graphs of [@rc1; @rc3; @gamenets] to be lost, and which may make them more suitable for modelling in areas such as social networks; the authors of [@ILT] particularly suggest their model as a model for online social networks, mentioning Facebook and Twitter among other examples.
We will show that our model, which depends on three parameters $\alpha,\beta$ and $\gamma$, which are to be thought of as probabilities, exhibits a number of phase transitions as the parameters vary; for example for some values of the parameters we will show that the degree distribution of the graph converges to a limiting probability distribution, while for other choices of the parameters the degree of a randomly chosen (in an appropriate sense) vertex in $G_n$ can be shown to tend to infinity as $n\to\infty$. We will also show that for certain choices of the parameter values the model exhibits a power-law-like decay of the degree distribution, which is a property reported for many “real world” networks, and is also associated with other random graph models such as preferential attachment.
We start with a graph $G_0$, and form a new graph $G_{n+1}$ by adding a “child” vertex for every vertex of $G_n$. As in [@rc1; @rc3; @gamenets] we denote the vertices by binary strings, writing $v0$ for the “child” of vertex $v\in V(G_n)$ and $v1$ for the continuation of vertex $v$ as a vertex of $G_{n+1}$. The edges of $G_{n+1}$ are then obtained according to the following mechanism. For each $n$ define independent (of each other and of the random variables at other stages of the construction) Bernoulli random variables ${a^{(n)}_{\{u,v\}}}\sim Ber(\alpha)$ for each unordered pair $\{u,v\}$ of vertices of $G_n$, ${b^{(n)}_{u}}\sim Ber(\beta)$ for each vertex in $G_n$, ${c^{(n)}_{(u,v)}}\sim Ber(\gamma)$ for each ordered pair $(u,v)$ of vertices of $G_n$, and connect vertices as follows:
(a)
: $u1$ is connected to $v1$ in $G_{n+1}$ if and only if $u$ and $v$ are connected in $G_n$, that is existing edges are retained, and no further edges are formed between existing vertices.
(b)
: $u0$ is connected to $u1$ in $G_{n+1}$ if and only if ${b^{(n)}_{u}}=1$, so each child is connected to its parent with probability $\beta$.
(c)
: $u0$ is connected to $v1$ in $G_{n+1}$ if and only if ${c^{(n)}_{(u,v)}}=1$ and $u$ and $v$ are connected in $G_n$, so each child is connected to each of its parent’s neighbours with probability $\gamma$.
(d)
: $u0$ is connected to $v0$ in $G_{n+1}$ if and only of ${a^{(n)}_{\{u,v\}}}=1$ and $u$ and $v$ are connected in $G_n$, so each child is connected to each of its parent’s neighbours’ children with probability $\alpha$.
The models introduced in [@rc1; @rc3; @gamenets] have $\alpha,\beta,\gamma\in\{0,1\}$, so are deterministic. Additionally the case where $\alpha=0,\beta=1,\gamma=1$ is the ILT model, introduced in [@ILT] as a model for online social networks. The ILT$(p)$ model introduced in [@ILT] as a stochastic generalisation of the ILT model adds extra random edges between the child vertices without regard to whether their parents were connected, and thus cannot be seen as a special case of our model. In addition as defined in [@ILT] the ILT$(p)$ model always has at least the edges found in the basic ILT model, so is not suited to producing relatively sparse graphs.
The model differs from duplication graphs, for example those considered in [@biodupe; @jorddupe], in that in those models only one vertex, chosen at random, duplicates at any one time step, whereas in the models considered here and in [@ILT; @rc1; @rc3; @gamenets] all vertices simultaneously duplicate.
Our main results concern the degree distribution. We will deal with the cases where $\beta=0$ and $\beta>0$ separately, as the behaviour of the model when $\beta=0$ is potentially quite different, with large numbers of isolated vertices.
\[beta0\] Let $\beta=0$. Then, if $(1+\gamma)(\alpha+\gamma)\leq 1$, the probability that a randomly chosen vertex in the graph $G_n$ is isolated tends to $1$ as $n\to\infty$, and the proportion of vertices in the graph with degree zero tends to $1$, almost surely. If $(1+\gamma)(\alpha+\gamma)>1$, then the probability that a randomly chosen vertex in the graph $G_n$ is isolated converges to some value strictly less than $1$.
\[betanot0\] Assume $\beta>0$, and let ${p^{(n)}_{d}}$ be the proportion of vertices in $G_n$ with degree $d$. Then
(a)
: If $(1+\gamma)(\alpha+\gamma)< 1$ there exists a random variable $X$ such that ${p^{(n)}_{d}}\to P(X=d)$ as $n\to\infty$, almost surely.
(b)
: Under the conditions of (a), the random variable $X$ has a finite $p$th moment if $(1+\gamma)^p+(\alpha+\gamma)^p<2$, and does not have a finite $p$th moment if $(1+\gamma)^p+(\alpha+\gamma)^p>2$.
(c)
: If $(1+\gamma)(\alpha+\gamma)> 1$ then ${p^{(n)}_{d}}\to 0$ as $n\to\infty$, almost surely.
Note that Theorem \[betanot0\](b) implies that if $(1+\gamma)^p+(\alpha+\gamma)^p=2$ the tail of the degree distribution is asymptotically close to a power law degree distribution with index given by $-(p+1)$, in the sense that $q$th moments exist for $q<p$ but not for $q>p$.
In [@ILT], it is shown that the ILT model exhibits a “densification power law”, which is defined to mean that, if $E_n$ is the number of edges of $G_n$ and $V_n$ the number of vertices, then $E_n$ is proportional to $(V_n)^a$ for some $a\in(1,2)$. The following result shows that our model exhibits a phase transition in this respect, with the transition occurring where $2\gamma+\alpha=1$. Note that in our model, as in the ILT model, $V_n=2^nV_0$ for all $n$.
\[dense\]
(a)
: If $2\gamma+\alpha>1$ then $W_n=\frac{E_n}{(1+2\gamma+\alpha)^n}$ converges to a positive limit, so that the model has a densification power law as defined by [@ILT] with exponent $\frac{\log(1+2\gamma+\alpha)}{\log 2}$.
(b)
: If $2\gamma+\alpha<1$ then $$\frac{E_n}{2^n}\to \frac{V_0\beta}{1-2\gamma-\alpha},$$ almost surely, as $n\to\infty$, so that the number of edges grows at the same rate as the number of vertices
(c)
: If $2\gamma+\alpha=1$ then $$\frac{E_n}{2^n n}\to \frac{V_0\beta}{2},$$ almost surely, as $n\to\infty$.
Note that the combination of Theorems \[dense\] and \[betanot0\] implies that when $2\gamma+\alpha>1$ but $(1+\gamma)(\alpha+\gamma)<1$ the process exhibits both a densification power law in the sense of [@ILT] and an approximately power law limit for the degree distribution.
A further result in [@ILT] on the ILT model concerns the spectral gap. They show that the normalised graph Laplacian $\mathcal{L}$, as defined by Chung [@chung], of the ILT model has a large spectral radius, defined as $\max\{|\lambda_1-1|,|\lambda_{n-1}-1|\}$, where $\lambda_1$ is the second smallest eigenvalue (the smallest being $\lambda_0=0$ for any graph) and $\lambda_{n-1}$ is the largest eigenvalue and thus that the graph has relatively poor expansion properties. The following results show that the same is also true for our model. We concentrate on the case $\beta=1$, where the graphs are connected; otherwise $\lambda_1$ will be zero. The proofs use the Cheeger constant and its relationship to $\lambda_1$, as defined in Chapter 2 of Chung [@chung].
\[gap\]Let $\beta=1$ and assume that $G_0$ is connected, so that $G_n$ will also be connected for all $n$. Let $\lambda_1(G_n)$ be the smallest non-negative eigenvalue of the Laplacian of $G_n$. Then
(a)
: If $2\gamma+\alpha\leq1$ then $\lambda_1(G_n)\to 0$ as $n\to\infty$.
(b)
: If $2\gamma+\alpha> 1$ then there exists a (random) $\Lambda$, with $\Lambda<1$ almost surely, such that $\limsup_{n\to\infty} \lambda_1=\Lambda$.
Pictures
========
This section shows a few examples of graphs of this type, which were generated using a script written with the igraph package, [@igraph], in R. The first two pictures show examples with $n=7$, $\alpha=0$, $\beta=1$ and $\gamma=0.2$ and $0.49$ respectively.
-- --
-- --
The next two pictures have $\alpha=0, \gamma=0.366 (\approx
\frac{\sqrt{3}-1}{2})$, and again $n=7$. The first has $\beta=0.5$, the second $\beta=0.8$.
-- --
-- --
Proofs of Theorems
==================
We start with a lemma on the conditional expectation and variance of the number of edges in $G_n$, $E_n$. This lemma will be useful for obtaining the mean of the stationary distribution of a Markov chain which we will use to prove Theorems \[beta0\] and \[betanot0\]. We define ${\mathcal{F}}_n$ to be the $\sigma$-algebra generated by the graphs $G_m$ for $m\leq n$.
\[subgraphs\] The conditional expectation and variance of $E_{n+1}$ satisfy $$\begin{aligned}
{\mathbb E}(E_{n+1}|{\mathcal{F}}_n) &=& (1+2\gamma+\alpha)E_n + 2^n\beta V_0 \\ {\mathop{\rm Var}\nolimits}(E_{n+1}|{\mathcal{F}}_n) &=& E_n(2\gamma(1-\gamma)+\alpha(1-\alpha))+2^nV_0\beta(1-\beta)\end{aligned}$$
This follows from the fact that the $E_{n+1}$ can be written $E_n+E_{n+1,1}+E_{n+1,2}+E_{n+1,3}$ where $E_{n+1,1}$, $E_{n+1,2}$ and $E_{n+1,3}$ are independent, $E_{n+1,1}$ represents the edges between parents and children of their neighbours and, conditional on ${\mathcal{F}}_n$ has a $Bi(2E_n,\gamma)$ distribution, $E_{n+1,2}$ represents the edges between children of neighbouring vertices and, conditional on ${\mathcal{F}}_n$ has a $Bi(E_n,\alpha)$ distribution, and $E_{n+1,3}$ represents the edges between parents and their children and, conditional on ${\mathcal{F}}_n$ has a $Bi(V_n,\beta)$ distribution. As $V_n=2^n V_0$ the result follows.
Proof of Theorem \[dense\]
--------------------------
We start with (a). The following approach is based on that in [@athreya] for multitype branching processes, the idea being that the vertices and edges in the graph $G_n$ can be thought of as the two types in a population undergoing branching. However the resulting multitype branching process is not irreducible, so the results in [@athreya] cannot be used directly.
Given an edge in $G_m$ between vertices $u$ and $v$, there will be an edge in $G_{m+1}$ between $u1$ and $v_1$, and in addition there will be edges between $u1$ and $v0$ and $v1$ and $u0$ each with probability $\gamma$ and an edge between $u0$ and $v0$ with probability $\alpha$. We can consider these edges as offspring of the edge between $u$ and $v$, and thus consider the set of edges in $G_n$ (for $n>m$) which are descendants of the edge between $u$ and $v$ in $G_m$ as a generation in a Galton-Watson branching process with offspring mean $1+2\gamma+\alpha$, and where the extinction probability is zero and the number of offspring bounded. Treating the descendants of a given edge in $G_m$ as a subset of the edge set of $G_n$, this shows that $\liminf \frac{E_n}{(1+2\gamma+\alpha)^n}$ is a positive random variable.
Now define $$W_n=\frac{V_n+\frac{2\gamma+\alpha-1}{\beta}E_n}{(1+2\gamma+\alpha)^n}.$$ Then ${\mathbb E}(W_{n+1}|{\mathcal{F}}_n)$, so $(W_n)_{n\in {\mathbb N}}$ is a non-negative martingale, and thus almost surely has a non-negative limit $W$. The above conclusion on $\liminf \frac{E_n}{(1+2\gamma+\alpha)^n}$ shows that $P(W=0)=0$, giving the result.
For (b), Lemma \[subgraphs\] shows that ${\mathbb E}(E_n)=\frac{V_0\beta}{1-2\gamma-\alpha}2^n+o(2^n)$ and $${\mathop{\rm Var}\nolimits}(E_{n})=2^{n-1}\frac{V_0\beta}{1-2\gamma-\alpha}(2\gamma(1-\gamma)+o(2^n)+\alpha(1-\alpha))+2^nV_0\beta(1-\beta)+(1+2\gamma+\alpha)^2
{\mathop{\rm Var}\nolimits}(E_{n-1}),$$ which shows that $${\mathop{\rm Var}\nolimits}\left(\frac{E_n}{2^n}\right)=O\left(\left(\max\left({\frac{1}{2}},\frac{(1+2\gamma+\alpha)^2}{4}\right)\right)^n\right),$$ which implies the result via Chebyshev’s inequality and the Borel-Cantelli Lemmas.
For (c), an iterative use of Lemma \[subgraphs\] shows that ${\mathbb E}(E_n)=2^n\left(E_0+\frac{\beta V_0}{2}n\right)$ and ${\mathop{\rm Var}\nolimits}(E_{n+1})=2^{n-1}n\beta V_0 (2\gamma(1-\gamma)+\alpha(1-\alpha)+4)+O(2^n)$. Then $${\mathop{\rm Var}\nolimits}\left(\frac{E_n}{2^n n}\right)=O\left(\frac{1}{n^2 2^n}\right),$$ allowing the Chebyshev/Borel-Cantelli argument again.
Proof of Theorem \[gap\]
------------------------
We consider some small $m$, and find the Cheeger constant of $G_m$. By the definition in [@chung], this will be $e(S_m,\bar{S_m})/{\mathop{\rm vol}\nolimits}(S_m)$ for some $S_m\subseteq V(G_m)$, where for two subsets of the vertex set $S$ and $S'$ $e(S,S')$ is the number of edges between a vertex in $S$ and one in $S'$, and ${\mathop{\rm vol}\nolimits}(S)$ is the sum of the degrees of vertices in $S$. (Note that ${\mathop{\rm vol}\nolimits}(S)=2e(S,S)+e(S,\bar{S}$.) Now consider the descendants of $S_m$ in $G_n$ as a subset $S_n\subseteq V(G_n)$. Then the same arguments as in the proof of Theorem \[dense\], applied to the subgraphs descending from $S_m$ and $\bar{S_m}$, show that if $2\gamma+\alpha<1$ then the $e(S_n,S_n)$ and $e(\bar{S_n},\bar{S_n})$ both grow at rate $2^n$, in the sense that $\frac{e(S_n,S_n)}{2^n}$ and $\frac{e(\bar{S_n},\bar{S_n})}{2^n}$ converge almost surely to positive constants as $n\to\infty$, and similarly if $2\gamma+\alpha=1$ $e(S_n,S_n)$ and $e(\bar{S_n},\bar{S_n})$ both grow at rate $2^n n$, and if $2\gamma+\alpha>1$ $e(S_n,S_n)$ and $e(\bar{S_n},\bar{S_n})$ both grow at rate $(1+2\gamma+\alpha)^n$.
Next, again as in the proof of Theorem \[dense\], $(e(S_n,\bar{S_n}))_{n\in{\mathbb N}}$ forms a Galton-Watson branching process with mean of the offspring distribution $1+2\gamma+\alpha$, and extinction probability zero, so $e(S_n,\bar{S_n})$ will grow at rate $(1+2\gamma+\alpha)^n$. So for $2\gamma+\alpha> 1$ $\frac{e(S_n,\bar{S_n})}{\min({\mathop{\rm vol}\nolimits}(S_n),{\mathop{\rm vol}\nolimits}(\bar{S_n}))}$, which by the definition in [@chung] is greater than the Cheeger constant of $G_n$, converges to a constant (less than $1$, as ${\mathop{\rm vol}\nolimits}(S_n)=2e(S_n,S_n)+e(S_n,\bar{S_n})$) and this constant bounds the lim sup of the Cheeger constant of $G_n$ above.
In the case where $2\gamma+\alpha<1$ $\frac{e(S_n,\bar{S_n})}{\max({\mathop{\rm vol}\nolimits}(S_n),{\mathop{\rm vol}\nolimits}(\bar{S_n}))}=O\left(\left(\frac{1+2\gamma+\alpha}{2}\right)^n\right)\to 0$ as $n\to\infty$, and hence so is the Cheeger constant. Similarly if $2\gamma+\alpha=1$ $\frac{e(S_n,\bar{S_n})}{\max({\mathop{\rm vol}\nolimits}(S_n),{\mathop{\rm vol}\nolimits}(\bar{S_n}))}=O\left(\frac{1}{n}\right)$, and again so is the Cheeger constant.
Hence by the Cheeger inequality (Lemma 2.1 and Theorem 2.2 of [@chung]), $\limsup_{n\to\infty}\lambda_1(G_n)<1$ almost surely in the case $2\gamma+\alpha> 1$, and $\lambda_1(G_n)$ tends to zero in the case $2\gamma+\alpha\leq 1$.
Proofs of Theorems \[beta0\] and \[betanot0\]
---------------------------------------------
The proofs of Theorems \[beta0\] and \[betanot0\] will rely on defining a certain Markov chain whose value $X_n$ represents the degree of a random vertex in the graph $G_n$. We construct this by letting $v_0$ be a vertex of $G_0$ chosen uniformly at random, and then, using the binary string notation for the vertices described above, for $n\geq 1$ let $v_{n}=v_{n-1}1$ with probability $1/2$ and letting $v_n=v_{n-1}0$ with probability $1/2$. We then let $X_n$ be the degree of $v_n$ in $G_n$.
Then $$\label{xrec}X_{n+1}=\xi_{n+1}X_n+(1-\xi_{n+1})W_{n+1}+
Y_{n+1}+Z_{n+1},$$ where, conditional on $G_n$, $Y_{n+1}\sim Bin(X_n,\gamma)$, $W_{n+1}\sim Bin(X_n,\alpha)$, $Z_{n+1}\sim Bin(1,\beta)$ and $\xi_{n+1}\sim Bin(1,{\frac{1}{2}})$, with all these variables being conditionally independent given $G_n$.
Here, $\xi_{n+1}=1$ if our vertex in $G_{n+1}$ is a parent and $0$ if it is a child, $W_{n+1}$ represents child-child connections (so does not appear if $\xi_{n+1}=1$), $Y_{n+1}$ represents connections between a child and its parents’ neighbours, and $Z_{n+1}$ represents the connection between the child and its parent.
As defined above, $(X_n)_{n\in {\mathbb N}}$ is a discrete time Markov chain on the natural numbers (including zero if $\beta<1$). It is irreducible and aperiodic if $\beta>0$, $\alpha<1$ and $\gamma<1$. (If $\beta=0$ then zero is an absorbing state, and if either $\alpha$ or $\gamma$ is $1$ then $X_n$ is increasing in $n$ and so the chain is certainly not irreducible, but otherwise $P(X_{n+1}=1|{\mathcal{F}}_n)$ is always positive.)
If $2\gamma+\alpha<1$ the distribution of $X_n$ converges in the Wasserstein-$1$ metric to a unique fixed point with finite mean $\frac{2\beta}{1-2\gamma-\alpha}$.
Note that if we have another random variable $\hat{X}_n$ with a different distribution on ${\mathbb N}_0$, we can apply to it by defining, conditional on $\hat{X}_n$, $\hat{Y}_{n+1}\sim Bin(\hat{X}_n,\gamma)$ and $\hat{W}_{n+1}\sim Bin(\hat{X}_n,\alpha)$ using the same set of Bernoulli trials as for $Y_{n+1}$ and $W_{n+1}$ respectively, and letting $$\hat{X}_{n+1}=\xi_{n+1}\hat{X}_n+(1-\xi_{n+1})\hat{W}_{n+1}+
\hat{Y}_{n+1}+Z_{n+1},$$ in which case ${\mathbb E}(|X_{n+1}-\hat{X}_{n+1}||{\mathcal{F}}_n)={\frac{1}{2}}(2\gamma+\alpha+1)|X_{n}-\hat{X}_{n}|$, so that we have a contraction in the Wasserstein metric if $2\gamma+\alpha<1$. Hence in this case there is convergence in the Wasserstein-$1$ metric of the degree distributions to a unique fixed point with finite mean.
We can calculate the mean of this distribution by using Lemma \[subgraphs\]: letting $m=2$ we get $${\mathbb E}({k^{(n+1)}_{2}})=(2\gamma+\alpha+1){k^{(n)}_{2}}+2^n\beta v_0,$$ where $v_0$ is the number of vertices in the initial graph, and solving this we find that the expected number of edges in $G_n$ is $$\frac{\beta v_0(2^n-(1+2\gamma+\alpha)^n)}{1-2\gamma-\alpha},$$ so (as the number of vertices in $G_n$ is $2^n v_0$) the expected average degree is $$\frac{\beta(2^n-(1+2\gamma+\alpha)^n)}{2^{n-1}(1-2\gamma-\alpha)},$$ which converges to $\frac{2\beta}{1-2\gamma-\alpha}$ as $n\to\infty$.
To go further than this we use Foster-Lyapunov techniques, as described in Meyn and Tweedie [@meyntweedie] in the more general case of an uncountable state space. The following lemma on the conditional moments of $X_{n+1}$ (including negative and fractional moments) will be useful.
\[moments\]Let $p\in{\mathbb R}$. Then as $x\to\infty$, $${\mathbb E}\left(\left(\frac{1+x+Y_{n+1}+Z_{n+1}}{1+x}\right)^p|X_n=x\right)\to (1+\gamma)^p,$$ and $${\mathbb E}\left(\left(\frac{1+W_{n+1}+Y_{n+1}+Z_{n+1}}{1+x}\right)^p|X_n=x\right)\to (\alpha+\gamma)^p.$$
In the case where $p<0$ this is a special case of Theorem 2.1 of García and Palacios in [@negativemoments]. When $p>0$ we can adapt the argument of Theorem 2.1 in [@negativemoments] with the additional condition that the random variables are bounded above by a constant multiple of their mean, which is satisfied in this case. We assume that $(A_n)_{n\in{\mathbb N}}$ is a sequence of positive random variables with associated sequences of constants $\mu_n$ and $\sigma_n$ such that $A_n\leq k\mu_n$ with probability $1$ for some $k$ and for all $n$, and that the conditions of Theorem 2.1 in [@negativemoments] hold.
For the upper bound, $$\begin{aligned}
{\mathbb E}(Y_n^p) &=& {\mathbb E}(Y_n^pI_{\{Y_n<\mu_n+\mu_n^\delta\}})+{\mathbb E}(Y_n^pI_{\{Y_n\geq \mu_n+\mu_n^\delta\}}) \\ &\leq & (\mu_n+\mu_n^\delta)^p+(k\mu_n)^p{\mathbb P}(Y_n\geq \mu_n+\mu_n^\delta,\end{aligned}$$ with order of magnitude $\mu_n^p$.
For the lower bound (which is obvious by Jensen’s inequality if $p\geq 1$), $${\mathbb E}(Y_n^p)\geq (\mu_n-\mu_n^{\delta})^p{\mathbb P}(X_n>\mu_n-\mu_n^\delta),$$ again with order of magnitude $\mu_n^p$.
\[posrec\]If $(1+\gamma)(\alpha+\gamma)<1$, the Markov chain is positive recurrent, and thus the distribution of $X_n$ converges to a stationary distribution.
This uses Theorem 11.0.1 of [@meyntweedie].
We choose $p\in(0,1)$ such that $(1+\gamma)^p+(\alpha+\gamma)^p<2$. Because $\frac{{\mathrm{d}}}{{\mathrm{d}}p}((1+\gamma)^p+(\alpha+\gamma)^p)$ is negative at $p=0$ if $\log(1+\gamma)+\log(\alpha+\gamma)<0$, it will be possible to find such a $p$ if $(1+\gamma)(\alpha+\gamma)<1$.
We now let $V(x)=x^p$. In [@meyntweedie], the drift $\Delta V(x)$ is defined as $$\Delta V(x)={\mathbb E}(V(X_{n+1})-V(X_n)|X_n=x),$$ and by Theorem 11.0.1 of [@meyntweedie] the chain will be positive recurrent if (for some $V$) $\Delta V(x)\leq -1$ for $x$ large enough. Now $$\begin{aligned}
{\mathbb E}(X_{n+1}^p|X_n=x) &=& \frac{x^p}{2}\left({\mathbb E}\left(\left(1+
\frac{Y_{n+1}}{x}+\frac{Z_{n+1}}{x}
\right)^p|G_n\right)+{\mathbb E}\left(\left(\frac{W_{n+1}}{x}+
\frac{Y_{n+1}}{x}+\frac{Z_{n+1}}{x}
\right)^p|G_n\right)\right)\\ & \leq & \frac{x^p}{2}\left(\left(1+\gamma\right)^p+ \left(\alpha+\gamma\right)^p\right)+o(x^p) \\ & & \mbox{(by Lemma \ref{moments})},\end{aligned}$$ so $$\Delta V(x)\leq x^p\left(\frac{\left(1+\gamma\right)^p+ \left(\alpha+\gamma\right)^p}{2}-1\right)+o(x^p),$$ which will be less than $-1$ for $x$ large enough, giving the result.
We now investigate the tail behaviour of the stationary distribution, in the case where Proposition \[posrec\] shows one exists.
\[tail1\] Let $p>0$. If $(1+\gamma)^p+(\alpha+\gamma)^p<2$, then a random variable $X$ with the stationary distribution of the chain has finite $p$th moment ${\mathbb E}(X^p)$, and we have convergence of $p$th moments, ${\mathbb E}(X_n^p)\to {\mathbb E}(X^p)$ as $n\to\infty$.
Again this uses a Foster-Lyapunov type technique, in this case Theorem 14.0.1 of [@meyntweedie] which states that if, for a given function $f\geq 1$, we can find $V$ such that $\Delta V(x)<-f(x)$ for $x$ large enough then $f$ has a finite integral with respect to the stationary distribution and that ${\mathbb E}(f(X_n))$ converges to this integral. We will set $f(x)=x^p+1$.
Let $V(x)=kx^p$, where $k$ is chosen so that $$k\left(\frac{\left(1+\gamma\right)^p+ \left(\alpha+\gamma\right)^p}{2}-1\right)<-1.$$ Then, by Lemma \[moments\], $$\Delta V(x)\leq kx^p\left(\frac{\left(1+\gamma\right)^p+\left(\alpha+\gamma\right)^p}{2}-1\right)+o(x^p),$$ and so $\Delta V(x)\leq -f(x)$ for $x$ large enough, giving the result.
\[tail2\] Let $p>0$. If $(1+\gamma)^p+(\alpha+\gamma)^p>2$, then a random variable $X$ with the stationary distribution of the chain does not have finite $p$th moment ${\mathbb E}(X^p)$.
As $Z_{n+1}\geq 0$, we have $${\mathbb E}(X_{n+1}^p|X_n=x)\geq \frac{x^p}{2}\left({\mathbb E}\left(\left(1+\frac{Y_{n+1}}{x}\right)^p|X_n=x\right)+{\mathbb E}\left(\left( \frac{W_{n+1}}{x}+\frac{Y_{n+1}}{x}\right)^p|X_n=x\right)\right),$$ so by Lemma \[moments\] $${\mathbb E}(X_{n+1}^p|X_n=x)\geq \frac{x^p}{2}\left((1+\gamma)^p+(\alpha+\gamma)^p\right)+o(x^p).$$ Hence the $p$th moment of $X_n$ tends to infinity as $n\to\infty$, so by Theorem 14.0.1 of [@meyntweedie], again applied to $f(x)=x^p+1$, the stationary distribution cannot have a finite $p$th moment.
If $\beta>0$ and $(1+\gamma)(\alpha+\gamma)>1$ the Markov chain is transient.
By Lemma \[moments\], $$\label{negmom}\frac{E\left(\left(1+X_{n+1}\right)^p|X_n=x\right)}{(1+x)^p}\to \frac{(1+\gamma)^p+(\alpha+\gamma)^p}{2},$$ so we apply Theorem 8.0.2 (i) of [@meyntweedie] with $V(x)=1-(1+x)^p$ for some $p<0$ such that $(1+\gamma)^p+(\alpha+\gamma)^p<2$. With this choice of $V$, shows that $\Delta V(x)>0$ for $x$ large enough, and as $V$ is bounded and positive on the natural numbers Theorem 8.0.2 (i) of [@meyntweedie] gives the result.
The case where $\beta=0$ is something of a special case as the chain is not irreducible. However we can show that when $(1+\gamma)(\alpha+\gamma)\leq 1$ the probability that a randomly chosen vertex is isolated tends to $1$, while there is positive probability that a randomly chosen vertex is not isolated when $(1+\gamma)(\alpha+\gamma)>1$.
\[beta0p\] If $\beta=0$, then
1. if $(1+\gamma)(\alpha+\gamma)\leq 1$ then almost surely $X_n=0$ for $n$ sufficiently large, and the proportion of isolated vertices in $G_n$ tends to $1$ almost surely as $n\to\infty$;
2. if $(1+\gamma)(\alpha+\gamma)> 1$ then there is $q>0$ such that the probability that $X_n\to \infty$ as $n\to\infty$ is $q$ and the probability that $X_n\to 0$ as $n\to \infty$ is $1-q$.
We note that $(X_n)$ follows a Smith-Wilkinson branching process in random environment, [@swbpre]. The environmental variables which determine the random environment are the random variables $\xi_n$, with the offspring distribution of the branching process at time $n$ having mean $1+\gamma$ if $\xi_{n+1}=1$ and $\alpha+\gamma$ if $\xi_{n+1}=0$. Hence, by Theorem 3.1 of [@swbpre], the branching process dies out with probability $1$ if ${\frac{1}{2}}\log(1+\gamma)+{\frac{1}{2}}\log(\alpha+\gamma)\leq 0$, i.e. if $(1+\gamma)(\alpha+\gamma)\leq 1$, and the branching process dies out with probability strictly less than $1$ otherwise, hence there is positive probability that $X_n\to \infty$ as $n\to\infty$. To see that the proportion of isolated vertices tends to $1$ almost surely when $(1+\gamma)(\alpha+\gamma)\leq 1$, note that the proportion of isolated vertices is increasing (as if a vertex $v$ is isolated in $G_n$ both $v0$ and $v1$ are isolated in $G_{n+1}$) and therefore must converge to some value, which cannot be less than $1$ as the degree of a random vertex converges to zero almost surely.
\[as\]
(a)
: If $\beta>0$ and $(1+\gamma)(\alpha+\gamma)<1$, the degree distribution of the graph converges to the stationary distribution of the Markov chain in the sense that if we let ${p^{(n)}_{d}}$ be the proportion of vertices in $G_n$ with degree $d$, and let $X$ be a random variable with the stationary distribution of the Markov chain, then ${p^{(n)}_{d}}\to P(X=d)$ as $n\to\infty$, almost surely, for all $d\in{\mathbb N}_0$.
(b)
: If $\beta>0$ and $(1+\gamma)(\alpha+\gamma)>1$ then ${p^{(n)}_{d}}\to 0$ as $n\to\infty$, almost surely, for all $d\in{\mathbb N}_0$.
The graph at stage $r$ contains $2^rv_0$ vertices. We then consider the edges of $G_{r+s}$ in two sets: those which are between descendants of the same vertex in $G_r$, and those which are between descendants of different vertices in $G_r$. For the former, the appearance of edges between descendants of one given vertex is independent of what happens to the descendants of the other vertices, so we can model these edges of $G_{r+s}$ as consisting of $2^rv_0$ independent copies of $\tilde{G}_s$, where $(\tilde{G}_n)_{n\in{\mathbb N}}$ represents the process as evolved from a single vertex with no edges, $\tilde{G}_0$. Then, by Chebyshev’s inequality and a Borel-Cantelli argument, as $r\to\infty$ proportions of vertices in $G_{r+s}$ with degree $d$ excluding connections to descendants of different vertices in $G_r$ converge, almost surely, to $P(X_s=d)$. In the $(1+\gamma)(\alpha+\gamma)>1$ case we know that $P(X_s=d)\to 0$ as $s\to\infty$ for all $d$, and the actual degree of a vertex is bounded below by the degree excluding some connections, so this is enough to prove (b).
To complete the proof of (a), we need to consider edges between vertices which are descendants of different vertices in $G_r$. We couple the process, starting from $G_r$, with a process with $\beta=0$ by removing all edges between a vertex and its offspring, and all edges descended from such edges. The edges thus removed from $G_{r+s}$ will all be between vertices descended from the same vertex in $G_r$, so all edges between vertices descended from different vertices in $G_r$ are present in the $\beta=0$ version. But by Proposition \[beta0p\] the proportion of vertices in $G_{r+s}$ which have non-zero degree in the $\beta=0$ version tends to zero as $s\to\infty$, and so this also applies to the proportion of vertices in $G_{r+s}$ which have edges connecting them to vertices with a different ancestor in $G_r$.
Hence as both $r$ and $s$ $\to\infty$ the proportion which have degree $d$ converges to $P(X_s=d)$.
Finally we can put the Propositions above together to deduce Theorems \[beta0\] and \[betanot0\].
*Proof of Theorem \[beta0\].* Theorem \[beta0\] follows from Proposition \[beta0p\]; in the supercritical case where $(1+\gamma)(\alpha+\gamma)>1$ the probability that a randomly chosen vertex in the graph is isolated tends to $1-q<1$.
*Proof of Theorem \[betanot0\].* Theorem \[betanot0\] follows immediately from Propositions \[posrec\], \[tail1\], \[tail2\] and \[as\].
| {
"pile_set_name": "ArXiv"
} |
=1
lect.def =-1.0 truecm =12.5pt plus 1.7pt
=-1
COMPLEX CURVES IN ALMOST-COMPLEX
MANIFOLDS AND MEROMORPHIC HULLS
Sergei IVASHKOVICH – Vsevolod SHEVCHISHIN
[**Preface**]{}
[Chapter I. Local Properties of Complex Curves.]{}
=11.5pt
[Chapter II. Compactness Theorem.]{}
[**Lecture 10. Envelopes of Meromorphy of Two-Spheres. pp. 160–168**]{}
[**Appendix IV. Complex Points and Stein Neighborhods. pp. 169–172**]{}
[**Appendix V. Seiberg-Witten Invariants and Envelopes. pp. 173–181**]{}
[**References. pp. 182–186**]{}
=0
**P r e f a c e**
This are the notes of a course, given by the first author for the Graduiertenkollegs (= graduate students) at the Ruhr-University of Bochum, in December, 1997. Previously, shorter courses were given in Göteborg, Warsaw and Bonn.
These lectures pursued two main tasks:
[**First**]{} — to give a systematic and self-contained introduction to the Gromov theory of so-called pseudoholomorphic curves (a term which will be completely eliminated in these notes). This is handled in Chapters I,II,III.
[**Second**]{} — to explain our joint results on envelopes of meromorphy of real two-spheres in complex two-dimensional manifolds, which we obtained using Gromov theory. We do this in Chapter IV.
It was impossible, of course, to omit the original motivation of Gromov in the development of the subject:
- in Appendix III, §A3.5, we explain the Gromov’s idea how to attach an analytic disk to a Lagrangian submanifold in $\cc^n$;
- in §8.5 we give the proof of his “non-squeezing theorem";
- in §8.6 we study exeptional spheres in symplectic $4$-manifolds.
In Appendix V, following S. Nemirovski, we discuss one more approach to the problem of constructig envelopes of meromorphy of spheres. This approach uses Seiberg-Witten theory. As an example if his results
- we prove that a complex disk cannot be attached from the outside to the strictly pseudoconvex domain in $\cc^2$, which is diffeomorphic to the ball.
The authors would like to give their thanks to Alan Huckleberry, the head of Graduiertenkolleg program at the Ruhr-University and to SFB-236 for the constant support of our joint research. We would like also to give our thanks to Barbara Huckleberry for the uncountable number of grammatical corrections.
[Chapter I. Local Properties of Complex Curves.]{}
In this chapter we shall concentrate on local properties of complex curves in almost-complex manifolds.The first lecture is very basic and mainly defining. Its aim is to introduce a couple of notions and recall a few standard facts. In the second one we prove, following Vekua and Sikorav, the existence of $J$-complex curves (locally through any point and any direction) for any almost-complex structure $J$ of smoothness $C^{1,\alpha}$, $\alpha >0$. One more principal result here is the so-called [*first a priori estimate*]{} in §2.4. The third lecture, as well as [*Appendix 2*]{}, is devoted to the deeper study of complex curves in real 4-dimensional manifolds. We shall prove the positivity of intersections and the Adjunction Formula, showing that local properties of complex curves in nonintegrable almost-complex structures are similar to that in integrable ones.
[Lecture 1]{}
[Complex Curves in Almost-Complex Manifolds.]{}
[1.1. Almost-Complex Manifolds, Hermitian Metrics, Associated (1,1)-forms.]{}
Let $X$ be a real manifold, and $TX$ its tangent bundle.
Definition 1.1.1. [*A continuous endomorphism $J:TX\to TX$ such that $J^2=-\id $ is called an almost-complex structure on $X$. A pair $(X,J)$ is called an almost-complex manifold.* ]{}
Exercise. Prove that locally one can always find vector fields $e_1,e_2,...,e_{2n-1},e_{2n}$ such that $Je_{2k-1}=e_{2k}, k=1,...,n$. In particular, $X$ is even-dimensional.
Examples. $1^0.$ $(\rr^{2n},J\st )$. Denote by $e_1,...,e_{2n}$ the standard basis in $\rr^{2n}$. Define a [*standard*]{} complex structure $J\st $ in $\rr^{2n}$ by $J\st (e_{2k-1})=e_{2k}$.
If $(V,J)$ is an $\rr $-linear space endowed with an endomorphism $J$, such that $J^2=-\id $, then a structure of a $\cc $-linear space on $V$ is defined by $z\cdot v=(x+iy)\cdot v:=x\cdot v +y\cdot Jv$. In the case $(\rr^{2n},J\st )$ this gives $\cc^n$.
$2^0.$ Let $x_0\in (X,J)$. By an affine change of coordinates in some chart $U\ni x_0$ we can always suppose that $x_0=0$ and $J(0)=J\st $ in basis ${\d \over \d x_1},...,{\d \over \d x_{2n}}$. One can now put $y_j=x_{2j}$ in order to have $J(0) {\d \over \d x_j}={\d \over \d y_j}$.
*$3^0.$ Complex manifolds. By definition, a complex manifold is an almost-complex manifold $(X,J)$ such that in the neighborhood of any point $x_0\in X$ a coordinate system $x_1,y_1,...,x_n,y_n$ can be found such that $J {\d \over \d x_j}={\d \over \d y_j}$ everywhere in this neighborhood.*
Equivalently there is in a neighborhood of each point a $C^1$-diffeomorphism $\phi $ into $(\rr^{2n},J\st )$ such that its differential is $(J,J\st )$-linear. The latter means that $d\phi \scirc J=J\st \scirc d\phi $ as mappings $TX\to \rr^{2n}$.
Definition 1.1.2. [*Such an almost-complex structure is called complex.* ]{}
*$4^0.$ Riemann surfaces. Riemann surfaces are oriented, connected two-dimensional manifolds. Fix some Riemannian metric $g$ on such $S$. For $v\in TS$ define $Jv=\{$ rotation of $v$ in positive direction onto $90^\circ \} $. $J$ is an almost-complex structure on $S$. If $g_1=\lambda g$ is another, conformally equivalent to $g$ metric, then this construction leads to the same almost-complex structure.*
Vice-versa, given $J$ on $S$ , denote by $J^{\cc }$ the $\cc $-linear extension of $J$ onto the complexified tangent bundle $T^{\cc }S:=TS\otimes
\cc $. Here $J^{\cc }$ is defined by $J^{\cc }(v\otimes \lambda )=Jv\otimes
\lambda $ for $v\in TS$ and $\lambda \in \cc $. Take an eigenvector $w$ of $J^{\cc }$ in $T^{\cc }S$ with an eigenvalue $-i$. Write $w=u+iv$, with $u,v\in TS$ (more accurately $w=u\otimes 1 + v\otimes i$), and put $g(u,v)=0,
g(u)=g(v)=1$. This defines a Riemannian metric, which corresponds to $J$. The statement $J^{\cc }w=-iw=v-iu$ means that $Ju=v, Jv=-u$.
The freedom in choosing $w$ here permits us to define $g$ up to a real factor, , $J$ corresponds to a class of conformally equivalent metrics.
Exercise. Prove the last assertion. More precisely: if $\lambda w$ is another eigenvector of $J^{\cc }$ with an eigenvalue $-i$ and a metric $g_{\lambda }$ is constructed from $\lambda w$ as $g$ from $w$ above, then $g=\vert \lambda \vert^2g_{\lambda }$.
We shall see later in [*Corollary 2.2.2*]{} that any almost-complex structure on the Riemann surface is complex.
Definition 1.1.3.
*$J$-Hermitian metric on $X$ is a $\rr $-bilinear form $h:TX\times TX\to \cc $ such that*
\(a) $h(Ju,v)=\overline h(v,Ju)=ih(u,v)$;
\(b) $h(u,u)>0$ for $u\not= 0$.
Decompose $h(u,v)=g_h(u,v) -i\omega_h(u,v)$. Then :
Exercise. (a) $g_h$ is a Riemannian metric on $X$ and $\omega_h$ is an exterior 2-form.
\(b) $g_h(u,v)=\omega_h(u,Jv)$ and thus $h(u,v)=\omega_h(u,Jv) -
i\omega_h(u,v)$. More over, one has $\omega_h (Ju,Jv)=\omega_h (u,v)$.
\(c) $\omega_h(u,v)=-g_h(u,Jv)$.
Definition 1.1.4. The form [*$\omega_h$ is called a (1,1)-form associated to $h$.*]{}
Conversely,
Definition 1.1.5.
*An exterior $2$-form $\omega $ on $X$ is called $J$-calibrated if*
\(a) $\omega (Ju,Jv)=\omega (u,v)$;
\(b) $\omega (u,Ju)>0$ for $u\not= 0$.
Thus, $\omega $ clearly defines a $J$-Hermitian metric $h^{\omega }(u,v)=\omega (u,Jv) - i\omega (u,v)$.
A triple $(X,J,h)$, where $h$ is $J$-Hermitian is called an almost-Hermitian manifold. We shall not use this term in these notes.
Example. To explain the coefficient $-i$ chosen in the decomposition $h(u,v)=g(u,Jv) - i\omega (u,v)$ consider $(\rr^{2n}, \omega\st, J\st)$. Here $\omega\st = \sum_{j=1}^ndx_j\wedge dy_j$ is a standard symplectic form in $\rr^{2n}$. The Hermitian metric here is $h\st = \sum_{j=1}^ndz_j\otimes d\bar z_j = \sum_{j=1}^n (dx_j\otimes
dx_j + dy_j\otimes dy_j) - i\sum_{j=1}^n(dx_j\otimes dy_j - dy_j\otimes dx_j)
= g\st -i\omega\st $.
Thus $\omega\st = -\im h\st $ is a canonical expression for a Kähler form of the Hermitian metric.
Definition 1.1.6. [*We say that an exterior $2$-form $\omega $ on $X$ tames an almost-complex structure $J$ if $\omega (v,Jv)>0$ for $v\not= 0$.* ]{}
Remark. [Here and later we will use the convention that for forms $f$ and $h$ on the vector space $V$ the exterior product is defined as $f\wedge g: = f\otimes g - g\otimes f$, , without coefficient ${1\over 2}$. ]{}
1.2. Existence of Calibrating and Tame Structures.
Here we shall prove here the following simple but very important statement:
Proposition 1.2.1. [*Let $X$ be a real manifold and $\omega $ a nowhere degenerate exterior $2$-form on $X$, , $\omega^n\not=0$, where $2n=\dim_{\rr } X$. On the open subset $U\subset X$ a $\omega $-calibrating almost-complex structure $J$ is given. Then for any relatively compact open $U_1\comp U$ there exists a $\omega $-calibrating almost-complex structure $J_1$ on the whole $X$ such that $J_1\mid_{U_1}=J\mid_{U_1}$.* ]{}
Proof. Consider a Riemannian metric $g(u,v):=\omega (u,Jv)$ on $U$. Find a Riemannian metric $g_1$ on $X$ such that $g_1\mid_{U_1}=g\mid_{
U_1}$. Since $\omega $ is not degenerate, there exists a (unique!) isomorphism $A_1:TX\to TX$ such that $$\omega (u,v) = g_1(A_1u,v)\eqno(1.2.1)$$ for all $u,v\in TX$. Further, $\omega $ is antisymmetric, so $$g_1(A_1u,v) = \omega (u,v) = -\omega (v,u) = -g_1(A_1v,u) = -g_1(u,A_1v).$$ Thus $A_1^*=-A_1$. So $A_1^*A_1=A_1A_1^*=-A_1^2$ is positively definite and self-adjoint $g_1$. Let $Q_1:=\sqrt{-A_1^2}$ be a positive square root of $-A_1^2$. Put $$J_1=A_1Q_1^{-1}.\eqno(1.2.2)$$ $J_1$ is an almost-complex structure. Indeed $$J_1^2=A_1Q_1^{-1}A_1Q_1^{-1}=A_1^2(Q_1^{-1})^2=A_1^2(-A_1^{-2})=-\id .$$ Let us check that on $U_1$ one has $J_1=J$. In fact, on $U_1$ we have $g=g_1$ and so $$\omega (Ju,Jv) = \omega (u,v) = g(A_1u,v) = \omega (A_1u,Jv),$$ and thus $A_1=J$ on $U_1$. So $Q_1=\sqrt{-A_1^2}=\id $. From here we have $J_1=A_1Q_1^{-1}=A_1=J$ on $U_1$.
Remark. Note that at this point of the proof we construct a correspondence $P:g\to J$, which maps a Riemannian metric $g$ to an $\omega $-calibrated a.-c. structure $J$ with $g(u,v)=\omega (u,Jv)
$. This correspondence is obviously a continuous map from the space of metrics to the space of structures. Moreover, note that if $g$ appears as $g=g_J(u,v)=\omega (u,Jv)$ for some $\omega $-calibrated $J$, then $P(g_J)=J$.
Let us now check that $\omega $ is $J_1$-calibrated. First: $$\omega (J_1u,J_1v)=g_1(A_1J_1u,J_1v)=g_1(A_1^2Q_1^{-1}u,J_1v)=
g_1(Q_1u,-J_1v)=$$ $$=-g_1(Q_1u,J_1v)=-g_1(u,Q_1J_1v)=-g_1(u,A_1v)=-g_1(A_1v,u)=-\omega (v,u)=$$ $$=\omega (u,v).$$ Second: $$\omega (u,J_1u)=g_1(A_1u,J_1u)=g_1(A_1u,A_1Q_1^{-1}u)=$$ $$=g_1(u,u)>0$$ for nonzero $u$.
This remark leads to the following
Corollary 1.2.2.
*Let $(X,\omega )$ be as above. Then:*
1\. There exists a $\omega $-calibrating almost-complex structure on $X$.
2\. The space of $\omega $-calibrating almost-complex structures is contractible.
Proof. The first statement of this corollary is a particular case of the proposition when $U=\emptyset $.
To prove the second one, we fix some calibrating structure $J_0$. Denote by $g_{J_0}$ the corresponding Riemannian metric on $X$, , $g_{J_0}(u,v)=\omega (u,J_0v)$. The space of Riemannian metrics is a convex cone $\calc$ in $\hom_{\rr }(TX,T^*X)$. Therefore there exists a contraction $\Psi : \calc \times [0,1]\to \calc $ to $g_{J_0}$, , $\Psi (\cdot , 0)=\id $, $ \Psi (\cdot ,1 )\equiv g_{J_0}$ and $\Psi (g_{J_0},t)\equiv g_{J_0}$.
Consider the following map $\Phi :\calj_{\omega }^c\to \calc $ from the space of $\omega $-calibrating structures to the space of metrics: $$\Phi (J)(u,v)=g_{J}(u,v)=\omega (u,Jv).$$ In the proof of Proposition 1.2.1 (see Remark there) we showed that $$P\scirc \Phi =\id :\calj_{\omega }^c\to \calj_{\omega }^c.$$
Now $P\scirc \Psi (\cdot ,t)\scirc \Phi $ will be a contraction of $\calj_{\omega }^c$ to $J_0$.
Let us prove the following
Proposition 1.2.3. [*For any symplectic manifold $(X,\omega)$, the set $\calj_\omega$ of $\omega$-tame almost complex structures on $X$ is a non-empty contractible manifold*]{}.
The proposition follow immediately from the following result from linear algebra.
Lemma 1.2.4. [*Let $V$ be a (finite-dimensional) real vector space and $\omega$ a linear symplectic form on $V$. Then the set $\jj_\omega$ of $\omega$-tame linear complex structures on $V$ is a non-empty open contractible subset in the set $\jj$ of all linear complex structures on $V$*]{}.
Proof. [*Step 1*]{}. The linear version of the Darboux theorem states that there exists a basis $(\bfx_1, \ldots, \bfx_n, \bfy_1, \ldots, \bfy_n)$ of $V$ such that for corresponding linear coordinates $(x_1, \ldots, x_n,
y_1, \ldots, y_n)$ on $V$ $$%\textstyle
\omega=\sum_{i=1}^n dx_i \wedge dy_i.
\eqno\rom(1.2.3)$$ Set $J_0 (\bfx_i) \deff \bfy_i$, $J_0 (\bfy_i) \deff -\bfx_i$. It is easy to see that $J_0$ is a linear complex structure on $V$. Moreover, the $\cc$-valued linear functionals $z_i \deff x_i + \isl\, y_i$ induce the isomorphism $\phi
:(V,J_0) \cong (\cc^n, J\st)$, $J\st \deff \isl$, of [*complex*]{} linear spaces, such that the given symplectic form $\omega$ on $V$ is mapped into the standard sympletic form $\omega\st$ on $\cc^n$ given by the same formula (A.T.1), $\phi^*\omega\st = \omega$. Thus we can conclude that there exists a $J_0$-Hermitian metric $h$ on $V$ with the following properties:
$g(v,w) \deff \re h(v,w)$ is a symmetric positively definite form on $V$ such that $h(v,w)= g(v,w) + \isl\,\omega(v,w)$ for any $v,w \in V$;
$g(J_0 v, J_0w) = g(v,w)$ and $\omega(J_0 v, J_0w) = \omega(v,w)$.
$g(v,w) = \omega(v, J_0w)$ and $\omega(v,w)=g(J_0v,w)$ for any $v,w
\in V$.
In particular, $\omega(v, J_0w)$ is positive definite, $J_0 \in
\jj_\omega$. This shows that $\jj$ is non-empty.
[*Step 2. For any $J \in \jj_\omega$ the operator $J + J_0$ is invertible*]{}. Really, otherwise we have $Jv = - J_0v$ for some non-zero $v\in V$ and then $0< \omega(v, Jv) = -\omega(v, J_0v) <0$, a contradiction.
For $J \in \jj$ with $J + J_0$ invertible we set $\sfL(J) \deff -(J - J_0) (J+
J_0)\inv$. The equivalent definitions are $$\matrix\format \r&\,\c\, & \l \\
\sfL (J) &=& -(J - J_0) (J+J_0)\inv =
-\bigl((J - J_0)J_0\bigr) \cdot \bigl( (J+J_0) J_0 \bigr)\inv \cr
\vph&=&(\bfone + JJ_0 )(\bfone - JJ_0 )\inv
= (\bfone - JJ_0)\inv (\bfone + JJ_0).
\endmatrix
\eqno\rom(1.2.4)$$
[*Step 3. If $J + J_0$ is invertible, then $W= \sfL(J)$ is $J_0$-antilinear, $W J_0= - J_0 W$*]{}. Really, $$\matrix \format \c &\,\c\, &\r &\,\c \,&\r &\,\c\, &\c
\vbox to 11pt{} \\
W J_0 &=& -(J - J_0) (J+J_0)\inv J_0 &=&
(\bfone - JJ_0)\inv (\bfone + JJ_0) J_0 &=&
\cr
&&
(\bfone -JJ_0)\inv (J_0- J)&=&
- (\bfone -JJ_0)\inv J(\bfone + JJ_0) &=&
\cr
&&
(\bfone -JJ_0)\inv J\inv (\bfone +JJ_0) &=&
(J - J^2 J_0)\inv (\bfone +JJ_0) &=&
\cr
&&
(J_0 + J)\inv (\bfone +JJ_0) &=&
\bigl( (\bfone -JJ_0) J_0 \bigr)\inv (\bfone +JJ_0) &=&
\cr
&&
J_0\inv(\bfone -JJ_0)\inv (\bfone +JJ_0) &=&
- J_0(\bfone -JJ_0)\inv (\bfone +JJ_0) &=& - J_0 W
\endmatrix$$
[*Step 4. $J \in \jj_\omega$ implies $\bfone - W^\sft W \gg 0$ for $W=
\sfL(J)$*]{}, the operator $\bfone - W^\sft W$ is positively definite, $g(v,v) > g(Wv,Wv)$ for any non-zero $v \in V$. The conjugation of $W \in
\endo(V)$ is done using the metric $g$.
First we note that $J_0^\sft = J_0\inv =-J_0$. Since $\bfone -JJ_0$ is invertible, the positivity of $\bfone - W^\sft W$ is equivalent to the positivity of $(\bfone -JJ_0)^\sft(\bfone - W^\sft W) (\bfone -JJ_0)$. Computing this operator we obtain $$\matrix\format \c \vbox to 11.5pt{} \\
(\bfone -JJ_0)^\sft(\bfone - W^\sft W) (\bfone -JJ_0) =
(\bfone -JJ_0)^\sft(\bfone -JJ_0) -
\cr
(\bfone -JJ_0)^\sft\bigl( (\bfone +JJ_0) (\bfone -JJ_0)\inv \bigr)^\sft
\bigl( (\bfone +JJ_0) (\bfone -JJ_0)\inv \bigr)(\bfone -JJ_0)=
\cr
(\bfone -JJ_0)^\sft(\bfone -JJ_0) -
(\bfone +JJ_0)^\sft (\bfone +JJ_0) =
\cr
(\bfone +J_0 J^\sft)(\bfone - JJ_0) - (\bfone - J_0 J^\sft)(\bfone +JJ_0)=
\cr
\bfone +J_0 J^\sft - JJ_0 - J_0 J^\sft JJ_0 -
\bfone + J_0 J^\sft - JJ_0 + J_0 J^\sft JJ_0 =
\cr
2(J_0 J^\sft - JJ_0).
\endmatrix$$
Now, writing any non-zero $v\in V$ as $v= J_0w$, we obtain $$\matrix\format \l &\,\c\,\;& \l \vbox to 12.5pt{} \\
g(v, (J_0 J^\sft - JJ_0)v) &=& g(J_0w, (J_0 J^\sft - JJ_0)J_0w)=
\cr
g(J_0w, J_0J^\sft J_0w) + g(J_0w, Jw) &=&
g(w, J^\sft J_0w) + g(J_0w, Jw) =
\cr
g(Jw, J_0w) + g(J_0w, Jw) &=& 2\omega(w, Jw)>0.
\endmatrix$$
Thus $\sfL$ maps $\jj_\omega$ into the set $$\calw \deff
\{\, W \in \endo(V)\;:\; W J_0 =- J_0W,\; \bfone -W^\sft W \gg0 \,\}.
\eqno\rom(1.2.5)$$
It is easy to see that $\calw$ is contractible. Thus it is sufficient to show that $\sfL: \jj_\omega \to \calw$ is a diffeomorphism. Explicit calculations show that the inverse map should be given by $$J= \sfK(W) \deff J_0 (\bfone + W) (\bfone - W)\inv=
J_0 (\bfone - W)\inv (\bfone + W) =
J_0 {\bfone + W \over \bfone - W}
\eqno\rom(1.2.6)$$
[*Step 5. Let $W\in \calw$. Then $J= \sfK(W) = J_0 {\bfone + W \over
\bfone - W}$ is well-defined and $J^2 = -\bfone$*]{}. Really, the condition $\bfone -W^\sft W \gg0$ implies that $\bfone \pm W$ is invertible, and then $$\matrix\format \c &\,\c\,&\l \vbox to 11.5pt{} \\
J^2 &=&
-J_0 (\bfone + W) (\bfone - W)\inv J_0\inv (\bfone + W) (\bfone - W)\inv
\cr
&=& - (\bfone + J_0 W J_0\inv) (\bfone - J_0 W J_0\inv)\inv
(\bfone + W) (\bfone - W)\inv
\cr
&=& - (\bfone - W ) (\bfone + W )\inv
(\bfone + W) (\bfone - W)\inv \qquad =- \bfone
\endmatrix$$
[*Step 6. Let $W\in \calw$ and $J= \sfK(W)$. Then $J+ J_0$ is invertible and $J \in \jj_\omega$*]{}. In fact, $J + J_0 = J_0\left( \bfone + {\bfone +
W \over \bfone - W} \right) = {2 J_0 \over \bfone - W}$ is invertible. Now, repeating argumantations from [*Step 4*]{} the positivity of $\bfone -
W^\sft W$ is equivalent to the positivity of $J_0 J^\sft - JJ_0$, which means the tameness of $J$.
The lemma is proved.
Remark. One can consider $\sfK:\calw \to \jj_\omega \subset \jj$ a generalized Cayley transformation defined on a bounded domain $\calw$ in the complex linear space $\barr \endo_\cc(V,J_0)$ of $J_0$-antilinear endomorphisms of $V$.
[1.3. Almost-Complex Submanifolds, Complex Curves, Energy and Area.]{}
Let $J$ be some linear complex structure on $\rr^{2n}$, , $J\in
\endo(\rr^{2n}), J^2=-\id $. Let $h$ be some $J$-Hermitian metric on $\rr^{2n}$, $\omega =\omega_h$ its associated $(1,1)$-form and $g=g_h$ corresponding Riemannian metric. Put $\sigma_k = {1\over k!}\omega^k$.
Wirtinger Inequality. [*For any $g$-orthonormal system $v_1,...,v_{2k}$ in $\rr^{2n}$ one has $$\vert \sigma_k(v_1,...,v_{2k})\vert \le 1, \eqno(1.3.1)$$ with equality taking place iff the subspace $\<v_1,...,v_{2k}\>$ is $J$ - invariant.* ]{}
Remark. We shall also call $J$-invariant subspaces $J$-complex.
Proof. We shall prove (1.3.1) by induction on $k$.
Let $k=1$. Then $\vert \omega (v_1,v_2)\vert =\vert g(v_1,Jv_2)\vert
\le \Vert v_1\Vert_g\cdot \Vert Jv_2\Vert_g=$ $\Vert v_1\Vert_g\cdot \Vert v_2
\Vert_g$, with equality iff $v_1$ and $Jv_2$ are collinear. Thus the subspace $\<v_1,v_2\>$ is $J$-invariant.
Now let $V$ be a subspace of dimension $2k$. Put $\omega'=\omega\mid_V$. Find a $g$-orthonormal base $e_1,...,e_{2k}$ of $V$, s.t. $\omega'=
\lambda_1e^1\wedge e^2 + ... +\lambda_ke^{2k-1}\wedge e^{2k}$. By the case $k=1$ we have $\vert \omega' (e_{2p-1},e_{2p})\vert=
\vert \omega (e_{2p-1},e_{2p})\vert =\vert \lambda_p\vert \le 1$, with equality taking place iff $e_{2p}={+\over -}Je_{2p-1}$. So, for $\sigma'_k
=\sigma_k\mid_V$ we obtain $$\vert \sigma_k'(e_1,...,e_{2k})\vert = \vert {1\over k!}\omega^k(e_1,...,
e_{2k})\vert = \vert \lambda_1...\lambda_k\vert \le 1,$$ with equality iff $e_{2p}={+\over -}Je_{2p-1}$ for all $1\le p\le k$.
A submanifold $Y$ of an almost-complex manifold $(X,J)$ is called an almost-complex (or $J$-complex) if the tangent spaces to $Y$ are $J$-invariant.
Definition 1.3.1. [*Nowhere degenerate exterior two-form $\omega$ on $X$ is called symplectic if $d\omega =0$.*]{}
A pair $(X,\omega )$ is called a symplectic manifold.
Corollary 1.3.1. [*Let $(X,w,J)$ be a symplectic manifold with $w$ being $J$-calibrated. Then $J$-complex submanifolds are minimal and their volume is given by $$vol_{2k}(Y) = {1\over k!}\int_Yw^k. \eqno(1.3.2)$$*]{}
Proof. Let $Y$ be a $J$-complex submanifold of $X$, and $Y_1$ some submanifold of dimension $2k$ as well as $Y$. Suppose $\d Y=\d Y_1$ and $Y\cup (-Y_1) \sim 0$. Denote by $dy$ and $dy_1$ the volume forms of $Y$ and $Y_1$ with respect to metric $g(u,v)=\omega (u,Jv)$. Then we have $$0= \int_{Y\cup (-Y_1)}\sigma_k = \int_{Y}\sigma_k - \int_{Y_1}\sigma_k =
\int_{Y}\sigma_k (TY)dy - \int_{Y_1}\sigma_k(TY_1)dy_1\ge$$ $$\ge \int_{Y}dy - \int_{Y_1}dy_1=vol(Y)-vol(Y_1),$$ with equality taking place if and only if $\sigma_k(T_pY_1)=1$ for all $p\in Y_1$, , when $Y_1$ is also $J$-complex.
Definition 1.3.2. [*A $C^0\cap L^{1,2}$-map $u :(Y,J_Y)
\to (X,J_X)$ is called holomorphic if for a.-a. $x\in Y$ $$du_x\scirc J_Y(x)= J_X(u(x))\scirc du_x\eqno(1.3.3)$$ as mappings $T_xY\to T_{u(x)}X$.* ]{}
Exercise. [Check that for complex valued functions, , $u :(\rr^2,i)\to (\rr^2,i)$, (1.3.3) is a Cauchy-Riemann equation. ]{}
Sometimes one calls such $u$ a $(J_Y,J_X)$-holomorphic map. In the special case when $(Y,J_Y)$ is a Riemann surface $(S,J_S)$, one calls $(S,J_S,u)$ a parameterized $J_X$-complex curve. Its image $u(S)$ one simply calls a $J_X$-complex curve.
If some $J$-calibrated exterior $2$-form $\omega $ on an almost-complex manifold $(X,J)$ is chosen, one defines the $\omega $-area of a $J$-complex curve $u:(S,J_S)\to (X,J)$ as $$area[u(S)] = \int_Su^*\omega .\eqno(1.3.4)$$ Remark that if $\omega $ is symplectic, then by [*Corollary 1.3.1*]{} this area coincides with $g$-area for the metric $g(\cdot ,\cdot ) =
\omega (\cdot ,J\cdot )$ and $J$-complex curves are $g$-minimal surfaces.
Remark also that for a $J$-complex curve $u:(\Delta ,J\st )\to (X,J)$, parameterized by a standard unit disk, one has $$\int_{\Delta }u^*\omega = \int_{\Delta }\omega (du({\d \over \d x}),
du({\d \over \d y}))dx\wedge dy = -\int_{\Delta }{1\over 2}g(du({\d \over
\d x}),Jdu(i{\d \over \d x}))dx\wedge dy =$$ $$=-\int_{\Delta }{1\over 2}g(du({\d \over \d x}),J^2du({\d \over \d x}))dx
\wedge dy = \int_{\Delta }{1\over 2}g({\d u\over \d x}, {\d u \over \d x}))
dx\wedge dy = {1\over 2}\Vert {\d u\over \d x}\Vert^2_{L^2(\Delta ,X)}.$$ Thus $$\int_{\Delta }u^*\omega = \Vert {\d u\over \d x}
\Vert^2_{L^2(\Delta ,X)}+\Vert {\d u\over \d y}\Vert^2_{L^2(\Delta ,X)}.
\eqno(1.3.5)$$ The right hand side of (1.3.5) is called an energy of a $C^0\cap L^{1,2}$-map (not necessary holomorphic!) from $\Delta $ to a Riemannian manifold $(X,g)$. Thus, (1.3.5) tells us that, for holomorphic maps, energy and the area of the image, taken with multiplicities, are equal.
Remark. Let $g$ be a Riemannian metric on $C$ compatible with $j_C$, $h$ a Riemannian metric on $X$, and $u:C
\to X$ a $J$-holomorphic immersion. Then $\norm{du}^2_{L^2(C)}$ is independent of the choice of $g$ and coincides with the area of the image $u(C)$ the metric $h_J(\cdot, \cdot) \deff \half(h(\cdot, \cdot) + h(J\cdot, J\cdot)
)$. The metric $h_J$ here can be seen as a “Hermitization” of $h$ $J$. The independence of $\norm{du}^2_{L^2(C)}$ of the choice of a metric $g$ on $C$ in the same conformal class is a well-known fact, see, , \[S-U\]. Thus we can use the flat metric $dx^2 + dy^2$ to compare the area and the energy. For a $J$-holomorphic map we get $$\norm{du}^2_{L^2(C)}= \int_C |\d_x u|_h^2 + |\d_y u|_h^2 =
\int_C |\d_x u|_h^2 + |J \d_x u|_h^2 = \int_C |du|_{h_J}^2 =
\area_{h_J}(u(C)),$$ where the last equality is another well-known result, see, , \[G\]. Since we consider changing almost complex structures on $X$, it is useful to know that we can use any Riemannian metric on $X$ having a reasonable notion of area.
[1.4. Symplectic Surfaces.]{}
First we recall some elementary facts about orthogonal complex structures in $\rr^4$.
In $\rr^4$ with coordinates $x_1, y_1, x_2, y_2$ consider the standard symplectic form $\omega\st = dx_1\wedge dy_1 + dx_2\wedge dy_2 $ and the standard complex structure $J\st$ defined by the operator $$J\st =
\left( \matrix
0&-1&0&0\cr
1&0&0&0\cr
0&0&0&-1\cr
0&0&1&0\cr
\endmatrix \right).$$
Let $\jj$ denote the set of all orthogonal complex structures in $\rr^4$ giving $\rr^4$ the same orientation as $J\st$. Orthogonality here means just that $J$ is an orthogonal matrix. The same orientaion means that for any pair $x_1, x_2$ of $J$-independent vectors the basis $x_1, Jx_1, x_2, Jx_2$ gives the same orientation of $\rr^4$ as $J\st$.
Exercise. [Prove that this orientation does not depend on the particular choice of $x_1, x_2$ and coincides with the orientation given by $\omega\st^2$. ]{}
The following lemma summarizes the elementary facts which we need for the sequel.
Lemma 1.4.1. *The elements of $\jj$ have the form $$J_{s} =
\left(
\matrix
0&-s &c_1&c_2\cr
s &0&c_2&-c_1\cr
-c_1&-c_2&0&-s \cr
-c_2&c_1&s &0\cr
\endmatrix \right),\eqno(1.4.1)$$ with $c_1^2 + c_2^2 + s^2 = 1$. One also has for $x\in \rr^4$ $$\omega\st (x, J_s x) = s \Vert x\Vert^2 . \eqno(1.4.2)$$*
We remark that the set $\jj$ is a unit two-dimensional sphere $S^2$ in $\rr^3$ with coordinates $c_1, c_2,s $. We note also that the number $\omega\st (x, J_s x)$ does not depend on the choice of a *unit vector $x$. One also remarks that the standard structure corresponds to the north pole of $S^2$ and structures tamed by $\omega\st $ constitute the upper half-sphere.*
Exercise. 1. Note that $\jj = \{ J\in O(4): J^2=-E\} = \{ J:
J^t=J^{-1}=-J\} $. Prove that condition $J^t=-J$ implies that $$J = \left(
\matrix
0&-s &c_1&c_2\cr
s &0&c_3&c_4\cr
-c_1&-c_3&0&-t \cr
-c_2&-c_4&t &0\cr
\endmatrix \right),$$ for some $s,t,c_i, i=1,...,4.$
[**2.**]{} Denote by $C$ the matrix $\matrix c_1& c_2 \cr c_3& c_4\endmatrix$. Prove that condition $J^2=-E$ implies that:
a\) $$C = \left(
\matrix
\pm c_1&c_2\cr
\mp c_2&c_1\cr
\endmatrix \right)$$ with $c_1^2+c_2^2=1-s^2$;
b\) $\vert t\vert =\vert s\vert $;
c\) and, finally, that $s=t$ and for $s\not= \pm 1$ the matrix ${1\over \sqrt{1-s^2}}\cdot C$ lies in $O_-(2)$ (the set of orthogonal $2\times 2$-matrices with determinant $-1$) iff the structure $$J_s :=
\left(
\matrix
0&-s &c_1&c_2\cr
s &0&c_3&c_4\cr
-c_1&-c_3&0&-s \cr
-c_2&-c_4&s &0\cr
\endmatrix \right)$$ defines the same orientation of $\rr^4$ as $J\st = J_1$.
[**3.**]{} Check finally that $\omega\st (x,J_sx)=s\cdot \Vert x\Vert^2$.
Let $M_1$ and $M_2$ be two smooth oriented surfaces in the unit ball $B\subset
\rr^4$ with zero as a common point. Let $v_1, w_1$ and $v_2, w_2$ be oriented bases of $T_0M_1$ and $T_0M_2$, respectively. Suppose that $M_1$ and $M_2$ intersect transversally at zero, [**]{}, $v_1, w_1, v_2, w_2$ is the basis of $\rr^4$. We say that they intersect positively if this basis gives the same orientation of $\rr ^4$ as the standard one.
Let $(X,\omega )$ be a manifold with nowhere degenerate exterior $2$-form $\omega $.
Definition 1.4.1. [*An immersion $u:S\to X$ of a real surface $S$ into $X$ is called $\omega $-positive if $u^*\omega $ never vanishes.* ]{}
If $\omega $ is symplectic we call such immersions [*symplectic*]{}.
Definition 1.4.2. [*An almost-complex structure $J$ is said to be tamed by an exterior $2$-form $\omega $ if $\omega (u,Ju)>0$ for any nonzero $u\in TX$*]{}.
In the following lemma we suppose for simplicity of proof that $\dim_{\rr }X
=4$.
Lemma 1.4.2.
*Let $M$ be a $\omega $-positive compact surface immersed into $(X,\omega )$ with only double positive local self-intersections, and let $U_1\subset \subset U$ be neighborhoods of $M$. Then for any given $\omega $-tamed a.-c. structure $J$ there exists a smooth family $\{ J_t\} _{t\in [0, 1]} $ of almost-complex structures on $X$ such that:*
a\) $J_0$ is the given structure $J$ on $X$;
b\) for each $t\in [0, 1]$ the set $\{ x\in X: J_t(x) \not= J_0(x) \} $ is contained in $U_1$;
c\) $M$ is $J_1$-holomorphic;
d\) all $\{J_t\}$ are tamed by the given form $\omega $, , $\omega(v,J_tv) > 0$ for every nonzero $v\in TX$.
Proof. Let $N$ be a normal bundle to $M$ in $X$ and $V_1$ a neighbourhood of the zero section in $N$. Shrinking $V_1$ and $U_1$ we can assume that $U_1$ is an image of $V_1$ under an $\exp$-map in $J$-Hermitian metric $h$ associated to $\omega $. More precisely, one should take the Riemannian metric associated to $h$, $g = \re h$. Shrinking $V_1$ and $U_1$ once more, if necessary, we can extend the distribution of tangent planes to $M =$ (zero section of $N$) to the distribution $\{ L_x\} _{x\in V_1}$ of $\omega $-positive planes on $V_1$. Here we do not distinguish between $\omega$ and its lift onto $V_1$ by $\exp$. Denote by $N_x$ the subspace in $T_xV_1$ which is $g$-orthogonal to $L_x$. Perturbing $g$ we can choose the distribution $\{ L_x\}$ in such a way that if $\exp(x)= \exp(y)$ for some $x\not= y$ from $V_1$ then $d\exp_x(L_x)=d\exp_y(N_y)$ and $d\exp_y(L_y)=d\exp_x(N_x)$. In particular, we suppose that $M$ intersects itself $g$-orthogonally. For every $x\in V_1$ choose an orthonormal basis $e_1(x), e_2(x)$ of $L_x$ such that $\omega_x(
e_1(x), e_2(x)) > 0$, and an orthonormal basis $e_3(x), e_4(x)$ of $N_x$ such that the basis $e_1(x), e_2(x), e_3(x), e_4(x)$ gives the same orientation of $T_xV_1$ as $\omega^2$.
Define an almost complex structure $J^1$ on $U_1$ by $J^1e_1 = e_2$, $J^1 e_3 =
e_4$. Then $J^1$ depends smoothly on $x$, even when $e_j(x)$ are not smooth in $x$. Furthermore, we have $\omega _x(e_1(x), J_x^1e_1(x)) > 0$ and thus from [*Lemma 1.1.1*]{} we see that the relation $\omega _x(e_3(x), J_x^1e_3(x)) > 0$ is also satisfied. This means that our $J^1$ is tamed by $\omega $. Note also that $M$ is $J^1$-holomorphic.
Denote by $\jj_x$ the sphere of $g$-orthogonal complex structures on $T_x
X$ as in [*Lemma 1.4.1*]{}. Let $\gamma_x$ be the shortest geodesic on $\jj_x$ joining $J(x)$ - our given integrable structure, with $ J_x^1$. Put $J_t(x)
= J^1_{\gamma _x(t\cdot \Vert \gamma _x\Vert \cdot \phi (x ))}$. Here $\Vert
\gamma_x \Vert $ denotes the length of $\gamma _x$, $\phi$ is smooth on $X$ with support in $U_1$ and identically one in the neighborhood of $M$. The curve $\{ J_t \}$ satisfies all the conditions of our lemma.
Lemma 1.4.3. [*Under the conditions of Lemma 1.4.2 the structure $J_1$ can be made complex in the neighborhood of $M$.* ]{}
The proof is left to the reader.
[1.5. Adjunction Formula for Immersed Symplectic Surfaces.]{}
Let us now prove the Adjunction Formula for *immersed symplectic surfaces. Let $u:\bigsqcup_{j=1}^dS_j\to (X,\omega)$ be a reduced compact symplectic surface (see Definition 1.4.1) immersed into a symplectic four-dimensional manifold. Let $g_j$ denote the genus of $S_j$ and $M_j=u(S_j)$. Put $M:=\bigcup_{j=1}^dM_j$ and denote by $[M]^2$ the homological self-intersection number of $M$. Define a geometrical self-intersection number $\delta $ of $M$ in the following way. Perturb $M$ to obtain a symplectic surface $\widetilde M$ with only transversal double points. Then $\delta$ will be the sum of indices of the intersection over those double points. Those indices can be equal to $1$ or $-1$.*
Exercise. Let $P$ be an oriented plane in $\rr^4$. Call $P$ symplectic if $\omega\st (P)>0$. Find two symplectic planes $P_1,P_2\in \rr^4$ which intersect transversally at origin with intersection index $-1$.
Let $J$ be some almost complex structure which is tamed by $\omega$, [*,*]{} $\omega(\xi, J\xi)>0$ for any nonzero $\xi \in TX$. Denote by $c_1(X, J)$ the first Chern class of $X$ with respect to $J$. Since, in fact, $c_1(X, J)$ does not depend on continuous changes of $J$ and since the set of $\omega$-tamed almost complex structures is contractible, we usually omit the dependence of $c_1(X)$ on $J$.
Lemma 1.5.1. [*Let $M=\bigcup_{j=1}^d M_j$ be a compact immersed symplectic surface in four-dimensional symplectic manifold $X$. Then $$\sum_{j=1}^d g_j = {[M]^2 - c_1(X)[M]\over2} + d - \delta .\eqno(1.5.1)$$*]{}
Proof. By replacing every $M_j$ by its small perturbation, we can suppose that $M_j$ has only transversal double self-intersection points. Let $N_j$ be a normal bundle to $M_j$ and let $\widetilde M_j$ denote the zero section of $N_j$. Also let $\exp_j $ be the exponential map from a neighborhood $V_j$ of $\widetilde M_j\subset N_j$ onto the neighborhood $W_j$ of $M_j$. Lift $\omega$ and $J$ onto $V_j$. Since $\widetilde M_j$ is embedded to $V_j$, we can apply [*Lemma 1.4.2*]{} to obtain the $\omega$-tame almost complex structure $J_j$ on $V_j$ such that $\widetilde M_j$ is $J_j$-holomorphic.
For every $j$ we now have the following exact sequence of complex bundles:
$$0\longrightarrow TS_j {\buildrel du\over \longrightarrow }
E_j \buildrel \pr \over \longrightarrow
N_j \longrightarrow 0 \eqno(1.5.2)$$
Here $E_j=(u^*TX)\ogran_{S_j}$ is endowed with complex structure given by $J_j$. Since $du$ is nowhere degenerate [*complex*]{} linear morphism, $N_j \deff
E_j/du(TM_j)$ is a complex line bundle over $S_j$. From (1.5.2) we get
$$c_1(E_j) = c_1(TS_j) + c_1(N_j). \eqno(1.5.3)$$
Observe now that $c_1(E_j) = c_1(X)[M_j]$ and that $c_1(TS) = \sum_{j=1}^d
c_1(TS_j) = \sum_{j=1}^d(2-2g_j) = 2d - 2\sum_{j=1}^d g_j$. Furthermore, $c_1(N_j)$ is the algebraic number of zeros of a generic smooth section of $N_j$. To compare this number with the self-intersection of $M_j$ in $X$, note that if we move $M_j$ generically to obtain $M_j'$, then the intersection number $int(M_j,
M_j')$ is equal to the algebraic number of zeros of generic section of $N_j$ plus two times the sum of intersection numbers of $M_j$ in self-intersection points, [*,*]{} $[M_j]^2 = c_1(N_j) + 2\delta_j$. So $$c_1(X)[M_j] = 2 - 2 g_j + [M_j]^2 - 2\delta _j . \eqno(1.5.4)$$ Now it only remains to take the sum over $j=1,\ldots, d$ and to remark that the intersection points of $M_i$ with $M_j$ for $i\not=j$ make a double contribution to $[M]^2$.
[Appendix 1]{}
[Chern Class and Riemann-Roch Formula]{}
[A1.1. First Chern Class.]{}
Let $L\to M$ be a complex line bundle over a real manifold $M$. One of the possible definitions consists of taking a real rank two bundle over $M$ with an operator $J\in \End (L)$, satisfying $J^2=-\id_L$. One can than locally find a frame $e_1(x), e_2(x)$ with $Je_1(x)=e_2(x)$. This gives a covering $\{ U_\alpha \} $ of $M$ together with isomorphisms of complex line bundles $\phi_{\alpha }:L\mid_{U_{\alpha }}\to U_{\alpha }
\times \cc $, , a standard definition of a complex line bundle. Sometimes we shall mark as $e_1^J, e_2^J, \phi_{\alpha }^J$ the corresponding objects to underline their dependence on $J$.
Denoting by $\cala=\cala_M$ and by $\cala^*=\cala_M^*$ the sheaves of complex valued (resp. complex valued nonvanishing) functions on $M$, we observe the following exact sequence
$$0\to \zz \buildrel{i}\over {\longrightarrow} \cala \buildrel{exp(2\pi i
\cdot )}\over {\longrightarrow} \cala^* \to 0. \eqno(A1.1.1)$$
Here $i$ is an imbedding of the sheaf of locally constant integer valued functions into $\cala$, end $exp(2\isl\pi \cdot ):f\to e^{2\isl\pi f}$. The sequence(A1.1.1) gives rise to the following long exact sequence of Čech cohomologies $$0=\sfh^1(M,\cala)\buildrel{\exp(2\isl\pi \cdot)}\over {\longrightarrow}
\sfh^1(M,\cala^*\buildrel{\delta }\over {\longrightarrow} \sfh^2(M,\zz )
\to \sfh^2(M,\cala)=0. \eqno(A1.1.2)$$ Equalities $\sfh^1(M,\cala)=\sfh^2(M,\cala)=0$ follow from the fact that the sheaf $\cala$ admits a partition of unity.
Classes from $\sfh^1(M,\cala^*)$ are the defining cocycles of complex line bundles - one more possible definition of the line bundle. In terms of local trivializations $\{ \phi_{\alpha }\} $ such cocycles can be obtained as $\phi_{\alpha ,\beta }=\phi_{\alpha }\scirc \phi_{\beta }^{-1}
:(U_{\alpha }\cap U_{\beta })\times \cc \to (U_{\alpha }\cap U_{\beta })
\times \cc $, , $\phi_{\alpha ,\beta }\in \cala^*_{U_{\alpha }\cap
U_{\beta }}$.
Definition A1.1.1. [*If $\{ \phi_{\alpha ,\beta }\}
\in \sfh^1(M, \cala^*)$ is a defining cocycle of a complex line bundle $L$, then $\delta (\{ \phi_{\alpha ,\beta }\} )\in \sfh^2(M, \zz)$ is called the first Chern class of $L$ and is usually denoted as $c_1(L)$.* ]{}
For the complex bundle $E$ of complex rank $r$ the first Chern class is defined as $c_1(\Lambda^rE)$. If $E=TX$, the tangent bundle to an almost-complex manifold $X$, then one simply writes $c_1(X)$ or $c_1(X,J)$ if an almost-complex structure is needed to be specified.
If an almost-complex structure $J$ on the real bundle $E$ varies continuously, then the corresponding trivializations $\{ \phi_{\alpha }\} $ above (on $\Lambda^r E$) can be obviously chosen to also vary continuously. Thus $c_1(E,J)$ varies continuously. But $c_1(E,J)\in \sfh^2(M, \zz )$, , takes values in a discrete group. So, it does not change at all. This simple but important observation together with [*Corollary 1.2.2*]{} leads to the following
Corollary A1.1.1. [*Let $\omega $ be a nondegenerate exterior two-form on the even-dimensional real manifold $X$. Then $c_1(X,J)$ does not depend on the choice of $\omega $-calibrating (and even $\omega $-compatible) almost-complex structure $J$.* ]{}
*A1.2. Riemann-Roch Formula and index of $\dbar $-type operators.*
For the holomorphic bundle $E$ of complex rank $r$ over a compact Riemann surface $S$ denote by $\calo_E$ the sheaf of its holomorphic sections. In a usual way one defines the cohomology groups $\sfh^0(S, \calo_E$ and $\sfh^1(S, \calo_E)$. Denote $h^i=dim_{\cc }\sfh^i$. By $g$ we denote the genus of $S$ and $$c_1 = \int_Sc_1(E).$$ These numbers are related by the classical
Riemann-Roch Formula.
*For a holomorphic bundle $E$ over a Riemann surface $S$ one has*
$$h^0-h^1 = c_1 + r\cdot (1 - g). \eqno(A1.2.1)$$
This formula can be interpreted as the formula for the index of some operators acting on the spaces of smooth sections of $E$. On the sheaf of smooth sections of a holomorphic bundle over a Riemann surface, or more generally over a complex manifold, one can define $\dbar $-operators. Those are $\cc $-linear operators $\dbar :\Gamma^{1,p}(S,E)\to \Gamma^p_{0,1}(S,E)$ satisfying
$$\dbar (f\cdot \sigma ) = \dbar_Sf\otimes \sigma + f\cdot \dbar \sigma .
\eqno(A1.2.2)$$
Here by $\Gamma^{1,p}(S,E)$ we denote the Sobolev space of $(1,p)$- smooth sections of $E$, and by $\Gamma^p_{0,1}(S,E)$ the space of $(0,1)$ $L^p$-integrable forms with coefficients in $E$. $\dbar_S$ is a canonical $\dbar $-operator on $S$.
If one additionally fixes some Hermitian metric on $E$, then such an operator is determined uniquely if one imposes the additional condition to preserve the scalar product, see \[G-H\] Ch.0 for details.
The operator $\dbar $ being elliptic is Fredholm and its index is defined as $\ind \dbar := \dim \ker \dbar \allowbreak - \dim \coker \dbar $. Remark that $\ind \dbar = h^0-h^1$ and so by the Riemann-Roch formula $$\ind \dbar = c_1 + r\cdot (1 - g). \eqno(A1.2.3)$$
Definition A1.2.1. [*An $\rr $-linear operator $D:\dbar :\Gamma^{1,p}(S,E)\to \Gamma^p_{0,1}(S,E)$ which can be represented as $\dbar + R$ with $R\in C^0(S, \hom _{\rr }(E,\Lambda^{0,1}
\otimes E))$ shall be called a $\dbar $-type operator.* ]{}
This is again an elliptic (Fredholm) operator, which is homotopic to $\dbar $. So by the homotopy invariance of the index we have that for any $\dbar $-type operator $D$ $$\ind_\rr D = 2\cdot (c_1 + r\cdot (1 - g)). \eqno(A1.2.4)$$ The reader should take into account that since $D$ is real, the real dimensions in the last formula are considered. That is why the number $2$ appears.
[Lecture 2]{}
[Local Existence of Curves]{}
[2.1. Sobolev Imbeddings, Cauchy-Green Operators, Calderon-Zygmund Inequality.]{}
For a natural $k$ and a real $p\ge 1$ the Sobolev space $L^{k,p}
(\Delta ,\cc^n)$ consists of functions $f\in L^p(\Delta ,\cc^n)$ such that their derivatives up to the order $k$ are also in $L^p(\Delta ,\cc^n)$. One puts $\Vert f\Vert_{k,p}:= \Sigma_{0\le \vert i\vert \le k}\Vert D^if\Vert_p$.
For $0<\alpha <1$ one considers also the Hölder spaces $C^{k,\alpha }(\Delta , \cc^n)$. $C^{k,\alpha }(\Delta , \cc^n)$ is the space of $f\in C^k(\Delta )$ such that $$\Vert f\Vert_{k,\alpha }:=
\Vert f(x)\Vert_{C^k} + \sup_{x\not= y}{\Vert D^kf(x)-D^kf(y)\Vert
\over \vert x-y\vert^{\alpha }}<\infty.$$ Where $D^kf$ denotes the vector of derivatives of $f$ of order $k$.
One has the following important
Sobolev Imbeddings.
**
\(a) For $1\le p<2$ and $1\le q\le {2p\over 2-p}$
$$L^{k,p}(\Delta , \cc^n) \subset L^{k-1,q}(\Delta , \cc^n); \eqno(2.1.1)$$
\(b) for $2<p\le \infty $ and $0\le \alpha \le 1-{2\over p}$
$$L^{k,p}(\Delta , \cc^n) \subset C^{k-1,\alpha }(\Delta , \cc^n).
\eqno(2.1.2)$$ Moreover, the imbedding (2.1.1) is a bounded operator and imbedding (2.1.2) is compact.
Existence of such imbeddings means, in particular, the existence of universal constants $C_q$ and $C_{\alpha }$ s.t. $\Vert f\Vert_{L^{k-1,q}(
\Delta )}\le C_q\cdot \Vert f\Vert_{L^{k,p}(\Delta )}$ and $\Vert f\Vert_{C^{k-1,\alpha }(\Delta )}\le C_{\alpha }\cdot
\Vert f\Vert_{L^{k,p}(\Delta )}$. This is the most frequently used form of Sobolev Imbedding Theorem.
In the Schwartz spaces $\cals (\cc ,\cc^n)$ and $\cals'(\cc ,\cc^n)$ consider the operators $\d ={\d \over \d z}$, $\dbar = {\d \over \d \bar z}$ and $T=T_{CG}={1 \over 2\pi iz}*(\cdot )$, $\bar T={1 \over 2\pi i\bar z}*
(\cdot )$. Note that Cauchy-Green operators $T$ and $\bar T$ act from $\cals $ only to $\cals'$. Nevertheless one has the following identities on $\cals $, on $L^p(\cc )$ and on $\cals'$: $$\dbar \scirc T=T\scirc \dbar = \id , \eqno(2.1.3)$$ and $$\d \scirc \bar T=\bar T\scirc \d = \id .$$
Recall also the
Calderon-Zygmund Inequality. [*For all $1<p<\infty $ there is a constant $C_p$ such that for all $f\in L^p(\cc ,\cc^n )$ $$\Vert (\d \scirc T)(f)\Vert_{L^p}\le C_p\cdot \Vert f\Vert_{L^p}
\eqno(2.1.4)$$ and $$\Vert (\dbar \scirc \bar T)(f)\Vert_{L^p}\le C_p\cdot
\Vert f\Vert_{L^p}.$$* ]{}
This implies that for $f\in L^p(\cc ,\cc^n )$ and $g=Tf$ (or $g=\bar Tf$) one has $g\in L^p_\loc(\cc ,\cc^n )$ and $\Vert dg\Vert_{L^p}(\cc ,\cc^n )\le (1+C_p)
\Vert f\Vert_{L^p(\cc ,\cc^n)}$. For the proof of (2.1.4) we refer to \[Mc-Sa\].
Properties (2.1.3) and (2.1.4) also imply that the Cauchy-Green operator $T$ is a bounded linear operator from $L^{k,p}(\Omega , \cc^{n})$ to $L^{k+1,p}(\Omega , \cc^{n})$ if $\Omega \comp \cc $. The same is true in Hölder spaces $C^{k,\alpha }$.
We shall repeteadly use
Hölder Inequality.
*Let $p,q>1$ and let ${1\over p}+{1\over q}=
{1\over r}$. Then for all $f\in L^p(\Delta )$, $g\in L^q(\Delta )$ we have that $fg\in L^r(\Delta )$ and*
$$\Vert fg\Vert_{L^r(\Delta )}\le \Vert f\Vert_{L^p(\Delta )}\cdot
\Vert g\Vert_{L^q(\Delta )}.\eqno(2.1.5)$$
The behavior of $L^p$-norms under dilatations is also frequently used.
Lemma 2.1.1.
*Let $h\in L^p(\Delta )$, $\tau >0$ and $\pi_{\tau }:
\Delta\to \Delta $ denotes the contraction $\pi_{\tau }:z\to \tau z$. Put $\pi_{\tau }^*h(z)=h(\pi_{\tau }(z))=h(\tau z)$ and $\Delta_{\tau }=
\pi_{\tau }(\Delta )=\{ z\in \cc : \vert z\vert <\tau \} $. Then*
$$\Vert \pi_{\tau }^*h\Vert _{L^p(\Delta )} = \tau^{-{2\over p }}\cdot
\Vert h\Vert_{L^p(\Delta_{\tau })}.\eqno(2.1.6)$$
Proof. [ $$\Vert \pi_{\tau }^*h\Vert _{L^p(\Delta )} = \bigl(\int_{\Delta }\vert h(\tau z)
\vert^pdz\wedge d\bar z\bigr)^{1\over p} = \bigl(\int_{\Delta_{\tau } }
\vert h(w)\vert^p{1\over \tau^{2}}dw\wedge d\bar w\bigr)^{1\over p} =
\tau^{-{2\over p}}\cdot \Vert h\Vert_{L^p(\Delta_{\tau })},$$ where $w=\tau z$. ]{} [2.2. Local Existence of Curves]{}
Proposition 2.2.1. [*(Local existence of curves)*]{} [*Let $(X,J)$ be an almost-complex manifold with $J$ of class $C^{k,\alpha }, k\ge 1, \alpha
>0$, and $x_0\in
X$. Then for every $v\in T_{x_0}X$ small enough there exists a $J$-complex curve $u:(\Delta ,0)\to (X,x_0)$ such that $du(0)({\d \over \d x})=v$.* ]{}
Proof.
Take a chart $B\ni x_0$ with $x_0=0$ and $J(0)=J\st$. Denote by $z=x+iy$ the coordinate in $\Delta $, and by $u_1,...,u_{2n}$ the coordinates in $B$. The Cauchy-Riemann equation for $u:(\Delta ,J_{\Delta })
\to (B,J)$ has, in our local coordinates, the form $${\d u\over \d y} + J(u)du(J_{\Delta }{\d \over \d y}) = 0,$$ where $J_{\Delta }$ denotes the canonical a.-c. structure in $\Delta $, , a multiplication by $i$. Using the fact that $J_{\Delta }({\d \over \d y})=-{\d \over
\d x}$, one obtains $${\d u\over \d y} - J(u){\d u\over \d x} = 0.$$ This is equivalent to $${\d u\over \d \bar z} - Q(u){\d u\over \d z} = 0, \eqno(2.2.1)$$ where $Q(u)=[J\st + J(u)]^{-1}\scirc [J\st - J(u)]$.
After rescaling, , considering $Q_t(u):=Q(tu)$ and $u_{t,\tau }(z)
:= t^{-1}u(\tau z)$, we can assume that $\Vert Q\Vert_{C^1}$ is sufficiently small. Note also, that $(2.2.1)$ is equivalent to the holomorphicity of $h=(\id - T[Q(u){\d \over \d z}])u$.
Consider a $C^1$-mapping $$\Phi :]-1,1[ \times C^{1,\alpha }(\Delta ,B)\to C^{1,\alpha }(\Delta ,\cc^n)$$ given by $$\Phi (t,u) = (\id - T[Q(tu){\d \over \d z}])u. \eqno(2.2.2)$$ Note that $\Phi (0,u) = \id (u) = u$. Thus, the Implicit Function Theorem tells us that there exists a $t_0>0$, such that for $\vert t\vert <t_0$ the map $\Phi (t,\cdot )$ is a $C^1$-diffeomorphism of the neighborhood of zero in $C^{1,\alpha }(\Delta ,B)$ onto the neighborhood of zero $V$ in $C^{1,\alpha }(\Delta , \cc^n)$.
For $w\in \cc^n$ small enough, a holomorphic function $h_w(z):=z\cdot w$ belongs to $V$. Put $u_{t,w}=\Phi (t,\cdot )^{-1}(h_w)$. Then $tu_{t,w}$ is $J$-holomorphic. In fact, $h_w=\Phi (t,\cdot )[{1\over t}u_{t,w}]=
{1\over t}[\id - TQ(u_{t,w}){\d \over \d z}]u_{t,w}$. Moreover $u_{0,w}=
h_w$, so than $du_{0,w}(0)=w$-linear map from $\cc $ to $\cc^n$. This shows that for $t>0$ small enough $w\to du_{t,w}(0)({\d \over \d x})$ is a diffeomorphism between the neighborhoods of zero in $\cc^n$.
We immediately obtain the following
Corollary 2.2.2. [*Every almost complex structure on a Riemann surface is complex.*]{}
[2.3. Generalized Calderon-Zygmund Inequality. ]{}
Consider a continuous linear complex structure $J(z)$ on the trivial bundle $\cc \times \rr^{2n}\to \cc $, , $J(z)$ is a continuous family of endomorphisms $\rr^{2n}\to \rr^{2n}$ with $J(z)^2=-Id$. Define an operator $\dbar_{J}:\cals' (\cc ,\rr^{2n})\to \cals'(\cc ,\rr^{2n})$ by the formula $$(\dbar_{J}f)(z)={1\over 2}[\d_xf(z) + J(z)\d_yf(z)].$$ If $J_0$ is another continuous complex structure on $\cc \times \rr^{2n}$, then for $f\in L^p(\cc ,\rr^{2n})$ it holds that $$\Vert (\dbar_J\scirc T - \dbar_{J_0}\scirc T)f\Vert_{L^p(\cc )}\le
\Vert J - J_0\Vert_{L^{\infty }(\cc )}\cdot \Vert d(Tf)\Vert_{L^p(\cc )}
\le$$ $$\le \Vert J - J_0\Vert_{L^{\infty }(\cc )}(1+C_p)\Vert f\Vert_{L^p(\cc )}.
\eqno(2.3.1)$$ If we take $J_0(z)\equiv J\st $, the standard structure in $\cc^n$, then as was remarked above, $\dbar_{J_0}\scirc T: L^p(\cc ,\cc^n)\to L^p(\cc,\cc^n)$ is an identity. So from (2.3.1) we see that there exists $\eps_p={1\over 1+C_p}$ such that if $\Vert J - J\st \Vert <\eps_p$, then $\dbar_{J}\scirc T: L^p(\cc,
\cc^n)\to L^p(\cc , \cc^n)$ is an isomorphism. Moreover, since $\dbar_J\scirc T = \dbar_{J\st}\scirc T + (\dbar_J-\dbar_{J\st})\scirc T$, we have $$(\dbar_J\scirc T)\inv = (\id + (\dbar_J-\dbar_{J\st})\scirc T)^{-1} =
\Sigma_{n=0}^{\infty }(-1)^n[(\dbar_J -
\dbar_{J\st})\scirc T]^n.\eqno(2.3.2)$$ This shows, in particular, that $(\dbar_J\scirc T)\inv $ does not depend on $p>1$. Now we shall prove the following statement, which can be viewed as a generalization of the Calderon-Zygmund estimate.
Lemma 2.3.1. [*For any $u\in L^{1,2}(\cc ,\rr^{2n})$ with compact support, any continuous $J$ with $\Vert J-J\st \Vert_{L^{\infty }(\cc ,End(\rr^{2n}))}<
\eps_p$ the condition $\dbar_Ju\in L^p(\cc ,\rr^{2n})$ implies $$\Vert du \Vert_{L^p(\cc )}\le C\cdot \Vert \dbar_Ju\Vert_{L^p(\cc )}
\eqno(2.3.3)$$ with some $C=C(p,\Vert J-J\st \Vert_{L^{\infty }(\cc )})$.* ]{}
Proof. Put $v=u-T\scirc \dbar_{J\st}u $. Then $\dbar_{J\st}v=0$. So $v$ is holomorphic and descends at infinity. Thus $v=0$, which implies $u=(T\scirc \dbar_{J\st})u$. By the Calderon-Zygmund inequality, to estimate $\Vert du\Vert_{L^p(\cc )}$ it is sufficient to estimate $\Vert \dbar_{J\st}u
\Vert_{L^p(\cc )}$. We have $(\dbar_J\scirc T)\scirc \dbar_{J\st}u =
\dbar_Ju
\subset L^p(\cc )\cap L^2(\cc )$. From (2.3.2) we obtain $\dbar_{J\st}u\in
L^p(\cc )\cap L^2(\cc )$ with an estimate $$\Vert \dbar_{J\st }u\Vert_{L^p(\cc )}\le \sum_{n=0}^{\infty }\Vert
(\dbar_J-\dbar_{J\st})\scirc T\Vert_p^n\cdot \Vert \dbar_Ju\Vert_{L^p(\cc )}
\le
C\cdot \Vert \dbar_Ju\Vert_{L^p(\cc )},$$ where $C=C(p,\Vert J-J\st\Vert_{L^{\infty }})=\sum_{n=1}^{\infty }
\left(\vert J-J\st\vert_{L^{\infty }}(1+C_p)\right)^n$, provided that $\Vert J-J\st\Vert_{L^{\infty }}<\eps_p={1\over 1+C_p}$.
This yields (2.3.3).
Corollary 2.3.2. [*If $u\in (L^{1,2}\cap C^0)(\Delta ,\rr^n)$ is $J$-holomorphic for a continuous a.-c. structure $J$ belongs to $L^{1,p}_\loc
(\Delta ,\rr^2)$ for all $2<p<\infty $.* ]{}
[2.4. First A priori Estimate.]{}
Let $(X,J_0)$ be an almost complex manifold, and the tensor $J_0$ is supposed to be of class $C^0$ (continuous) only. Let $K\comp X$ be some compact in $X$. We fix some Riemannian metric $h$ on $X$ and denote by $\mu (J_0)$ the module of continuity of $J_0$. All norms and distances are taken with respect to $h$.
Recall that for a map $u:X\to Y$ into a metric space an oscillation of $u$ on a subset $D\subset X$ is defined as $\osc (D,u):=\sup \{ d_Y(u(x),u(y)):
x,y\in D\} $.
Lemma 2.4.1.
*(First a priori estimate). For every $2< p<\infty $ there exists an $\eps_1 =\eps_1(p,\mu (J_0),K,h)$ and $C_p=C(p,\mu (J_0),K,h)$ such that for any $J\in C^0$, $\vert\vert J - J_0\vert\vert_{C^0(K)}<\eps_1 $ and every $J$-holomorphic map $u\in C^0\cap L^{1,2}(\Delta ,X)$ with $u(\Delta )\subset K$, satisfying $\Vert du\Vert_{L^2(
\Delta )}<\eps_1 $ the following holds $$\Vert du\Vert_{L^p({1\over 2}\Delta )}\le C_p\cdot
\Vert du\Vert_{L^2(\Delta )}.\eqno(2.4.1)$$*
Proof.
This will be done in several steps.
*Step 1. We prove first the inequality (2.4.1) for the case when $K\comp U\subset \cc^n$, $h$ Euclidean metric, $J_0$ the standard complex structure in $\cc^n=\rr^{2n}$, and $\Vert J-
J_0\Vert_{L^{\infty }}< \eps_p$ for $\eps_p$ from [*Lemma 2.3.1*]{}.*
Note that in this case the condition that $\Vert du\Vert_{L^2(\Delta )}$ should be small is not needed.
To prove the first step consider a $J$-holomorphic map $u:\Delta \to (\rr^{2n},J)$, $u(\Delta )\subset K$ and $\Vert J-J\st \Vert
<\eps_p$. Define on $\Delta \times \rr^{2n}$ a linear complex structure $J(z)=(u^*J)(z)$. Then $u$ defines a $J$-holomorphic section of $(\Delta \times \rr^{2n}, J)$ with $\Vert J-J\st \Vert_{L^{\infty } (\Delta )}
< \eps_p$. Extend $J$ onto $\cc \times \rr^{2n}$ with the same estimate.
Let $\psi $ be a non-negative cut-off function supported in $\Delta_{3/4}$ and equal to one on $\Delta_{1/2}$. Put $u_1=u\psi $. Then $u_1\in
L^{1,2}(\Delta)$ (because $u\in L^{1,2}(\Delta)$) and $\dbar_Ju_1=u\dbar_J\psi
\in L^p(\cc )$ with $\Vert \dbar_Ju_1\Vert_{L^p(\Delta )}=\Vert u\dbar_J\psi
\Vert_{L^p(\Delta )} \le C\Vert du\Vert_{L^2(\Delta )}$ for any $p$ by the Sobolev imbedding $L^{1,2}(\Delta ,\cc )\to L^p(\Delta ,\cc )$ for all $p<\infty $. Now the generalized Calderon-Zygmund estimate (2.3.3) applies in obtaining an estimate of *Step 1.*
Using the Sobolev imbedding $L^{1,p}\subset C^{1-{2\over p}}$ and obvious properties of $L^p$-norms with respect to dilatations, one easily derives from [*Step 1*]{} the following
*Step 2. *Fix $2<p<\infty $. There exists $\eps_2=\eps_2(p,\mu (J_0),K,h)>
0$ such that for every $J$ with $\Vert J-J_0\Vert_{L^{\infty }}<\eps_p$ and any $J$-holomorphic $u$ with $\osc(u,\Delta (x,r))\le \eps_2$ on any $\Delta (x,r)\subset \Delta$ it holds that $$\osc(u,\Delta (x,{r\over 2}))\le Cr^{{2\over p}-1}\cdot
\vert \vert du\vert \vert_{L^2(\Delta (x,r))},\eqno(2.4.2)$$**
and $$\Vert du\Vert_{L^p(\Delta (x,r/2))}\le C_p\cdot
\Vert du\Vert_{L^2(\Delta (x,r))}.$$
In fact, $$\osc(u,\Delta (x,{r\over 2}))\le C_1\vert \vert du\vert \vert_
{L^p(\Delta (x,r/2))}=C(\int_{\Delta (0,{r\over 2})}\Vert du\Vert^pdxdy)^{1
\over p}=$$ $$=C(\int_{\Delta (0,{1\over 2})}\Vert du(rw){1\over r}\Vert^p)^{1\over p}=
d(rw)d\overline{(rw)})^{1\over p}=r^{2-p\over p}\Vert du_r\Vert_{L^p(\Delta
(0,{1\over 2}))}$$ for $u_r(w)=u(rw)$.
Furthermore, if $\eps_2>0$ is such that $\Vert J(x) - J(y)\Vert
<\eps_p$ for $h(x,y)<\eps_2$, $x,y\in K$, then applying Step 1 to $u_r$ we get
$$r^{2-p\over p}\Vert du_r\Vert_{L^p(\Delta (0,{1\over 2}))}\le Cr^{2-p\over p}
\Vert du_r\Vert_{L^2(\Delta (0,1))} =
Cr^{2-p\over p}\Vert du\Vert_{L^2(\Delta (0,r))}.$$
Let $\alpha \ge 0$ be the following continuous function: $\alpha \equiv 1$ for $t\le {1\over 2}$ and $\alpha (t)\equiv 0$ for $t\ge 3/4$. On the interval ${1\over 2}\le t\le {3\over 4}, \alpha (t)=3-4t$. Put for $x\in \Delta $
$$f(x):=\max \{ t: 0\le t \le {1\over 8}, \osc(u,\bar\Delta(x,t\cdot \alpha
(\vert x\vert )))\le \eps_2\}.$$
Clearly $f$ is continuous and $f\equiv {1\over 8}$ for $ {3\over 4}\le
\vert x \vert <1$.
*Step 3. *There exists an $\eps_1=\eps_1(p,J_0,K,h)>0$ such that $f(x)\equiv {1\over 8}$.**
Suppose that there is an $x_0$ with $f(x_0)=\min \{ f(x):x\in \Delta \}
< {1\over 8} $. It is clear that $f(x_0)>0$.
Take the disc $\Delta (x_0,a), a:=f(x_0)\alpha (\vert x_0\vert )$. Note that $$\osc(u,\Delta (x_0,a))=\eps_2\eqno(2.4.3)$$ Condition (2.4.2) together with the Sobolev embedding $L^{1,4}(\Delta )\subset C^{0,{1\over 2}}(\Delta )$ tells us (because $\osc(u,\Delta (x_0,a))=\eps_2$) that $\osc(u,\Delta (x_0,{a\over 2}))$ $\le C\cdot \vert \vert du\vert \vert_{L^2(\Delta (x_0,a))}$. Take a point $x_1 \in \Delta (x_0,a)$ on the distance not more then ${3\over 4}a$ from $x_0$. We have from $f(x_0)={a\over \alpha (\vert x_0\vert )}$ that $f(x_1)\ge {a\over \alpha (\vert x_0\vert )}$ and thus $f(x_1)\alpha (\vert x_0\vert )\ge a$. At the same time $\alpha (\vert
x_1\vert )\ge \alpha (\vert x_0\vert ) - 3a$, so $f(x_1)\alpha (\vert x_1
\vert )\ge a-3a\cdot f(x_1)\ge {a\over 2}$, because $f(x_1)\le {1\over 8}$. That means that $\osc(u,\Delta(x_1,{a\over 2}))\le \eps_2$ and thus $\osc(u,\Delta(x_1,{a\over 4}))\le C\cdot \vert \vert du\vert \vert_{L^2
(\Delta (x_1,{a\over 2}))}$. Thus, $\osc(u,\Delta(x_0,a))\le 4C\cdot \vert \vert du\vert \vert_{L^2(\Delta )}$. If $\eps_1 $ were taken smaller than ${\eps_2\over 4C}$, then we would obtain a contradiction with (2.4.3). Step 3 is proved.
This means that $\osc(u,\Delta(x,{1\over 8}))\le \eps_2$ for any $x\in
{1\over 2}\Delta $. Therefore, Step 2 with $r={1\over 8}$ gives us the conclusion of the Lemma.
This statement can be used to prove once more the following statement, which was already proved in Corollary 2.3.2:
Corollary 2.4.2. [*A $J$-holomorphic map $u:\Delta
\to (X,J)$ is $L^{1,p}$-continuous for any $p<\infty $, provided $J$ is of class $C^0$.* ]{}
Proof. Note that $u_{\eps }(z):=u(\eps z)$ is also $J$-holomorphic and $\Vert du_{\eps }\Vert_{L^2(\Delta (0,1))} =
\Vert du\Vert_{L^2(\Delta (0,\eps ))}$. For $\eps $ small enough we shall have $\Vert du_{\eps }\Vert_{L^2(\Delta (0,1))} =
\Vert du\Vert_{L^2(\Delta (0,\eps ))}<\eps_1 $ - from Lemma 2.4.1. Now estimate (2.4.1) gives us $L^{1,p}$-continuity of $u_{\eps }$ and thus of $u$ in the neighborhood of zero.
2.5. Convergence outside of a finite set of points.
Another immediate consequence of the a priori estimate (2.4.1) is the following
Corollary 2.5.1. [*Let $\{ J_n\}$ be a sequence of continuous almost complex structures on $X$ such that $J_n\to J$ in $C^0$-topology on $X$. Let $u_n\in C^0\cap L^{1,2}_{loc }(\Delta ,X)$ be a sequence of $J_n$-holomorphic maps such that $\Vert du_n\Vert_{L^2(\Delta )}\le \eps_1$. Then there exists a subsequence $u_{n_k}$ which $L^{1,p}_\loc$-converge to a $J$-holomorphic map $u_\infty$ for all $p\ge 2$.*]{}
Proof. The main estimate together with the Sobolev imbedding $L^{1,p}
\subset C^{1-{2\over p}}$ gives us a subsequence $\{ u_{n_k}\} $ which converges to $u_\infty$ in $C^{\alpha }(\Delta )$ for any $\alpha
<1$. Take $\phi \in C_0^{\infty }(\Delta )$, $\phi\mid_{\Delta (r)}
\equiv 1$ and consider $\dbar_{J_{n_k}}(\phi u_{n_k})=\d_x(\phi u_{n_k})
+ J_{n_k}(u_{n_k})\d_y(\phi u_{n_k})=$ $(\d_x\phi +J_{n_k}(u_{n_k})
\d_y\phi)u_{n_k}$, which is $C^0$ and thus $L^p$-convergent for any $p\ge 2$. The a priori estimate for $\dbar_J$-operator (Calderon-Zygmund inequality 2.3.1) gives us $L^{1,p}_\loc$-convergency of $\{ u_{n_k}\} $.
This corollary implies another one:
Corollary 2.5.2. [*Let $\{ J_n\}$ be a sequence of continuous almost-complex structures on $X$ such that $J_n\to J\in C^0$ in $C^0$-topology on $X$. Let $u_n\in C^0\cap L^{1,2}_{loc }(\Delta ,X)$ be a sequence of $J_n$- holomorphic maps such that $\Vert du_n\Vert_{L^2(\Delta )}\le C$. Then there exists a subsequence $\{ u_{n_k}\} $ and a finite set of points $\{ x_1,...,x_l\} $ in $\Delta $, such that $\{ u_{n_k}\} $ is $L^{1,p}_\loc$-convergent on compacts in $\Delta \setminus \{ x_1,...,x_l\} $ for all $p\ge 2$. Moreover $l\le {3C\over \eps_1}$.* ]{}
Proof. For any $n\in \nn $, cover our disk $\Delta $ by a finite number of disks of radii ${1\over n}$ in such a way that no point in $\Delta $ belongs at the same time to more than $3$ of them. Then there is at most $[{3C\over \eps_1}$ disks, on which the energy of our maps (after going to a subsequence) is more then $\eps_1$. Thus, on the complement to those disks the [*Corollary 2.5.1*]{} applies.
Taking $n$ bigger and bigger (and passing to a subsequence) we obtain the desired result.
One can prove also the following regularity statements about $j$-holomorphic maps:
Lemma 2.5.3.
*a) If the structure $J$ is Lipschitz then $J$-holomorphic maps are of class $C^1$;*
2\) If $J\in C^k$ then $J$-holomorphic maps are in $C^{k,\alpha }$ for all $\alpha <1$.
[Lecture 3]{}
[Positivity of Intersections of Complex Curves.]{}
*3.1. Unique Continuation Lemma.*
We start with a unique continuation-type lemma for $\dbar$-unequalities, namely with [*Lemma 3.1.1*]{} (compare with \[Ar\] and \[Hr-W\]). In the proof we shall use the following special version of the theorem of Harvey and Polking \[Ha-Po\]:
Theorem of Harvey-Polking. [*Let $f:\Delta
\to \cc^n$ be locally $L^2$-integrable. Assume that for some $g \in
L^1_\loc (\Delta, \cc^n)$ the equation $\dbar f =g$ holds (in the weak sence) in the punctured disc $\check\Delta$. Then $\dbar f =g$ holds in the whole disc $\Delta$.* ]{}
Lemma 3.1.1.
*Suppose that the function $u\in L^2_\loc(\Delta,
\cc^n)$ is not identically $0$, $\dbar u\in L^1_\loc(\Delta,\cc^n)$ and satisfies [*a.e.*]{} the inequality $$\vert \dbar u \vert \le h\cdot \vert u \vert\eqno(3.1.1)$$ for some nonnegative $h\in L^p_\loc(\Delta)$ with $2<p<\infty$. Then*
$u\in L^{1, p}_\loc(\Delta)$, in particular $u\in C^{0,\alpha}_\loc
(\Delta)$ with $\alpha\deff1-{2\over p}$;
for any $z_0\in\Delta$ such that $u(z_0)=0$ there exists $\mu\in\nn$—the multiplicity of zero of $u$ in $z_0$—such that $u(z)=(z-z_0)^\mu \cdot g(z)$ for some $g\in L^{1, p}_\loc (\Delta)$ with $g(z_0)\not=0$.
Proof. Statement is easily obtained by increasing the smoothness argument. Let us for the methodological reasons give it in full details.
*Step 0. $u$ is in $L^{p_0}_\loc$ for $p_0:=p>2$.*
From (3.1.1) and the Hölder inequality we see that $h\vert u\vert \in L_\loc^{{1\over {1\over 2}+{1\over p}}}=
L^{{2p\over p+2}}_\loc$. Therefore one obtains $\bar\partial u\in L^{2p\over p+2}_\loc(\Delta)$. Consequently, $u\in L^{1, {2p\over p+2}}_\loc(\Delta)$ due to ellipticity of $\dbar$ and because ${2p\over p+2} >1$. By the Sobolev imbedding we have that $u\in L_\loc^{{2{2p\over p+2}\over 2-{2p\over p+2}}}=L^p_\loc(\Delta)$.
Thus we proved that $u\in L^{p_0}_\loc$ for $p_0:=p>2$.
*Step 1. $u$ is in $L^{p_1}_\loc$ for $p_1:={2p\over 4-p}>p_0$.*
Again, $\vert \dbar u\vert \le h\vert u\vert\in
L_\loc^{{1\over {1\over p}+{1\over p}}}=L_\loc^{{p\over 2}}$. Therefore $u\in L_\loc^{1,{p\over 2}}\subset L_\loc^{{p\over 2-{p\over 2}}}=
L_\loc^{{2p\over 4-p}}$.
*Step n. There exists $r>1$ such that $u\in L_\loc^{p_n}$, where $p_n\ge rp_{n-1}$ for all $n$.*
$u\in L^{p_{n-1}}$ implies $\vert \dbar u\vert \le h\vert u\vert
\in L^{{1\over {1\over p}+{1\over p_{n-1}}}}=L^{{pp_{n-1}\over p+p_{n-1}}}$. Therefore $u\in L^{1,{pp_{n-1}\over p+p_{n-1}}}\subset
L^{{2pp_{n-1}\over 2p+2p_{n-1}-pp_{n-1}}}=L^{p_n}$ with $p_n={1\over {1\over p}
+{1\over p_{n-1}}-{1\over 2}}$. One easily checks that ${p_n\over p_{n-1}}\ge
r>1$, which does not depend on $n$.
*Step $\infty $. $u$ is in $C^{0,\alpha }$ with $\alpha =1-{2\over p}$.*
When $p_n\to \infty $ $\vert \dbar u\vert \le h\vert u\vert \in
L^{{1\over {1\over p}+{1\over p_n}}}=L^q$, where $q\sim p$, because ${1\over
p_n}\sim 0$. Therefore $u\in L^{1,q}$ with $q>2$ close to $p$. Sobolev Imbedding tells in this case that $L^{1,q}\subset C^{0,\beta }$ with $\beta
=1-{1\over q}$. Finally we see that $\dbar u\in L^p$, therefore $u\in C^{0,
\alpha }$.
Now suppose that $u(z_0)=0$. Then, due to the Hölder continuity, we have $\vert u(z) \vert \le C\vert z-z_0 \vert^\alpha$ for $z$ close enough to $z_0$ and consequently $u_1(z) \deff u(z) / (z-z_0)$ is from $L^2_\loc(\Delta)
$. The theorem of Harvey-Polking provides that $\dbar u_1\in
L^1_\loc(\Delta)$ and $u_1$ also satisfies inequality (3.1.1). In particular, $u_1$ is also continuous. Iteration of this procedure gives the possibility of defining the multiplicity of zero of $u$ in $z_0$ provided we show that after a finite number of steps we obtain the function $u_N$ with $u_N(z_0)\not= 0$. To do this we may assume that $z_0=0$. Let $\pi_\tau (z)\deff\tau\cdot z$ for $0<\tau<1$. Then $u_\tau\deff
\pi_\tau^*(u)$ satisfies the inequality $\vert \dbar u_\tau \vert \le
\tau\pi_\tau^*h \cdot \vert u_\tau \vert$. Since $$\Vert \tau\pi_\tau^*h \Vert_{L^p(\Delta)} =
\tau^{1-2/p}\cdot \Vert h \Vert_{L^p(\pi_\tau(\Delta)),}$$ we can also assume that $\Vert h \Vert_{L^p(\Delta)}$ is small enough. Fix a cut-off function $\varphi\in C^{\infty }_0(\Delta)$ which is identically $1$ in ${1\over2}\Delta$, the disk of radius $1\over2$. Then $$\Vert \varphi u \Vert_{L^p(\Delta)}\le
C_1\cdot \Vert \varphi u \Vert_{L^{1, {2p\over2+p}}(\Delta)}\le
C_2\cdot \Vert \dbar(\varphi u) \Vert_{L^{2p\over2+p}(\Delta)}\le$$ $$\le C_2\cdot (\Vert \varphi \dbar u \Vert_{L^{2p\over2+p}(\Delta)}+
\Vert \dbar\varphi u \Vert_{L^{2p\over2+p}(\Delta)})\le
%%$$ $$\le
C_2\cdot ( \Vert \varphi h u \Vert_{L^{2p\over2+p}(\Delta)}
+ \Vert \dbar\varphi u \Vert_{L^{2p\over2+p}(\Delta)} \le$$ $$\le C_3(\Vert \varphi u \Vert_{L^p(\Delta)}\cdot \Vert h \Vert _
{L^p(\Delta )} + \Vert u\Vert _{L^p(\Delta \setminus {1\over2}\Delta )}
\cdot \Vert \dbar\varphi \Vert _{L^2(\Delta \setminus {1\over2}\Delta )}).$$
Here we used the fact that the support of $\dbar\varphi$ lies in $\Delta\bss{1\over2}\Delta$. Since $\Vert h \Vert_{L^p(\Delta)}$ is small enough we obtain the estimate $$\Vert u \Vert_{L^p({1\over3}\Delta)}
\le C\cdot \Vert u \Vert_{L^p(\Delta\setminus {1\over2}\Delta)}$$ with the constant $C$ independent of $u$. Thus if the multiplicity of zero of $u$ in $z_0=0$ is at least $\mu$, then $$\Vert z^{-\mu}u \Vert_{L^p({1\over3}\Delta)}
\le C\cdot \Vert z^{-\mu}u \Vert_{L^p(\Delta\bss{1\over2}\Delta))}$$ which easily gives $$\Vert u \Vert_{L^p({1\over3}\Delta)} \le C
\left(\hbox{$2\over3$}\right)^{-\mu}
\cdot \Vert u \Vert_{L^p(\Delta\bss{1\over2}\Delta))}.$$ Now one can easily see that either $u$ has isolated zeros of finite multiplicity, or $u$ is identically zero. Yet the last case is excluded by the hypothesis of the lemma.
Lemma 3.1.2. [*Under the hypothesis of [*Lemma 3.1.1*]{} suppose additionally that $u$ satisfies [*a.e.*]{} the inequality $$\vert \dbar u(z) \vert \le \vert z-z_0 \vert^\nu
h(z)\cdot \vert u(z) \vert,\eqno(3.1.2)$$ with $z_0\in \Delta$, $\nu\in\nn$, and $h\in L^p_\loc(\Delta)$, $2<p<\infty$. Then $$u(z)=(z-z_0)^\mu\bigl(P^{(\nu)}(z) + (z-z_0)^\nu g(z)\bigr),
\eqno(3.1.3)$$ where $\mu\in\nn$ is the multiplicity of $u$ in $z_0$, defined above, $P^{(\nu)}$ is a polynomial in $z$ of degree $\le \nu$ with $P^{(\nu)}(z_0)
\not=0$, $g\in L^{1, p}_\loc(\Delta,\cc^n)\hookrightarrow C^{0,\alpha}$, $\alpha=1-{2\over p}$, and $g(z)=O(\vert z-z_0\vert^\alpha)$.* ]{}
**Proof. The proof uses the same idea as in the previous lemma. Define $u_0(z)
\deff \msmall{u(z)\over (z-z_0)^\mu}$ and $h_1(z) \deff h(z)\cdot \vert u_0(z)
\vert$. Due to [*Lemma 3.1.1*]{}, $u_0\in C^{0,\alpha}$, $u_0(z_0)\not=0$, $h_1\in L^p_\loc$, and $u_0$ satisfies [*a.e.*]{} the inequality $\vert \dbar
u_0(z) \vert \le \vert z-z_0 \vert^\nu h_1(z)$.**
Set $a_0=u_0(z_0)$. Since $u_0(z)-a_0 = O(\vert z-z_0 \vert^\alpha)$, we have $u_1\deff\msmall{u_0(z) - a_0 \over z-z_0}\in L^2_\loc$. Applying the theorem of Harvey-Polking once more, we obtain $\vert \dbar u_1 \vert
\le \vert z-z_0 \vert^{\nu-1} h_1$, and consequently $u_1\in C^{0,\alpha}$, $u_1(z)-u_1(z_0)=O(\vert z-z_0 \vert^\alpha)$. Repeating this procedure $\nu$ times, we obtain the polynomial $$P^{(\nu)}(z)=a_0 + (z-z_0)a_1 + \cdots + (z-z_0)^\nu a_\nu$$ with $$a_k\deff\lim_{z\to z_0}
\msmall{ u(z)-\sum_{i=0}^{k-1} (z-z_0)^i a_i \over (z-z_0)^k },
\qquad 0\le k\le\nu,$$ and the function $$g(z)\deff \msmall{{u(z)-P^{(\nu)}(z) \over (z-z_0)^\nu }},$$ which satisfies the conclusion of the lemma.
Corollary 3.1.3. [*Let $J$ be a Lipschitz-continuous almost complex structure in a neighborhood $U$ of $0\in \cc^n$ such that $J(0)=J\st$, the standard complex structure in $\cc^n$. Suppose that $u:\Delta\to U$ is a $J$-holomorphic $C^1$-map with $u(0)=0$. Then there exist uniquely defined $\mu\in\nn$ and a (holomorphic) polynomial $P^{(\mu-1)}$ of degree $\le\mu-1$ such that $u(z)=z^\mu\cdot P^{(\mu-1)} +z^{2\mu -1}v(z)$ with $v(0)=0$ and $v\in L^{1,p}(\Delta,\cc^n)$ for any $p<\infty$.* ]{}
Proof.
In fact, by the hypothesis of the lemma $du + J(u)
\scirc du \scirc J_\Delta =0$ and hence $$\vert \dbar u \vert =
\bigl\vert \msmall{1\over2}( du + J\st \scirc du \scirc J_\Delta)\bigr\vert =
\bigl\vert \msmall{1\over2} (J\st - J(u))\scirc du \scirc J_\Delta \bigr\vert
\le \Vert J \Vert_{C^{0,1}(B)} \cdot \vert u \vert \cdot
\vert du \vert.
\eqno(3.1.4)$$
Thus by [*Lemma 3.1.1*]{}, $u(z) = z^\mu w(z)$ with some $\mu \in\nn$ and $w\in L^{1,p}_\loc(\Delta, \cc^n)$ for any $p<\infty$. If $\mu =1$ we are done. Otherwise $ du(z)/z^{\mu-1} =
\mu\,w\,dz + z\,dw \in L^p_\loc(\Delta, \cc^n)$ for any $p<\infty$. Thus, the corollary follows now from [*Lemma 3.1.2*]{}.
Corollary 3.1.4. [*Let $u:S\to (X, J)$ be a $J$-holomorphic map. Then for any $p\in X$ the set $u\inv(p)$ is discrete in $S$, provided that $J$ is Lipschitz-continuous.* ]{}
Proof.
Take an [*integrable*]{} complex structure $J_1$ in some neighborhood $U$ of $p\in X$, such that $J(p)=J_1(p)$. Take also $J_1$-holomorphic coordinates $w_1,\ldots, w_n$ in $U$, such that $w_i(p)=0$. Then the statement follows easily from [*Corollary 3.1.3.* ]{}
*3.2. Inversion of $\dbar_{J}+R$ and $L^{k,p}$-topologies.*
Now suppose that $J\in C^0(\Delta, \endo_\rr(\rr^{2n}))$ satisfies the identity $J^2\equiv -1$,, $J$ is a continuous almost complex structure in the trivial $\rr^{2n}$-bundle over $\Delta$. Define the $\rr$-linear differential first order operator $\dbarj:L^{1, p}(\Delta,\rr^{2n}) \to L^p(\Delta,\rr^{2n})$ by setting $$\dbarj(f)=\msmall{1\over2}\left(
\msmall{\d f \over \d x} + J \msmall{\d f\over \d y} \right).\eqno(3.2.1)$$ For example, for $J\equiv J\st$, the standard complex structure in $\rr^{2n}=
\cc^n$, the operator $\dbarj$ is a usual Cauchy-Riemann operator $\dbar$. The operator $\dbarj$ is elliptic and possesses nice regularity properties in Sobolev spaces $L^{k, p}$ with $1<p<\infty$ and Hölder spaces $C^{k,\alpha}$ with $0<\alpha<1$. The following two statements are typical for (nonlinear) elliptic PDE and produces a result which we need for the purpose of this paper.
Lemma 3.2.1.
*Let $J$ be $C^k$-continuous with $k\ge0$ and $\dbarj$ be as above. Also let $R$ be an $\endo (\rr^{2n})$-valued function in $\Delta$ of class $L^{k,p}$ with $1<p<\infty$. If $k=0$, we also assume that $p>2$. Suppose that for $\dbarj f +Rf \in L^{k,p}$ for some $f\in L^{1,
1}(\Delta,\rr^{2n})$. Then $f\in L^{k+1,p}_\loc (\Delta,\rr^{2n})$ and for $r<1$ $$ i) f \_[L\^[k+1, p]{}((r))]{} C\_1(J, R\_[L\^[k,p]{}]{}, k, p, r) ( f + Rf \_[L\^[k, p]{}()]{} + f \_[L\^1()]{}) $$*
If, in addition, $J$ and $R$ are $C^{k,\alpha}$-smooth with $0<\alpha<1$, then $f\in C^{k+1,\alpha}_{\loc }(\Delta,\rr^{2n})$ and $$ ii) f \_[C\^[k+1,]{}((r))]{} C\_2( J , R \_[C\^[k,]{}]{},k, , r)( f + Rf \_[C\^[k,]{}()]{}+ f \_[L\^1().]{}) $$ If, additionally, $\norm{ J-J\st }_{C^k(\Delta)} + \norm{R }\Vert_{L^{k,p}(\Delta)}$ (resp., $\Vert J-J\st \Vert_{C^{k,\alpha}(\Delta)}+
\Vert R \Vert_{C^{k,\alpha}(\Delta)}$) is small enough, then there exists a linear bounded operator $T_{J, R}:L^{k,p}(\Delta,\rr^{2n}) \to L^{k+1, p}(\Delta,\rr^{2n})$ $($resp., $T_{J, R}:C^{k, \alpha}(\Delta,\rr^{2n}) \longrightarrow
C^{k+1,\alpha}(\Delta,\rr^{2n})$ such that $(\dbarj + R)\scirc T_{J, R}\equiv \id$ and $T_{J,
R}(f)\vert_{z=0}=0$.
Proof. The estimates and of the theorem are obtained from the ellipticity of $\dbarj$ by standard methods. See \[Mo\] for the case of general elliptic systems or \[Sk\] for the direct proof. We note only that for $J$ being only continuous the constant $C_1$ in crucially depends on the modulus of continuity of $J$, whereas in all the other cases of the theorem the dependence is essentially only on the corresponding norm of $J$.
Let $T$ be a composition ${\d\over\d z} \scirc G$ where $G(f)$ is the solution of the Poisson equation $\Delta u=f$ with the Dirichlet boundary condition on $\d\Delta$. Then for $J\equiv J\st$ and $R\equiv0$ we can define $T_{J\st,0}(f) \deff T(f)- T(f)(0)$. In the general case we set $$T_{J, R}=\sum_{n=0}^\infty (-1)^nT_{J\st, 0}\scirc
\bigl((\dbarj-\dbar_{J\st}+R)\scirc T_{J\st, 0}\bigr)^n.$$ Due to the Sobolev imbedding theorem, the series converges in an appropriated norm, provided $\Vert J-J\st \Vert + \Vert R \Vert$ is small enough.
Remark.
The last statement of this lemma Implies that for $\Vert J-J\st \Vert + \Vert R\Vert $ small enough $L^{k+1,p}(\Delta ,\rr^{2n})
=H_{\dbar_J+R}\oplus T_{J,R}(L^{k,p}(\Delta ,\rr^{2n}))$ where $H_{\dbar_J+R}=\ker (\dbar_J+R)$. This will be used later in the proof of Lemma 6.3.4.
Corollary 3.2.2.
*Let $k\in \nn$, $q>2$, and $J$ a $C^k$-smooth almost complex structure in $X$. Also let $(S, J_S)$ be a complex curve. Suppose that $L^{1,q}$-map $u:S\to X$ satisfies the equation $$du + J \scirc du \scirc J_S = 0.$$ Then $u$ is $L^{k+1,p}$-smooth for any $p<\infty$. If, in addition, $J$ is $C^{k,\alpha}$-smooth with $0<\alpha<1$, then $u$ is $C^{k+1,\alpha}$-smooth.*
Let $J^{(n)}$ (resp. $J^{(n)}_S$) be a sequence of almost complex structures on $X$ (resp. on $S$), which $C^k$-converges to $J$ (resp. to $J^{(n)}_S$) and let $u_n:S \to X$ be a sequence of $(J^{(n)}_S, J^{(n)})$-holomorphic maps. Then the $C^0$-convergence $u_n \longrightarrow u$ implies the $L^{k+1,p}$-convergence for any $p<\infty$, and $C^{k+1, \alpha}$-convergence if $(J^{(n)}_S, J^{(n)})$ converge to $(J_S, J)$ in $C^{k,\alpha}$, $0<\alpha<1$.
Proof. The map $u$ is continuous and in local coordinates $w_1,\ldots, w_{2n}$ on $X$ and $z=x+iy$ on $S$ the equation has the form $$du(z) + J(u(z)) \scirc du \scirc J_M = 0,$$ which is equivalent to $\dbar_{J\scirc u} u =0$. Using [*Lemma 3.2.1*]{} and induction in $k'=0\ldots k$, one can obtain the regularity of $u$.
Similarly, for $J^{(n)}$ and $u^{(n)}$ satisfying the hypothesis of the corollary one gets $$\dbar_{J\scirc u}(u^{(n)}-u)=
(\dbar_{J\scirc u} - \dbar_{J^{(n)} \scirc u^{(n)}})u^{(n)}
\longrightarrow 0 \quad\hbox{in $L^{k',p}_\loc$
(resp.\ in $C^{k',\alpha}_\loc$)}$$ by induction in $k'=0\ldots k$.
Remark. The corollary implies that for a compact Riemann surface $S$ the topology in the space $\calp$ of $(J_S, J)$-holomorphic maps $u:S\to
X$ is independent of the particular choice of the functional space $L^{k', p}(S,X)\supset \calp$ with $1\le k' \le k+1$, $1< p<\infty$, and $kp>2$, provided $J_S$ and $J$ are changing $C^k$-smoothly. In the same way, [*Lemma 3.2.1*]{} implies that for $J\in C^k$ with $k\ge1$ the differential structure on $\calp$ is also independent of the particular choice of a functional space $L^{k',p}(S,X)$ with $1\le k' \le k$, $1<p<\infty$ and $k'p>2$.
Lemma 3.2.3.
*For any natural numbers $\mu>\nu\ge0$ and real numbers $p>2$, $\alpha<{2\over p}$ and $\gamma>0$ there exists $C =
C(\mu, \nu, \alpha, p, \gamma)>0$ with the following property. Let $J$ be an almost complex structure in $B \subset \cc^n$ with $J(0)=J\st$ and let $u:\Delta \to B$ be a $J$-holomorphic map of the form $u(z)=z^\mu(P(z) +z^\nu v(z))$, where $P(z)$ is some (holomorphic) polynomial of the degree $\nu$ and $v\in L^{1,p}(\Delta, \cc^n)$ with $v(0)=0$. Suppose that $\Vert J-J\st
\Vert_{C^1(B)} \le \gamma$ and $\Vert u\Vert_{L^{1,p}(\Delta)} \le \gamma$. Then for any $0<r<{1\over2}$ $$\bigl\Vert z dv \bigr\Vert_{C^0(\Delta(r))} +
\bigl\Vert z dv \bigr\Vert_{L^{1,p}(\Delta(r))}
\le
C \cdot r^\alpha\cdot \Vert u \Vert_{L^{1,p}(\Delta)},
\eqno(3.2.2)$$*
In particular, $du(z) = d(z^\mu P(z)) + o(|z|^{\mu+ \nu -1 +\alpha})$ for any $\alpha<1$.
Proof. For $0<r<{3\over4}$ we define the map $\pi_r:B\to B$, setting $\pi_r(w) \deff r^\mu w$. We also set $J^{(r)} \deff \pi_r^*J$, $u^{(r)}(z)
\deff \pi_r^{-1} \scirc u(rz)$, $P^{(r)}(z)\deff P(rz)$ and $v^{(r)}(z)=
r^\nu v(rz)$. Then $J^{(r)}$ is an almost complex structure in $B$ with $\Vert J^{(r)} -J\st \Vert_{C^1(B)} \le r^\mu
\Vert J-J\st\Vert_{C^1(B)}$, and $u^{(r)} \equiv z^\mu (P^{(r)}(z) +
z^\nu v^{(r)}(z))$ is a $J^{(r)}$-holomorphic.
Without losing generality, we can suppose that $\alpha>0$. Set $\beta \deff
1+\alpha - {2\over p}$ and $q\deff {2\over1-\beta}$. Then $\alpha<\beta<1$, $\beta + {2\over p} -1=\alpha$, and $q>2$. [*Lemma 3.1.3*]{} implies that $\Vert v \Vert_{C^{0,\beta}(\Delta(2/3))} + \Vert dv \Vert_{L^q(\Delta(2/3))}
\le C_1 \cdot \Vert u \Vert_{L^{1,p}(\Delta)}$. Here the constant $C_1$, as well as the constants $C_2,\ldots,C_6$ below, depend only on $\mu, \nu, p, \alpha$, and $\gamma$, but are independent of $r$. Consequently, $$\Vert u^{(r)} - z^\mu P^{(r)}(z) \Vert_{L^{1,q}(\Delta)}
\le C_2\cdot r^{\nu + \beta} \cdot
\Vert u \Vert_{L^{1,p}(\Delta)}.$$ Furthermore, due to [*Corollary 3.2.2*]{}, we have $\Vert u^{(r)}
\Vert_{L^{2,p}(\Delta)} \le C_3\cdot \Vert u \Vert_{L^{1,p}(\Delta)}$. Thus $$\Vert \dbar_{\!J\st}( u^{(r)} - z^\mu P^{(r)}(z) ) \Vert_{L^{1,p}(\Delta)} =
\Vert (\dbar_{\!J\st} -\dbar_{\!J^{(r)} \scirc u^{(r)} } )
( u^{(r)}) \Vert_{L^{1,p}(\Delta)}
\le$$ $$\le C_4\cdot r^\mu \cdot
\Vert u \Vert_{L^{1,p}(\Delta)}.$$ Applying [*Lemma 3.2.1*]{}, we obtain $$\bigl\Vert z dv^{(r)} \bigr\Vert_{L^{1,p} (\Delta(2/3))} \le
C_5 \cdot r^{\nu + \beta} \cdot \Vert u \Vert_{L^{1,p}(\Delta)},
\eqno(3.2.3)$$ which is equivalent to $$\bigl\Vert z dv \bigr\Vert_{L^{1,p}(\Delta(2r/3))} \le
C_5 \cdot r^{\beta + 2/p -1} \cdot \Vert u \Vert_{L^{1,p}(\Delta)}.$$ On the other hand, $(3.2.3)$ implies that $$\bigl\Vert z dv^{(r)} \bigr\Vert_{C^0 (\Delta(2/3))} \le
C_6 \cdot r^{\nu + \beta} \cdot \Vert u \Vert_{L^{1,p}(\Delta)},$$ and consequently $$\bigl\Vert z dv \bigr\Vert_{C^0(\Delta(2r/3))} \le
C_6 \cdot r^\beta \cdot \Vert u \Vert_{L^{1,p}(\Delta)}.$$
*3.3. Perturbing Cusps of Complex Curves.*
For understanding the rest of this lecture the reader should consult §6.2 and §6.3 for the definition and some elementary properties of Gromov’s $\dbar $-operator.
Lemma 3.3.1. [*For a given $p$, $2<p<\infty$, and $\gamma>0$ there exist constants $\epsi=\epsi(p,\gamma)$ and $C=C(p,\gamma)$ with the following property. Suppose that $J$ is a $C^1$-smooth almost complex structure in the unit ball $B \subset \cc^n$ with $\Vert J-J\st \Vert_{C^1(B)}
\le \epsi$ and $u\in L^{1,p}(\Delta, B(0,{1\over2}))$ is a $J$-holomorphic map with $\Vert u \Vert_{L^{1,p}(\Delta)}
\le \gamma$. Then for every almost complex structure $\tilde J$ in $B$ with $\Vert \tilde J-J \Vert_{C^1(B)} \le \epsi$ there exists a $\tilde J$-holomorphic map $\tilde u: \Delta \to
B$ with $$\Vert \tilde u-u \Vert_{L^{1,p}(\Delta)} \le
C\cdot \Vert \tilde J-J \Vert_{C^1(B)}$$ such that $\tilde u(0) = u(0)$.* ]{}
Proof. Let $J_t$ be a curve of $C^1$-smooth almost complex structures in $B$ starting at $J_0=J$ and depending $C^1$-smoothly on $t\in [0,1]$. Consider an ordinary differential equation for $u_t \in L^{1,p} (\Delta, B)$ with the initial condition $u_0=u$ and
$$\msmall{du_t\over dt} = - T_{J_t \scirc u_t,\, R_t}
\biggl(\msmall{dJ \over dt} \scirc du_t \scirc J_\Delta \biggr),$$
where $R_t$ is defined by the relation $D_{J_t, u_t}=\dbar_{J_t, u_t}
+ R_t$. (See [*paragraph 6.2*]{}). Since $J_0$ and $R_0$ satisfy the hypothesis of [*Lemma 3.2.1*]{} and $R_t$ depends $L^p$-continuously on $J_t\in C^1$ and $u_t \in
L^{1,p}$, the solution exists in some small interval $0\le t \le
t_0$. For such a solution using (6.3.1) one has $$\left\Vert \msmall{du_t\over dt} \right\Vert_{L^{1,p}} +
\left\Vert \msmall{dR_t\over dt} \right\Vert_{L^p}
\le C\cdot (\Vert u_t\Vert_{L^{1,p}} +
\Vert J_t - J\st \Vert_{C^1} )
\cdot \left\Vert \msmall{dJ_t\over dt} \right\Vert_{C^1}.$$ This implies the existence of the solution of our equation for all $t\in [0,1]$, provided $$\int_{t=0}^1
\left\Vert \msmall{dJ_t\over dt} \right\Vert_{C^1}
dt \le \epsi.$$
In this paragraph we suppose that some $p$ with $2<p<\infty$ is fixed.
Lemma 3.3.2.
*For every $\gamma >0$ and every pair of integers $\nu\ge1$, $\mu\ge1$ there exists an $\epsi=\epsi(\mu,\nu,\gamma)>0$ such that for an almost complex structure $J_0$ in $B$ with $\Vert J_0-J\st \Vert_{C^1(B)}\le \epsi$, $J_0(0)=J\st$, and for a $J_0$-holomorphic map $u_0: \Delta \to B(0, {1\over2})$ with $\Vert u_0
\Vert_{L^{1,p}(\Delta)} \le\gamma $ and with the multiplicity $\mu$ at $0\in \Delta$ the following holds:*
If $\nu\le2\mu-1$, then for every almost complex structure $J$ in $B$ with $J(0)=J\st$ and for every $v\in L^{1,p}(\Delta,\cc^n)$ with $\Vert v \Vert_{L^{1,p}(\Delta)} + \Vert J-J\st \Vert_{C^1(\Delta)}
\le\epsi$ there exists $w\in L^{1,p}
(\Delta,\cc^n)$ with $w(0)=0$, satisfying $$\dbarj(u_0+z^\nu(v+w))=0\eqno(3.3.1)$$ and $$\Vert w \Vert_{L^{1,p} (\Delta)}\le
C\cdot \left(
\Vert v \Vert_{L^{1,p}(\Delta)} +
\Vert J-J_0 \Vert_{C^1(\Delta)}
\right).\eqno(3.3.2)$$
If $\nu\ge2\mu$, then for every $v\in L^{1,p}(\Delta,\cc^n)$ with $\Vert v \Vert_{L^{1,p}(\Delta)}\le\epsi$ there exists $w\in
L^{1,p}(\Delta,\cc^n)$ with $w(0)=0$, satisfying $$\bar\partial_{J_0}(u_0+z^\nu(v+w))=0\eqno(3.3.3)$$ and $$\Vert w \Vert_{L^{1,p} (\Delta)}\le
C\cdot \Vert v \Vert_{L^{1,p}(\Delta)}.
\eqno(3.3.4)$$
Proof. If $\nu\le 2\mu -1$, we fix a curve $J_t$, $t\in [0,1]$, of $C^1$-smooth almost complex structures in $B$, which starts in $J_0$, finishes in $J$ and satisfies the conditions $J_t(0)=J\st$ and $\int_{t=0}^1\Vert \dot J_t \Vert_{C^1(B)} dt \le 2\epsi$, where $\dot J_t \deff dJ_t/dt$. If $\nu\ge2\mu$, we simply set $J_t\equiv J_0$.
As in the previous lemma, we want to find a needed $w$ by solving for $t\in [0,1]$ and $w_t\in L^{1,p}(\Delta,\cc^n)$ the equation $$z^{-\nu}\dbar_{J_t}(u_0 + z^\nu(t\cdot v + w_t)) =0.$$ However, this time we need to consider the fact that now we are dealing with different (almost) complex structures on $B$, namely $J\st$ and $J_t$ for any fixed $t\in [0,1]$. Thus, we write the last equation in the more correct form: $$(x + y J\st)^{-\nu}\dbar_{J_t}
(u_0 + (x + y J\st)^\nu(t\cdot v + w_t)) =0.
\eqno(3.3.5)$$
The differentiation of (3.3.5) with respect to $t$ gives $$(x + y J\st)^{-\nu}D_{u_t, J_t}
((x + y J\st)^\nu(v + \dot w_t))
+(x + y J\st)^{-\nu}\dot J_t
\scirc du_t \scirc J_\Delta
=0,
\eqno(3.3.6)$$ where $u_t\deff u_0+ (x + y J\st)^\nu(t\cdot v + w_t)$ and $J\st$ also denotes the pull-back of the standard complex structure on $E\deff u_t^*TB\cong \cc^n$.
First we show that the operator $(x + y J\st)^{-\nu} \scirc
D_{u_t, J_t} \scirc (x + y J\st)^\nu$ has the form $\dbar_{J^{(\nu)}_t} + R^{(\nu)}_t$ for an appropriate (almost) complex structure $J^{(\nu)}_t$ in $E\cong \cc^n$ and $\rr$-linear operator $R^{(\nu)}_t$. In fact, the explicit formula (6.2.5) for $D_{u, J}$ shows that taking the standard connection $\nabla$ in $TB\cong \cc^n$ and identifying $\Lambda^{(0,1)}\Delta
\cong \cc$ one gets $$D_{u_t, J_t}(v)=\msmall{1\over2}\left( \msmall{\d \over \d x}v +
J_t \msmall{\d \over \d x}v \right) + R^{(0)}_t(v)$$ with $R^{(0)}_t\in C^0(\Delta, \endo_\rr(\cc^n))$. Moreover, the formula (6.3.1) implies that $$\Vert R^{(0)}_t \Vert_{L^p(\Delta)} \le \Vert J_t \Vert_{C^1(B)}
\cdot \Vert du_t \Vert_{L^p(\Delta)}.$$ Hence $$2(x + y J\st)^{-\nu} \scirc D_{u_t, J_t}
\scirc (x + y J\st)^\nu = \Bigl[\msmall{ \d \over \d x} +
(x + y J\st)^{-\nu} \scirc J_t \scirc (x + y J\st)^\nu
\msmall{\d \over \d y} \Bigr]+$$ $$+ \Bigl[ \nu \cdot(x + y J\st)^{-\nu} \scirc
(1 + J_t \scirc J\st) \scirc (x + y J\st)^{\nu -1}
+ 2(x + y J\st)^{-\nu} \scirc R^{(0)}_t \scirc (x + y J\st)^\nu
\Bigr]=$$ $$=2\dbar_{J^{(\nu)}_t} + 2\,R^{(\nu)}_t.$$ One has the obvious identities $(J^{(\nu)}_t)^2 \equiv -1$, $$\Vert J^{(\nu)}_t -J\st \Vert_{C^0(\Delta)}=
\Vert (x + y J\st)^{-\nu} \scirc (J_t -J\st)
\scirc (x + y J\st)^\nu \Vert_{C^0(\Delta)} =
\Vert J_t -J\st \Vert_{C^0(B)},$$ and $$\Vert (x + y J\st)^{-\nu} \scirc R^{(0)}_t \scirc (x + y J\st)^\nu
\Vert_{L^p(\Delta)} =\Vert R^{(0)}_t \Vert_{L^p(\Delta)} .$$ Furthermore, $1 + J_t(0) J\st=0$; hence $\bigl\Vert \> |z|^{-1}(1 +J_t J\st)
\bigr\Vert_{C^0(\Delta)} \le \Vert J_t -J\st \Vert_{C^1(B)}$. This gives the estimate $$\Vert R^{(\nu)}_t \Vert_{L^p(\Delta)} \le \bigl(C_1 \cdot \nu +
C_2 \cdot \Vert du_t \Vert_{L^p(\Delta)} \bigr) \cdot
\Vert J_t -J\st \Vert_{C^1(B)}.$$
To show the existence of the solution of (3.3.5) for all $t\in [0,1]$, we need the estimate $$\left\Vert z^{-\nu}\cdot \dot J_t (u_0 + z^\nu w)
\scirc d(u_0 + z^\nu w) \right\Vert_{L^p(\Delta)}
\le C_1(\mu, \nu) \cdot \Vert u_0 \Vert_{L^{1,p}(\Delta)}^2
\cdot \Vert \dot J_t \Vert_{C^1(B)}
\eqno(3.3.7)$$ for all sufficiently small $w\in L^{1,p}(\Delta, \cc^n)$. The estimate trivially holds for $\nu\ge 2\mu$, because in this case $\dot J_t\equiv 0$. Otherwise for $\lambda \deff \min\{\mu, \nu\}$ from [*Corollary 3.1.3.*]{} one obtains $$\left\Vert z^{-\lambda}(u_0 + z^\nu w)
\right\Vert_{L^\infty (\Delta)} +
\left\Vert z^{-\lambda +1} d(u_0 + z^\nu w)
\right\Vert_{L^p(\Delta)}
\le C_2(\mu, \nu)\cdot
\left\Vert u_0 \right\Vert_{L^{1,p}(\Delta)}
\eqno(3.3.8)$$ which gives the estimate (3.3.7).
Now consider the ordinary differential equation for $w_t \in L^{1,p}
(\Delta, B)$ with the initial condition $w_0\equiv 0$ and $$\msmall{dw_t\over dt} = - T_{J^{(\nu)}_t,\, R^{(\nu)}_t}
\Bigl(z^{-\nu}\cdot \dot J_t \scirc du_t \scirc J_\Delta + v \Bigr).
\eqno(3.3.9)$$ As in [*Lemma 3.3.1*]{} one has the estimate $$\biggl\Vert \msmall{dw_t\over dt} \biggr\Vert_{L^{1,p}(\Delta)} +
\biggl\Vert \msmall{dR_t^{(\nu)}\over dt} \biggr\Vert_{L^p(\Delta)}
\le$$ $$\le C\cdot (\Vert u_t\Vert_{L^{1,p}(\Delta)} +
\Vert J_t - J\st \Vert_{C^1(B)} )
\cdot \left( \bigl\Vert \msmall{dJ_t\over dt} \bigr\Vert_{C^1(B)}
+ \bigl\Vert v \bigr\Vert_{L^{1,p}(\Delta)} \right),$$ which implies the existence of the solution of (3.3.9) for all $t\in [0,1]$.
*3.4. Primitivity.*
We need to study first the notions of primitivity and distinctness of complex curves in $C^1$-smooth nonintegrable structures. Let us start with the following
Definition 3.4.1. Two $J$-holomorphic maps $u_1: (S_1,j_1) \to X$ and $u_2: (S_2,j_2) \to X$ with $u_1(a_1)=u_2(a_2)$ for some $a_i\in S_i$ are called [*distinct in $(a_1,a_2)$*]{} if there are no neighborhoods $U_i\subset S_i$ of $a_i$ with $u_1(U_1) =
u_2(U_2)$. We call $u_i: (S_i,j_i) \to X$ [*distinct*]{} if they are distinct for all pairs $(a_1,a_2)\in S_1 \times S_2$ with $u_1(a_1)=u_2(a_2)$.
Definition 3.4.2. A $J$-holomorphic map $u: (S,j) \to X$ is called [*primitive*]{} if there are no disjoint non-empty open sets $U_1, U_2\subset S$ with $u(U_1) =u(U_2)$. Note that $u$ must be nonconstant.
An important regularity property of a $J$ - complex curve with $J\in C^1$ is contained in the following
**Theorem 3.4.1. *Let $(S_1, j_1)$ and $(S_2, j_2)$ be smooth connected complex curves and $u_i : (S_i,j_i) \to (X,J)$ non-constant $J$-holomorphic maps with $J \in C^1$. If there are non-empty open sets $U_i \subset S_i$ with $u_1(U_1)= u_2(U_2)$, then there exists a smooth [*connected*]{} complex curve $(S, j)$ and a $J$-holomorphic map $u: (S,j) \to (X,J)$ such that $u_1(S_1) \cup u_2(S_2) =u(S)$ and $u:S \to X$ is primitive.***
Moreover, maps $u_i :S_i \to X$ factorize through $u: S \to
X$, i.e., there exist holomorphic maps $g_i : (S_i,j_i) \to (S,j)$ such that $u_i = u \scirc g_i$.
**Proof.**
[*Step 1.*]{} *For every $J$-holomorphic imbedding $u:\Delta \to (X,J)$ with $J\in C^1$ and for every $p\in \Delta$ there exist a neighborhood $U \subset X$ of $u(p)$, a $C^1$-smooth integrable complex structure $J_1$ in $U$ and $J_1$-holomorphic coordinates $(w_1, \ldots, w_n)=(w_1,w')$ ($w'=(w_2, \ldots, w_n))$ in $U$ such that $u(\Delta) \cap U = \{ (w_1, w')\in U: w'=0\}$ and $J|_{u(\Delta)} = J_1|_{u(\Delta)}$. In particular $w_1|_{u(\Delta)}$ is a holomorphic coordinate in $u\inv(U) \subset \Delta$ and*
$$|J(w_1, w')-J_1(w_1, w')| \le C(|w'|).
\eqno(3.4.1)$$
For the proof take any $J$-Hermitian metric $h$ on $X$. Let $N_1$ denote the normal bundle $u_1^*TX/ du(TS_1)$ with fiber $N_{1,z}$ over $z\in S_1$. Then the exponential map $\phi: \xi \in N_{1,z} \mapsto
\exp_{u_1(z)} (\xi)$ is a local diffeomorphism in some neighborhood $V$ of the zero section of $N_1$ over $S_1$. Note that along the zero section we can identify $\phi$ with $u_1$.
Equip $N_1$ with a canonical holomorphic structure of the quotient $u_1^*TX/
du(TS_1)$. Since $h$ was chosen $J$-Hermitian, the operator of the complex structure $J_{N_1}$ coincides along zero section with the structure $\phi^*J
\equiv u_1^*J$. Note that $J_{N_1}$ is [*integrable*]{} in $V$.
[*Step 2. (“Unique Continuation Lemma”).*]{} *Let the open subset $U \subset X$, $C^1$-smooth complex structure $J_1$ in $U$, and $J_1$-holomorphic coordinates $(w_1, \ldots, w_n)=(w_1,w')$ be as in the previous step. Set $C \deff \{ (w_1, w')\in U: w'=0\}$. Let $v: \Delta \to U$ be a $J$-holomorphic map. Then either $v(\Delta) \subset
C$ or $v\inv(C)$ is discrete.*
To show this, we use $(w_1, w')$ as complex coordinates and consider $U$ as an open subset of $\cc^n$, so that $J_1= J\st$. Set $v' \deff w' \scirc v$, i.e., $v'$ is obtained from $v$ by forgetting the first component. Then $v\inv(C)=v^{\prime\,-1}(0)$. Furthermore, $$|\dbar_{J\st}v'(z)| \le |\dbar_{J\st}v(z)| = |\dbar_{J_1}v(z)| =
|\dbar_Jv(z)-\dbar_{J_1}v(z)| \le$$ $$\le \norm{dv}_{L^\infty} \cdot
|J(v'(z))- J_1(v'(z))|,$$ which means that locally $|\dbar_{J\st}v'| \le C |v'|$. By [*Lemma 3.1.1*]{} either $v^{\prime\,-1}(0)$ is discrete, or $v'$ is identically zero.
Remarks. 1. Note that, in the latter case, $v:\Delta \to C$ is holomorphic.
**2. The “Unique Continuation Lemma” remains true for Lipschitz $J$ (i.e., $J\in C^{0,1}$) but not for Hölder and continuous $J$ (i.e., $J\in C^{0,\alpha}$ with $0\le\alpha<1$). To see this, we note that if $J\in C^{0,\alpha}$ with $0<\alpha\le1$, then any $J$-holomorphic map $u: \Delta \to (X,J)$ is $C^1$-smooth by standard regularity for the elliptic equation $\dbar_J u=0$. If $u$ is an immersion, then $du: T\Delta \to u^*TX$ is an imbedding of complex (not holomorphic!) bundles, and $N \deff u^*TX/ du(T\Delta)$ has a canonical complex structure. Equip $N$ with any holomorphic structure compatible with the complex one. Then [*Step 1*]{} from above goes through and we can locally construct $J_1$, which coincides with $J$ along some neighborhood of a given point $p\in \Delta$. Instead of (1) we obtain $$|J(w_1, w')-J_1(w_1, w')| \le C(|w'|^\alpha).
\eqno(3.4.2)$$ Thus the “similarity principle” and [*Step 2*]{} are applicable if $\alpha=1$ (Lipschitz case) but fail if $\alpha<1$ (Hölder case).**
Exercise. [Construct a counter example to “unique continuation lemma” for the Hölder $J$ in the following way. Take two different smoothly imbedded disks which coincide along some open subset, and try to construct a complex structure $J$ to make both $J$ - complex . It appears that for appropriately chosen disks one succeeds with $J \in C^{0,\alpha}$ for at least some $0<\alpha<1$. ]{}
[*Step 3*]{}. [*Proof of the Theorem in the case when $u_i: (S_i, j_i)
\to
(X,J)$ are immersions.*]{}
Consider set $\cala$ of those $V$ such that $V$ is an open subset in one of $S_i$ and $u_i: V \to X$ is injective. Equip every $V \in \cala$ with the complex structure and with a $J$-holomorphic map $\ti u:
V\to X$ induced from $S_i$. Define the equivalence relation $\sim$ on the disjoint union $\sqcup_{V\in \cala}V$, identifying points $p'\in V'$ and $p'' \in V''$ if there exist $V\in \cala$ and $p\in V$ such that $\ti u(V) \subset \ti u(V') \cap \ti u(V'')$ and $\ti u(p) = \ti u(p') =
\ti u(p'')$.
Set $S \deff \bigl(\sqcup_{V\in \cala}V\bigr) / \sim$ with projection $\pi :
\sqcup_{V\in
\cala}V \to S$. Induce the quotient topology on $S$ declaring $\{ \pi(V) :V\in \cala\}$ as the basis of the topology. Then $\ti u$ induces the map $u: S \to X$ which is holomorphic. We shall show that $S$ and $u$ obey the desired properties. The main point at this step is to show that the quotient topology on $S$ is Hausdorff.
Let $p'$ and $p''$ be two distinct points on $S$. If $u(p') \not= u(p'')$, then there exist disjoint neighborhoods $u(p') \in U' \subset X$ and $u(p'') \in U'' \subset X$ and their pre-images $V' \deff u\inv(U')$ and $V'' \deff u\inv(U'')$ are disjoint neighborhoods of $p'$ and $p''$. It remains to consider the case when $u(p') = u(p'')$.
Let $\ti p' \in \ti V'$ and $\ti p''\in \ti V''$ be representatives, i.e., $\ti V', \ti V'' \in \cala$, $\pi(\ti p')=p'$, and $\pi(\ti p'')=p''$. Shrinking $\ti V'$, if needed, we may assume that for $\ti u(\ti V')$ there exists a neighborhood $U \subset X$, an (integrable) complex structure $J_1$ in $U$ and $J_1$-holomorphic coordinates $(w_1, w')$ in $U$ with the properties of [*Step 1*]{}. Shrinking $\ti V''$, we may assume that $\ti u(\ti V'')$ is also contained in $U$. Using [*Step 2*]{}, we may additionally assume that $u(p')=u(p'')$ is the only intersection point of $\ti u(\ti V')$ and $\ti u(\ti V'')$. Then $\pi(\ti V')$ and $\pi(\ti V'')$ are disjoint neighborhoods of $p'$ and $p''$. This shows that $S$ is Hausdorff.
By assumption, maps $u_i: S_i \to X$ are immersions. Thus every $S_i$ can be covered by open sets $V \subset S_i$ such that restrictions $u_i|_V$ are imbeddings. Every such $V$ belongs to the atlas $\cala$, and the canonical projection $\pi: V \to S$ is a holomorphic imbedding. By the definition of the equivalence relation $\sim$, these local maps $\pi: V\subset S_i \to S$ can be glued together to holomorphic maps $g_i: S_i \to S$. One can see that the constructed $(S,j)$, $u:S \to X$, and $g_i: S_i \to S$ have the desired properties. **Non immersed case. [*Here we consider the case where $u_i: (S_i,j_i) \to (X,J)$ are not necessarily immersions.*]{}**
From [*Proposition 6.3.2*]{} of [*Lecture 6*]{}, maps $u_i$ define bundle homomorphisms $du_i : TS_i \to E_i :=
u_i^*(TX)$ which are holomorphic w.r.t. the canonical holomorphic structure on $E_i$. This implies that the zero set $A_i:=\{y \in S_i: du_i(y)=0 \}$ is discrete in $S_i$. In particular, every $u_i$ is an immersion outside $A_i$.
Set $\check S_i \deff S_i \bs A_i$. Then $u_i: \check S_i \to X$ are immersions, and the previous arguments remain valid. Let $(\check S, \check \jmath) $, $\check u:\check S \to X$, and $\check g_i: \check S_i \to \check S$ be the corresponding objects.
Numerate points in $A_1$ in any order, $A_1 = \{\, a_1, a_2, \ldots \,\}$.
[*Step 4.*]{} [*We state that for $k=1,2,\ldots\,$ there exist Riemann surfaces $S^{(k)}$ and holomorphic maps $g_1^{(k)}: S_1^{(k)} \to S^{(k)}$ such that $S_1^{(k)} \deff
\check S_1 \cup \{a_1, \ldots a_k\}$, $g_1^{(k)}|_{\check S} =\check g_1$, and $S^{(k)} = \check S \cup g_1^{(k)}(\{a_1, \ldots a_k\})$.* ]{}
In other words, at the $k$-th step we add to $S_1^{(k-1)}$ the point $a_k$, and extend $g_1^{(k-1)}
: S_1^{(k-1)} \to S^{(k-1)}$ through $a_k$ holomorphically, adding, if needed, some new point to $S^{(k-1)}$.
The proof of this statement proceeds inductively. For a given $k$ take a sufficiently small neighborhood $V_k \subset S_1$ of $a_k$ which contains no other points of $A_1$. Set $\check V_k \deff V_k
\bs \{a_k\}$. If the restriction $g_1^{(k-1)}|_{\check V_k}: \check V_k \to
S^{(k-1)}$ extends to a holomorphic map $g_1^{(k)}|_{V_k}: \check V_k \to
S^{(k-1)}$, one has nothing to do but set $S^{(k)}\deff S^{(k-1)}$.
Otherwise we set $S^{(k)}\deff S^{(k-1)} \sqcup \{b_k\}$ and $g_1^{(k)}(a_k)
= b_k$. We must prove existence of appropriate topology and complex structure on the constructed $S^{(k)}$. We give a proof of these properties in full detail.
Choose $C^2$-smooth local complex (not holomorphic) coordinates $(w_1, \ldots, w_n)$ in a neighborhood of $u_1(a_k) \in X$ such that a complex structure $J'$ defined by $(w_1, \ldots,
w_n)$ coincides with $J$ in $u_1(a_k)$. Let $z$ be a local complex coordinate on $S_1$ in a neighborhood of $a_k$. We may assume that $(w_1, \ldots, w_n)$ (resp. $z$) are centered in $u_1(a_k)$ (resp. in $a_k$). [*Lemma 3.1.1*]{} provides that in these coordinates the map $u_1 : S_1 \to X$ has the form $u_1(z) = z^l \cdot (v_1(z), \ldots,v_n(z))$ with $\vec v=(v_1,
\ldots, v_n) \in L^{1,p}$ and $\vec v(0) \not=0$.
Choose a coordinate $w_p$ such that the corresponding component $v_p$ does not vanish at $z=0$, $v_p(0) \not=0$. Then there exists the $l$-th root of this $v_p$, i.e., $v_p(z) = (f(z))^l$ for some continuous non-vanishing function $f(z)$. Since $v_p(z)$ is $L^{1,p}$-smooth, so is $f(z)$. Shrinking our neighborhood $V_k$, if needed, we may assume that $f(z)$ is not vanishing in $V_k$. Thus for an appropriate neighborhood $U$ of $u_1(a_k)$ there exists a $C^2$-smooth map $w_p:U \to \cc$ such that $h_p \deff w_p\scirc u_1 (z)=
(zf(z))^l$ for some $L^{1,p}$-smooth non-vanishing function $f(z)$ defined in a neighborhood of $a_k$. We may additionally assume that $$|f(z)-f(0)|\le \eps <\!\!<1 \qquad\hbox{for all $z\in V_k$}.
\eqno(3.4.3)$$ Note that for every sufficiently small $b \in \cc$ we have at least one solution $z\in V_k$ of the equation $h_p(z)=b$ with $|z f(0)|^l \le
2|b|$. In order to see this, one considers $h_p$-images of circles $\gamma_t \deff
\{\, |zf(0)|^l = t|b|\, \}$. Then due to (3.4.3), the image $h_p(\gamma_{1\over2})$ lies in the disk $\Delta(|b|)$, whereas $h_p(\gamma_2)$ lies outside of $\Delta(|b|)$. Homotopy argument shows that $b\in \psi(\gamma_t)$ for some $t\in ]{1\over2}, 2[$. Similarly, one obtains the estimate $$c\cdot |h_p(z)|^{1\over l} \le |z| \le C\cdot |h_p(z)|^{1\over l}
\qquad\hbox{for $z\in V_k$}
\eqno(3.4.4)$$ with some constants $c$ and $C$. Take a sufficiently small disk $\Delta(r)$ in $\cc$ (range of $h_p(z)= \omega_p
\scirc
u_1(z)$) and consider the set $V'_k \deff \{ z\in V_k: |z|^l < {2r \over
|f(0)|^l}, \, |h_p(z)|< r\}$. Consider a sequence $z_i \in V'_k$ such that $h_p(z_i)$ converges to some $b\in \Delta(r)$. Then some subsequence of $z_i$ converges to $z^*$ with $|z^*| \le {2r \over |f(0)|^l}$. But then $|h_p(z^*)|=
|z^*f(z^*)|^l =|b|$ which implies $|z^*|^l = {|b| \over |f(z^*)|^l} <{2r
\over |f(0)|^l}$, so that $z^* \in V'_k$. This shows that the map $h_p: V_k
\to \Delta(r)$ is proper. By (4), $h_p(z)=0$ for $z\in V'_k$ implies $z=0$. Thus, for $\check V'_k \deff \{z\in V'_k : z\not=0\} = V'_k \bs\{a_k\}$ the map $h_p: \check V'_k \to \check\Delta(r)$ is also proper.
Recall that the bundle homomorphism $du_1: TS_1 \to E_1\deff u_1^*TX$ is holomorphic with respect to the canonical holomorphic structure on $E_1$. Thus, in $V'_k$ we can represent $du_1$ in the form $du_1(z)= z^{l-1}\cdot
s(z)$ with $s(z)$ also holomorphic and $s(0)\not=0$. This implies that for $z\not=0 \in V'_k$ sufficiently close to $0$ the image $du_1(T_zS_1)$ of the tangent space $T_zS_1$ is close to a real 2-dimentional space $\cc s(0)\subset E_1|_0=T_{u_1(a_k)}X$. This is a complex line in $T_{u_1(a_k)}X$ generated by $s(0)$. We may assume that the coordinates $(w_1, \ldots, w_n)$ were chosen in such a way that the $p$-th component of the vector $s(0)$ is non-zero. This means that the linear map $dw_p|_{du_1(a_k)}: T_{u_1(a_k)}X \to \cc$ is not degenerated $\cc s(0) \subset T_{u_1(a_k)}X$. Thus, for any $b$ sufficiently close to $u_1(a_k)$, and for any real 2-dimensional space $L \subset T_bX$ sufficiently close to $\cc s(0)$ the linear map $dw_l|_b: T_bX \to \cc$ is not degenerated on $L$.
Hence we can conclude that $h_p: \check V'_k \to \check \Delta(r)$ is not degenerated provided $r$ was chosen sufficiently small. This means that $h_p: \check V'_k \to \check \Delta(r)$ is a covering. Note that $\pi_1(\check
\Delta(r)) = \zz$ and that $\pi_1(\check V'_k)$ contains $\zz$ as a subgroup generated by a small circle $\{ |z| = \rho\}$. From inclusion $(u_1)_*:
\pi_1(\check V'_k) \hookrightarrow \pi_1(\check \Delta(r))$ we conclude that $\pi_1(\check V'_k) =\zz$, and that $h_p: \check V'_k \to \check
\Delta(r)$ is a finite covering. Since a small circle $\{ |z| = \rho\}$ around $a_k \in V'_k$ is a generator of $\pi_1(\check V'_k)$, we conclude that $V'_k$ is a disk and $\check V'_k$ is a punctured disk.
Now recall that the map $h_p:\check V'_k \to \check \Delta(r)$ can be seen as a composition of $g^{(k-1)}: \check V'_k \to S^{(k-1)}$ with a coordinate function $w_p$ restricted on $\check u_1(\check V'_k)$. Set $\check W_k \deff
g^{(k-1)} (\check V'_k)$ and consider a holomorphic map $g^{(k-1)}: \check
V'_k \to \check W_k$. Then this map is again proper and non-degenerate and therefore is also a finite covering. This implies that $\check W_k$ is also a punctured disk.
Let $\check\psi: \check W_k \to \check \Delta$ be a biholomorphism which can be seen as a local chart on $\check W_k \subset S^{(k-1)}$. Define holomorphic structure on $S^{(k)}= S^{(k-1)} \sqcup \{b_k\}$ in the following way. Interpret $b_k$ as a puncture point of $\check W_k$, set $W_k \deff \check W_k\sqcup \{b_k\}$, extend $\check \psi$ to the map $\psi: W_k \to \Delta$ setting $\psi(b_k) \deff 0$. Extend the topology on $S^{(k)}$ in such a way that $\psi: W_k \to \Delta$ becomes a homeomorphism. Declare $\psi$ to be a holomorphic coordinate on $S^{(k)}$. Note that the composition $\check\psi \scirc g^{(k-1)}: \check V'_k \to \check \Delta$ is bounded; thus it extends to a holomorphic map from $V'_k$ to $\Delta$. Consequently, $\check\psi\inv: \check \Delta \to \check W_k \subset S^{(k-1)}$ cannot be holomorphically extended to $\psi\inv: \Delta \to S^{(k-1)}$, because otherwise this would mean holomorphic extensibility of $g^{(k-1)}: S_1^{(k-1)}
\to S^{(k-1)}$ through $a_k$. This property insures that $S^{(k)}$ is Hausdorff. Thus constructed, $S^{(k)}$ and $g^{(k)}: S_1^{(k)} \to S^{(k)}$ obey the desired properties.
Setting $S' \deff \cup_k S^{(k)}$ we obtain the minimal extension of $\check S$ such that the map $\check g_1: \check S_1 \to \check S$ extends holomorphically to $g'_1: S_1 \to S'$. A similar construction should be done to extend $\check g_2 : \check S_2 \to S'$ to a holomorphic map $g_2: S_2 \to S$. This finishes the proof of the [*theorem* ]{}.
In the spirit of [*Theorem 3.4.1*]{} one obtains the following result.
Proposition 3.4.2. *Let $(X,J)$ be an almost complex manifold with $J \in C^1$. Then any primitive $J$-holomorphic map $u: (S,j) \to X$ with connected irreducible $(S,j)$ is injective outside some countable subset $A\subset S$. Moreover, $A$ can be chosen in such a way that for any compact $K\Subset S$ the intersection $A \cap K$ is finite.*
Proof. We have shown that $A_1 \deff \{a\in S: du(a)=0\}$ is discrete; therefore it is countable. Set $\check S:= S\setminus A_1$. Consider the set $A_2 \deff \{ (a,b) \in \check S\times \check S : u(a)=u(b) \}$. We state that $A_2$ is discrete in $\check S\times \check S\setminus \Delta $, where $\Delta $ is a diagonal. Otherwise there must exist a sequence $(a_n,b_n)$ converging to some pair $(a^*,b^*) \in \check S\times
\check S$. Take a sufficiently small neighborhood $V \subset \check S$ of $a^*$ and find a neighborhood $U \subset X$ of $u(a^*)$, which obeys the properties of [*Step 1*]{} from the previous proof. Note that obviously $u(b^*) = u(a^*)$. Now we obtain a contradiction with argumentations from [*Step 2*]{} above. This shows that $A_2$ must be discrete in $\check S\times \check S\Delta $.
Thus it is also countable. Let $A_3$ be the projection of $A_2$ onto the first $\check S$. Then $A:=A_1\cup A_2$ obeys the desired properties.
Remark. In fact, positivity of the intersection of $J$ - complex curves with $J\in C^1$ provides that there exists a [*discrete*]{} subset $A\subset
S$ fulfilling the condition of [*Proposition 3.4.2*]{}. We will prove this in next section.
*3.5. Positivity of Intersection.*
Let us first recall the notion of the intersection number (index) of two surfaces in $\rr^4$. Let $M_1$ and $M_2$ be two-dimensional, oriented, smooth surfaces in $\rr^4$ passing through the origin. We suppose further that both $M_1,M_2$ intersect the unit sphere $S^3$ transversally by curves $\gamma_1$ and $\gamma_2$, respectively, and that $\gamma_1$ and $\gamma_2$ do not meet one another. Let $\tilde M_i$ be small perturbations of $M_i$ making them intersect transversally.
Definition 3.5.1. [*The intersection number of $M_1$ and $M_2$ is defined to be the algebraic intersection number of $\tilde M_1$ and $\tilde M_2$. If $M_1,M_2$ intersect only at zero, we also say that the number just defined is the intersection index of $M_1$ and $M_2$ at zero. It will be denoted by $\delta_0(M_1,M_2)$ or $\delta_0$.* ]{}
Remark. [This number is independent of the particular choice of perturbations $\tilde M_i$. We shall use the fact that the intersection number of $M_1$ and $M_2$ is equal to the [*linking number*]{} $l(\gamma_1,
\gamma_2)$ of the (reducible in general) curves $\gamma_i$ on $S^3$, see \[Rf\]. ]{}
In the following theorem the structure $J$ is supposed to be of class $C^1$.
Theorem 3.5.1. *Let $u_i:\Delta\to(\rr^4, J)$, $i=1, 2$ be two primitive distinct $J$ - complex disks such that $u_1(0)=u_2(0)$. Set $M_i \deff u_i
(\Delta)$. Let $Q = M_1\cap M_2$ be the intersection set of the disks. If $J$ is $C^1$-smooth, then the following is true.*
The set $\{\, (z_1,z_2) \in \Delta \times \Delta:
u_1(z_1)= u_2(z_2)\,\}$ is a discrete subset of $\Delta\times \Delta $. In particular, $u_1(\Delta) \cap
u_2(\Delta)$ is a countable set;
The intersection index in any such point of $Q$ is strictly positive. Moreover, if $\mu_1$ and $\mu_2$ are the multiplicities of $u_1$ and $u_2$ in $z_1$ and $z_2$, respectively, with $u_1(z_1)=u_2(z_2)=p$, then the intersection number $\delta_p$ of branches of $M_1$ and $M_2$ at $z_1$ and $z_2$ is at least $\mu_1\cdot \mu_2$;
$\delta_p=1$ iff $M_1$ and $M_2$ intersect at $p$ transversally.
Proof. [*Case 1. The map $u_1$ has no critical points*]{}. Thus $u_1: \Delta \to \rr^4$ is an immersion. Take any $(a_1,a_2) \in \Delta \times \Delta$ with $u_1(a_1)=u_2
(a_2)$. In a neighborhood of $u_1(a_1)\in \rr^4$ we find local coordinates $(w_1, w_2)$ with the properties from [*Step 1*]{} of the proof of [*Theorem 3.4.1*]{}. We may assume that in these coordinates the map $u_1$ has the form $u_1(z)=(z,0)$ and that $a_1$ is the point with $z=0$. Choose the local coordinate $z$ on the second disk such that $z=0$ is the point $a_2$. Consider representation of the map $u_2: \Delta \to \rr^4$ in coordinates $(w_1,w_2)$ such that $u_2(z)= (w_1(z),w_2(z))$. From the inequality (3.4.1) and [*Lemma 3.1.1*]{} we conclude that $w_2(z)=
z^\nu b(z)$ for some $b\in L^{1,p}$ with $b(0)\not =0$. This implies that the only intersection point of images of small neighborhoods of $a_1$ and $a_2$ is $u_1(a_1)=u_2(a_2)=p$, and its intersection index is $\nu$. This provides the properties – for the case when one of the maps $u_1$ or $u_2$ is an immersion.
[*Case 2. Both $u_1$ and $u_2$ are immersions outside $0\in \Delta$, but $du_1(0)=0=du_2(0)$. Moreover, $u_1(0)=u_2(0)$.*]{}
First we collect some information about the local behavior of $u_i$ at $0\in
\Delta$. We consider only the map $u_1$; the same procedure should be done for $u_2$.
Without loss of generality we may assume that the point $u_1(0)=u_2(0)$ is the center of coordinates in $\rr^4$. We may also assume that the standard complex structure $J\st$ in $\rr^4$ coincides with the given structure $J$ in the point $P$. Let $(w_1,w_2)$ be the standard complex coordinate in $(\rr^4,
J\st) =\cc^2$. To avoid possible confusion, denote by $\ti u=\ti u(z)$ the representation of $u$ in coordinates $(w_1,w_2)$. Then $\ti u_1(z)= z^{\mu_1}
v_1(z)$ with $v_1 \in L^{1,p}$ and $v_1(0) \not=0$ by the “similarity principle”.
Now consider the bundle $E_1:= u_1^*T\rr^4$ over $\Delta$. Two complex structures in $\rr^4$, $J$ and $J\st$, induce two complex structures in $E_1$, $u_1^*J$ and $u_1^*J\st$. We always equip $E_1$ with the co-structure $u_1^*J$; otherwise the contrary is noted explicitly. In particular, $E_1$ equipped with $u_1^*J$ has the canonical holomorphic structure such that the homomorphism $du_1: T\Delta \to E_1$ is holomorphic. Consequently, $du_1(z)=
z^{\nu_1} s_1(z)$ for some holomorphic section $s_1$ of $E_1$ with $s_1(0)
\not =0$.
On the other hand, since the tangent bundle $(T\cc^2, J\st)$ is trivial, the complex bundle $(E_1, u_1^*J\st)$ has natural trivialization. Denote this trivialization by $\psi: (E_1, u_1^*J\st) \to \cc^2$, where $\cc^2$ stands for the trivial bundle over $\Delta$. Then the homomorphism $\psi(z): E_1|_z
\to \cc^2$ is an $\rr$-linear isomorphism in general, and complex linear exactly for those $z\in\Delta$, for which $J(u_1(z))= J\st(u_1(z))$. In particular, this holds for $z=0$. This implies $\psi(z^\nu s(z))= z^\nu
\psi(s(z)) + O(|z|^{\nu+1})$. From our construction of $psi$ follows the equality $\psi \scirc du_1(z) = d\ti u_1(z)$. Thus $d\ti u_1(z)= z^{\nu_1}
\cdot \psi \scirc s_1(z) + O(z^{\nu_1+1})$. Comparing this asymptotic expansion with $\ti u_1(z)= z^{\mu_1} v_1(z)$ we conclude that $\nu_1=
\mu_1-1$ and $s_1(0)= \mu_1 v_1(0)$.
From now on we do not distinguish the map $u_1$ from its representation $\ti u_1$ in coordinates $(w_1, w_2)$. Note that we can use the same coordinates $(w_1, w_2)$ and the structure $J\st$ considering the map $u_2$. Thus we get the asymptotic relations $$\eqalignno{
u_i(z) &= z^{\mu_i} v_i(0) + O(|z|^{\mu_i+ \alpha}) &(3.5.1)
\cr
du_i(z) &= z^{\mu_i-1} v_i(0) + O(|z|^{\mu_i-1+\alpha}) &(3.5.2)
}$$ with some Hölder exponent $\alpha>0$. This implies the transversality of small $J$ - complex disks $u_i(\Delta
(\rho))$ to small spheres $S^3_r$. More precisely, there exist radii $\rho>0$ and $R>0$ such that for any $0<r<R$ the $J$-curves $u_i(\Delta (\rho))$ intersect the sphere $S^3_r\deff \{ |w_1|^2+ |w_2|^2=r^2 \}$ transversally along smooth immersed circles $\gamma_i(r)$. In fact, the asymptotic relation (4) provides that for any $\theta \in [0,2\pi]$ there exists at least one solution of the equation $\bigl|u_i(\rho_{i} e^{\isl\theta}) \bigr| =r$ with $\rho_i< \rho$, and that any such solution $\rho_i$ must be close to $\left(
{r \over |v_1(0)|}\right)^{1/\mu_i}$. Then one uses (5) to show that the set $\ti\gamma_i(r) \deff \{ z\in\Delta: \bigl|u_i(z) \bigr| =r \}$ is, in fact, a smooth imbedded curve in $\Delta$, parameterized by $\theta\in [0,2\pi]$, and that $u_i: \ti\gamma_i(r) \to S^3_r$ is an immersion with the image $\gamma_i(r)$.
Taking an appropriate small subdisk and rescaling, we may assume that $\rho=1
=R$. Note that the points of the self- (resp. mutual) intersection of $\gamma_i(r)$ are self-( or resp. mutual) intersection points of $u_i
(\Delta)$. Let us call $r \in ]0,1[$ non-exceptional if curves $\gamma_i(r)
\subset S^3_r$ are imbedded and disjoint. Thus $r^* \in ]0,1[$ is exceptional if $S^3_{r^*}$ contains intersection points of $u_i (\Delta)$.
The “unique continuation lemma” of Step 2 of the proof of [*Theorem 3.4.1*]{} provides that any such intersection point is isolated in the punctured ball $\check B \deff \{ 0<
|w_1|^2 + |w_2|^2 <1 \}$. This implies that either there exist finitely many exceptional radii $r^*\in ]0,1[$, or that they form a sequence $r^*_n$ converging to $0$.
Denote $M_i(r) \deff u_i (\Delta) \cap B(r)$. For non-exceptional $r$ we can correctly define the intersection index of $M_1(r)$ with $M_2(r)$.
By [*Corollary 3.1.3*]{}, the maps $u_i$ can be represented in the form $u_i(z)=z^{\mu_i}\cdot v_i(z)$ with $v_i\in L^{1,p}(\Delta,\cc^2)$ such that $v_1(0) \not= 0 \not= v_2(0)$. As above, we consider two cases.
[*Case 1. The vectors $v_1(0)$ and $v_2(0)$ are not collinear.*]{}
It is easy to see that, in this case, $0\in\cc^2$ is an isolated intersection point of $u_1(\Delta)$ and $u_2(\Delta)$ with multiplicity exactly $\mu_1\cdot \mu_2$. The asymptotic formula (3.4.4) provides that if vectors $v_1(0)$ and $v_2(0)$ are complex linear independent, then $0\in \rr^4$ is an isolated intersection point of $Q =u_1(\Delta) \cap u_2(\Delta)$ with the index $\mu_1 {\cdot}
\mu_2$. Thus we have only finitely many intersection points $p\in Q$. Since all other points $p\in Q$ are smooth, the intersection index in every such point is positive. Thus in the case of non-collinear $v_1(0)$ and $v_2(0)$ for any non-exceptional $r>0$ the intersection index of $M_1(r)$ and $M_2(r)$ is positive.
[*Case 2. The second case is when the vectors $v_1(0)$ and $v_2(0)$ are collinear.*]{}
The idea is to “turn” the map $u_2$ a little and to reduce the case to the previous one. So let $T\in \endc(\cc^2)$ be a linear unitary map which is close enough to identity such that $T(v_2(0))$ is not collinear to $v_1(0)$. Define $J^T\deff T\inv \scirc J \scirc T$ so that $\Vert J^T -J \Vert_{C^1(B)} \le \Vert T - \id \Vert$. Applying [*Lemma 3.3.1*]{} with $u_0=z^{\mu _2}v_2$ and $v=0$, we find $w\in L^{1,p}(
\Delta,\cc^2)$ with $w(0)=0$ such that $\dbar_{J^T} (z^{\mu_2}(v_2+w)) =0$. The map $\tilde u_2\deff T(z^{\mu_2}(v_2+w))$ is the required “turned” $J$-holomorphic map. Since such a “turn” can be made small enough, the intersection number of $u_i(\Delta)$ in the ball $B_{r^-}$ does not change.
This implies that the intersection number of $u_1(\Delta)$ and $u_2(\Delta)$ in any ball $B_r$ is not less than $\mu_1\cdot \mu_2$. Another conclusion is that $0\in\cc^2$ is an isolated intersection point of $u_j(\Delta)$. Otherwise we could find a sequence $r_i\msmall{\searrow}0$ with at least one intersection point of $u_j(\Delta)$ in every spherical shell $B_{r_i} \!\backslash B_{r_{i+1}}$; therefore the intersection number of $u_j(\Delta)$ in the balls $B_{r_i}$ would be strictly decreasing in $i$.
Thus the statements (i) and (ii) are proved. The proof of (iii) is now obvious and follows from the observation that $\mu _1 = \mu _2 =1$ in this case.
Corollary 3.5.2. [*Let $u_i:S_i\to(X, J)$, $i=1, 2$ be compact irreducible $J$ - complex curves such that $u_1(S_1)=M_1\not=u_2(S_2)=M_2$. Then they have finitely many intersection points and the intersection index in any such point is strictly positive. Moreover, if $\mu_1$ and $\mu_2$ are the multiplicities of $u_1$ and $u_2$ in such a point $p$, then the intersection number of $M_1$ and $M_2$ in $p$ is at least $\mu_1\cdot \mu_2$.* ]{}
[Appendix 2]{}
[The Bennequin Index and Genus Formula.]{}
*A2.1. Bennequin Index of a Cusp.*
Let $u : (\Delta , 0)\to (\cc^2, J, 0) $ be a germ of a nonconstant complex curve in zero. Without loss of generality we always suppose that $J(0)=J\st$. Taking into account that zeros of $du$ are isolated, we can suppose that $du$ vanishes only at zero. Furthermore, let $w_1, w_2$ be the standard complex coordinates in $(\cc^2, J\st)$. According to [*Lemma 3.1.3*]{} we can write our curve in the form
$$u(z) = z^\mu\cdot a + o(\vert z\vert^{\mu +\alpha}),
\qquad a\in \cc^2 \setminus \{ 0\},\> 0<\alpha<1. \eqno(A2.1.1)$$
For $r>0$ define $F_r\deff TS^3_r \cap J(TS^3_r)$ to be the distribution of $J$-complex planes in the tangent bundle $TS^3_r$ to the sphere of radius $r$. $F_r$ is trivial, because $J$ is homotopic to $J\st = J(0)$. By $F$ we denote the distribution $\cup_{r>0}F_r\subset \cup_{r>0}TS^3_r \subset
TB^*$, where $TB^*$ is the tangent bundle to the punctured ball in $\cc^2$.
**Lemma A2.1.1. *The (possibly reducible) curve $\gamma_r = M\cap S^3_r$ is transversal to $F_r$ for all sufficiently small $r>0$.***
**Proof. In fact, due to [*Lemma 3.2.3*]{} one has $du(z) = \mu
z^{\mu-1} \cdot a\,dz + o(\vert z\vert^{\mu -1+\alpha})$. Since $J\approx J\st$ for $r$ sufficiently small, $T\gamma_r$ is close to $J\st n_r$, where $ n_r$ is the field of normal vectors to $S^3_r$.**
On the other hand, for sufficiently small $r$, the distribution $F_r$ is close to the one of $J\st$ complex planes in $TS^3_r$, which is orthogonal to $J\st n_r$.
This fact permits us to define the Bennequin index of $\gamma_r$. Namely, take any nonvanishing section $\vec v$ of $F_r$ and move $\gamma_r$ along the vector field $v$ to obtain a curve $\gamma'_r$. We can make this move for a small enough time, so that $\gamma'_r $ does not intersect $\gamma_r$. Following Bennequin \[Bn\], we have
Definition A2.1.1. Define the *Bennequin index $b(\gamma_r)$ to be the linking number of $\gamma_r$ and $\gamma'_r$.*
This number does not depend on $r>0$, taken sufficiently small, because $\gamma_r$ is homotopic to $\gamma_{r_1}$ for $r_1<r$ within the curves transversal to $F$, see \[Bn\]. It is also independent of the particular choice of the field $\vec v$. Thus for the standard complex structure $J\st$ in $B\subset \cc^2$ we use $\vec v\st(w_1,w_2)\deff (-\bar w_2, \bar w_1)$ for calculating the Bennequin index of the curves on sufficiently small spheres. For an arbitrary almost complex structure $J$ with $J(0)=J\st$ we can find the vector field $\vec v_J$, which is defined in a small punctured neighborhood of $0\in B$, is a small perturbation of $\vec v\st$, and lies in the distribution $F$ defined by $J$.
Denote by $B_{r_1, r_2}$ the spherical shell $B_{r_2}\setminus B_{r_1}$ for $r_1<r_2$.
Lemma A2.1.2. *Let $\Gamma $ be an immersed $J$ - complex curve in the neighborhood of $\overline B_{r_1, r_2}$ such that all self intersection points of $\Gamma $ are contained in $B_{r_1, r_2}$ and for every $r_1\le r\le r_2$ all components of the curve $\gamma_r = \Gamma \cap
S^3_r$ are transversal to $F_r$. Then $$b(\gamma_{r_2}) = b(\gamma_{r_1}) + 2\cdot \sum\nolimits
_{x\in Sing(\Gamma) }\delta_x,
\eqno(A2.1.2)$$ where the sum is taken over self-intersection points of $\Gamma $.*
Proof. Move $\Gamma $ a little along $v_J$ to obtain $\Gamma^\varepsilon$. By $\gamma^\varepsilon_{r_1},
\gamma^\varepsilon_{r_2}$ denote the intersections $\Gamma^\varepsilon \cap
S^3_{r_1},\Gamma^\epsi\cap S^3_{r_2}$, which are of course the moves of $\gamma_{r_j}$ along $v_J$. We have $l(\gamma_{r_2},\gamma
_{r_1}^\epsi) - l(\gamma_{r_1},\gamma_{r_1}^ \epsi) =
{\sf int}(\Gamma ,\Gamma^\epsi)$, where $l(\cdot ,\cdot )$ is the linking number and ${\sf int}(\cdot ,\cdot )$ is the intersection number.
Now let us calculate $int(\Gamma ,\Gamma^\epsi)$. From [*Theorem 3.5.1*]{} we know that there are only a finite number $\{ p_1,\ldots, p_N\} $ of self-intersection points of $\Gamma $. Take one of them, say $p_1$. Let $M_1,
\ldots, M_d$ be the disks on $\Gamma $ with a common point $p_1$ and otherwise mutually disjoint. More precisely we take $M_j$ to be irreducible components of $\Gamma \cap B_{p_1}(\rho)$ for $\rho >0$ small enough. Remark that $M_j$ are transversal to $v_J$, so their moves $M_j^\varepsilon$ don’t intersect them, [*i.e.,*]{} $M_j\cap M_j^\epsi = \emptyset $. Note also that ${\sf int}(M_k, M_j) = {\sf int}(M_k, M^\epsi_j)$ for $\varepsilon >0$ sufficiently small. So $int(M_k, M_j) =
int(M_k, M^\epsi_j) + int(M^\epsi_k, M_j)$. Finally $\delta_{p_1} = \sum_{1\le k<j\le d} int(M_k, M_j) =
int(\Gamma \cap B_{p_1}(\rho ),\Gamma^\epsi\cap
B_{p_1}(\rho ))$. This means that $int(\Gamma ,\Gamma^\epsi)
= 2\cdot \sum_{j=1}^N\delta_{p_j}$.
*A2.2. Proof of Adjunction Formula.*
In $\S \S 3.4, 3.5$ we have proved that compact $J$ - complex curve with a finite number of irreducible components $M = \bigcup_{i=1}^d M_i$ has only a finite number of nodes ([*i.e.,*]{} self-intersection points) points, provided $J$ is of class $C^1$.
For each such point $p$ we can introduce, according to [*Definition 3.5.1*]{}, the self-intersection number $\delta_p(M)$ of $M$ at $p$. Namely, let $S_j$ be a parameter curve for $M_j$, [**]{} $M_j$ is given as an image of the $J$-holomorphic map $u_j : S_j\to M_j$. We always suppose that the parameterization $u_j$ is primitive, [*i.e.,*]{} they cannot be decomposed like $u_j=v_j\scirc r$ where $r$ is a nontrivial covering of $S_j$ by another Riemann surface. Denote by $\{ x_1,\ldots, x_N\}$ the set of all pre-images of $p$ under $u : \bigsqcup_{i=1}^d S_j\to X$, and take mutually disjoint disks $\{ D_1,\ldots, D_N\}$ with centers $x_1,\ldots, x_N$ such that their images have no other common points different from $p$. For each pair $D_i,
D_j$, $i\not= j$, define an intersection number as in [*Definition 3.5.1*]{} and take the sum over all different pairs to obtain $\delta_p(M)$.
Now put $\delta = \sum_{p\in D(M)}\delta_p(M)$, where the sum is taken over the set $D(M)$ of all nodes of $M$, [*i.e.,*]{} points which have at least two pre-images.
Consider now the set $\{ p_1,\ldots, p_L\} \subset \bigcup_{j=1}^d S_j = S$ of all cusps of $M$, [*i.e.,*]{} points where the differential of the appropriate parameterization vanishes. Take a small ball $B_r$ around $u(p_i)$ and a small disk $D_{p_i}$ around $p_i$. Let $\gamma^i_r\deff u(\Delta_{p_i}) \cap \partial B_r$ and $b_i$ be the Bennequin index of $\gamma^i_r$, defined in [*Definition A2.1.1*]{}.
Definition A2.2.1. The number $\varkappa_i\deff (b_i +1)/2$ is called the [*conductor of the cusp $a_i=u(p_i)$*]{}.
Also let $\varkappa\deff \sum_{i=1}^L \varkappa_i$. We are now able to state the genus formula.
**Theorem A2.2.1. *Let $M =\cup_{j=1}^d M_j$ be a compact $J$ - complex curve in an almost complex surface $(X, J)$ with the distinct irreducible components $\{M_j\}$, where $J$ is of class $L^{1,p}$. Then***
$$\sum_{j=1}^d g_j = {[M]^2 - c_1(X, J)[M]\over2} + d -\delta - \varkappa ,
\eqno(A2.2.1)$$
where $g_j$ are the genera of parameter curves $S_j$.
Proof. The main line of the proof is the reduction of a general case to the case where $M$ is immersed, which was proved in $\S 1$.
Let $u: \bigsqcup_{j=1}^d S_j \longrightarrow X$ be a $J$-holomorphic map, which parameterizes the curve $M$. Also let $\{x_1,\ldots,x_n\}$ be the set of cusp-points of $M$, [*i.e.,*]{} the images of critical points of $u$. Our reduction procedure is local, and we make our constructions in a neighborhood of every point $x_j$ separately. Therefore, from now on we fix such a point $x$. Due to [*Corollaries 2.2.3*]{} and [*3.4.2*]{} there exists a neighborhood $U$ of $x$ which contains no other cusp-points and no other self-intersection points. [*Theorem 3.5.1 i)*]{} implies that by taking the neighborhood $U$ small enough, we may assume that any component of $M\cap U$ goes through $p\in U$. Without losing generality we may also assume that $U$ is the unit ball $B
\subset \cc^2$, that $x$ corresponds to the center $0$ of $B$, and $J(0)=J\st$.
Denote by $\Gamma_j$ the irreducible components of $M \cap B$ and let $u_j :\Delta \to B$ be a parameterization of $\Gamma_j$ such that $u_j(0)=0
\in B$. Denote $\mu_j \deff \ord_0 du_j +1$; thus $0\in \Delta$ is the cusp-point for $u_j$ $\mu_j\ge2$.
[*Step 1. Rescaling procedure*]{}.
Take some cusp-component $\Gamma_j$. Due to [*Corollary 3.1.3*]{} the map $u_j$ has the form $u_j(z)= z^{\mu_j} \cdot a_j + O\bigl(|z|^{\mu_j+\alpha}
\bigr)$ with the constant $a_j\not=0\in\cc^2$. Moreover, [*Lemma 3.2.3*]{} provides that $$\bigl\vert du_j(z) - \mu_j z^{\mu_j-1} a_j \cdot dz \bigr\vert
\le C(\Vert u_j \Vert_{L^{1,p}(\Delta)}, \alpha, p) \cdot
|z|^{\mu_j-1+\alpha},$$ for any $0<\alpha<1$ and $2<p<\infty$.
For $0<r\le1$ we consider the maps $\pi_r:B\to B$, $\pi_r(w)\deff r^{\mu_j}
\cdot w$, the rescaled maps $u^{(r)}_j:\Delta \to B$, $\pi_r\scirc
u^{(r)}_j(z) = u_j (rz)$, and the rescaled almost complex structures $J^{(r)}
\deff \pi_r^*J$ in $B$. The map $u^{(r)}_j$ is a parameterization of $J^{(r)}$-holomorphic curves $\pi_r^{-1}\Gamma_j$. One can see that $\Vert u^{(r)}_j(z) - z^{\mu_j} a_j \Vert_{C^0(\Delta)} \le C\cdot
r^{\mu_j+\alpha}$ and $\Vert du^{(r)}_j(z) - \mu_j\cdot z^{\mu_j-1} a_j \cdot
dz \Vert_{C^0(\Delta)} \le C\cdot r^{\mu_j-1+\alpha}$. In particular, there exists $r_j>0$ such that for all $0<r<r_j$ the maps $u^{(r)}$ are transversal to all spheres $S^3_1$. On the other hand, $\Vert J^{(r)} - J\st \Vert_{C^1(B)}
= O(r^{\mu_j})$, and for any $\epsi>0$ we can choose sufficiently small $r_j>0$ such that $\Vert J^{(r)} - J\st \Vert_{C^1(B)} \le \epsi$ for all $0<r<r_j$. Thus, by replacing $B$, $J$ and all the maps $u_j$ by their rescaled counterpart, we may assume that the almost complex structure $J$ satisfies the estimate $$\Vert J - J\st \Vert_{C^1(B)} \le \epsi.
\eqno(A2.2.2)$$
[*Step 2. Reduction to the case of holomorphic cusp-points*]{}.
Recall that in [*Lemma 1.4.1*]{} a natural diffeomorphism between the space $\jj$ of orthogonal complex structures in $\rr^4$ and the unit sphere $S^2\deff \{\,(c_1, c_2, s)\in \rr^3 : c_1^2 + c_2^2
+ s^2 =1\, \}$ was established. Define the function $\Phi :\jj \to \rr^2$ by setting $\Phi(J)
=(c_1, c_2)$. The map $\Phi$ defines a diffeomorphism between the upper half-sphere in $S^2$ and the unit disk $\rr^2$ such that the north pole $(0,0,1)\in S^2$ corresponds to the center of the disk.
Define the function $f:B \to \rr^2$ setting $f(w)\deff \Phi(J(w))$. Due to $(A2.2.3)$, $f$ parameterizes the given almost complex structure $J$, $\Vert f \Vert_{C^1(B)} \le C\epsi$, and $f(0)=0$. Fix a cut-off function $\chi$ in $B$ such that $0\le\chi\le1$, $\chi\bigm|_{B(1/2)}\equiv 1$, $\supp \chi \comp B$, and $\Vert d\chi \Vert_{C^1(B)} \le 3$. For $0\le t \le1$ and $0<\sigma<1$ set $f_{\sigma, t}(w) \deff (1-t\chi(w/\sigma))\cdot f(w)$ and define $J_{\sigma, t}\deff \Phi^{-1}(f_{\sigma, t})$. One can easily see that $\Vert f_{\sigma, t} \Vert_{C^1(B)} \le 4\cdot\Vert f \Vert_{C^1(B)}$, and consequently $\Vert J_{\sigma, t} - J\st \Vert_{C^1(B)} \le C_1\cdot
\Vert J - J\st \Vert_{C^1(B)}\le C_1\cdot \epsi$. Here the constant $C_1$, as well as the constants $C_2,\ldots, C_5$ below in the proof, are independent of $\epsi$, $t$ and $\sigma$. For fixed $\sigma$, the curve $J_{\sigma, t}$, $0\le t\le1$ is a homotopy between $J\equiv J_{\sigma, 0}$ and an almost complex structure $J_\sigma \deff J_{\sigma, 1}$ such that $J_{\sigma, t} \equiv J$ in $B\backslash B(\sigma)$ and $J_\sigma \equiv J\st$ in $B({\sigma\over2})$. Moreover, we have $$\Vert J_{\sigma, t} - J\st \Vert_{C^1(B)} +
\Vert \msmall{ dJ_{\sigma, t} \over dt} \Vert_{C^1(B)} \le C_2\cdot \epsi.
\eqno(A2.2.3)$$
Now we fix some $p>p'>2$ and set $\alpha\deff {2\over p'} -{2\over p}$. Due to (4.2.3) we can apply [*Lemma 3.3.1*]{} to the map $u_j$ with $\nu_j=\mu_j$, $v_j\equiv0$, and with the curve of $C^1$-smooth almost complex structures $J_{\sigma, t}$. As result we obtain a curve of maps $u_{j, \sigma, t} = u_j + z^{\mu_j}w_{j, \sigma, t}$ with $w_{j, \sigma, t}(0)
=0$ and $\Vert w_{j, \sigma, t} \Vert_{L^{1,p}(\Delta)} \le C_3\cdot \Vert u_j
\Vert_{L^{1,p}(B)}$. The condition $\supp \dot J_{\sigma, t} \subset B(\sigma)$ together with [*Corollary 3.1.3*]{} provide that $\supp (\dot J_{\sigma, t}
\scirc du_{j, \sigma, t}) \subset \Delta(\rho)$ with $\rho\sim\sigma^{1/\mu_j}$. Due to (3.3.8) and the Hölder inequality we obtain $$\left\Vert z^{-\mu_j}\cdot \dot J_{\sigma, t} (u_{j, \sigma, t})
\scirc du_{j, \sigma, t} \right\Vert_{L^{p'}_{\vphantom{1}}(\Delta)}
\le C_4 \cdot \sigma^{\alpha/\mu_j}\cdot \Vert u_j \Vert_{L^{1,p}(\Delta)}^2
\cdot \Vert \dot J_{\sigma, t} \Vert_{C^1(B)},
\eqno(A2.2.4)$$ and consequently
$$\Vert w_{j, \sigma, 1} \Vert_{L^{1,p'}_{\vphantom{1}}(\Delta)}
\le C_5 \cdot \sigma^{\alpha/\mu_j}\cdot \Vert u_j \Vert_{L^{1,p}(\Delta).}^2
\eqno(A2.2.5)$$ [*Lemma 3.2.3*]{} provides that for $\sigma$ small enough the $J_\sigma$-holomorphic maps $u_{j, \sigma} \deff u_{j, \sigma, 1}$ are transversal to all spheres $S^3_r$, $0<r<{1\over2}$, have no self-intersection points in $B({1\over2}) \backslash B({1\over4})$, and the Bennequin index of $u_{j, \sigma}(\Delta)\cap S^3_r$, ${1\over4}<r<{1\over2}$, coincides with the one of $\Gamma_j\cap S^3_r$, $0<r<1$. Moreover, we may match $u_{j, \sigma}$ to the rest of $M$, changing in an appropriate way the almost complex structure $J$, see [*Lemma A2.3.1*]{}.
As a result, we conclude that for appropriate small enough neighborhoods $U_1
\comp U$ of any cusp-point $p\in M$ there exist a perturbed almost complex structure $\tilde J$ and $\tilde J$-holomorphic curve $\widetilde M$, which is parameterized by $\tilde u:
\bigsqcup_{j=1}^d S_j \longrightarrow X$ and has the following properties:
$\tilde J$ and $\widetilde M$ coincide on $X\backslash U$ with $J$ and $M$ correspondingly;
$U_1$ is a ball centered in $p$ and $\widetilde M \cap U_1$ is obtained from $M\cap U_1$ by perturbing its components in the above described way;
the (possibly reducible) curve $\tilde \gamma \deff \d U_1\cap \wt M$ is isotopic to $\gamma \deff \d U_1\cap M$, in particular all the corresponding components $\tilde \gamma_j$ of $\tilde \gamma$ and $\gamma_j$ of $\gamma$ have the same Bennequin index, and the linking number $l(\tilde \gamma_i,
\tilde \gamma_j)$ is equal to $l(\gamma_i,\gamma_j)$;
$\tilde u$ is homotopic to $u$;
the cusp-points of $\widetilde M$ coincide with the ones of $M$ and $\tilde\delta + \tilde\varkappa= \delta +\varkappa$,
$\tilde J$ is integrable in a neighborhood.
The last equality of follows from due to [*Lemma 4.1.2*]{}. Thus the formulas (4.2.1) for $M$ and $\widetilde M$ are equivalent.
[*Step 3. Final reduction to the case of an immersed curve*]{}.
This step is rather obvious and uses the following fact, shown in \[Bn\]. Let $B$ be the unit ball in $\cc^2$ and let $\Gamma_0$ be an irreducible holomorphic curve in $B$, which is transversal to $\d B$ and is defined as a zero-divisor of a holomorphic function $f$. Then for any sufficiently small nonzero $\epsi\in\cc$ the curve $\Gamma_\epsi$, defined as the zero-divisor of the function $f+\epsi$, are smooth and of the same genus $g$. Moreover, all $\Gamma_\epsi$ are transversal to $\d B$, and the Bennequin index of $\gamma_\epsi \deff \Gamma_\epsi \cap \d B$ equals $2g-1$. In particular, the conductor of a single cusp-point of a holomorphic curve in $B$ can be defined as a genus of general small perturbation to a smooth curve.
In general, let $M$ be a $J$ - complex curve in $X$, such that $J$ is integrable in a neighborhood of every cusp-point of $M$. One can now see that we can perturb $J$ and $M$ to an almost complex structure $\tilde J$ and a $\tilde J$-holomorphic curve $\widetilde M$, satisfying the conditions [*i)*]{}–[*iv)*]{} from above, and the desired condition
[*v$'$)*]{} $\widetilde M$ has no cusp-points, $\sum\tilde g_j =\sum g_j +
\varkappa$, $\tilde \delta =\delta$.
An important corollary of the proof of [*Theorem 3*]{} is the estimate from below of the conductor number of a cusp point.
Corollary A2.2.2. *Let $J$ be an almost complex structure in the unit ball $B\subset \cc^2$ with $J(0)=J\st$, and let $u:\Delta \to B$ be a $J$ - complex curve with the cusp-point $u(0)=0$. Then $\varkappa_0$ is an integer, $\varkappa_0 \ge \ord_0du$, or equivalently, for all sufficiently small $r>0$ the Bennequin index of $\gamma_r\deff u(\Delta) \cap S^3_r$ is odd and satisfies the inequality $$b(\gamma_r)\ge 2\cdot \ord_0du - 1. \eqno(A2.2.6)$$*
**Proof. Rescaling $u$ as in the [*Step 1*]{} of the proof of [*Theorem 3*]{} we may assume that $\Gamma\deff u(\Delta)$ has no nodes and cusps, excepting $0\in \Delta$ , and is transversal to all spheres $S^3_r$, $0<r<1$, so that the Bennequin index $b(\gamma_r)$ is the same and equals $2\varkappa_0 -1$. We may also assume that for the rescaled almost complex structure $J$ the estimate $\Vert J - J\st \Vert_{C^1(B)} \le \epsi$ with the appropriate $\epsi$ is fulfilled. Applying [*Step 2*]{} with a sufficiently small $\sigma$, we can deform $u$ into a $\tilde J$ holomorphic map $\tilde u$ which has the following properties:**
$\widetilde \Gamma \deff \tilde u(\Delta)$ is transversal to all spheres $S^3_r$, $0<r<1$;
$\widetilde \Gamma$ has no self-intersection points in $B\backslash
B({1\over2})$, and the Bennequin index of $\tilde\gamma_r \deff \widetilde
\Gamma\cap S^3_r$, ${1\over2}<r<1$, coincides with the one of $\gamma_r$;
$\tilde J$ coincides with $J$ in $B\backslash B({1\over2})$, is integrable in the neighborhood of $0\in B$, and $\ord_0 d\tilde u =
\ord_0 du $.
Now the corollary follows from [*Lemmas A2.1.2*]{} and the fact that for [*integrable*]{} complex structures the same statement is true, see \[Bn\].
Another modification of the proof of [*Theorem 3*]{} leads to the following consequence.
Lemma A2.2.3 *Let $J$ be a $C^1(X)$-almost complex structure on $X$ and $M\subset X$ a compact $J$ holomorphic curve parameterized by $u: S \to X$. Then*
$u$ can be $L^{2,p}$-approximate by $J_n$-holomorphic immersions $u_n: S \to X$ with $J_n \longrightarrow J$ in $C^1(X)$.
there exists a $C^1(X)$-approximation $J_n$ of $J$ and a sequence of $J_n$-holomorphic imbedded curves $M_n$ which converge to $M$ in the Gromov topology.
Proof. We do not need this result for the purpose of this paper; therefore, we give only a sketch of the proof.
In the first step, one applies the rescaling procedures in order to find appropriate small neighborhoods of the cusp-points of $M$.
In the second step, one applies [*Lemma 3.1.3*]{} to the chosen neighborhoods, taking $\nu=1$, $v$ sufficiently small, and $J$ unchanged. After matching procedure we obtain the statement [*i)*]{}. To obtain the statement , one must first deform all the nodes of $M$ into simple transversal ones and then find an appropriate small neighborhood $U$ of every node. In some complex coordinates $(w_1, w_2)$ in $U$, the curve $M$ is defined by the equation $w_1\cdot w_2=0$. It remains to replace a node $M\cap U$ by a “small handle” $M_\epsi\deff \bigl\{\,
(w_1, w_2) \in U : w_1\cdot w_2=\epsi\,\}$ with $\epsi$ sufficiently small and to use the matching procedure once more.
*A2.3. A matching Structures Lemma.*
Let $B(r)$ be a ball of radius $r$ in $\rr^4$ centered at zero, and $J$ a $C^1$-smooth almost-complex structure on $B(2)$, $J(0)=J\st$. Further, let $M = u(\Delta )$ be a closed primitive $J$ - complex disk in $B(2)$ such that $u(0)=0$ and $M$ transversally meet $S^3_r$ for $r\ge 1/2$. Here $S^3_r = \d B(r) $ and transversality are understood with respect to both $TS^3_r$ and $F_r$, see [*paragraph 4.1.*]{}
By $B(r_1,r_2)$ we shall denote the spherical shell $\{ x\in \rr^4:r_1<\Vert
x\Vert <r_2 \}$. In the lemma below denote by $D_{\delta }$ the pre-image of $B(1+\delta )$ by $u$.
Lemma A2.3.1. *For any positive $\delta >0$ there exists an $\epsi >0$ such that if an almost complex structure $\tilde J$ in $B(1+\delta )$ and a closed $\tilde J$-holomorphic curve $\tilde u : D_{\delta }\to B(1+\delta )$ satisfy $\Vert \tilde J - J\Vert_{C^1(\bar B(1+\delta ))} < \epsi $ and $\Vert \tilde u - u\Vert_{L^{1,p}(\bar D_{\delta })} < \epsi $, then there exists an almost-complex structure $J_1$ in $B(2)$ and $J_1$-holomorphic disk $M_1$ in $B(2)$ such that:*
a\) $J_1\ogran_{B(1-\delta )} = \tilde J\ogran_{B(1-\delta )}$ and $J_1\ogran_{B(1+\delta ,2)} = J\ogran_{B(1+\delta ,2)}$.
b\) $M_1\ogran_{B(1-\delta )} = \tilde u(D_{\delta })\cap B(1-\delta )$ and $M_1\cap B(1+\delta ,2) = M\cap B(1+\delta ,2)$.
Proof.
We have chosen the parameterization of $M$ to be primitive. Thus, $u$ is an imbedding on $D_{-\delta ,\delta } = u^{-1}(B_{1-\delta ,1+\delta })$. Let us identify a neighborhood $V$ of $u(D_{-\delta ,\delta })$ in $B_{1-
\delta ,1+\delta }$ with the neighborhood of the zero-section in the normal bundle $N$ to $u(D_{-\delta ,\delta })$. Now $\tilde u\ogran_{D_{-\delta ,\delta
}}$ can be viewed as a section of $N$ over $u(D_{-\delta ,\delta })$ which is small [*i.e.,*]{} contained in $V$. Using an appropriate smooth function $\phi $ on $D_{-\delta ,\delta }$ (or equivalently on $u(D_{-\delta ,\delta} )$), $\phi
\ogran_{B(1-\delta )\cap D_{\delta }}\equiv 1$, $\phi\ogran_{\partial D_{\delta
}
}\equiv 0$, $0\le \phi \le 1$ we can glue $\tilde u$ and $u$ to obtain a *symplectic surface $M_1$ which satisfies (b).*
Patching $J$ and $\tilde J$ and simultaneously making $M_1$ complex can be done in an obvious way.
\#1
[Chapter II. Compactness Theorem.]{}
The goal of this chapter is to give a proof of the Gromov Compactness Theorem for continuously varying almost complex structures and for the sequences of complex curves parameterized by some fixed real surface.
More precisely, we consider a sequence $J_n$ of continuous (of class $C^0$) almost complex structures on a manifold $X$ which converge uniformly to some $J_\infty$, again of class $C^0$. Further, let $(C_n,j_n)$ be a sequence of Riemann surfaces with boundaries of fixed topological type. This means that each $(C_n,\d C_n)$ can be parameterized by the same real surface $(\Sigma, \d\Sigma ) $ (see §2 for details). Denote by $\delta_n:\Sigma
\to C_n$ some parameterizations. However, the complex structures $j_n$ on $C_n$ may vary in an arbitrary way. Finally, let some sequence of $(j_n,J_n)$-holomorphic maps $u_n:C_n\to X$ be given.
Theorem 2.1.
*If the areas of $u_n(C_n)$ are uniformly bounded (with respect to some fixed Riemannian metric on $X$) and the structures $j_n$ do not degenerate at the boundary (see [*Definition 1.7*]{}), then there exists a subsequence, denoted $(C_n, u_n)$, such that*
*1) *$(C_n,j_n)$ converge to some nodal curve $(C_\infty,j_\infty)$ in an appropriate completion of the moduli space of Riemann surfaces of given topological type, i.e., there exists a parameterization map $\sigma_\infty:
\Sigma \to C_\infty$ by the same real surface $\Sigma$;**
*2) *one can choose a new parameterizations $\sigma_n$ of $C_n$ in such a way that $\sigma_n^*j_{C_n}$ will converge to $\sigma_\infty^*j_{C_\infty}$ in the $C^\infty$-topology on compact subsets outside of the finite set of circles on $\Sigma$, which are pre-images of the nodal points of $C_\infty$ by $\sigma_\infty$, where $j_{C_n}$ denote the complex structures on $C_n$;**
[*3)*]{} the maps $u_n\scirc \sigma_n$ converge, in the $C^0$-topology on $\Sigma$ and in the $L^{1,p}_\loc$-topology (for all $p<\infty$) outside of the pre-images of the nodes of $C_\infty$ to a map $u_\infty \scirc
\sigma_\infty$ such that $u_\infty$ is a $(j_\infty, J_\infty) $-holomorphic map $C_\infty\to X$.
This description of convergence is precisely the one given by Gromov in \[G\]. Our statement is slightly more general in two directions. First, we consider not only the case of closed curves, but also the case where $C_n$ is open and of a “fixed topological type”. Second, we note that the Gromov compactness theorem is still valid for continuous and continuously varying almost complex structures. This could have an interesting application, because now one can consider $C^0$-perturbations of complex structures being assured that at least the compactness theorem still holds true. For the definitions involved and the formal statement see §4.1 and [*Theorem 4.1.1*]{}.
We also prove the compactness theorem for curves with boundaries on a totally real submanifold. This “boundary” result needs appropriate generalizations of all “inner” constructions and estimates. The related considerations are shown in Appendix III.
Another result of this paper, which we would like to mention in the introduction, is the improvement of the removable singularity theorem in two directions.
First we prove (see [*Corollary 5.2.1*]{}) the following generalization of the removability theorem for the point singularity.
Theorem 2.2. [*If the area of the image of $J$-holomorphic map $u:(\check\Delta, J\st)\to (X,J)$ from the punctured disk into a compact almost complex manifold “is not growing too fast", i.e., if $\area(u(R_k))\le \eps $ for all annuli $R_k \deff \{ z\in \cc :{1\over e^{k+1}}\le | z| \le {1 \over
e^k}\}$ with $k>\!>1$, then $u$ extends to the origin.* ]{}
The positive constant $\eps $ here depends on the Hermitian structure $(J,h)$ of $X$. This theorem, under the stronger assumption $\sum_k
\area(u(R_k)) \equiv \norm{du}^2_{L^2(\check\Delta)} < \infty$, was proved by Sacks and Uhlenbeck \[S-U\] for harmonic maps, and by Gromov \[G\] for $J$-holomorphic maps.
This fact (which is proved here for continuous $J$’s) is new even in the integrable case. In fact, it measures the “degree of non-hyperbolicity" (in the sense of Kobayashi) of $(X,J,h)$.
Another (see [*Corollary A3.3.5*]{}) is a generalization of Gromov’s result about removability of the boundary point singularity, see \[G\]. An improvement is that the statement remains valid also when one has [*different*]{} boundary conditions to the left and to the right of a singular point. Let us explain this in more detail.
Define the (punctured) half-disk by setting $\Delta^+ \deff \{ z\in \Delta :
\im(z)>0\}$ and $\check\Delta^+ \deff \Delta^+\bs \{0\}$. Define $I_- \deff
]{-1},0[ \subset \d\check\Delta^+$ and $I_+ \deff ]0,+1[ \subset \d\check
\Delta^+$. Let a $J$-holomorphic map $u:(\check\Delta^+, J\st)\to (X,J)$ be given, where $J$ is again continuous. Suppose further that $u(I_+)\subset
W_+$ and $u(I_-) \subset W_-$, where $W_+,W_-$ are totally real submanifolds of dimension $n= \half \dim_\rr X$ and intersect transversally.
Theorem 2.3. [*There is an $\eps^b>0$ such that if for all half-annuli $R^+_k\deff\{ z\in \Delta^+ :e^{-(k+1)} \le | z| \le e^{-k} \}$ one has $\area(u(R^+_k))\le \eps^b$, then $u$ extends to the origin $0\in
\Delta^+$ as an $L^{1,p}$-map for some $p>2$.* ]{}
As in the “inner” case, the necessary condition is weaker than the finiteness of energy. But unlike “inner" and smooth boundary cases, it is possible that the map $u$ in the last statement is $L^{1,p}$-regular in the neighborhood of the “corner point” $0\in \Delta^+$ only for some $p>2$. For example, the map $u(z) = z^\alpha$ with $0<\alpha <1$ satisfies the totally real boundary conditions $u(I_+) \subset \rr$, $u(I_-) \subset e^{\alpha\pi\isl}\rr$ and is $L^{1,p}$-regular only for $p< p^* \deff {2\over1-\alpha}
\cdot$ Note also that by the Sobolev imbedding $L^{1.p}\subset C^{1,\alpha }$ with $\alpha = 1-{2\over p}$, $u$ extends to zero at least continuously. Thus $u(0)\in W_+ \cap W_-$.
One can see such a point $x$ as a [*corner point*]{} for a corresponding complex curve. A typical example appears in symplectic geometry when one takes Lagrangian submanifolds as boundary conditions.
The organization of the chapter is the following. In Lecture 4 we present, for the convenience of the reader, the basic notions concerning the topology on the space of stable curves and complex structure on the Teichmüller space of Riemann surfaces with boundary. In Lecture 5 we give the necessary a priori estimates for the inner case, and the proof of [*Theorem 2.1*]{}, related to curves with free boundary. This includes the case of closed curves. In Appendix III we consider curves with totally real boundary conditions, obtain necessary a priori estimates at a “totally real boundary”, and prove the compactness theorem for such curves. In particular, we prove [*Theorem 2.3*]{} there.
[Lecture 4]{} [Space of Stable Curves]{}
[4.1. Stable Curves and Gromov Topology.]{}
Before stating the Gromov compactness theorem, we need to introduce an appropriate category of complex curves. Since the limit of a sequence smooth curves can be singular,i.e., a cusp-curve in Gromov’s terminology, we need to allow certain types of singularities of curves. On the other hand, it is desirable to have singularities as simple as possible.
A similar problem appears in looking for a “good” compactification of moduli spaces $\calm_{g,m}$ of abstract complex smooth closed curves of genus $g$ with $m$ marked points. The Deligne-Mumford compactification $\barr \calm
_{g, m}$, obtained by adding [*stable curves*]{}, gives a satisfactory solution to this problem and suggests a possible way of generalizing to other situations. In fact, the only singularity type one should allow are nodes, or nodal points. An appropriate notion for curves in a complex algebraic manifold $X$ was introduced by Kontsevich in \[K\]. Our definition of stable curves over $(X,J)$ is simply a translation of this notion to almost complex manifolds. The change of terminology from [*stable maps*]{} to [*stable curves over $(X,J)$*]{} is motivated by the fact that we want to consider our objects as curves rather than maps.
Recall that a [*standard node*]{} is the complex analytic set $\cala_0 \deff
\{ (z_1,z_2)\in \Delta^2 : z_1\cdot z_2 =0\}$. A point on a complex curve is called a [*nodal point*]{} if it has a neighborhood biholomorphic to the standard node.
Definition 4.1.1. [*A [*nodal curve*]{} $C$ is a complex analytic space of pure dimension 1 with only nodal points as singularities*]{}.
In other terminology, nodal curves are called [*prestable*]{}. We shall always suppose that $C$ is connected and has a “finite topology”, i.e., $C$ has finitely many irreducible components, finitely many nodal points, and that $C$ has a smooth boundary $\d C$ consisting of finitely many smooth circles $\gamma_i$, such that $\barr C \deff C \cup \d C$ is compact.
Definition 4.1.2.
*We say that a real connected oriented surface with boundary $(\Sigma, \d\Sigma)$ [*parameterizes*]{} a complex nodal curve $C$ if there is a continuous map $\sigma :\barr\Sigma \to \barr C$ such that*
if $a\in C$ is a nodal point, then $\gamma_a = \sigma\inv(a)$ is a smooth imbedded circle in $\Sigma \bs \d \Sigma $, and if $a\not= b$ then $\gamma_a \cap \gamma_b= \emptyset$;
$\sigma :\barr\Sigma \bs \bigcup_{i=1}^N\gamma_{a_i}\to \barr C \bs \{
a_1,\ldots ,a_N\} $ is a diffeomorphism, where $a_1,\ldots ,a_N$ are the nodes of $C$.
=12.5pt plus 1.5pt
Note that such a parameterization is not unique: if $g:\barr\Sigma \to
\barr\Sigma$ is any orientation preserving diffeomorphism, then $\sigma \scirc
g: \barr\Sigma \to \barr C$ is again a parameterization.
A parameterization of a nodal curve $C$ by a real surface can be considered as a method of “smoothing” of $C$. An alternative method of “smoothing” — the normalization — is also useful for our purposes.
Consider the normalization $\hat C$ of $C$. Mark on each component of this normalization the pre-images (under the normalization map $\pi_C: \hat C \to
C$) of nodal points of $C$. Let $\hat C_i$ be a component of $\hat C$. We can also obtain $\hat C_i$ by taking an appropriate irreducible component $C_i$, replacing nodes contained in $C_i$ by pairs of disks with marked points, and marking remaining nodal points. Since it is convenient to consider components in this form, we make the following
Definition 4.1.3. *A component $C'$ *of a nodal curve $C$ is a normalization of an irreducible component of $C$ with marked points selected as above.**
This definition allows us to introduce the Sobolev and Hölder spaces of functions and (continuous) maps of nodal curves. For example, a continuous map $u: C \to X$ is Sobolev $L_\loc ^{1,p}$-smooth if its restrictions on every component of $C$ is $L_\loc^{l,2}$-smooth as well. The most interesting case is, of course, the one of continuous $L_\loc ^{1,2}$-smooth maps. In this case the energy functional $\norm{du}^2_{L^2(C)}$ is defined. The definition of the energy $\norm{du}^2
_{L^2(C)}$ involves Riemannian metrics on $X$ and $C$ which are supposed to be fixed.
Let $C$ be a nodal curve and $(X,J)$ an almost complex manifold with continuous almost complex structure $J$.
Definition 4.1.4. [*A continuous map $u:C \to X$ is $J$-holomorphic if $u\in L^{1,2}_\loc(C,X)$ and $$du_x + J\scirc du_x\scirc j_C=0
\eqno(4.1.1)$$ for almost all $x\in C$. Here $j_C$ denotes the complex structure on $C$.* ]{}
Recall that the area of a $J$-holomorphic map is defined as $$\area(u(C)) \deff \norm{du}^2_{L^2(C)}.$$
See the end of §1.3. We shall show later that every $J$-holomorphic $u$ is, in fact, $L^{1,p}_\loc(C,X)$-smooth for all $p< \infty$, see [*Corollary 2.4.2*]{}. The following notion of stability was introduced by Kontsevich in \[K\].
Definition 4.1.5. [*A [*stable curve over $(X,J)$*]{} is a pair $(C,u)$, where $C$ is a nodal curve and $u:C\to X$ is a $J$-holomorphic map, satisfying the following condition. If $C'$ is a compact component of $C$, such that $u$ is constant on $C'$, then there exist finitely many biholomorphisms of $C'$ which preserve the marked points of $C$.* ]{}
Remark. One can see that the stability condition is nontrivial only in the following cases:
[*1)*]{} some component $C'$ is biholomorphic to $\cc\pp^1$ with 1 or 2 marked points; in this case $u$ should be non-constant on any such component $C'$;
[*2)*]{} some irreducible component $C'$ is $\cc\pp^1$ or a torus without nodal points.
Since we consider only connected nodal curves, case [*2)*]{} can happen only if $C$ is irreducible, so that $C'=C$. In this case $u$ must be non-constant on $C$.
Now we are going to describe the Gromov topology on the space of stable curves over $X$ introduced in \[G\]. Let a sequence $J_n$ of continuous almost complex structures on $X$ be given. Suppose that $\{J_n\}$ converges to $J_\infty$ in the $C^0$-topology. Furthermore, let $(C_n,
u_n)$ be a sequence of stable curves over $(X, J_n)$ such that all $C_n$ are parameterized by the same real surface $\Sigma$.
Definition 4.1.6.
*We say that $(C_n,u_n)$ [*converges to a stable $J_\infty$-holomorphic curve $(C_\infty,u_\infty)$ over $X$*]{} if the parameterizations $\sigma_n: \barr\Sigma \to \barr C_n$ and $\sigma_\infty:
\barr \Sigma \to \barr C_\infty$ can be chosen in such a way that the following hold:*
$u_n\scirc \sigma_n$ converges to $u_\infty\scirc \sigma_\infty$ in the $C^0( \Sigma, X)$-topology;
if $\{ a_k \}$ is the set of nodes of $C_\infty$ and $\{\gamma_k\}$ are the corresponding circles in $\Sigma$, then on any compact subset $K\comp
\Sigma \bs \cup_k \gamma_k$ the convergence $u_n\scirc \sigma_n \to u_\infty
\scirc \sigma_\infty$ is $L^{1,p}(K, X)$ for all $p< \infty$;
for any compact subset $K\comp \barr\Sigma \bs \cup_k\gamma_k$ there exists $n_0=n_0(K)$ such that $ \sigma_n(K) \subset C_n\setminus \{ nodes\} $ for all $n\ge n_0$ and the complex structures $\sigma_n^*j_{C_n}$ converge smoothly to $\sigma_0^*j_{C_0}$ on $K$;
the structures $\sigma_n^*j_{C_n}$ are constant in $n$ near the boundary $\d\Sigma$.
The reason for introducing the notion of a curve stable over $X$ is similar to the one for the Gromov topology. We are looking for a completion of the space of smooth imbedded complex curves which has “nice” properties, namely: 1) such a completion should contain the limit of some subsequence of every sequence of smooth curves, bounded in an appropriate sense; 2) the same should also hold for every sequence in the completed space; 3) such a limit should be unique. Gromov’s compactness theorem insures us that the space of curves stable over $X$ enjoys these nice properties.
Condition is trivial if $\Sigma $ is closed, but it is important when one considers the “free boundary case”, i.e., when $\Sigma$ (and thus all $C_n$) are not closed and no boundary condition is imposed. However, we would like to point out that in our approach the “free boundary case” is essentially involved in the proof of the compactness theorem also in the case of closed curves. On the other hand, in the case of curves with boundary on totally real submanifolds such a condition is unnecessary.
Recall that a complex annulus $A$ has a conformal radius $R>1$ if $A$ is biholomorphic to $A(1,R) \deff \{ z\in \cc \,:\, 1 <|z| < R \}$. An annulus $A$ is said to be [*adjacent*]{} to a circle $\gamma$, if $\gamma$ is one of its boundary components.
Definition 4.1.7. [*Let $C_n$ be a sequence of nodal curves, parameterized by the same real surface $\Sigma$. We say that complex structures on $C_n$ [*do not degenerate near the boundary*]{}, if there exists $R>1$ such that for any $n$ and any boundary circle $\gamma_{n, i}$ of $C_n$ there exists an annulus $A_{n,i} \subset C_n$ adjacent to $\gamma_{n, i}$ such that all $A_{n,i}$ are mutually disjoint, do not contain nodal points of $C_n$, and have the same conformal radius $R$.*]{}
Since the conformal radii of all the $A_{n, i}$ are the same, we can identify them with $A(1,R)$. This means that all changes of complex structures of $C_n$ take place away from the boundary. The condition is trivial if $C_n$ and $\Sigma$ are closed, $\d\Sigma = \d C_n = \emptyset$.
Remark. Changing our parameterizations $\sigma_n: \Sigma \to C_n$, we can suppose that for any $i$ the pre-image $\sigma_n\inv (A_{n,i} )$ is the same annulus $A_i$ independent of $n$.
Now we state our main result. Fix some Riemannian metric $h$ on $X$ and some $h$-complete set $A\subset X$.
Theorem 4.1.1.
*Let $\{(C_n,u_n)\}$ be a sequence of stable $J_n$-holomorphic curves over $X$ with parameterizations $\delta_n: \Sigma \to
C_n$. Suppose that*
[*a)*]{} $J_n$ are continuous almost complex structures on $X$, $h$-uniformly converging to $J_\infty$ on $A$ and $u_n(C_n)\subset A$ for all $n$;
[*b)*]{} there is a constant $M$ such that $\area [u_n (C_n)]\le M$ for all $n$;
[*c)*]{} complex structures on $C_n$ do not degenerate near the boundary.
Then there is a subsequence $(C_{n_k},u_{n_k})$ and parameterizations $\sigma
_{n_k}: \Sigma \to C_{n_k}$ such that $(C_{n_k}, u_{n_k}, \sigma_{n_k})$ converges to a stable $J_\infty$-holomorphic curve $(C_\infty, u_\infty,
\sigma_\infty)$ over $X$.
Moreover, if the structures $\delta_n^*j_{C_n}$ are constant on the fixed annuli $A_i$, each adjacent to a boundary circle $\gamma_i$ of $\Sigma$, then the new parameterizations $\sigma_{n_k}$ can be taken equal to $\delta_{n_k}$ on some subannuli $A'_i \subset A_i$, also adjacent to $\gamma_i$.
Remarks. 1. In the proof, we shall give a precise description of convergence with estimates in neighborhoods of the contracted circles $\gamma_i$. The convergence of curves with boundary on totally real submanifolds will be studied in §5.5. 2. In applications, one uses a generalized version of the Gromov compactness theorem for nodal curves with a marked point. This version is an immediate consequence of [*Theorem 1.1*]{} due to the following construction. Consider a nodal curve $C$ and let a $J$-holomorphic map $u: C \to X$. Let $\bold
x \deff \{x_1,\ldots,x_m\}$ be the set of marked points on $C$ which are supposed to be distinct from the nodal points of $C$. Define a new curve $C^+$ as the union of $C$ with disks $\Delta_1,\ldots, \Delta_m$ such that $C
\cap \Delta_i = \{x_i\}$ and any $x_i$ becomes a nodal point of $C^+$. Extend $f$ to a map $f^+: C^+ \to X$ by setting $f^+\ogran_{\Delta_i}$ to be constant and equal to $f(x_i)$. An appropriate definition of stability, used for triples $(C, \bold x, f)$, is equivalent to stability of $(C^+, f^+)$. Similarly, the Gromov convergence $(C_n, \bold x_n, f_n) \to (C_\infty, \bold x_\infty, f_\infty)$ is equivalent to the Gromov convergence $(C^+_n, f^+_n) \to (C^+ _\infty, f^+_\infty)$. Thus the Gromov compactness for curves with marked points reduces to the case considered in our paper. However, we shall consider curves with marked points as well.
plus 30pt
[4.2. Fenchel-Nielsen Coordinates on the Space of Nodal Curves.]{}
In the rest of this section we shall describe topology and conformal geometry of nodal curves and compute the set of moduli parameterizing deformations of a complex structure. As a basic reference we use the book of Abikoff \[Ab\].
Let $C$ be a complex nodal curve parameterized by $\Sigma$.
Definition 4.2.1.
*A component $C'$ of $C$ is called [*nonstable*]{} if one of the following two cases occurs:*
[ 1)]{} $C'$ is $\cc\pp^1$ and has one or two marked points;
[ 2)]{} $C'$ is $\cc\pp^1$ or a torus and has no marked points.
This notion of stability of abstract closed curves is due to Deligne-Mumford, see \[D-M\]. It was generalized by Kontsevich \[K\] for the case of maps $f:C
%%%%%%%%%%%%%%%%%%Question%%%%%%%%%%%%%%%%%%%
\to X$, i.e., for curves over $X$ in our terminology. As was already noted, the last case can happen only if $C=C'$. Strictly speaking, this case should be considered separately. However since such considerations require only obvious changes, we just skip them and suppose that case [*2)*]{} does not occur.
Our first aim is to analyze the behavior of complex structures in the sequence $(C_n,u_n)$ of $J_n$-holomorphic curves stable over $X$ with uniformly bounded areas, which are parameterized by the same real surface $\Sigma$. At the moment, the uniform bound on the area of $u_n(C_n)$ is needed only to show that the number of components of $C_n$ is bounded. Passing to a subsequence, we can assume that all $C_n$ are homeomorphic. This reduces the problem to a description of complex structures on a fixed nodal curve $C$.
To obtain such a description, it is useful to cut the curve into pieces where the behavior of a complex structure is easy to understand. Such a procedure is a [*partition into pants*]{} which is well- known in the theory of moduli spaces of complex structures on curves, see, e.g., \[Ab\], p.93. Here we shall make use of a related but slightly different procedure. Namely, we shall choose a special covering of $\Sigma$ instead of its partition. Further, as blocks for our construction we shall use not only pants, but also disks and annuli. The reasons are that, first, the considered curves can have unstable components and, second, it is convenient to use annuli for a description of the deformation of the complex structure on curves. We start with
Definition 4.2.2. *An annulus $A$ *on a real surface or on a complex curve is a domain which is diffeomorphic (resp. biholomorphic) to the standard annulus $A(r,R) \deff \{ z\in \cc \;:\; r<|z|<R \}$ such that its boundary consists of smoothly imbedded circles. *Pants *(also called a *pair of pants*) on a real surface or on a complex curve is a domain which is diffeomorphic to a disk with 2 holes.******
The boundary of pants consists of three components, each of them being either a smoothly imbedded circle or a point. This point can be considered as a puncture of pants or as a marked point. An annulus or pants is [*adjacent to a circle $\gamma$*]{} if $\gamma$ is one of its boundary components.
Let $C$ be a nodal curve parameterized by a real surface $\Sigma$. We shall associate with every such curve $C$ a certain graph $\Gamma_C$ which determines $C$ topologically in a unique way. In fact, $\Gamma_C$ will also determine a decomposition of some components of $C$ into pants.
By definition, a compact component $C'$ is stable if it contains only a finite number of automorphisms preserving marking points. In this case $C'
\bs\mapo$ possesses a unique so-called intrinsic metric.
Definition 4.2.3.
*The [*intrinsic metric*]{} for a smooth curve $C$ with marked points $\{ x_i \}$ and with boundary $\d C$ is a metric $g$ on $C\bs \mapo$ satisfying the following properties:*
$g$ induces the given complex structure $j_C$;
the Gauss curvature of $g$ is constantly -1;
$g$ is complete in a neighborhood of every marked point $x_i$;
every boundary circle $\gamma$ of $C$ is geodesic $g$.
Note that such a metric, if it exists, is unique, see .
Now consider a component $C'$ of $C$ adjacent to some boundary circle of $C$. Then $C' \bs \mapo$ is one of the following:
a) a disk $\Delta$, or
b) an annulus $A$, or
c) a punctured disk $\check\Delta$, or else d) $C'\bs \mapo$ admits the intrinsic metric.
Note that if a component $C'$ is a disk or an annulus (both without marked points), then $C'$ is the whole curve $C$. We shall consider cases a) and b) later. Now we assume, for simplicity, that cases a) and b) do not occur.
Definition 4.2.4 *A component $C'$ of a nodal curve $C$ is called [*non-exceptional*]{} $C' \bs \mapo$ admits the intrinsic metric.*
In particular, nonstable components are exceptional compact ones, and exceptional non-compact components are those of types a)–c) above.
Take some non-exceptional component $C'$ of $C$. There is a so- called maximal partition of $C'\bs \mapo$ into pants $\{ C_1,\ldots C_n\}$ such that all boundary components of these pants are either simple geodesic circles in intrinsic metric or marked points, see \[Ab\]. Let us fix such a partition and mark the obtained geodesic circles on $C'$.
Now let $\sigma: \Sigma \to C$ be some parameterization of $C$. This defines the set $\bfgamma'$ of the circles on $\Sigma$ which correspond to the nodes of $C$. Let $\bfgamma''$ be the set of $\sigma $-pre-images of the geodesics chosen above. Then $\bfgamma\deff \bfgamma' \sqcup \bfgamma''$ forms a system of disjoint “marked” circles on $\Sigma$, which encodes the topological structure of $C$. Now the graph $\Gamma_C$ in question can be constructed as follows.
Define the set $V_C$ of vertices of $\Gamma_C$ to be the set $\{ S_j \}$ of connected components of $\Sigma \bs \cup_{\gamma \in \bfgamma} \gamma =
\sqcup_j S_j$. Any $\gamma \in \bfgamma$ lies between 2 components, say $S_j$ and $S_k$, and we draw an edge connecting the corresponding 2 vertices. Further, any boundary circle $\gamma$ of $\Sigma$ has the uniquely defined component $S_j$ adjacent to $\gamma$. For any such $\gamma$ we draw a [*tail*]{}, i.e., an edge with one end free, attached to vertex $S_j$. Finally, we mark all edges which correspond to the circles $\bfgamma'$, i.e., those coming from the nodes.
Having the graph $\Gamma$, which characterizes uniquely the topological structure of $C$, we are now going to describe the set of parameters defining (uniquely) the complex structure of the curves $C$. This is equivalent to determining the complex structure and marked points on all components of $C$. If such a component $C'$ is a sphere with 1 or 2 marked points or a disk with 1 marked point, then its structure is defined by its topology uniquely up to diffeomorphism. Otherwise, the component $C'$ is non-exceptional. In this case the complex structure and the marked points can be restored by the so- called [*Fenchel-Nielsen coordinates*]{} on the Teichmüller space $\ttt_{g,m,b}$. Recall that the space $\ttt_{g,m,b}$ parameterizes the complex structures on a Riemann surface $\Sigma$ of genus $g$ with $m$ punctures (marked points) and with a boundary consisting of $b$ circles, see \[Ab\].
Let $C$ be a smooth complex curve with marked points of non-exceptional type, so that $C$ admits the intrinsic metric. Fix some parameterization $\sigma:
\Sigma \to C$. Consider the pre-images of the marked points on $C$ as marked points on $\Sigma$ or, equivalently, as punctures of $\Sigma$. Let $C \bs
\mapo = \cup_j C_j$ be a decomposition of $C$ into pants and $\Sigma\bs
\mapo = \cup_j S_j$ the induced decomposition of $\Sigma$.
Let $\{\gamma_i\}$ be the set of boundary circles of $\Sigma$. The boundary of every pants $S_j$ has three components, each of them being either a marked point of $\Sigma$ or a circle. In the last case this circle is either a boundary component of $\Sigma$ or a boundary component of another pants, say $S_k$. In this situation we denote by $\gamma _{jk}$ the circle lying between the pants $S_j$ and $S_k$. Fix the orientation on $\gamma _{jk}$, induced from $S_j$ if $j<k$ and from $S_k$ if $k<j$. For any such circle $\gamma
_{jk}$, fix a boundary component of $S_j$ different from $\gamma_{jk}$ and denote it by $\d_k S_j$. In the same way fix a boundary component $\d_j S_k$. Make similar notations on $C$ using primes to distinguish the circles on $C$ from those on $\Sigma$, i.e., set $\gamma'_i \deff \sigma(\gamma_i)$ and $\gamma'_{jk} \deff \sigma(\gamma_{jk})$.
By our construction, $\gamma'_{jk} = \sigma( \gamma_{jk})$ is a geodesic the intrinsic metric in $C$.
Using the same procedure in $C_k$, we obtain another point $x^*_{k,j} \in
\gamma' _{jk}$. Denote by $\ell_{jk}$ (resp. by $\ell_i$) the intrinsic length of $\gamma'_{jk}$ (resp. of $\gamma'_i \deff \sigma( \gamma_i)$) in $C$. For $j<k$ define $\lambda_{jk}$ as the intrinsic length of the arc on $\gamma' _{jk}$, which starts at $x^* _{j,k}$ and goes to $x^* _{k,j}$ in the direction determined by the orientation of $\gamma _{jk}$. Set $\vartheta
_{jk} \deff {2\pi \lambda _{jk} \over \ell_{jk} }$. We shall consider $\vartheta _{jk}$ as a function of the complex structure $j_C$ on $C$ with values in $S^1 \cong \rr/ 2\pi \zz$.
The parameters $\bfell \deff (\ell_i, \ell_{jk} )$ and $\bfvartheta \deff(\vartheta_{jk})$ are called [*Fenchel-Nielsen coordinates*]{} of the complex structure $j_C$. The reason is that these parameters determine up to isomorphism the complex structure $j_C$ on the smooth complex curve with marked points parameterized by a real surface $\Sigma$. In other words, $(\bfell, \bfvartheta)$ can be considered as coordinates on $\ttt_{g,m,b}$. More precisely, one has the following
Proposition 4.2.2. *Let $\Sigma$ be a real surface of genus $g$ with $m$ marked points and with the boundary consisting of $b$ circles, so that $2g+m +b\ge3$. Let $\Sigma \bs \mapo = \cup_j S_j$ be its decomposition into pants. Then*
[*i)*]{} for any given tuples $\bfell= (\ell_i, \ell_{jk})$ and $\bfvartheta = (\vartheta_{jk})$ with $\ell_i, \ell_{jk} >0$ and $\vartheta_{jk} \in S^1$ there exists a complex structure $j_C$ on $\Sigma$ such that boundary circles of all $S_j$ are geodesic the intrinsic metric on $\Sigma\bs \mapo$ defined by $j_C$ and the given $(\bfell, \bfvartheta)$ are Fenchel-Nielsen coordinates of $j$; moreover, such a structure $j_C$ is unique up to a diffeomorphism preserving the pants $S_j$ and the marked points;
[*ii)*]{} let $C$ be a smooth complex curve with parameterization $\sigma:
\Sigma \to C$ which has $m$ marked points; then there exists a parameterization $\sigma_1: \Sigma \to C$ isotopic to $\sigma$, which maps boundary components and marked points of $\Sigma$ onto the ones of $C$ in prescribed order such that the boundary circles of $\sigma(S_j)$ are geodesic the intrinsic metric on $C\bs \mapo$.
Proof. See \[Ab\].
[4.3. Complex Structure on the Space $\ttt_\Gamma$.]{}
Let $\Sigma$ be a real surface of genus $g$ with $m$ marked points and with the boundary consisting of $b$ circles. Assume that $2g+m +b\ge3$. Then there exists a decomposition of $\Sigma \bs \mapo$ into pants, which is in general not unique. The topological type of such a decomposition can be encoded in graph $\Gamma$, associated with the decomposition. It is constructed in a similar way to that above, but this time we must draw a tail for every marked point, and then mark all those tails on the graph.
Let such a graph $\Gamma$ be fixed. We call two complex structures $J_1$ and $J_2$ on $\Sigma$ isomorphic if there exists a biholomorphism $\phi: (\Sigma,
J_1) \cong (\Sigma, J_1)$ preserving the marked points of $\Sigma$ and the decomposition of $\Sigma$ into pants given by graph $\Gamma$. Denote by $\ttt' _\Gamma$ the space of isomorphism classes of complex structures on $\Sigma$. By [*Proposition 1.2*]{}, Fenchel-Nielsen coordinates identify $\ttt'_\Gamma$ with the real manifold $\rr_+^{3g-3+m+2b} \times (S^1)^{3g-3+m+b}$.
It is desirable to equip $\ttt'_\Gamma$ with some natural complex structure. In so doing, the main difficulty is that the real dimension of $\ttt'_\Gamma$ can be odd. A possible explanation of this fact is that not all relevant information (parameters) about a complex structure has been taken into consideration. Note that for any “inner circle” $\gamma_{jk}$ which appears after the decomposition into pants, we have obtained a pair of coordinates, mainly the length $\ell_{jk}$ and the angle $\vartheta_{jk}$. On the other hand, for any boundary circle $\gamma_i$ of $\Sigma$ we have only the length $\ell_i$. An obvious way to produce additional angle coordinates is to introduce an additional marking of every boundary circle.
Definition 4.3.1. *A real surface $\Sigma$ or a nodal complex curve $C$ is said to have a [*marked boundary*]{} if on every boundary circle of $\Sigma$ (resp. $C$) a point is fixed.*
Remark. Later in §5 we shall consider complex curves with several marked points on boundary circles. But now we shall assume that on every boundary circle exactly one point is marked.
“Missed” angle coordinates $\vartheta_i$ can be now introduced similarly to $\vartheta_{jk}$. For a boundary circle $\gamma_i$ we consider the adjacent pants $S_j$. Fix a boundary component $\d_i S_j$ different from $\gamma_i$. Let $J$ be a complex structure on $\Sigma$ such that the boundary circles of all pants $S_k$ are geodesic the intrinsic metric defined by $J$. Using constructions from above, find a geodesic (resp. a ray) $\alpha_i$ starting at point $x^*_i \in \gamma_i$ and ending at the boundary circle $\d_i S_j$ (resp. approaching marked point $\d_i S_j$ of $\Sigma$). Take the marked boundary point $\zeta_i$ on $\gamma_i$ and consider the length $\lambda_i$ of the geodesic ark on $\gamma_i$, starting at $x^*_i$ and going to $\zeta_i$ in the direction defined by the orientation of $\gamma_i$. Define $\vartheta_i
\deff {2\pi \lambda_i \over \ell_i }$, $\vartheta_i \in S^1\cong \rr/ 2\pi\zz$. We include the coordinates $\vartheta_i$ in the system of angle coordinates $\bfvartheta$. Denote by $\ttt_\Gamma$ the set of isomorphism classes of complex structures on $\Sigma$ with marked boundary and with a given decomposition into pants.
Let $C$ be a smooth complex curve with marked points and a marked boundary, $C
\bs \mapo = \cup_j C_j$ its decomposition into pants, $\sigma: \Sigma \to C$ a parameterization, and $\Sigma \bs \mapo = \cup_j S_j$ the induced decomposition of $\Sigma$. To define a complex structure on $\ttt_\Gamma$, we introduce special local holomorphic coordinates in a neighborhood of boundary of pants on $C$. Consider some pants $S_j$ and its boundary circle $\gamma^*$. It can be a boundary circle of $\Sigma$, $\gamma_i$ in our previous notation, or a circle $\gamma_{jk}$ separating $S_j$ from another pants $S_k$. Let $\ell^*$ be the intrinsic length of $\gamma^*$. Fix some small $a>0$ and consider the annulus $A$ consisting of those $x\in S_j$ for which the intrinsic distance $\dist(x,\gamma^*) <a$. The universal cover $\ti A$ can be imbedded into the hyperbolic plane $\hh$ as an infinite strip $\Theta$ of constant width $a$ such that one of its borders is a geodesic line $L$. The action of a generator of $\pi_1(A) \cong \zz$ on $\ti A$ is defined by the shift of $\Theta$ along $L$ by distance $\ell^*$.
Now consider the annulus $A' \deff [0, {\pi^2\over \ell^*}] \times S^1$ with coordinates $\rho, \theta$, $0\le\rho < {\pi^2\over \ell^*}$, $0\le \theta \le
2\pi$ and with the metric $({\ell^* \over 2\pi} / \cos {\ell^* \rho\over 2\pi}
)^2 (d\rho^2 + d\theta^2)$. A direct computation shows that this metric is of constant curvature $-1$ and that the boundary circle $\d_0A_1 \deff S^1 \times
\{0\}$ is geodesic of length $\ell^*$, whereas $A'$ is complete in a neighborhood of the other boundary circle. Consequently, the universal cover $\ti A{}'$ of $A'$ can be imbedded in the hyperbolic plane $\hh$ as a hyperbolic half-plane $\hh^+_L$ with a boundary line $L$ such that $\Theta \subset \hh^+_L$. Moreover, the action of $\pi_1(A') \cong \zz$ on $\ti A{}' \cong \hh^+_L$ is the same as for $\ti A \cong \Theta$. This shows that there exists an [*isometric*]{} imbedding of $A$ into $A'$ which maps $\gamma^*$ onto $\d_0 A'$. Moreover, such an imbedding is unique up to rotations in the coordinate $\theta$. This leads us to the following
Proposition 4.3.1. *Let $C_j$ be pants with a complex structure and $\gamma^*$ its boundary circle of the intrinsic length $\ell^*$. Let $x^*$ be a point on $\gamma$. Then some collar annulus $A$ of $\gamma^*$ possesses the uniquely defined conformal coordinates $\theta\in S^1\cong \rr/
2\pi\zz$ and $\rho$ such that the intrinsic metric has the form $({\ell^*
\over 2\pi} / \cos {\ell^* \rho\over 2\pi})^2 (d\rho^2+ d\theta^2)$, $\rho|_{
\gamma^*} \equiv0$, $\theta(x^*) =0$ and the orientation on $S_j$ is given by $d\theta\wedge d\rho$.*
We shall represent $\rho$ and $\theta$ also in the complex form $\zeta \deff
e^{-\rho + \isl \theta}$ and call $\zeta$ the [*intrinsic coordinate*]{} of the pants $C_j$ at $\gamma^*$. An important corollary of the description of the intrinsic metric in a neighborhood of the boundary circle is the following statement about non-degenerating complex structures in pants, see [*Definition 1.7*]{}.
Lemma 4.3.2. *Let $C$ be a smooth complex curve with marked points admitting the intrinsic metric and let $\gamma^*$ be a boundary circle of $C$ of length $\ell^*$.*
If there exists an annulus $A \subset C$ of conformal radius $R$ ($A
\cong \{\,z \in \cc \;:\; 1<|z| < R \;\})$, adjacent to $\gamma^*$ and containing no marked points, then $\log R \le {\pi^2 \over \ell^*}$.
There exists a universal constant $a^*$ such that the condition $\ell^*
\le 1$ implies that there exists an annulus $A \subset C$ of conformal radius $R$ with $\log R \ge { \pi^2\over \ell^*} -{2\pi\over a^*}$, which is adjacent to $\gamma^*$, has area $a^*$ and contains no marked points of $C$.
Let $\gamma\subset C$ be a simple geodesic circle of the length $\ell$ and $A \subset C\bs\mapo$ annulus of conformal radius $R$ homotopy equivalent to $\gamma$. Then $\log R \le {2\pi^2 \over \ell}$.
Proof. Let $\Omega$ be the universal cover of $C\bs \mapo$ equipped with the intrinsic metric lifted from $C$. Then $\Omega$ can be isometrically imbedded into the hyperbolic plane $\hh$ as a domain bounded by geodesic lines such that each of these lines covers some boundary circle $\gamma_i$. Take some (not unique!) line $L$ covering the circle $\gamma^*$ and fix a hyperbolic half-plane $\hh^+_L$ with a boundary line $L$, so that $\Omega \subset
\hh^+_L$.
Now consider the universal cover $\ti A$ of the annulus $A$ and provide it with the metric induced from $C$. Then we can isometrically imbed $\ti A$ in $\hh^+_L$ in such a way that the line covering $\gamma^* \subset \d A$ will be mapped onto $L$. The action of a generator of $\pi_1(A) \cong \zz$ on $\ti A$ is defined by the shift of $\hh$ along $L$ onto a distance $\ell^*$. Consequently, $A$ can be isometrically imbedded into $\hh^+_L/ \pi_1(A)$, which is the annulus $A'= [0, {\pi^2\over \ell^*}] \times S^1$ with coordinates $\rho, \theta$, $0\le\rho < {\pi^2\over \ell^*}$, $0\le \theta \le 2\pi$ and with a metric $({\ell^* \over 2\pi} / \cos {\ell^* \rho\over 2\pi})^2
(d\rho^2 +d\theta^2)$. Note that the conformal radius of $A'$ is $e^{\pi^2/\ell^*}$. The monotonicity of the conformal radius of annuli (see , Ch.II, §1.3) yields the inequality $R\le e^{\pi^2/\ell^*}$ which is equivalent to first assertion of the lemma.
Part of the lemma can be proved by same argument. More precisely, under the hypothesis of part we imbed the annulus $A$ into the annulus $A''= ]-{\pi^2\over \ell}, {\pi^2\over \ell}[ \times S^1$ with coordinates $\rho, \theta$, $- {\pi^2\over \ell} < \rho < {\pi^2\over \ell}$, $0\le \theta \le 2\pi$ and with a metric $({\ell \over 2\pi} / \cos {\ell \rho \over 2\pi})^2 (d\rho^2 +d\theta^2)$. The conformal radius of $A$ is now estimated by the conformal radius of $A''$, which is equal to $e^{2\pi^2\over \ell}$.
The second part of our lemma follows from results of Ch.II, §3.3 of \[Ab\]. [*Lemma 2*]{} says that there exists a universal constant $a^*$ with the following property: If $\ell^* \le 1$, then there exists a collar neighborhood $A$ of constant width $\rho^*$ and of area $a^*$, which is an annulus imbedded in $C$ and contains no marked points of $C$. In particular, we can extend the intrinsic coordinates $\rho$ and $\theta$ in $A$. Using these coordinates, we present $A$ in the form $\{ (\rho,\theta): 0\le \rho \le \rho^*\}$ and compute the area, $$a^*=\area A = 2\pi \int_{\rho=0}^{\rho^*} \left(\msmall{\ell^*/2\pi \over
\cos(\ell^*\rho/2\pi) }\right)^2 d\rho =
\ell^* \tan\left(\msmall{\ell^*\rho^* \over 2\pi}\right).$$ Consequently, $\tan\left({\pi\over2}-{\ell^*\rho^* \over 2\pi}\right) =
\cotan\left({\ell^*\rho^* \over 2\pi}\right) = {\ell^*\over a^*}$. This implies ${\pi\over2}-{\ell^*\rho^* \over 2\pi} \le {\ell^*\over a^*}$, which is equivalent to $\rho^* \ge { \pi^2\over \ell^*} -{2\pi\over a^*}$. To finish the proof we note that the conformal radius $R$ of $A$ is equal to $e^{\rho^*}$.
Let $C$ be a smooth complex curve with marked points, $C_j$ a piece of a decomposition of $C\bs \mapo$ into pants and $\gamma^*$ its boundary circle. Then as a “base point” $x^*=\{\theta=0=\rho\}$ for the definition of the intrinsic coordinate we shall use the point $x^*_{j,k}$ if $\sigma( \gamma^*
)$ is the geodesic separating $C_j$ from another pants $C_k$, or respectively, the point $x^*_i$ if $\gamma^*$ is a boundary circle of $C$. We denote these coordinates $\zeta_{j,k}= e^{-\rho_{j,k} + \isl \theta_{j,k}}$ and $\zeta_i=
e^{-\rho_i + \isl \theta_i}$. Note that $\vartheta_i$ is exactly the $\theta$-coordinate of the marked boundary point $x_i\in \gamma_i$ with respect to $x^*_i$, and $\vartheta_{jk}$ is the $\theta$-coordinate of $x^*_{k,j}$ with respect to $x^*_{j,k}$.
Note also that any intrinsic coordinate of a pair $(\zeta_{j,k}, \zeta_{k,j}
)$ extends canonically from one collar neighborhood of $\gamma_{jk}$ to another side in such a way that the formula for the intrinsic metric remains valid. This extension possesses the property $\zeta_{j,k}\cdot \zeta
_{k,j} \equiv e^{\isl \vartheta_{jk}}$, where $\vartheta_{jk}$ is a constant function. We can view this relation as the transition function from $\zeta_{j,k}$ to $\zeta_{k,j}$.
A similar construction is possible in the case of a boundary circle $\gamma_i$. Namely, allowing $\rho_i$ to change also in the interval $(-{\pi^2 \over
\ell_i}, 0]$ and maintaining the formula $({\ell_i \over 2\pi} / \cos {\ell_i
\rho_i\over 2\pi})^2 (d\rho_i^2+ d\theta_i^2)$ for the metric, we can glue to $\Sigma$ an annulus $(-{\pi^2\over \ell_i}, 0] \times S^1$ and extend the coordinate $\zeta_i = e^{-\rho_i + \isl \theta_i}$ there.
Making such a construction with every boundary circle $\gamma_i$, we obtain a complex curve $C^{(N)}$ with the following properties. $C$ is relatively compact in $C^{(N)}$ and the intrinsic metric of $C$ extends to a complete Riemannian metric on $C^{(N)}$ with constant curvature $-1$. Such an extension and the metric are unique. $C^{(N)}$ is called the [*Nielsen extension of $C$*]{}, see \[Ab\]. Note that the complex coordinate $\zeta_i$ can be extended further to the unit disk $\{|\zeta_i|<1\}$.
Using the introduced complex coordinates $\zeta_i$ and $\zeta_{j,k}$, we define a deformation family of complex structures on the curve $C$ with marked boundary. Let $\lambda_i$ and $\lambda_{jk}$ be complex parameters changing in small neighborhoods of $e^{\isl\theta_i}$ and $e^{\isl \theta
_{jk}}$, respectively. Having these data $\bflambda =(\lambda_i, \lambda_{jk}
)$, construct a complex curve $C_\bflambda$ in the following way. Take the pants $\{C_j\}$ of the given decomposition of $C$ and extend all the complex coordinates $\zeta_i$ and $\zeta_{jk}$ outside the pants. Glue the pairs of coordinates $(\zeta_{j,k},\zeta_{k,j})$ with new transition relations $\zeta_{j,k}\cdot \zeta_{k,j}= \lambda_{jk}$ (constant functions). Move original boundary circles $\gamma_i= \{|\zeta_i|=1\}$ of $C$ to new positions defined by the equations $|\zeta_i|= |\lambda_i|$ and mark the points $\zeta_i = \lambda_i$ on them.
Theorem 4.3.3. *The natural map $F:\bflambda \to(\bfell,\bfvartheta)$ is non-degenerated. In particular, $\bflambda$ can be considered as the set of local complex coordinates on $\ttt_\Gamma$ and $\calc\deff \{ C_\bflambda
\}$ as a (local) universal holomorphic family of curves over $\ttt_\Gamma$.*
Proof. Write the functions $\bflambda = (\lambda_i, \lambda_{jk})$ in the form $\lambda_i= e^{-r_i + \isl\phi_i}$, $\lambda_{jk}= e^{-r_{jk} +
\isl\phi_{jk}}$. From the definition of the map $$\textstyle
F: (e^{-r_i + \isl\phi_i}, e^{-r_{jk} + \isl\phi_{jk}})
\mapsto (\ell_i, \ell_{jk}; \vartheta_i, \vartheta_{jk})$$ it is easy to see that ${\d (\vartheta_i, \vartheta_{jk}) \over \d(\phi_i, \phi_{jk})}$ is the identity matrix, whereas ${\d (\ell_i, \ell_{jk}) \over \d(\phi_i, \phi_{jk}
)}$ is equal to 0. So it remains to show that the matrix ${\d (\ell_i,
\ell_{jk}) \over \d(r_i, r_{jk})}$ is non-degenerate.
Consider a special case where $C$ is pants with the boundary circles $\gamma_i$ (at least one) and, possibly, with marked points $x_j$. We shall consider such points as punctures of $C$. Let $J$ denote the complex structure on $C$ and let $\mu_0$ be the intrinsic metric. Extend the coordinates $\zeta_i$ and the metric $\mu_0$ outside of $\gamma_i$ to some bigger complex curve $\wt C$ with $C \comp \wt C$.
Fix real numbers $v_i$ and consider the domains $C_t$ in $\wt C$ defined in local coordinates $\zeta_i= e^{-\rho_i + \isl\theta_i}$ by the inequalities $\rho_i \ge v_it$. This defines a family of deformations of $C= C_0$ parameterized by a real parameter $t$, corresponding to a real curve in the parameter space $\{ \bflambda \}$ given by $\lambda_i(t) = e^{v_it}$. Note that the deformation is made in such a way that the original complex structure $J$ and local holomorphic coordinates are preserved. Thus we can use them as an “invariable basis” in our calculations.
Let $\mu_t$ be the intrinsic metric of $C_t$. Without loss of generality we may assume that $\mu_t$ extends to $\wt C$ as a metric with constant curvature $-1$, which induces the original complex structure $J$ on $\wt C$. Since the intrinsic metric depends smoothly on the operator $J$ of a complex structure, $\mu_t$ are smooth in $t$. In a local holomorphic coordinate $z=x
+\isl y$ we can present $\mu_t$ in the form $e^{2\psi(t,z)}(dx^2 + dy^2)$. The condition ${\sf Curv}(\mu_t) \equiv -1$ is equivalent to the differential equation $$\d^2_{xx}\psi(t,\cdot) + \d^2_{yy} \psi(t, \cdot)= e^{2\psi(t, \cdot)},$$ where $\d_x$ denotes the partial derivation ${\d \over \d x}$ and so on. Differentiating it in $t$, we obtain $e^{-2\psi(t,\cdot)}(\d^2_{xx}+ \d^2_{yy})
\dot \psi(t,\cdot) = \dot\psi(t,\cdot)$, where $\dot\psi(t,\cdot)$ denotes the derivative of $\psi(t,\cdot)$ in $t$.
Note that $\d_t\mu_t = 2 \dot \psi(t,\cdot)\cdot \mu_t $, so $\dot \psi_t
(\cdot)= \dot\psi(t, \cdot)$ is independent of the choice of a local holomorphic coordinate $z= x+ \isl y$ and is defined globally. The equation $e^{-2\psi_z(t,\cdot)} (\d^2_{xx}+ \d^2_{yy})\dot\psi_z(t,\cdot) = \dot\psi_z
(t,\cdot)$ can be rewritten in the form $\Delta_t \dot\psi_t = 2 \dot \psi_t
$, with $\Delta_t$ denoting the Laplace operator for the metric $\mu_t$.
The condition that the circle $\gamma_i(t) \deff \{ \rho_i = v_it\}$ is $\mu_t $-geodesic means that the covariant derivative $\nabla_{\theta_i}
(\d_{\theta_i})$ of the vector field $\d_{\theta_i}$, the tangent vector field to $\gamma_i(t)$, must be parallel to $\d_{\theta_i}$ along $\gamma_i(t)$. Expressing this relation in local coordinates $\rho_i$ and $\theta_i$, we obtain $\d_{\rho_i}\psi_i(t; v_it, \theta_i)=0$, where $\mu_t =e^{2 \psi_i (t;
\rho_i, \theta_i)}(\d\rho_i^2 + \d\theta_i^2)$ is a local representation of the metric $\mu_t$. Deriving in $t$, we obtain $\d_{\rho_i}\dot \psi_i(t; v_it,
\theta_i) + v_i \d^2_{\rho_i\rho_i}\psi_i(t; v_it, \theta_i) =0$.
In the case where $t=0$ we have $\psi_i(0; 0, \theta_i) \equiv \log{\ell_i \over
2\pi}$, a constant. Hence $\d^2_{\theta_i\theta_i}\psi_i(0; 0, \theta_i) =
e^{2 \psi_i(0; 0, \theta_i)} $ and $\d_{\rho_i}\dot \psi_i(0; 0, \theta_i) =
- v_i e^{2 \psi_i(0; 0, \theta_i)}= - v_i \left({\ell_i \over2\pi} \right)^2
$. On the other hand, $\d_{\rho_i} = - {\ell_i \over 2\pi} \d_\nu$ on $\gamma_i(0) = \gamma_i$, where $\nu$ denotes the unit outer normal field to $C_0=C$. Consider the integral $\int_C |d\dot\psi {}_0|^2 + 2 \dot\psi{}_0^2
d\mu_0$. Integrating by parts, we get $$\eqalignno{
&\int_C |d\dot\psi{}_0|^2 + 2 \dot\psi{}_0^2 d\mu_0 =
\int_C \dot\psi_0(2 \dot\psi_0 - \Delta_0 \dot\psi_0) d\mu_0
+ \int_{\d C} \dot\psi_0 \d_\nu\dot\psi_0 dl=
\cr
\noalign{\vskip0pt\allowbreak}
=&
\sum_i \int_{\gamma_i} \dot\psi_0 \d_\nu\dot\psi_0
\msmall{\ell_i \over 2\pi}
d\theta_i =
\sum_i - \int_{\gamma_i} \dot\psi_0 \d_{\rho_i}\dot\psi_0
d\theta_i =
\sum_i \int_{\gamma_i} \dot\psi_0 v_i \left(\msmall
{\ell_i \over 2\pi}\right)^2 d\theta_i =
\cr
\noalign{\vskip0pt\allowbreak}
=&
\sum_i \msmall{v_i \ell_i \over 2\pi} \int_{\gamma_i} \dot\psi_0
e^{\psi_i(0; 0, \theta_i)} d\theta_i =
\sum_i \msmall{v_i \ell_i \over 2\pi} \int_{\gamma_i} (\dot\psi_0 +
v_i \d_{\rho_i} \psi_i(0; 0, \theta_i))e^{\psi_i(0; 0, \theta_i)} d\theta_i =
\cr
\noalign{\vskip0pt\allowbreak}
=&
\sum_i \msmall{v_i \ell_i \over 2\pi}
\left.\msmall{\d \over \d t}\right|_{t=0} \int_{\gamma_i}
e^{\psi_i(t; v_it, \theta_i)} d\theta_i =
\sum_i \msmall{v_i \ell_i \over 2\pi}
\left.\msmall{\d \over \d t}\right|_{t=0}\ell_i(t) =
\sum_i \msmall{ \ell_i \over 2\pi} v_i \dot\ell_i.
}$$ Here $\dot\ell_i$ denotes the derivative of the length parameter $\ell_i$ for the curve $C_t$ at $t=0$, so that $(\dot\ell_1,\dot\ell_2,\dot\ell_3)= dF (v_1, v_2, v_3)$. The obtained relation shows that the Jacobi matrix $dF= {\d(\ell_i, \ell_2, \ell_3) \over \d(r_1, r_2, r_3)}$ is non-degenerate. Otherwise there would exist a nonzero vector $(v_1, v_2, v_3)$ such that for the deformation constructed above we get $\dot\ell_i=0$. But then $\dot\psi_0
\equiv 0$, which is a contradiction.
Now consider a general situation. Let $\Sigma$ be a real surface with a marked boundary, $C$ a smooth curve with marked points, $\sigma:\Sigma \to C$ a parameterization, and $C \bs \mapo = \cup_j C_j$ a decomposition into pants with a given graph $\Gamma$. Let $\{\gamma_i\}$ be the set of boundary circles and $\{ \gamma_{jk} \}$ the set of circles lying between the pants $C_j$ and $C_k$, respectively. Consider these pants separately. Then for any circle $\gamma
_{jk} = \gamma_{kj}$ we obtain 2 distinguished ones, $\gamma_{j,k}$ considered as a boundary circle of $C_j$, and $\gamma_{k,j}$ considered as a boundary circle of $C_k$. Take real numbers $\bfv\deff (v_i, v_{j,k}, v_{k,j})$, where $v_i$ is associated with the circle $\gamma_i$, $v_{j,k}$ with $\gamma_{j,k}$, and $v_{k,j}$ with $\gamma_{k,j}$, respectively. Let $C_j(t\bfv)$ denote the pants obtained from $C_j$ by the above construction using the corresponding parameters $v_i$ and $v_{j,k}$. For $\bfv$ lying in a small ball $B = \{ |\bfv| < \eps \}$ all such families $C_j(t\bfv)$ can be extended for all $t \in [-1,1]$. Thus over $B$ we obtain a collection of deformation families $C_j(\bfv)$ of complex structure on pants $C_j$.
Let $\ell_i(\bfv)$, $\ell_{j,k}(\bfv)$, and $\ell_{k,j}(\bfv)$ denote the lengths of circles $\gamma_i$, $\gamma_{j,k}$, and $\gamma_{k,j}$ the obtained intrinsic metrics $\mu_j(\bfv)$ on $C_j(\bfv)$. Denote by $\dot\ell_i$ a linear functional $\d_t|_{t=0} \ell_i(t\bfv)$, and define $\dot\ell_{j, k}$ similarly. The explicit formula for an intrinsic metric near a boundary circle shows that $C_j(\bfv)$ can be glued to $C_k(\bfv)$ along $\gamma_{jk}$ exactly when $\ell_{j,k}(\bfv) = \ell_{k,j}(\bfv)$. Since the Jacobian $\d\bfell(\bfv) \over \d \bfv$ is non-degenerate, the conditions $\ell_{j,k}(\bfv) = \ell_{k,j}(\bfv)$ define a submanifold $V \subset B$ whose tangent space $T_0V$ is given by the relations $\dot\ell_{j, k}= \dot\ell
_{k,j}$. Note that this defines a deformation family of complex structures on $C$ over the base $V$ such that the map $\bfv \in V \mapsto \bfell(\bfv)$ is a diffeomorphism.
We state that the set $(v_i, v_{j,k} + v_{k,j})$ is a system of coordinates on $V$ in the neighborhood of $0\in V$. To prove this it is sufficient to show that the linear map $\bfv=(v_i, v_{j,k}, v_{k,j}) \in T_0V \mapsto (v_i,
v_{j,k} + v_{k,j})$ is non-degenerate. If it is not true, then there would exist a nontrivial $\bfv=(v_i, v_{j,k}, v_{k,j})\in T_0V$ with $v_i=0$ and $v_{j,k} + v_{k,j}=0$. Let $\dot\ell_i=\dot\ell_i(\bfv)$, $\dot\ell
_{j,k}= \dot\ell_{j,k}(\bfv)$ and $\dot \ell_{k,j} =\dot \ell_{k,j} (\bfv)$ be the corresponding derivatives of length. Then $\dot\ell_{j,k} = \dot
\ell_{k,j}$ and $$0< \sum_i \msmall{ \ell_i \over 2\pi} v_i \dot\ell_i +
\sum_{j<k} \msmall{ \ell_{jk} \over 2\pi} v_{j,k} \dot\ell_{j,k} +
\sum_{j<k} \msmall{ \ell_{jk} \over 2\pi} v_{k,j} \dot\ell_{k,j}=
\sum_{j<k} \msmall{ \ell_{jk} \over 2\pi} (v_{j,k}+ v_{k,j}) \dot\ell_{j,k}
=0.$$ The obtained contradiction leads us to the following conclusion. The functions $v_i$ and $v_{j,k}+ v_{k,j}$ define a coordinate system on $V$ equivalent to $\bfell=(\ell_i, \ell_{jk})$.
Let us return to the holomorphic deformation family of complex structures on $C$, defined by complex parameters $\lambda_i= e^{-r_i + \isl\phi_i}$ and $\lambda_{jk}= e^{-r_{jk} + \isl\phi_{jk}}$. It is easy to see that the Jacobian $\d(v_i, v_{j,k}+ v_{k,j}) \over \d( r_i , r_{jk})$ at the point $(r_i , r_{jk})=0$ is the identity matrix. This fact proves the statement of the lemma.
Remark. At this point we give a possible reason why the complex (holomorphic) structure introduced by the complex coordinates $\bflambda$ can be regarded as natural. Let $C$ be a complex curve with marked points and a nonempty marked boundary. In the case where $C$ is a disk or an annulus, assume additionally that at least one inner point of $C$ is marked. Then in a neighborhood of every boundary circle $\gamma_i$ of $C$ we can construct the intrinsic coordinate $\zeta_i$. Take 2 copies $C^+$ and $C^-$ of $C$ and denote by $\tau$ the natural holomorphic map $\tau: C^\pm
\to C^\mp$ interchanging the copies. Denote by $\zeta_i^\pm$ the local intrinsic coordinate on $C^\pm$ at boundary circles $\gamma_i^\pm$, both corresponding to $\gamma_i$. Now we can glue $C^+$ and $C^-$ together along every pair of circles $(\gamma_i^+, \gamma_i^-)$ by setting $\zeta_i^+ \cdot
\zeta_i^- =1$ as transition relations. We obtain a closed complex curve $C^d$ which admits a natural holomorphic involution $\tau: C^d \to C^d$. For the constructed family $\{C_\bflambda \}$, the corresponding family $\{C^d
_\bflambda \}$ will be holomorphic. In fact, the statement of [*Theorem 2.3*]{} means that $\{ C^d _\bflambda \}$ is a minimal complete family of deformation of $C^d$ in the class of curves with holomorphic involution. This construction of doubling should not be confused with another construction of the [*Schottky double*]{} $C^{Sch}$ of $C$ which provides an [*antiholomorphic*]{} involution $\tau^{Sch}: C^{Sch} \to C^{Sch}$. We shall use the Schottky double $C^{Sch}$ in [*Section 5*]{}, we will consider curves with totally real boundary conditions.
[4.4. Invariant Description of the Holomorphic Structure on $\ttt_\Gamma$.]{}
The construction of (holomorphic) double $C^d$ shows how to give an invariant description of a holomorphic structure on $\ttt_\Gamma$. Let $C$ be a smooth complex curve with marked points and marked boundary, and $x\in \ttt_\Gamma$ the corresponding point on the moduli space. Denote by $D$ the divisor of marked points. If the curve $C$ is not closed and $C^d$ is its double with holomorphic involution $\tau$, we denote by $D^d \deff D + \tau(D)$ the double of $D$.
Lemma 4.4.1. *If $C$ is closed, then the tangent space $T_x \ttt
_\Gamma$ is naturally isomorphic to $\sfh^1(C, \calo(TC)\otimes \calo(-D) )$.*
If $C$ is not closed, then the space $T_x\ttt_\Gamma$ is naturally isomorphic to the space $\sfh^1(C^d, \calo(TC^d)\otimes \calo(-D^d))^\tau$ of $\tau$-invariant elements in $\sfh^1(C^d, \calo(TC^d)\otimes \calo(-D^d))$.
In both cases the complex structure on $T_x\ttt_\Gamma$ induced by local complex coordinates $\bflambda$ coincides with those from $\sfh^1(C,
\calo(TC) \otimes \calo(-D))$ (resp. $\sfh^1(C^d, \calo(TC^d) \otimes
\calo(-D^d) )^{(\tau)}$). In particular, this defines a global complex structure on space $\ttt_\Gamma$.
Proof. The part concerning closed curves is well-known. In fact, the natural isomorphism $\psi: T_x\ttt_\Gamma \to \sfh^1(C, \calo(TC)\otimes
\calo(-D))$ is a Kodaira-Spencer map. Its description is very simple in the introduced local coordinates $\zeta_{j,k}$ on $C$ and $\bflambda =( \lambda
_{jk})$ on $\ttt_\Gamma$. Let $C\bs \mapo = \cup C_j$ be the decomposition of $C$ into pants with the graph $\Gamma$. For every pants $C_j$ choose an open set $\ti C_j$, containing a closure $\barr C_j= C_j \cup \d C_j$. Without loss of generality we may assume that $\ti C_j$ are chosen not too big, so that the covering $\calu\deff \{ \ti C_j \}$ is acyclic for the sheaf $\calo(TC)$ and the local coordinates $\zeta_{j,k}$ are well-defined in the intersections $\ti C_j \cap \ti C_k$. Then vector $v \in T_x\ttt_\Gamma$ with local representation $v= \sum_{j<k} v_{jk} {\d \over \d \lambda_{jk}}$ is mapped by the Kodaira-Spencer map $\psi$ to the Čech 1-cohomology class $$\psi(v) \in \sfh^1(C, \calo(TC)\otimes \calo(-D)) \cong
\check \sfh^1(\calu, \calo(TC) \otimes \calo(-D)),$$ represented by the 1-cocycle $$\left(v_{jk} \zeta_{j,k} \msmall{\d \over \d \zeta_{j,k}} \right)
\in \prod_{j<k} \Gamma(\ti C_j \cap \ti C_k, \calo(TC)\otimes \calo(-D)).$$ For more details see \[D-G\].
Using this description of the Kodaira-Spencer map for closed curves with marked points, it is easy to handle the case of curves with boundary. Let $C$ be a non-compact curve with marked points and with decomposition $C \bs \mapo
= \cup_j C_j$. Take its double $C^d$ with the involution $\tau$. Then the decomposition of $C$ induces a $\tau$-invariant decomposition $C^d =\bigcap_j(C_j \cap \tau C_j)$. The corresponding covering $\calu^d$ of $C^d$ can also be chosen to be $\tau$-invariant.
The local coordinates $\zeta_i$, corresponding to boundary circles $\gamma_i$ of $C$, can now be extended to a two-sided neighborhood of $\gamma_i$ in $C^d$. The coordinates $\zeta_{j,k}$, corresponding to inner circles $\gamma_{jk}$, induce local complex coordinates $\zeta^\tau_{j,k}\deff
\zeta_{j,k} \scirc \tau$ in $\tau( \ti C_j \cap \ti C_k)$.
Any deformation of the complex structure on $C$ induces a deformation of the complex structure on $C^d$. This defines a map $\phi: \ttt_\Gamma \to
\ttt_{\Gamma^d}$, with $\Gamma^d$ denoting the graph corresponding to the $\tau$-invariant decomposition of $C^d$ into pants. Using introduced coordinates, we present a tangent vector $v \in T_x \ttt_\Gamma$ in the form $$v= \sum_i v_i \msmall{\d \over \d \lambda_i} +
\sum_{j<k} v_{jk} \msmall{\d \over \d \lambda_{jk}}.$$ Then the composition of the Kodaira-Spencer map $\psi^d$ of $C^d$ with the differential of $\phi: \ttt_\Gamma \to \ttt_{\Gamma^d}$ maps $v$ to $$\psi^d \scirc d\phi(v) \in \sfh^1(C^d, \calo(TC^d)\otimes \calo(-D^d)) \cong
\check \sfh^1(\calu^d, \calo(TC^d) \otimes \calo(-D^d)),$$ represented by the Čech 1-cocycle $$\check v \deff \left(v_i \zeta_i \msmall{\d \over \d \zeta_i},
v_{jk} \zeta_{j,k} \msmall{\d \over \d \zeta_{j,k}},
v_{jk} \zeta^\tau_{j,k} \msmall{\d \over \d \zeta^\tau_{j,k}} \right)
\in \prod_i \Gamma(\ti C(i) \cap \tau C(i), \calo(TC^d)\otimes \calo(-D^d))
\times$$ $$\prod_{j<k} \Gamma(\ti C_j \cap \ti C_k, \calo(TC^d)\otimes \calo(-D^d))
\times \prod_{j<k} \Gamma(\tau(\ti C_j \cap \ti C_k),
\calo(TC^d)\otimes \calo(-D^d)),$$ where $C(i)$ denotes the pants of $C$ adjacent to circle $\gamma_i$. It is obvious that if all $v_i$ vanish, then this Čech 1-cocycle is $\tau
$-invariant. On the other hand, the relation $\zeta_i \cdot (\zeta_i \scirc
\tau) \equiv \lambda_i= \const$ implies that $\tau_*(\zeta_i {\d\over \d
\zeta_i}) = -\zeta_i {\d \over \d \zeta_i}$. The additional change of the sign of the corresponding part of cocycle $\check v$ comes from the fact that $\tau$ interchange $C(i)$ with $\tau(C(i))$. This shows that $\check v$ is $\tau
$-invariant and the statement of the lemma follows.
Now we study the connection between the geometry of $\ttt_\Gamma$ and the degeneration of complex structures on a real surface $\Sigma$ with marked points and marked boundary. Let $\Sigma \bs \mapo= \cup_j S_j$ be a decomposition into pants with graph $\Gamma$. The Fenchel-Nielsen coordinates on $\ttt_\Gamma$ define a map $(\bflambda, \bfvartheta): \ttt_\Gamma \to (\rr
\times S^1)^{3g-3+ m+2b}$, which is a diffeomorphism by [*Proposition 1.2*]{}. So, if $\{ j_n\}$ is a sequence of complex structures on $\Sigma$, its degeneration means that the sequence of Fenchel-Nielsen coordinates of $\{
j_n\}$ is not bounded in $(\rr \times S^1)^{3g-3+ m+2b}$.
One can see that, in fact, we have two types of the degeneration. The first one occurs when the maximum of the length coordinates $\ell_i$ and $\ell_{jk}$ of $j_n$ increases infinitely, and the second one is present when a minimum of the length coordinates of $j_n$ vanishes. It should be pointed out that for an appropriate sequence one can have both types of degeneration.
Note that by [*Proposition 1.2*]{} the Fenchel-Nielsen coordinates of a complex structure $j$ on $\Sigma$ are defined by a choice of a topological type of decomposition of $\Sigma$ into pants, encoded in the graph $\Gamma$. Thus the introduced notion of degeneration also depends on the choice of $\Gamma$. Possibly, the best choice of such decomposition is established by the following statement, proved in \[Ab\], Ch.II, §3.3.
Proposition 4.4.2. *Let $C$ be a complex curve with parameterization $\sigma: \Sigma \to C$. Then*
[*a)*]{} there exists a universal constant $l^*>0$ such that any two geodesic circles $\gamma'$ and $\gamma''$ on $C$ satisfying $\ell( \gamma') < l^*$ and $\ell(\gamma'') < l^*$ are either disjoint or they coincide;
[*b)*]{} there exists pants decomposition $C \bs \mapo= \cup_j S_j$ such that the lengths of inner boundary circles $\gamma_{jk} = \barr S_j \cap \barr
S_k$ are bounded from above by a constant $L$ which depends only on the topology of $\Sigma$ and the maximum $M$ of the lengths of the boundary circles of $C$; moreover, any simple geodesic circle $\gamma$ on $C$ with $\ell(\gamma) <l^*$ occurs as a boundary circle of some $C_j$.
Corollary 4.4.3. *Let $C_n$ be a sequence of nodal curves parameterized by a real surface $\Sigma$ with uniformly bounded number of components. Suppose that the complex structures of $C_n$ do not degenerate near the boundary. Then, passing to a subsequence, one can find a decomposition $\Sigma =\cup_j
S_j$ and new parameterizations $\sigma'_n : \Sigma \to C_n$, such that*
the decomposition $\Sigma =\cup_j S_j$ induces a decomposition of every non-exceptional component of $C_n$ into pants whose boundary circles are geodesics;
the intrinsic length of these geodesics are bounded uniformly in $n$.
Proof. By [*Lemma 4.3.2*]{}, the intrinsic lengths of boundary circles of non-exceptional components of $C_n$ are bounded uniformly in $n$. Find a decomposition into pants of every non-exceptional component of $C_n$ satisfying the conditions of part [*b)*]{} of [*Proposition 4.4.2*]{}. Let $\Gamma_n$ denote the obtained graph of the decomposition of $C_n$. Since the number of components of $C_n$ is uniformly bounded, we obtain a subsequence $C_{n_k}$ with the same graph $\Gamma$ for all $n_k$.
It follows from the proof in \[Ab\] that the constant $L$ from part [*b)*]{} of [*Proposition 4.4.2*]{} depends continuously on the maximum $M$ of the lengths of the boundary circles of $C_n$. This implies condition [*i$\!$i)*]{}. Applying [*Proposition 4.2.2*]{}, we complete the proof.
[4.5. Example: Degeneration to a Half-cubic Parabola on the Language of Stable Curves.]{}
[*Example.*]{} Consider a sequence of curves $T_n$ in $\cc\pp^2$ given by $$T_n= \{\; y^2z-x^3=n^{-6}z^3\;\} \subset \cc\pp^2$$ in homogeneous coordinates. In the affine chart $U_0=\{ z\not= 0\}$ the curves are given by the equation $T_n=\{ y^2-x^3=n^{-6}\} $. In any reasonable sense these curves should converge to a half-cubic parabola $T_\infty:= \{ y^2=x^3\} $. We shall now explain in great detail how the $T_n$ converge to $T_\infty$ in Gromov topology.
Each curve $T_n$ is the image of the curve $$T = \{ y_1^2z_1-x_1^3=z_1^3\}\subset \cc\pp^2$$ w.r.t. an algebraic map $f_n: [x_1, y_1,z_1] \in C
\mapsto [x:y:z]:= [n^{-2} x_1: n^{-3} y_1: z_1]$.
We shall denote by $[x_1:y_1:z_1]$ the coordinates in the pre-image and by $[x:y:z]$ in the range.
[*Convergence of graphs.*]{} Consider the graphs $\Gamma_n$ of these mappings as subsets of $\cc\pp^2\times \cc\pp^2$. Then the $\Gamma_n$ converge to a [*reducible*]{} curve $\Gamma_\infty\in
\cc\pp^2\times \cc\pp^2$, which consists of the graph of a [*constant*]{} mapping $[x_1: y_1:z_1] \in T \mapsto [0:0:1]$, which is a “horizontal component”, plus a “vertical component” over $[x_1:y_1:z_1] =[0:1:0]$, which is a limit curve $\{y^2 = x^3\}$ of the sequence $T_n$.
The picture of the convergence is the following. Starting from $\Gamma_1$, the curves $\Gamma_n$ transform continuously (we can define $\Gamma_t$ for any real $t$), so that for $n>\!>1$ curve $\Gamma_n$ consists of an “almost horizontal” torus and an “almost vertical” sphere connected by a thin “neck”. As $n\longrightarrow \infty$ this neck is shrinking to a point $P:=([0:1:0], [0:0:1]) \in \pp^2 \times
\pp^2$ which is the only singular point of $\Gamma_\infty$ where its components intersect.
[*Explanation of the picture:*]{} $i_1,i_2,i_\infty$ are imbeddings of $C_n$’s into $\pp^2 \times \pp^2$ as the graphs $\Gamma_n$’s. The map $\pr:\pp^2 \times \pp^2 \to \pp^2$ is the natural projection onto the range $\pp^2$.
Thus, the picture of convergency on the language of parameterized curves should be the following: the surface parameterizing all $T_n$ (and also $T_\infty
=\{ y^2=x^3\} $) should be a torus $\Sigma $. Some fixed circle $\gamma $ on $\Sigma $ around some fixed point $p\in \Sigma $ should be mapped by parameterizations $\sigma_n :\Sigma \to C_n$ to “smaller circles". Finally the parameterization $\sigma_\infty :\Sigma \to C_\infty$ contracts $\gamma $ to a point. Thus $C_\infty$ should be a nodal curve which consists of a smooth torus and a smooth sphere intersecting at one point. The limit map $u_\infty$ should map the torus to the point $[0:0:1]$ and the sphere onto the limit curve $T_\infty$.
Now we must determine curves $C_n$, parameterizations $\sigma_n$ and mappings $u_n$ explicitly. Note that one cannot take as $C_n$ simply the graph $\Gamma_n$, because $\Gamma_\infty$ has a point $P=([0:1:0], [0:0:1]) \in \pp^2 \times \pp^2$ as a cusp and a node at the same time. However, by our definition of a nodal curve, $C_\infty$ should have at most nodes. This is not a big problem, because one should “holomorphically” parameterize $\Gamma_\infty$ by a nodal curve and at the same time one also needs to parameterize “coherently" $C_n$’s.
Let us do just that.
[*Explicit Parameterization.*]{}
Now we give a description of the convergence in terms of stable curves (terminology of \[KM\]; an alternative terminology speaks of “stable maps”). Consider homogeneous equations $$y_1^2z_1 -x_1^3=z_1^3
\hbox{\qquad and \qquad}
{x\over n^{-2}x_1}= {y \over n^{-3}y_1}= {z\over z_1},
\eqno(4.5.1)$$ defining $\Gamma_n$ in $\pp^2 \times \pp^2$. Let $Q=[x_1:y_1:z_1]$ be a point on $T$ with $z_1\not =0$. Then obviously $f_n(Q)= [n^{-2} x_1: n^{-3} y_1:
z_1]$ converge to the point $[0:0:1]$ in the range $\pp^2$. Moreover, for a sufficiently small neighborhood $U$ of $Q$ in $T$ we obviously have a uniform convergence of the restriction $f_n|_U$ to a constant map $f_\infty:
U \to \pp^2$, $f_\infty|_U \equiv[0:0:1]$.
The set of points $[x_1:y_1:z_1]$ on $C$ with $z_1=0$ consists of one point $[0:1:0]$. This means that for any compact $K \Subset C \bs [0:1:0]$ the sequence of restricted maps $f_n|_K : \to \pp^2$ converge uniformly to the constant map $f_\infty: K \to \pp^2$, $u_\infty|_K \equiv[0:0:1]$. In terms of graphs $\Gamma_n$ this means that $\Gamma_n \cap(K\times \pp^2)$ converge to $\Gamma_\infty \cap(K \times \pp^2)$, which lies in the “horizontal” part $T \times [0:0:1]$ of $\Gamma_\infty$.
Consider the behavior of $f_n$ in a neighborhood $V\subset T$ of the point $[0:1:0]$. Setting $y_1=1$ and considering $x_1, z_1$ as an affine coordinate on the first $\pp^2$ we obtain the relation $$z_1= x_1^3 + z_1^3.$$ One can easily see that $x_1$ can be chosen as a local holomorphic coordinate in a neighborhood of $[0:1:0]$, and that there exists a holomorphic function $\phi(x_1)$ in the disk $\Delta(a)=\{ |x_1| <a\}$ such that $\phi(0)=1$ and $V=\{ z_1=x_1^3\phi(x_1)\} $.
Fix $a>{1\over b}>0$ and take $R>{1\over b}$. Denote by $\pr:V\to \Delta (a)$ the natural projection, see [*Fig. 7*]{}. For every $n$ sufficiently big, cover $\pr(V)=\Delta(a)$ by sets $V_n$, where $V_n:=\pr^{-1}\Delta({R\over n})$ and $V_n':=\pr^{-1}(A_n)$ with $A_n = \{ {1\over bn} <|x_1| <a\} $. First we consider the behavior of appropriately rescaled maps $f_n\circ \pr^{-1}$ in $V_n$. For a complex coordinate $\xi \in \cc$ we consider maps $v_n(\xi): \Delta(R) \to \pp^2$ with $v_n(\xi):=
f_n\circ \pr^{-1}({\xi\over n})$ and the domain of definition $\Delta(R):=\{|\xi|<R\}$. This means that we rescale $f_n\circ \pr^{-1}$ in $\Delta({R\over n})=$ by setting $nx_1 =\xi$. Then $$\textstyle
v_n(\xi)=f_n\circ \pr^{-1}\left(\left[{\xi\over n}:1: {\xi^3\over n^3}
\phi\left({\xi\over n}\right) \right]\right)=
\left[{\xi\over n^3}:{1\over n^3}: {\xi^3\over n^3}
\phi\left({\xi\over n}\right) \right]=
\left[\xi :1 :\xi^3\phi\left({\xi\over n}\right) \right].$$ Thus, images $v_n(\Delta(R))=f_n(\Delta ({R\over n}))$ lie in the affine chart $y\not=0$ in the range $\pp^2$, and in the affine coordinates $x,z$ ($y=1$) we obtain a representation $v_n(\xi)= (\xi, \xi^3 \phi({\xi\over n}))$. We easily see that $\phi({\xi\over n})$ converge on compacts $\Delta(R)$ to the constant function 1. This implies that maps $v_n$ converge on $\Delta(R)$ to the map $v_\infty: \cc \to \pp^2$, $v_\infty(\xi)= [\xi:1:\xi^3]$. This map is a parameterization of the “vertical part” of $\Gamma_\infty$.
Finally, we want to describe the behavior of the “neck” of $\Gamma_n$ when $n\lrar \infty$. Again, take $x_1$ to be a local holomorphic coordinate on $T$ in a neighborhood of $[0:1:0]$ setting $y_1=1$. Consider imbeddings $h_n: A_n \to \Delta^2(a,b)= \{ |x_1| < a\} \times \{|t|<b\}$ with $h_n(x_1)
=(x_1, {1\over n x_1})$. Denote by $\cala_n$ the images $h_n(A_n)$. Then $\cala_n = \{ (x_1,t) \in \Delta^2(a,b):
x_1{\cdot}t= {1\over n}\}$. As $n\lrar \infty$, the annuli $\cala_n$ converge to a subset $\cala_\infty \subset \Delta^2(a,b)$, which is the union of two disks $\Delta(a) \times \{0\}\cup \{0\} \times
\Delta(b)$.
Note that, along $\cala_n \subset \Delta^2(a,b)$ with coordinates $x_1,t$, we can represent $f_n\circ \pr^{-1}\circ h_n^{-1}$ in the form $$f_n\circ \pr^{-1}\circ h_n^{-1}:(x_1, t) \in A_n \mapsto \left[{x_1 \over n^2}
: {1\over n^3}: x_1^3
\phi(x_1) \right] =[ nx_1 t^3: t^3: n^3x_1^3 t^3\phi(x_1)]=$$ $$=[t^2:t^3: \phi(x_1)],$$ where we use the relation $nx_1 t=1$ along $\cala_n$. Thus the restriction of $f_n\circ \pr^{-1}$ onto $A_n$ coincides with the composition of imbeddings $h_n$ of $A_n$ into $\Delta^2(a,b)$ with the map $F: \Delta^2(a,b) \to \pp^2$, $F(x_1,t)
= [t^2:t^3: \phi(x_1)]$. In the affine chart with $z\not=0$ we have $F(x_1,t) =
\left({t^2 \over \phi(x_1)}, {t^3 \over\phi(x_1)}\right)$, see [*Fig. 8*]{}.
Note that $$F\mid_{\cala_n} = f_n\circ \pr^{-1}\circ h_n^{-1}:\cala_n\to \cc\pp^2.
\eqno(4.5.2)$$
Remark. Denote by $P_n$ the compact curve $\{ x_1\cdot t= {1\over n}
\}$ in $\cc\pp^2$. For us it is important that $F$ is well- defined on $P_n\setminus \{ \vert x_1\vert >a\} $.
We interpret this picture in the following way. Annuli $\cala_n$ degenerate to a normal crossing of two disks, a “node” in the terminology of \[OM\], whereas maps $f_n\circ \pr^{-1}\circ h_n^{-1}:\cala_n \to \pp^2$ converge to a (holomorphic) map from $\cala_\infty$ to $\pp^2$ which is, in this case, a restriction of $F$ onto $\cala_\infty$, i.e., constantly $[0:0:1]$ on the one component $\{ t = 0\} $ of $A_\infty$, and, on the component $\{ x_1 = 0\} $, mapping $F$ is a parameterization $t\in \Delta(b) \mapsto (t^2, t^3)$ of curve $y^2=x^3$.
Note that on $A_n$ the contractiong circle is $\gamma_n = \{ \vert x_1\vert
= \sqrt{{1\over n}}$, because $h_n(\gamma_n) = \{ \vert x_1\vert = \vert t
\vert = \sqrt{{1\over n}}$.
Let us finally construct stable curves $(C_n,u_n)$ and $(C_\infty,
u_\infty)$. Put $W_1 = T_1\setminus V$ and $W_2 = P_n\setminus \{ \vert
x_1 \vert > a \} $. Identify the boundaries of $W_1$ and $W_2$ by $h_n\circ \pr : \partial W_1 \to \partial W_2 $ to get the curve $C_n :=
W_1\sqcup W_2/\partial \sim \partial W_2$. Note that $C_\infty$ is a nodal curve with one nodal point and two components: torus and sphere.
The corresponding maps are defined as follows: $$u(q) = \cases f_n (q) & \text{ if }q\in W_1 \\
F(q) & \text{ if }q\in W_2 .
\endcases$$ Note that $u_n$ is well- defined due to relation (2).
The rest is obvious, i.e., one should take the parameterizations $\sigma_n:\Sigma
\to C_n$ satisfying the three following conditions:
\(a) Some fixed circle $\gamma $ should be mapped by $\sigma_n$ onto $h_n(
\gamma_n)\subset \cala_n$ (also $\sigma_\infty(\gamma)= 0\in \cala_\infty$).
\(b) Some fixed disk $\tilde V_2\subset \Sigma $, see [*Fig. 10*]{}, should be mapped by $\sigma_n$ onto the domain $P_n\cap \{ \vert t\vert >1\} \subset
W_2$ (also $\sigma_\infty(\tilde V_2 = P_\infty\cap
\{ \vert t\vert >1\} $.
\(c) Finally, some other domain $\tilde V_1\subset \Sigma$, which contains a handle, should be mapped onto $W_1$.
One should also take care that the structures $\sigma_n^*j_n$ should converge on compact subsets in $\Sigma \setminus \gamma$.
[Lecture 5]{}
[Gromov Compactness Theorem]{}
[5.1. Second A priori Estimate.]{}
Let $(X,J)$ be an almost complex manifold. In what follows the tensor $J$ is supposed to be only continuous, i.e., of class $C^0$. Fix some Riemannian metric $h$ on $X$. All norms and distances will be taken with respect to $h$. In particular, we have the following
Definition 5.1.1. [*A continuous almost complex structure $J$ is called [*uniformly continuous on $A \subset X$ with respect to $h$*]{}, if $\norm{J}_{L^\infty(A)} < \infty$ and for any $\epsi >0$ there exists $\delta=\delta(J, A, h) >0$ such that for any $x\in A$ one can find a $C^1$-diffeomorphism $\phi: B(x, \delta) \to B(0, \delta)$ from the ball $B(x, \delta)\deff \{y\in X \;:\; \dist_h(x,y) <\delta \}$ onto the standard ball in $\cc^n$ with the standard metric $h\st$ such that $$\norm{J -\phi^*J\st}_{L^\infty(B(x,\delta)\cap A)} +
\norm{h -\phi^*h\st}_{L^\infty(B
(x,\delta)\cap A)}
\le\epsi.
\eqno(5.1.1)$$* ]{}
Roughly speaking, this means that on the set $A$ we can $\norm\cdot _{L^\infty}
$-approximate $J$ by an integrable structure in $h$-metric balls of a radius independent of $x\in A$. The function $\mu(J,A,h)$ whose value at $\epsi>0$ is the biggest possible $\delta\le1$ with the above property is called the [*modulus of uniform continuity of $J$ on $A$*]{}. Note that every continuous almost complex structure $J$ is always uniformly continuous on [*relatively compact*]{} subsets $K \Subset X$.
Let $J^*$ be a continuous almost complex structure on $X$ and $A\subset X$ a subset. Assume that $J^*$ is uniformly continuous on $A$ and denote by $\mu_{J^*}= \mu( J^*,A, h)$ the modulus of uniform continuity of $J^*$ on $A$. Recall that in [*Lemma 2.4.1*]{} we proved the following
Proposition 5.1.1. [*(First A priori Estimate). *For every $p$ with $2< p<\infty$ there exists an $\eps_1 =\eps_1(\mu_{J^*}, A, h)$ (independent of $p$) and $C_p=C(p,\mu_{J^*}, A, h)$ such that for any continuous almost complex structure $J$ with $\norm{ J- J^*} _{L^\infty(A)}<\eps_1 $ and for every $J$-holomorphic map $u\in C^0\cap L^{1,2}(\Delta ,X)$, satisfying $u(\Delta )\subset A$ and $\norm{du} _{L^2( \Delta )} <\eps_1 $, one has the estimate $$\norm{du}_{L^p({1\over 2}\Delta )}\le C_p\cdot
\norm{du}_{L^2(\Delta )}.
\eqno(5.1.2)$$** ]{}
Definition 5.1.2. [*Define a [*cylinder*]{} $Z(a,b)$ by $Z (a,b) \deff S^1 \times [a,b]$, equipping it with coordinates $\theta \in [0,2\pi]$, $t\in[a,b]$, with the metric $ds^2= d\theta^2 + dt^2$ and the complex structure $J\st({\d \over \d\theta}) ={\d \over \d t}$. Denote $Z_i\deff Z(i-1,1)= S^1 \times [i-1,i]$.*]{}
Let $J^*$ be some continuous almost complex structure on $X$ and $A$ a subset of $X$ such that $J^*$ is uniformly continuous on $A$. Let $\mu_{J^*}$ denote the modulus of uniform continuity of $J^*$ on $A$.
Lemma 5.1.2. [*(Second A priori Estimate). *There exist constants $\gamma \in \,]\,0,1\,[$ and $\eps_2=\eps_2(\mu_{J^*}, A,h)>0$ such that for any $J$ with $\norm{ J- J^*} <\eps_2$ and every $J$-holomorphic map $u:Z(0,5)
\to X$ with $u(Z(0,5)) \subset A$ the condition $\norm{du}_{ L^2(Z_i)}<\eps_2$ for $i=1,\ldots,5$ implies $$\norm{du}^2_{L^2(Z_3)}\le {\gamma \over 2}\bigl(\norm{du}^2_{L^2(Z_2)} +
\norm{du}^2_{L^2(Z_4)}\bigr).
\eqno(5.1.3)$$** ]{}
Proof. Take $\eps_2>0$ small enough such that $\mu_{J^*}(\epsi_2) <
\epsi_1$, where $\eps_1$ is the constant from [*Lemma 5.1.1*]{}. Then for any $A' \subset A$ the condition $\diam(A') \le \epsi_2$ implies that $\osc(J^*,
A') \le \eps_1$. Due to [*Lemma 5.1.1*]{}, we may assume that $u(Z_i)\subset
B$ for $i=2,3,4$, where $B$ is a small ball in $\rr^{2n} =\cc^n$ with the structure $J\st$. Moreover, we may assume that $\norm{ J^* - J\st}_{
L^\infty(B)} \le \eps_1$.
Find $v\in C^0\cap L^{1,2}(Z(1,4),\cc^n)$ such that $\dbar_{J\st}v=0$ and $\norm{du - dv}_{L^2(Z(1,4))}$ is minimal. We have $$\norm{ \dbar_{J\st}(u-v)}_{L^2(Z_i)} = \norm{(J\st-J(u))\d_y u
}_{L^2(Z_i)}
\le \norm{J\st - J}_{L^\infty(B)}\norm{du}_{L^2(Z_i)}.$$ So for $i=2,3,4$ we get $$\norm{du - dv}_{L^2(Z_i)}\le C\norm{J\st -J}_{L^\infty(B)}
\norm{ du
}_{L^2(Z(1,4))}.\eqno(5.1.4)$$ Now let us check the inequality (4.7) for $v$. Write $v(z)= \Sigma_{k=
-\infty} ^\infty v_k e^{k(t+i\theta)}$. Then $\norm{dv}^2_{L^2 (S\times \{
t\} )}=4\pi \Sigma_{k=-\infty}^\infty k^2| v_k|^2 e^{2kt}$. Since obviously $$\int_2^3e^{2kt}\le {\gamma_1\over 2}
\left( \int_1^2e^{2kt}dt+\int_3^4e^{2kt}dt \right)
%\eqno(3.8A)$$ for all $k \not=0$ with $\gamma_1={2\over e^2}$, one gets the required estimate for all holomorphic $v$.
Using (5.1.4) with $\norm{J\st - J}_{L^\infty}$ sufficiently small, we conclude that the estimate (5.1.3) holds for $u$ with appropriate $\gamma > \gamma_1$.
Corollary 5.1.3.
*Let $X$, $h$, $J^*$, $A$, and the constants $\eps_2$ and $\gamma$ be as in [*Lemma 5.1.2*]{}. Suppose that $J$ is a continuous almost complex structure on $X$ with $\norm{J-J^*}_{L^\infty(A)}
<\eps_2$ and $u \in C^0\cap L^{1,2}(Z(0,l),X)$ a $J$-holomorphic map such that $u(Z)\subset A$ and $\norm{du}_{ L^2(Z_i)}<\eps_2$ for any $i=1,\ldots
,l$. Let $\lambda>1$ be (the uniquely defined) real number with $\lambda = {\gamma \over 2} (\lambda^2+ 1)$.*
Then for $2\le k\le l-1$ one has $$\norm{du}^2_{L^2(Z_k)} \le \lambda^{-(k-2)} \cdot \norm{du}^2_{L^2(Z_2)}
+ \lambda^{-(l-1-k)} \cdot \norm{du}^2_{L^2(Z_{l-1})}.
\eqno(5.1.5)$$
Proof. The definition of $\lambda$ implies that for any $a_+$ and $a_-$ the sequence $y_k \deff a_+ \lambda^k + a_- \lambda^{-k}$ satisfies the recurrent relation $y_k = {\gamma\over2}(y_{k-1} + y_{k+1})$. In particular, so does the sequence $$A_k \deff
\msmall{ \lambda^{-(k-2)} - \lambda^{6-2l+ k-2}
\over 1- \lambda^{6-2l} } \norm{du}^2_{L^2(Z_2)}
+ \msmall{ \lambda^{-(l-1-k)} - \lambda^{6-2l+ l-1-k}
\over 1- \lambda^{6-2l} } \norm{du}^2_{L^2(Z_{l-1})},$$ which is determined by the values $A_2 = \norm{du}^2_{L^2(Z_2)}$ and $A_{l-1}
= \norm{du}^2_{L^2(Z_{l-1})}$.
We claim that for $2\le k\le l-1$ one has the estimate $\norm{du}^2 _{L^2(
Z_k)} \le A_k$, which is obviously stronger than (5.1.5). Suppose that there exists a $k_0$ such that $2\le k_0\le l-1$ and $\norm{du}^2 _{L^2(Z_{k_0})}>A_{k_0}$. Choose $k_0$ so that the difference $\norm{du }^2_{L^2(Z_{k_0})}-A_{k_0}$ is maximal. By [*Corollary 2.5.1*]{} and by our recurrent definition of $A_k$, we have $2< k_0 < l-1$ and $$\norm{du }^2_{L^2(Z_{k_0})}-A_{k_0}\le {\gamma \over 2}(
\norm{du }^2_{L^2(Z_{k_0+1})}-A_{k_0+1}+
\norm{du }^2_{L^2(Z_{k_0-1})}-A_{k_0-1})\le$$ $$\le {\gamma \over 2}2(\norm{du }^2_{L^2(Z_{k_0})}-A_{k_0}).$$ The second inequality follows from the fact that $\norm{du}^2_{L^2(
Z_{k_0})} - A_{k_0}$ is maximal. This gives a contradiction.
[5.2. Removal of Point Singularities.]{}
An immediate corollary of this estimate is the following improvement of the Sacks-Uhlenbeck theorem about removability of a point singularity, see \[S-U\] and \[G\].
Corollary 5.2.1.
*(Removal of Point Singularities). *Let $X$ be a manifold with a Riemannian metric $h$, $J$ a continuous almost complex structure, and $u:(\check\Delta, J\st)\to (X,J)$ a $J$ - holomorphic map from the punctured disk. Suppose that**
$J$ is uniformly continuous on $A \deff u(\check\Delta)$ $h$ and the closure of $A$ is $h$-complete;
there exists $i_0$ such that, for all annuli $R_i\deff\{ z\in \cc
:{1\over e^{i+1}}\le | z| \le {1\over e^i}\} $ with $i\ge i_0$, one has $\norm{du}^2_{L^2(R_i)}\le \eps_2$, where $\eps_2$ is defined in [*Lemma 5.1.2*]{}.
Then $u$ extends to the origin.
Condition is automatically satisfied if $A= u(\check\Delta)$ is relatively compact in $X$. Condition of “slow growth” is clearly weaker than just the boundedness of the area, see, e.g., \[S-U\], \[G\]. It is sufficient to have $\lim_{i\lrar\infty} \norm{du}^2_{L^2(R_i)} =0$, whereas boundedness of the area means $\sum_{i=1}^\infty \norm{du}^2_{L^2(R_i)} <\infty$.
Proof. The exponential map $\exp(t,\theta) \deff e^{-t+i\theta}$ defines a biholomorphism between the infinite cylinder $Z(0,\infty)$ and the punctured disk $\check\Delta$, identifying every annulus $R_i$ with the cylinder $Z(i,i+1)$. Applying [*Corollary 5.1.3*]{} to the map $u \scirc \exp$ on cylinders $Z(i_0,l)$ and setting $l\lrar \infty$, we obtain the estimate $$\norm{du}^2_{L^2(R_i)} \le \lambda^{-(i-i_0)} \cdot
\norm{du}^2_{L^2(R_{i_0})}, \qquad i>i_0.$$ Using this and [*Lemma 5.1.1*]{} we conclude that $\diam(u(R_i)) \le C\cdot
\lambda^{-i/2}$ for $i>i_0$. Since $\sum \lambda^{-i/2} <\infty$, $u$ extends continuously into $0\in \Delta$.
In the proof of the compactness theorem we shall use the following corollary of [*Corollary 2.5.1*]{}. Let $X$ be a manifold with a Riemannian metric $h$, $J$ a continuous almost complex structure on $X$, $A\subset X$ a closed $h$-complete subset such that $J$ is $h$-uniformly continuous on $A$. Furthermore, let $\{J_n\}$ be a sequence of almost complex structures uniformly converging to $J$, $\{l_n\}$ a sequence of integers with $l_n\to \infty$ and $u_n:Z(0,l_n)
\to X$ a sequence of $J_n$-holomorphic maps.
Lemma 5.2.2.
*Suppose that $u_n(Z(0,l_n))\subset A$ and $\norm{du_n}_{L^2 (Z_i)}\le
\eps_2$ for all $n$ and $i\le l_n$. Take a sequence $k_n\to \infty$ such that $k_n<l_n-k_n\to \infty$. Then*
[*1)*]{} $\norm{du_n}_{L^2(Z(k_n,l_n-k_n))}\to 0$ and $\diam\bigl(u_n
(Z(k_n,l_n-k_n))\bigr)\to 0$;
[*2)*]{} if, in addition, all images $u_n(Z(0,l_n))$ are contained in some bounded subset of $X$, then there is a subsequence $\{ u_n \}$, still denoted $\{u_n \}$ such that both $u_n |_{Z(0,k_n)}$ and $u_n |_{Z(k_n,l_n)}$ converge in $L^{1,p}$-topology on compact subsets in $ \check \Delta \cong Z(0, +\infty )$ to $J$-holomorphic maps $u^+_\infty: \check \Delta \to X$ and $u^-_\infty:
\check \Delta \to X$. Moreover, both $u^+_\infty$ and $u^-_\infty$ extend to the origin and $u^+_\infty(0)= u^-_\infty(0)$.
Remarks. 1. The punctured disk $\check\Delta$ with the standard structure $J_{\Delta } {\d \over \d r}={1\over r}{\d \over \d \theta }$ is isomorphic to $Z(0,\infty )$ with the structure $J_Z{\d \over \d t}=-{\d \over
\d \theta }$ under a biholomorphism $(\theta ,t)\mapsto e^{-t+ \isl\theta}$. Thus, statement (2) of this corollary is meaningful.
2\. [*Lemma 5.2.2*]{} describes explicitly how the sequence of $J_n$-holomorphic maps of the cylinders of growing conformal radii converges to a $J$-holomorphic map of the standard node.
Lemma 5.2.3. [*There is an $\eps_3 =\eps_3(\mu_{J_\infty},A,h)$ such that for any continuous almost-complex structure $J$ on $X$ with $\norm{J
- J_\infty}_{L^\infty}\le \eps_3 $ and any non-constant $J$-complex sphere $u: \cc\pp^1 \to X$, $u(\cc\pp^1)\subset A$ one has the inequalities $$\area(u(\cc\pp^1 ))\ge \eps_3
\qquad\text{and}\qquad
\diam(u(\cc\pp^1))\ge \eps_3.$$* ]{}
Proof. Let $\eps_1$ be the constant from [*Lemma 5.1.1*]{}. Suppose that $\area u(\cc\pp^1)=\norm{du}^2_{L^2(\cc\pp^2)}\le \eps_1^2$. Cover $\cc\pp^1$ by two disks $\Delta_1$ and $\Delta_2$. By (3.1) and the Sobolev imbedding $L^{1,p} \subset C^{0,1-{2\over p}}$, we obtain that $\diam(u(\Delta_1))$ and $\diam(u (\Delta_2) )$ are smaller than $const\cdot \eps_1$. Thus the diameter of the image of the sphere is smaller than $const\cdot \eps_1$.
Therefore, we can suppose that the image $u(S^2)$ is contained in the coordinate chart, i.e., in a subdomain in $\cc^n$, and the structures $J$ and $J_\infty$ are $L^\infty$-close to a standard one. Consider now $u:S^2\to U\subset \cc^n$ as a solution of the linear equation $$\d_xv(z) + J(u(z))\cdot\d_yv(z) = 0\eqno(5.2.1)$$ on the sphere. The operator $\dbar _J(v) = \d_xv(z) + J(u(z))\cdot
\d_yv(z)$ acts from $L^{1,p}(S^2, \cc^n)$ to $L^p(S^2, \cc^n)$ and is a small perturbation of the standard $\dbar$-operator. Note that the standard $\dbar$ is surjective and Fredholm. Thus small perturbations are also surjective and Fredholm, having the kernel of the same dimension. But the kernel of $\dbar$ consists of constant functions. Since all constants are in the kernel of (5.2.1), our $u$ should be a constant map.
We have proved that if the area or a diameter of a $J$-holomorphic map is sufficiently small then this map is constant.
Remark. The same statement is true for the curves of arbitrary genus $g$. In that case, in addition to the estimate (5.1.2), one should also use the estimate (5.1.3). This yields the existence of an $\eps $ which depends on $g$ (and, of course, on $X$, $J$, and $K$), but not on the complex structure on the parameterizing surface.
[5.3. Compactness for Curves with Free Boundary.]{}
In this section we give a proof of the Gromov compactness theorem for the curves with boundaries of fixed finite topological type and without boundary conditions on maps. The case of closed curves is obviously included in this.
Throughout this section we assume that the following setting holds.
*Let $X$ be a manifold with a Riemannian metric $h$, $J_\infty$ a continuous almost complex structure on $X$, $A\subset X$ an $h$-complete subset, $\{ C_n
\}$ a sequence of nodal curves parameterized by a real surface $\Sigma$ with parameterizations $\delta_n:\Sigma \to C_n$, and $u_n:(C_n,j_n) \to (X, J_n)$ a sequence of holomorphic maps. Further, $J_\infty$ is $h$-uniformly continuous on $A$, $J_n$ are also continuous and converge to $J_\infty$, $h$-uniformly on $A$, $u_n(C_n)\subset A$ for all $n$.*
Let us explain the main idea of the proof of [*Theorem 1.1*]{}. The Gromov topology on the space of stable curves over $X$ is introduced in order to recover natural convergence of sequences $(C_n,u_n)$ of complex curves of bounded area which do not converge in the “strong" (i.e., $L^{1,p}$-type) sense. The are two reasons for this. The first is that a sequence of (say, smooth) curves $C_n$ could diverge in an appropriate moduli space and the second is a phenomenon of “bubbling". In both cases one has to deal with the appearance of new nodes, i.e., with a certain degeneration of the complex structure on curves. The “model” situation of [*Lemma 5.2.2*]{} describes a convergence of “long cylinders” $u_n :Z(0, l_n) \to X$, $l_n \lrar \infty$, to a node $u_\infty: \cala_0 \to
X$. In our proof we cover curves $C_n$ by pieces which are either “long cylinders” converging to nodes or have the property that complex structures and maps “strongly” converge. Here the “strong” convergence means the usual one, i.e., the $C^\infty$-topology for complex structures, and the $L^{1,p}$-topology with some $p>2$ for maps. In fact, the strong convergence of maps is equivalent to the uniform one, i.e., the $C^0$-topology, and implies further regularity in the case when $J_n$ and $J_\infty$ have more smoothness. One consequence of this is that we remain in the category of nodal curves. Another is that we treat degeneration of a complex structure on $C_n$ and the “bubbling” phenomenon in a uniform framework of “long cylinders”.
For the proof we need some additional results.
Lemma 5.3.1. *For any $R>1$ there exists an $a^+= a^+(R)>0$ with the following property. For any cylinder $Z = Z(0,l)$ with $0 <l\le +\infty$ and any annulus $A \subset Z(0,l)$, which is adjacent to $\d_0 Z = S^1\times
\{0\}$ and has a conformal radius $R$, one has $Z(0,a^+) \subset A$.*
Proof. Without loss of generality we may assume that $l=+\infty$ and identify $Z$ with the punctured disk $\check\Delta$ via the exponential map $(t+ \isl \theta) \mapsto e^{ -t+ \isl \theta}$ such that $\d_0 Z$ is mapped onto $S^1= \d\Delta$.
Suppose that the statement is false. Then there would exist holomorphic imbeddings $f_n : A(1, R) \to \check\Delta$ and points $a_n \in \Delta \bs
f_n(A(1, R))$ such that $f_n(A(1, R))$ are adjacent to $\d\Delta$ and $a_n
\lrar a\in \d\Delta$. Passing to a subsequence, we may assume that $\{f_n\}$ converges uniformly on compact subsets in $A(1, R)$ to a holomorphic map $f: A(1, R) \to\Delta$.
If $f$ is not constant, then $f(A(1, R))$ must contain some annulus $\{ b
<|z|<1 \}$ with $b<1$. But then $\{ \sqrt{b} <|z|<1\} \subset f_n(A(1, R))$ for $n>\!>1$, which is a contradiction.
If $f$ is constant, then the diameter of images of the middle circle $\gamma\deff \{ |z|= \sqrt{R} \} \subset A(1, R)$ must converge to $0$. But $\diam(f_n(\gamma)) \ge \dist(0, a_n) \lrar1$. The obtained contradiction finishes the proof.
For the proof of [*Theorem 1.1*]{} we need a special covering of $\Sigma$ which will be constructed in the following theorem.
Theorem 5.3.2.
*Under the conditions of [*Theorem 1.1*]{}, after passing to a subsequence, there exist a finite covering $\calv$ of $\Sigma$ by open sets $V_\alpha$ and parameterizations $\sigma_n:\Sigma \to C_n$ such that*
[*(a)*]{} all $V_\alpha$ are either disks, or annuli or pants;
[*(b)*]{} for any boundary circle $\gamma_i$ of $\Sigma$ there is some annulus $V_\alpha$ adjacent to $\gamma_i$;
[*(c)*]{} $\sigma_n^*j_n\ogran_{V_\alpha}$ does not depend on $n$ if $V_\alpha$ is a disk, pants or an annulus adjacent to a boundary circle of $\Sigma$;
[*(d)*]{} all non-empty intersections $V_\alpha \cap V_\beta$ are annuli, where the structures $\sigma^*j_n$ are independent of $n$;
[*(e)*]{} if $a$ is a node of $C_n$ and $\gamma^n_a=\sigma_n\inv (a)$ the corresponding circle, then $\gamma_a^n=\gamma_a$ does not depend on $n$ and is contained in some annulus $V_\alpha$, containing only one such “contracting" circle for any $n$; moreover, the structures $\sigma_n^*j_n\ogran_{V_\alpha\bs \gamma_a}$ are independent of $n$;
[*(f)*]{} if $V_\alpha$ is an annulus and $\sigma_n( V_\alpha)$ doesn’t contains nodes, then the conformal radii of $\sigma_n( V_\alpha)$ converge to some positive $R_\alpha^\infty>1$ or to $+\infty$;
[*(g)*]{} if for initial parameterizations $\delta_n$ and fixed annuli $A_i$, each adjacent to the boundary circle $\gamma_i$ of $\Sigma$, the structures $\delta_n^* j_n\ogran_{A_i}$ do not depend on $n$, then the new parameterizations $\sigma_{n}$ can be taken equal to $\delta_{n}$ on some subannuli $A'_i \subset A_i$ also adjacent to $\gamma_i$.
Proof. We shall prove the properties [*(a)–(f)*]{}. The property [*(g)*]{} will follow from [*Lemma 5.3.3*]{} below.
There are four cases where the existence of such a covering is obvious. If all $C_n$ are disks or annuli without nodal points, there is nothing to prove. In the third case each $C_n$ is a sphere, and we cover it by two disks.
In the forth case each $C_n$ is a torus without marked points. Then any complex torus can be represented by the form $\cc{\bigm/}(\zz+
\tau \zz)$ with $|\re\tau| \le {1\over 2}$ and $\im\tau > {1\over 2}$. Considering the map $z\in \cc \mapsto e^{2\pi\isl z} \in \check\cc \deff
\cc\bs\{0\}$, we represent $(T^2,j)$ as the quotient $\cc\bigm/ \{ z \sim
\lambda^2 z\}$ with $\lambda = e^{\pi\isl\tau}$, so that $|\lambda|<
e^{-\pi/2}< {1\over 3}$. The annuli $\{ {|\lambda| \over2} <|z| <1\}$ and $\{{|\lambda|^2 \over2} <|z| <|\lambda| \}$ form the needed covering.
In all remaining cases we start with the construction of appropriate graphs $\Gamma_n$ associated with some decomposition of $C_n$ into pants. [*Lemma 5.1.2*]{} and a non-degeneration of the complex structure $j_n$ on $C_n$ shows that lengths of all boundary circles of all non-exceptional components $C_{n,i}$ of $C_n$ are uniformly bounded from above. At this point we make the following
Remark. The Collar Lemma from \[Ab\], Ch.II, §3.3 yields the existence of the universal constant $l^*$ such that for any simple geodesic circles $\gamma'$ and $\gamma''$ on $C_{n,i}$ the conditions $\ell(\gamma') <l^*$ and $\ell( \gamma'') <l^*$ imply $\gamma' \cap \gamma'' = \emptyset$. We shall call geodesic circles $\gamma$ with $\ell(\gamma) <l^*$ [*short geodesics*]{}.
The fact that $(C_n,u_n)$ are $J_n$-complex and of bounded area shows that $C_n$ have a uniformly bounded number of components. Indeed, the number of exceptional components, which are spheres and disks, is bounded by the energy (see [*Lemma 5.2.3*]{}), and the number of boundary circles of $C_n$ is equal to that of $\Sigma$. Furthermore, the operation of contracting a circle on $\Sigma$ to a nodal point either diminishes the genus of some component of $C_n$ or increases the number of components. Thus, the number of nodal points on $C_n$ and the total number of marked points on its components are also uniformly bounded. This implies that the number of possible topological types of components is finite.
In this situation the Teichmüller theory (see \[Ab\], Ch.II, §3.3) states the existence of decomposition of every non-exceptional component $C_{n,i}\bs
\mapo$ into pants with the following properties:
Every short geodesic is a boundary circle of some pants of the decomposition.
The intrinsic length of every boundary circle is bounded from above by a (uniform) constant depending only on an upper bound of lengths of boundary circles and possible topological types of $C_{n,i} \bs \mapo$.
Having decomposed all $C_{n,i} \bs \mapo$ into pants, we associate with every curve $C_n$ its graph $\Gamma_n$. As was noted above, the number of vertices and edges of $\Gamma_n$ is uniformly bounded. Thus, after passing to a subsequence, we can assume that all $\Gamma_n$ are isomorphic to each other (as marked graphs). Denote this graph by $\Gamma$. Now, the parameterizations $\sigma_n: \Sigma \to C_n$ can be found in such a way that the decompositions of $C_{n,i} \bs\mapo$ into pants define the same set $\bfgamma =\{ \gamma
_\alpha \}$ of circles on $\Sigma$ and induce the same decomposition $S\bs
\cup_\alpha \gamma_\alpha = \cup_j S_j$ with the graph $\Gamma$.
By our construction of the graph $\Gamma$, each edge of $\Gamma$ corresponds either to a circle in $\Sigma$ contracted by every parameterization $\sigma_n$ to a nodal point, or to a circle mapped by every $\sigma_n$ onto a geodesic circle separating two pants. Furthermore, each tail of $\Gamma$ corresponds to a boundary circle of $\Sigma$. Thus, we shall use the same notation $\gamma
_\alpha$ for an edge or a tail of $\Gamma$ and for the corresponding circle on $\Sigma$. If $\sigma_n(\gamma_\alpha)$ is a boundary circle of some pants $C_j$, then the intrinsic length $\ell_{n,\alpha}= \ell_n (\gamma_\alpha)$ of $\sigma_n(\gamma_\alpha)$ is well- defined. This happens in the following two cases:
[*a)*]{} $\sigma_n(\gamma_\alpha)$ separates two pants, or else
[*b)*]{} $\gamma _\alpha$ is a boundary circle $\Sigma$ and the irreducible component of $C_n$ attached to $\sigma_n(\gamma_\alpha)$ is not a disk with a single nodal point. Note that the appearance of these two cases is independent of $n$.
By our choice of $\gamma_\alpha$, the lengths $\ell_n (\gamma_\alpha)$ are uniformly bounded from above. Passing to a subsequence, we may assume that for any fixed $\alpha$ the sequence $\{ \ell_{n,\alpha} \}$ converges to $\ell_{\infty,\alpha}$.
As one can expect, the condition $\ell_{n, \alpha} \lrar 0$ means that the circle $\gamma_\alpha$ is shrunk to a nodal point on the limit curve. We shall prove the statement of the theorem by induction in the number $N$ of those circles $\gamma_\alpha$ for which $\ell_{\infty,\alpha} =0$.
The case $N=0$, where there are no such circles, is easy. Passing to a subsequence, we may assume that the Fenchel-Nielsen coordinates $(\bfell_n,
\bfvartheta_n)$ of any non-exceptional component $C_{n,i}$ of $C$ converge to the Fenchel-Nielsen coordinates $(\bfell_\infty, \bfvartheta_\infty)$ of some smooth curve $C_{\infty, i}$ with marked points. Gluing together appropriate pairs of marked points, we obtain a nodal curve $C_\infty$, which admits a suitable parameterization $\sigma_\infty: \Sigma \to C_\infty$ and has the same graph $\Gamma$. [*Lemma 5.1.3*]{} shows that for $n>\! >1$ the curves $C_n$ can be obtained from $C_\infty$ by deformation of the transition functions for the intrinsic local coordinates on non-exceptional components of $C_\infty$. Note that such a deformation can be realized as a deformation of the operator $j_\infty$ of a complex structure on $C_\infty$, localized in small neighborhoods of circles $\sigma_\infty(\gamma_\alpha)$, see Fig. 10. In the case where $\gamma_\alpha$ is a boundary circle we may additionally assume that the annulus, where $j_n$ changes, lies away from $\gamma_\alpha$. Now the existence of the covering with desired properties is obvious.
Let us now consider the general case where the number $N$ of “shrinking circles” is not zero. Take a circle $\gamma_\alpha$ with $\ell_\infty
(\gamma _\alpha) =0$. Let $S_j$ be pants adjacent to $\gamma_\alpha$. Consider the intrinsic coordinates $\rho_\alpha$ and $\theta_\alpha$ at $\sigma_n(\gamma_\alpha)$ and the annuli $$\eqalign{
A_{n,\alpha,j} &\deff \{\, (\rho
_\alpha, \theta_\alpha) \in \sigma_n(S_j) \;:\; 0\le \rho_\alpha \le {\pi^2
\over \ell_{n,\alpha} } - {2\pi \over a^*} \;\}
\cr
A^-_{n,\alpha,j} &\deff
\{\, (\rho_\alpha, \theta_\alpha) \in \sigma_n(S_j) \;:\;
0\le \rho_\alpha \le {\pi^2 \over \ell_{n,\alpha} } - {2\pi \over a^*}-1
\;\},
}$$ adjacent to $\sigma_n( \gamma_\alpha)$. Note that ${\pi^2 \over \ell_{n,
\alpha} }- {2\pi \over a^*}$ (resp. ${\pi^2 \over \ell_{n,\alpha} }- {2\pi
\over a^*}-1$) is the logarithm of the conformal radius of $A_{n, \alpha, j}$ (resp. of $A^-_{n, \alpha,j}$). Consequently, we can use [*Lemma 4.3.2*]{} to show that these annuli are well-defined.
If $\gamma_\alpha$ is a boundary circle, we set $C^-_n \deff C_n \bs A^-_{n,
\alpha,j}$. Otherwise $\gamma_\alpha$ separates two pants, say $S_j$ and $S_k$. Then we define in a similar way the annuli $A_{n,\alpha, k}$ and $A^-
_{n,\alpha,k}$, set $A_{n,\alpha} \deff A_{n, \alpha,j} \cup A_{n, \alpha,
k}$ and $A^-_{n,\alpha} \deff A^-_{n,\alpha,j} \cup A^-_{n, \alpha, k}$ and put $C^-_n \deff C_n \bs A^-_{n,\alpha}$.
The parameterizations $\sigma_n: \Sigma \to C_n$ can be chosen in such a way that the annuli $\sigma_n\inv(A^-_{n,\alpha,j})$ (resp. $\sigma_n\inv(A^-
_{n, \alpha,k})$) define the same annulus $A^-_{ \alpha,j}$ (resp. $A^-
_{\alpha, k}$) on $\Sigma$. Let $\gamma^-_{ \alpha,j}$ (resp. $\gamma^-
_{\alpha,k}$) denote its boundary circles different from $\gamma_\alpha$. Thus, the curves $C^-_n$ are parameterized by a real surface $\Sigma^- \deff
\Sigma \bs A^-_{ \alpha,j}$ (resp. $\Sigma^- \deff \Sigma \bs (A^-_{\alpha,
j} \cup A^- _{\alpha, k}$), and the restrictions of $\sigma_n$ can be chosen as parameterization maps. Thus, the decompositions of components of $C_n$ into pants define the combinatorial type of decompositions of components of $C^-_n$ into pants. Moreover, the corresponding graph $\Gamma^-$ will be the same for all $C^-_n$. It coincides with $\Gamma$ if $\gamma_\alpha$ is a boundary circle. Otherwise $\Gamma^-$ can be obtained from $\Gamma$ by replacing the edge corresponding to $\gamma_\alpha$ by two new tails for two new boundary components.
Take some non-exceptional component $C^-_{n,i}$ of $C^-_n$ and decompose it into pants according to graph $\Gamma^-$ in the canonical way, so that the boundary circles of the obtained pants are geodesic. Note that even if the constructed pants are in combinatorial one-to-one correspondence with the pants of $C_n$, the intrinsic metric on $C^-_{n,i}$ and the obtained geodesics circle $\gamma^-_\beta$ differ from the corresponding objects on $C_{n,i}$
Nevertheless, we claim that for the obtained decomposition of $C^-_n$ the intrinsic lengths are uniformly bounded from above (possibly by a new constant) and that the sequence $\{ C^-_n \}$ has fewer “shrinking circles” than $\{ C_n \}$. The meaning of the above construction is the following. The curves $C^-_n$ are obtained from $C^-_n$ by cutting off the annuli $A^-_{n,\alpha,j}$ (and resp. the annuli $A^-_{n,\alpha,j}$). These annuli are sufficiently long so that one “shrinking circle” disappears, but not too long so that the complex structures of the curves $C^-_n$ remain non-degenerating near the boundary.
Indeed, the complex structures on $C^-_n$ do not degenerate at the boundary circle $\gamma^-_{ \alpha,j}$ (resp. at $\gamma^-_{\alpha,k}$), because $C^-_n$ contain annuli $A_{n,\alpha,j} \bs A^-_{n,\alpha,j}$ (resp.$A_{n,\alpha,k} \bs A^- _{n,\alpha,k}$) of the constant conformal radius $R=e>1$. This means that the lengths $\ell^-_n(\gamma^-_{\alpha,j})$ of $\sigma_n( \gamma^- _{\alpha,j})$ (resp. $\ell^-_n(\gamma^-_{\alpha,k})$ of $\sigma_n(\gamma^- _{\alpha,k})$) with respect to the intrinsic metrics on $C^-_n$ are uniformly bounded.
On the other hand, the lengths $\ell^-_n(\gamma^-_{\alpha,j})$ (resp.$\ell^-_n(\gamma^-_{\alpha,k})$) are also uniformly bounded from below by a positive constant. Otherwise, by [*Lemma 3.4*]{}, after passing to a subsequence, there would exist annuli $A_n\subset C^-_n$ of infinitely increasing radii $R_n$, adjacent to $\sigma_n(\gamma^-_{\alpha,j})$ (resp.to $\sigma_n(\gamma^-_{ \alpha,k})$). The superadditivity of the logarithm of the conformal radius of annuli, see \[Ab\], Ch.II, §1.3, shows that the conformal radius $R^+_n$ of the annulus $A^-_{n,\alpha,j} \cup A_n$ satisfies the inequality $\log R^+_n \ge {\pi^2 \over \ell_{n,\alpha} }- {2\pi \over
a^*}-1 + \log R_n$, which contradicts [*Lemma 4.3.2*]{}, part [*i)*]{}.
Now we estimate the intrinsic lengths of boundary circles and the number of “shrinking circles” on $C^-_n$. Denote $\ell_n \deff \ell_n(\gamma_\alpha)$, where $\gamma_\alpha$ is the circle on $\Sigma$ used in the above constructions. Compute the width $L_n$ of $A_{n,\alpha,j} \bs A^- _{n, \alpha,
j}$ the intrinsic metric on $C_n$. Using $\ell_n \lrar0$, we obtain $$\eqalign{
L_n &= \int_{\rho={\pi^2 \over \ell_n }- {2\pi \over a^*}-1}
^{ {\pi^2 \over \ell_n }- {2\pi \over a^*}}
\left(\msmall{ {\ell_n \over 2\pi} \over \cos {\ell_n \rho \over 2\pi}}
\right) d\rho =
\left[ \log\;\cotan \left( {\pi \over 4} -{\ell_n \rho \over 4\pi} \right)
\right]_{\rho={\pi^2 \over \ell_n }- {2\pi \over a^*}-1}
^{{\pi^2 \over \ell_n }- {2\pi \over a^*}}
\cr
&= \log \msmall{ \cotan{\ell_n \over2a^*} \over
\cotan({\ell_n \over 2a^*} + {\ell_n \over4\pi}) }
\approx \log \msmall{ {\ell_n \over2a^*} + {\ell_n \over4\pi} \over
{\ell_n \over2a^* } } =
\log \left( 1 + \msmall{ a^* \over 2\pi } \right) >0.
}$$
Thus we can find annular neighborhoods $A_n\subset C^-_n$ of $\sigma_n(\gamma_\alpha)$ with constant conformal radius $R>1$. Let $\gamma^-_{n, \alpha} \subset C^-_n$ be the geodesic circle corresponding to $\gamma_\alpha$. It is the unique circle, which is homotopic to $\gamma_{n,
\alpha} = \sigma_n( \gamma_\alpha)$ and is geodesic the intrinsic metric on $C^-_n$. Let $\ell^-_{n,\alpha}$ denote its $C^-_n$-intrinsic length. Using the monodromy argument as in the proof of [*Lemma 5.1.2*]{}, we see that every $A_n$ can be isometrically imbedded into the annulus $A^+_n \deff (-{\pi^2\over \ell^-
_{n,\alpha}} , {\pi^2\over \ell^-_{n,\alpha}}) \times S^1$ with coordinates $
-{\pi^2\over \ell^-_{n,\alpha}} <\rho < {\pi^2\over \ell^-_{n,\alpha}}$, $0\le
\theta \le 2\pi$ and with metric $\Bigl({\ell^-_{n,\alpha} \over 2\pi} /
\cos{\ell^-_{n,\alpha} \rho\over 2\pi}\Bigr)^2 (d\rho^2 + d\theta^2)$ of conformal radius $e^{2\pi^2/\ell^-_{n,\alpha}}$. The monotonicity of conformal radius of annuli yields the upper bound $\ell^-_{n,\alpha} \le {2\pi^2 \over \log\,R}$.
The same argumentation shows that if $\ell_{\infty,\alpha} \not=0$, then $\ell^- _{n,\alpha}$ also do not vanish. Indeed, if $\ell^-_{n,\alpha}\lrar 0$, then we could find annular neighborhoods $A'_n\subset C^-_n \subset C_n$ of $\gamma_{n,\alpha}$ with infinitely increasing radii $R'_n$. Using the homotopic equivalence of $\gamma^-_{n,\alpha}$ with $\sigma_n(\gamma_\alpha)$ as above, we would obtain the estimate $\ell_{n,\alpha} \le {2\pi^2 \over \log\,R'_n}
\lrar 0$, which is a contradiction.
Thus, we have shown that $C^-_n$, with the intrinsic metric and defined by the $\Gamma^-$ decomposition, have uniformly bounded lengths of marked circles and less “shrinking circles” than $C_n$. Using induction, we may assume that the postulated covering of $\Sigma^-$ and parameterizations of $C^-_n$ by $\Sigma^-$ exist. Since $C_n \bs C^-_n$ is an annulus of increasing conformal radius, the statement of the theorem is valid for $C_n$.
Lemma 5.3.3. *Let $C_n$ be a sequence of complex annuli with structures $j_n$, $\Sigma$ some fixed annulus and $\delta_n: \Sigma \to C_n$ parameterizations. Suppose that for some fixed annuli $A_1, A_2 \subset
\Sigma$ adjacent to the boundary circles of $\Sigma$ restrictions $\delta_n^*j_n|_{A_i}$ do not depend on $n$.*
Then one can find parameterizations $\sigma_n: \Sigma \to C_n$ such that $\sigma_n$ coincide with $\delta_n$ on some (possibly smaller) annuli $A'_i$, are also adjacent to the boundary circles of $\Sigma$ restrictions, and that
if conformal radii $R_n$ of $C_n$ converge to $R_\infty <\infty$, then $\sigma_n^*j_n$ converge to some complex structure;
if conformal radii $R_n$ of $C_n$ converge to $\infty$, then for some circle $\gamma \subset \Sigma$ structures $\sigma_n^*j_n$ converge on compact subsets $K \Subset \Sigma \bs \gamma$ to a complex structure of disjoint union of two punctured disks. Moreover, such a $\gamma$, an arbitrary imbedded circle generating $\pi_1(\Sigma)$, can be chosen.
Proof. Without loss of generality we may assume that $\Sigma =
A(1,10)$, $A_1=A(7,10)$, $A_2=A(1,4)$ and that a given circle $\gamma$ lies in $A(3,7)$. Let $\delta_n : \Sigma \to C_n$ be the given parameterizations. There exist biholomorphisms $\phi_n: C_n \to A(r_n,1)$ with $r_n\inv=R_n$ being the conformal radii of $C_n$ such that $\phi_n (\delta_n(A_1))$ is adjacent to $\{ |z|=1\} = \d\Delta$ and $\phi_n (\delta_n(A_2))$ is adjacent to $\{ |z|=r_n\}$. Define $\phi'_n(z) \deff {r_n \over \phi_n(z)}$.
Recall that structures $\delta_n^*j_n|_{A_i}$ are independent of $n$. We call this structure $j$. Consider maps $\phi_n \scirc \delta_n : (A_1, j) \to
A(r_n,1) \subset \Delta$ and $\phi'_n \scirc \delta_n : (A_2, j) \to A(r_n,1)
\subset \Delta$. Passing to a subsequence we can suppose that $r_n \lrar
r_\infty <1$ and that maps $\phi_n \scirc \delta_n$, $\phi'_n \scirc \delta_n$ converge on $A_1\cup A_2$ to holomorphic maps $\psi: (A_1\cup A_2, j) \to
\Delta$ and $\psi': (A_1\cup A_2, j) \to \Delta$, respectively. This means that maps $(\phi_n \scirc \delta_n, \phi'_n \scirc \delta_n): (A_1 \cup A_2)
\to \Delta^2$ take values in $\{(z,z') \in \Delta^2 \;:\; z{\cdot} z'= r_n
\}$ and converge to the map $(\psi, \psi'): (A_1 \cup A_2) \to \Delta^2$ with values in $\{(z,z') \in \Delta^2 \;:\; z{\cdot}z' = r_\infty \}$.
Arguments from the proof of [*Lemma 5.3.1*]{} show that $\psi(A_1)$ and $\psi'(A_2)$ are annuli adjacent to $\d\Delta$. This implies that for $n>\!>
1$ there exist diffeomorphisms $(\psi_n, \psi'_n): \Sigma \to \{(z,z') \in
\Delta^2 \;:\; z{\cdot}z' = r_n \}$ such that $\psi_n \equiv \phi_n \scirc
\delta_n$ on $A(9,10)$, $\psi'_n(z) \equiv \phi'_n \scirc \delta_n(z)$ for $z\in A(1,2)$, and $(\psi_n, \psi'_n)$ converge to $(\psi, \psi')$ on $\Sigma$. Moreover, we may assume that $|\psi_n(t)|=|\psi'_n(t)|= \sqrt{r_n}$ for any $t\in \gamma$. This means that $(\psi_n, \psi'_n)(\gamma)$ lies on the middle circle $\{ (z,z'): |z|= \sqrt{r_n},\; z'= {r_n\over z} \}$ of $\{(z,z') \in \Delta^2 \;:\; z{\cdot}z' = r_n \}$.
Set $\sigma_n \deff \phi_n\inv \scirc \psi_n: \Sigma \to C_n$. Then, obviously, $\sigma_n \equiv \delta_n$ on $A(1,2)$ and on $A(9,10)$, and $\sigma_n^*j_n
= (\psi_n,\psi')_n^* J\st \lrar (\psi,\psi')^* J\st$, where $J\st$ denotes the standard complex structure on $\Delta^2$.
Proof of Theorem 4.1.1. Let $\{(C_n, u_n)\}$ be the sequence from the hypothesis of the theorem. Then the condition [*c)*]{} of the theorem and [*Lemma 5.3.1*]{} provide the existence of parameterizations $\delta_n: \Sigma \to
C_n$ and annuli $A_i$, adjacent to each boundary circle $\gamma_i$, such that $\delta_n^* j_{C_n}$ are constant in every $A_i$. Thus we may assume that such $\delta_n$ are given.
Take a covering $\calv= \{V_\alpha\}$ and parameterizations $\sigma_n$ as in [*Theorem 4.3*]{}. With every such covering we can associate the curves $C_{\alpha,n} \deff \sigma_n(V_\alpha)$, the parameterizations $\sigma_{\alpha,
n} \deff \sigma_n\ogran_{V_\alpha}: V_\alpha \to C_{\alpha,n}$ and the maps $u_{\alpha,n} \deff u_n \ogran_{C_{\alpha, n}} : C_{\alpha, n} \to X$. Consider the following type of convergence of sequences $\{ (C_{\alpha, n}, u_{\alpha,
n}, \sigma_{\alpha, n} )\}$ with the fixed $\alpha$:
**
[A)]{} $C_{\alpha, n}$ are annuli of infinitely growing conformal radii $l_n$ and the conclusions of [*Lemma 5.2.2*]{} hold;
[B)]{} every $C_{\alpha, n}$ is isomorphic to the standard node $\cala_0=
\Delta \cup_{ \{0\} } \Delta$ such that the compositions $V_\alpha \buildrel
\sigma _{\alpha, n} \over \lrar C_{\alpha, n} \buildrel \cong \over \lrar
\cala_0$ define the same parameterizations of $\cala_0$ for all $n$; furthermore, the induced maps $\ti u_{\alpha, n}: \cala_0 \to X$ strongly converge;
[C)]{} the structures $\sigma_n^*j\vph_n \ogran_{V_\alpha}$ and the maps $u_{\alpha, n}\scirc \sigma _{\alpha, n}: V_\alpha \to X$ strongly converge.
Here the strong convergence of maps is the one in the $L^{1,p}$-topology on compact subsets for some $p>2$ (and hence for all $p<\infty$), and the convergence of structures means the usual $C^\infty $-convergence.
Suppose that there is a subsequence, still indexed by $n\to \infty$ such that for any $V_\alpha$ we have one of the convergence types [*A)–C)*]{}. Then the sequence of global maps $\{ (C_n , u_n, \sigma _n )\}$ converges in the Gromov topology which gives us the proof, as well as a precise description of the convergence picture.
Otherwise, we want to find a refinement of our covering $\calv$ and parameterizations $\sigma_n$ which have the needed properties. We shall proceed by induction, estimating an area of pieces of coverings of $\Sigma$. To do this, we fix $\eps>0$ satisfying $\eps \le {\eps_1^2 \over2 }$ with $\eps_1$ from [*Lemma 5.1.1*]{}, $\eps \le {\eps_2^2 \over2 }$ with $\eps_2$ from [*Lemma 5.1.2*]{} and $\eps \le {\eps_3\over 3}$ with $\eps_3$ from [*Lemma 5.2.3*]{}. First consider the
[*Special case: $\area(u_n(\sigma_n(V_\alpha))) \le \eps$ for any $n$ and any $V_\alpha \in \calv$*]{}. We can consider every $V_\alpha$ separately. If the structures $\sigma_n^* j_n |_{V_\alpha}$ are constant, then some subsequence of $u_n \scirc \sigma_n$ strongly converges due to [*Corollary 2.5.1*]{}.
If structures $\sigma_n^* j_n |_{V_\alpha}$ are not constant, then $V_\alpha$ must be an annulus. Fix biholomorphisms $\phi_n: Z(0,l_n) \buildrel \cong
\over \lrar \sigma_n(V_\alpha)$. If $l_n \lrar \infty$, [*Lemma 5.2.2*]{} shows that (and describes how!) an appropriate subsequence of $u_n \scirc \phi_n$ converges to a $J_\infty$-holomorphic map of a standard node. Otherwise, we can find a subsequence, still denoted $(C_n, u_n)$, for which $l_n \lrar l_\infty
<\infty$ and $u_n \scirc \phi_n$ converge to a $J_\infty$-holomorphic map of $Z(0,l_\infty)$ in $L^{1,p}$-topology on compact subsets $K \Subset Z(0,l_\infty)$ for any $p<\infty$. To construct refined parameterizations $\ti\sigma _{\alpha,
n} : V_\alpha \to C_{\alpha, n}= \sigma_n(V_\alpha)$, we use property [*(d)*]{} from [*Theorem 5.3.2*]{} and apply [*Lemma 5.3.1*]{}.
Thus, we reach one of the convergence types [*A)–C)*]{} which gives the proof in this [*Special case*]{}.
[*General case.*]{} Suppose that the theorem is proved for all sequences of $J_n$-holomorphic curves $\{(C_n, u_n)\}$ with parameterizations $\delta_n: \Sigma \to C_n$ which satisfy the additional condition $\area(u_n(C_n)) \le (N-1) \eps$ for all $n$. We see this as the hypothesis of the induction in $N$, so that our [*Special case*]{} is the base of the induction.
Assume that there exists a subsequence, still indexed by $n\to \infty$ such that for every $V_\alpha$ and for the curves $C_{\alpha, n} = \sigma_n(
V_\alpha )$ the statement of the theorem holds. This means that there exist refined coverings $V_\alpha= \cup_i V _{\alpha, i}$ and new parameterizations $\ti\sigma _{\alpha, n}: V_\alpha \to C_{\alpha, n}$ such that the $\ti\sigma_n$ coincide with the $\sigma_n$ near the boundary of every $V_\alpha$ and such that for the curves $C_{\alpha, i,n} \deff \ti\sigma_n (V _{\alpha ,i})$ we have convergence of one of the types [*A)–C)*]{}. Then we can glue $\ti\sigma
_{\alpha, n}$ together to global parameterizations $\ti\sigma_n: \Sigma \to
C_n$ and set $\wt\calv \deff \{ V _{\alpha, i} \}$, obtaining the proof.
In particular, due to the inductive hypothesis, this is also true for any $V_\alpha$ such that $\area(u_n(\sigma_n(V_\alpha))) \le (N-1) \eps$ for all $n$.
This implies that it is sufficient to consider only those $V_\alpha$ for which $(N-1) \eps\le \area(u_n(\sigma_n( V_\alpha ))) \le N \eps$ for all $n$. Obviously, it is sufficient to show the desired property only for such a piece of covering, say for $V_1$. To construct the refined parameterizations $\ti\sigma_{1,n}$ and the covering $V_1 = \cup_i V _{1, i}$, we consider four cases.
. Then we can realize $(V_1, \sigma_n^*j_n)$ as a constant bounded domain $D$ in $\cc$. Hence we can consider $u_n \scirc \sigma_n: V_\alpha \to
X$ as holomorphic maps $u_n :D \to (X,J_n)$. Now we use the “patching construction” of Sacks-Uhlenbeck \[S-U\].
Fix some $a>0$. Denote $D_{-a} \deff \{ z\in D\;:\; \Delta(z,a) \subset D\}$. Find a covering of $D_{-a}$ by open sets $U_i \subset D$ with $\diam(U_i) <a$ such that any $z\in D$ lies in at most 3 pieces $U_i$. Then for any $n$ there exists at most $3N$ pieces $U_i$ with $\area( u_n(U_i)) > \eps$. Taking a subsequence, we may assume that the set of such “bad” pieces $U_i$ is the same for all $n$. Repeat successively the same procedure for ${a\over2}$, ${a\over4}$, and so on, and then take a diagonal subsequence. We obtain at most $3N$ “bad” points $y^*_1, \ldots, y^*_l$ such that a subsequence of $u_n$ converges in $D\bs \{y^*_1, \ldots,y^*_l\}$ strongly, i.e., in the $L^{1,p}$-topology on compact subsets $K \Subset D\bs \{y^*_1, \ldots,y^*_l\}$. These “bad” points $y^*_1, \ldots, y^*_l$ are characterized by the property $$\text{\ \ for any $r>0$\ \ }
\area(u_n(\Delta(y^*_i, r)) >\eps \text{\ \ for $n$ all sufficiently big}.
\eqno(5.3.1)$$
Remark. As we shall now see, every such point is the place where the “bubbling phenomenon” occurs. Therefore we shall call $y^*_i$ [*bubbling points*]{}. The characterization property of a bubbling point is (4.1).
If there are no bubbling points, i.e., $l=0$, then the chosen subsequence $u_n$ converges strongly and we can finish the proof by induction.
Otherwise, we consider the first point $y^*_1 \in D$. Take a disk $\Delta
(y^*_1,\varrho)$ which doesn’t contain any other bubbling points $y^*_i$, $i>1$.
Then for any $n$ we can find the unique $r_n$ such that
\(1) $r_n \le {\varrho \over 2}$ and $\area(u_n(\Delta(x, r_n)))\le\eps$ for any $x\in \barr\Delta (y^*_1,{\varrho \over 2})$;
\(2) $r_n$ is maximal (1).
Then $r_n\lrar 0$, because otherwise for $r^+ \deff {\sf lim\,sup}\; r_n >0$ and for some subsequence $n_k \lrar \infty$ with $r_{n_k} \lrar r^+$ we would have $$\area\bigl(u_{n_k}(\Delta(y^*_1, r^+))\bigr) \le \eps,$$ which would contradict (5.3.1).
Lemma 5.3.4. [*For every $n>\!>1$ there exists $x_n\in \barr
\Delta(y^*_1, {\varrho \over2})$, such that $x_n\to y^*_1$ and $\area( u_n(
\Delta(x_n, r_n))) =\eps$.*]{}
Proof. If not, then for some subsequence $n_k\lrar \infty$ and every $x\in \barr\Delta (y^*_1, {\varrho \over2})$ we would have $\area(u_{n_k}(
\Delta (x, r_{n_k})))< \eps $. Since $r_n \lrar0$, this would contradict the maximality of $r_n$. In particular, there exists the postulated sequence $\{ x_n \}$.
If $x_n$ do not converge to $y^*_1$, then after going to a subsequence we would find $y'=\lim_{n\to \infty } x_n\not= y^*_1$. By our construction, $y'$ does not coincide with any other bubbling point $y^*_i$. Take $a>0$ such that $\Delta(y', a)$ contains no bubbling point. Then by [*Corollary 2.5.1*]{} some subsequence $u_{n_k}$ would converge to some $u' \in L^{1,p}_\loc (\Delta(y',
a), X)$ in strong $L^{1,p}(K)$-topology for any compact subset $K \Subset
\Delta(y', a)$ and any $p<\infty$. In particular, for sufficiently small $b<a$ we would have $\area(u_{n_k}(\Delta(y', b))) \lrar \area(u'( \Delta(y',
b))) < \eps$, which would contradict the choice of $r_n$ and $x_n$.
Using constructed $r_n$ and $x_n$, define maps $v_n:\Delta (0,{\varrho \over
2r_n}) \to (X,J_n)$ by $v_n(z)\deff u_n(x_n +r_n z)$. By the definition of $r_n$ we have $$\area(v_n(\Delta (x ,1))\le \eps \ \ \text{for all $x\in \Delta
(0,{\textstyle{\varrho \over 2r_n}}-1)$}.
\eqno(5.3.2)$$ On the other hand, $\area(v_n(\Delta(0,1)) = \area(u_n( \Delta (x_n, r_n)) =
\eps$ by [*Lemma 5.3.4*]{}. Thus $v_n$ converge (after going to a subsequence) on compact subsets in $\cc$ to a nonconstant $J_\infty $-holomorphic map $v_\infty$ with finite energy. Consequently, $v_\infty$ extends onto $S^2$ by the removable singularity theorem of [*Corollary 5.2.1*]{}.
Since $v_\infty$ is nonconstant, $\norm{ dv_\infty }^2_ {L^2( S^2)}= \area(
v_\infty(S^2))\ge 3\eps$ by [*Lemma 5.2.3*]{} and the choice of $\eps$. Choose $b>0$ in such a way that $$\area(v_\infty (\Delta (0,b)) =
\norm{ dv_\infty}_{L^2(\Delta (0,b))}^2 \ge 2\eps.
\eqno(5.3.3)$$ By [*Corollary 2.5.1*]{} this provides that $$\norm{du_n}_{L^2(\Delta (x_n,br_n))}^2=
\norm{ dv_n}_{L^2(\Delta (0,b))}^2 \ge\eps.
\eqno(5.3.4)$$
For $n>\!>$ we consider the coverings of $V_1$ by 3 sets $$\textstyle\mathsurround=0pt
V^{(n)}_{1,1} \deff V_1 \bs \barr\Delta (y^*_1, {\varrho\over2}),
\qquad
V^{(n)}_{1,2}\deff \Delta (y^*_1, \varrho) \bs \barr \Delta (x_n, br_n),
\qquad
V^{(n)}_{1,3} \deff \Delta (x_n, 2br_n).$$ Fix $n_0$ sufficiently big. Denote $V_{1,1} \deff V^{(n_0)}_{1,1}$, $V_{1,2}
\deff V^{(n_0)}_{1,2}$, and $V_{1,3} \deff V^{(n_0)}_{1,3}$. There exist diffeomorphisms $\psi_n: V_1 \to V_1$ such that $\psi_n: V_{1,1} \to V^{(n)}
_{1,1}$ is identity, $\psi_n: V_{1,2} \to V^{(n)}_{1,2}$ is a diffeomorphism and $\psi_n: V_{1,3} \to V^{(n)}_{1,3}$ is biholomorphic the complex structures, induced from $C_{1,n}$.
Thus, we have constructed the covering $\{ V_{1,1}, V_{1,2}, V_{1,3} \}$ of $V_1$ and parameterizations $\sigma'_n \deff \sigma_{1,n} \scirc \psi_n: V_1
\to C_{1, n}$ such that the conditions of [*Theorem 5.3.2*]{} are satisfied. Moreover, $\area(u_n(\sigma'_n(V_{1,i}))) \le (N-1)\eps$ due to inequality (5.3.4). Consequently, we can apply the inductive assumptions for the sequence of curves $\sigma'_n(V_{1,i})$ and finish the proof by induction.
[*Case 2):*]{} $V_1$ is a cylinder, structures $\sigma_n^*j_n|_{V_1}$ vary with $n$, but conformal radii of $(V_1, \sigma_n^*j_n)$ are bounded uniformly in $n$. Applying [*Lemma 5.3.3*]{}, we can assume that structures $\sigma_n^*
j_n$ converge to a structure of an annulus with finite conformal radius. The constructions of [*Case 1)*]{} are used here with the following minor modifications. First, we find the set of the bubble points $ y^*_i\in V_1$, using the same patching construction and the characterization (5.3.1). Then we find diffeomorphisms $\phi_n : V_1 \to V_1$ such that [*a)*]{} $\phi_n$ converge to the identity map $\id: V_1 \to V_1$; [*b)*]{} $\phi_n$ are identical in fixed (independent of $n$) annuli adjacent to the boundary circles of $V_1$; [*ċ)*]{} $\phi_n$ preserve every bubble point, $\phi_n(
y^*_i) = y^*_i$; and finally [*d)*]{} for the “corrected” parameterizations $\ti\sigma_n \deff \sigma_n \scirc \phi_n$ the structures $\ti\sigma_n^*j_n|
_{V_1}$ are constant in a neighborhood of every bubble point $y^*_i$. Then we repeat the rest of the constructions of [*Case 1)*]{} using the new parameterizations $\ti\sigma_n$.
[*Case 3): Every $C_{1,n} = \sigma_n(V_1)$ is isomorphic to the standard node $\cala_0$*]{}. Fix identifications $C_{1,n} \cong \cala_0$ such that the induced parameterization maps $\sigma_{1, n} : V_1 \to \cala_0$ are the same for all $n$. Represent $\cala_0$, and hence every $C_{1,n}$, as the union of two discs $\Delta'$ and $\Delta''$ with identification of the centers $0\in
\Delta'$ and $0\in \Delta''$ into the nodal point of $\cala_0$, still denoted by $0$. Let $u'_n :\Delta' \to X$ and $u''_n :\Delta'' \to X$ be the corresponding “components” of the maps $u_{1,n} : C_{1,n} \to X$. Find the common collection of bubbling points $y^*_i$ for both maps $u'_n :\Delta' \to X$ and $u''_n :\Delta'' \to X$. If there are no bubble points, then we obtain the convergence type [*B)*]{} and the proof can be finished by induction. Otherwise consider the first such point $y^*_1$ which lies, say, on $\Delta'$. If $y^*_1$ is distinct from the nodal point $0 \in
\Delta'$, then we simply repeat all the constructions from [*Case 1)*]{}.
It remains to consider the case $y^*_1=0 \in \Delta'$. Now one should modify the argumentations of [*Case 1*]{} in the following way. Repeat the construction of the radii $r_n\lrar 0$, the points $x_n \lrar y^*_1=0$, and the maps $v_n: \Delta(0, {\varrho \over 2r_n}) \to X$, $v_\infty: S^2 \to
X$ from [*Case 1*]{}. Set $R_n \deff |x_n|$, so that $R_n$ is the distance from $x_n$ to point $0=y^*_1 \in \Delta'$. After rescaling $u_n$ to the maps $v_n$, the point $0\in \Delta'$ will correspond to the point $z^*_n \deff
-{x_n \over r_n}$ in the definition domain $\Delta(0, {\varrho \over 2r_n})$ of the map $v_n$. We will now consider two subcases.
[*Subcase 3$'$): The sequence ${R_n\over r_n}$ is bounded*]{}. This is equivalent to boundedness of the sequence $ z^*_n$. Going to a subsequence we may assume that the sequence $z^*_n$ converges to a point $z^*\in \cc$. This point will be a nodal one for $(S^2, v_\infty)$. As above, $v_\infty$ is nonconstant and $\norm{ dv _\infty} ^2 _{L^2( S^2)}=
\area( v_\infty( S^2) )\ge 3\eps$. Choose $b>0$ in such a way that $$\norm{ dv_\infty}_{L^2(\Delta (0,b))}^2 \ge 2\eps
\eqno(5.3.5)$$ and $b \ge 2|z^*|+2$. Due to [*Corollary 2.5.1*]{} for $n>\!>1$ we obtain the estimate $$\norm{du'_n}_{L^2(\Delta' (x_n, br_n))}^2 =
\norm{ dv_n}_{L^2(\Delta (0,b))}^2 \ge \eps.
\eqno(5.3.6)$$ Here $\Delta' (x, r)$ denotes the subdisc of $\Delta'$ with the center $x$ and the radius $r$. Furthemore, for $n>\!>1$ we have the relation $z^*_n \in \Delta( 0, b-1)$, or equivalently, $0 \in \Delta' (x_n, (b-1)
r_n))$.
Define the coverings of $\cala_0$ by four sets $$\mathsurround=0pt
\matrix\format\l\ \ &\l\\
W^{(n)}_1 \deff \Delta'
\bs \barr\Delta' (0, {\textstyle{\varrho\over2}}),
&
W^{(n)}_2 \deff \Delta' (0, \varrho)
\bs \barr \Delta' (x_n, br_n),
\cr
\noalign{\vskip5pt}
W^{(n)}_3 \deff \Delta' (x_n, 2br_n)
\bs \barr\Delta'(0,{\textstyle {r_n \over 2}}),
&
W^{(n)}_4 \deff \Delta' (0, r_n) \cup
\Delta'',
\endmatrix$$ and lift them to $V_1$ by putting $V^{(n)}_{1,i} \deff \sigma\inv_{1,n}(
W^{(n)}_i)$. Choose $n_0 >\!> 0$ such that $z^*_{n_0} \in \Delta( 0, b-1)$ and the relation (5.3.6) holds. Set $V_{1,i} \deff V^{(n_0)} _{1, i}$. Choose diffeomorphisms $\psi_n: V_1 \to V_1$ such that $\psi_n: V_{1,1} \to V^{(n)}
_{1,1}$ is the identity map, $\psi_n: V_{1,2} \to V^{(n)}_{1,2}$ and $\psi_n:
V_{1,3} \to V^{(n)}_{1,3}$ are diffeomorphisms, and $\psi_n: V_{1,4} \to
V^{(n)}_{1,4}$ corresponds to isomorphisms of nodes $W^{(n)}_4 \cong
\cala_0$. Set $\sigma'_n \deff \sigma_n \scirc \psi_n$. The choice above can be done in such a way that the refined covering $\{ V_{1,i} \}$ of $V_1$ and parameterization maps $\sigma'_n: V_1 \to C_{1,n}$ have the properties of [*Theorem 5.3.2*]{}. Moreover, relations (5.3.2) and (5.3.6) imply the estimate $\area(
u_n (\sigma'_n(V_{1,i})) \le (N-1)\,\eps$. This provides the inductive conclusion for [*Subcase 3$'$)*]{}.
[*Subcase 3$''$): The sequence $R_n \over r_n$ increases infinitely*]{}. This means that the sequence $ z^*_n$ is not bounded. Nevertheless $R_n \lrar
0$ since $x_n \lrar 0$. We proceed as follows. Repeat the construction of the raduis $b$ from [*Case 1)*]{}. For $n>\!>0$ define the coverings of $\cala_0$ by six sets $$\mathsurround=0pt
\matrix\format\l\ \ &\l\\
W^{(n)}_1 \deff \Delta'
\bs \barr\Delta' (0, {\varrho\over2}),
&
W^{(n)}_2 \deff \Delta' (0, \varrho)
\bs \barr \Delta' (x_n, 2R_n),
\cr\noalign{\vskip5pt}
W^{(n)}_3 \deff \Delta' (x_n, 4R_n)
\bs \bigl(\barr \Delta' (x_n, {R_n \over 6})
\cup \barr \Delta' (0,{R_n\over 6}) \bigr)
&
W^{(n)}_4 \deff \Delta' (0, {R_n\over 3}) \cup \Delta'',
\cr\noalign{\vskip5pt}
W^{(n)}_5 \deff \Delta' (x_n, {R_n \over 3})
\bs \barr\Delta'(x_n, br_n),
&
W^{(n)}_6 \deff \Delta' (0, 2br_n ),
\endmatrix$$ and lift them to $V_1$ by putting $V^{(n)}_{1,i} \deff \sigma\inv_{1,n}(
W^{(n)}_i)$. Choose $n_0 >\!> 0$ such that $R_{n_0} >\!> br_{n_0}$ , and set $V_{1,i} \deff V^{(n_0)} _{1, i}$. Choose diffeomorphisms $\psi_n: V_1 \to
V_1$ such that $\psi_n: V_{1,1} \to V^{(n)} _{1,1}$ is the identity map, $\psi_n: V_{1,2} \to V^{(n)}_{1,2}$, $\psi_n: V_{1,4} \to V^{(n)}_{1,4}$ and $\psi_n: V_{1,5} \to V^{(n)}_{1,5}$ are diffeomorphisms, and finally, $\psi_n: V_{1,6} \to V^{(n)}_{1,6}$ corresponds to isomorphisms of nodes $W^{(n)}_6 \cong \cala_0$. Set $\sigma'_n \deff \sigma_n \scirc \psi_n$. Again, this choice can be done in such a way that $\{ V_{1,i} \}$ and parameterization maps $\sigma'_n: V_1 \to C_{1,n}$ have the properties of [*Theorem 5.3.2*]{}. As above, we get the estimate $\area(u_n (\sigma'_n(
V_{1,i} )) \le (N-1)\,\eps$ due to (5.3.2). Thus we get the inductive conclusion for [*Subcase 3$''$)*]{} and can proceed further.
[*Case 4):*]{} $V_1$ is a cylinder, structures $\sigma_n^*j_n|_{V_1}$ vary with $n$, and conformal radii of $(V_1, \sigma_n^*j_n)$ converge to $+\infty$. Using [*Lemma 5.3.3*]{}, we can assume that structures $\sigma_n^*
j_n$ satisfy property of this lemma.
Fix biholomorphisms $\sigma_n(V_1) \cong Z(0, l_n)$. If $\area(u_n(Z(a-1,
a))) \le \eps$ for any $n$ and any $a \in [1, l_n]$, then [*Lemma 5.2.2*]{} shows that $u_n: \sigma_n(V_1) \to X$ converge to a $J_\infty$-holomorphic map from a node.
If not, then, after passing to a subsequence, we can find a sequence $\{a_n
\}$ with $a_n \in [1, l_n]$ such that $\area(u_n(Z(a_n \allowbreak -1, a_n)))
\ge \eps$. If $a_n$ is bounded, say $a_n \le a^+$, then we cover $Z(0,
l_n)$ by the sets $V_{1,1} \deff Z(0, a^+ +2)$ and $V_{1,2} \deff Z(a^+ +1,
l_n)$. If $l_n - a_n$ is bounded, say $l_n - a_n \le a^+$, then we cover $Z(0,
l_n)$ by the sets $V_{1,1} \deff Z(0, l_n - a^+ +2)$ and $V_{1,2} \deff
Z(l_n -a^+ +1, l_n)$. In the remaning case, when both $a_n$ and $l_n - a_n$ increase infinitely, we cover $Z(0, l_n)$ by 3 sets $V_{1,1}^{(n)} \deff
Z(0,a_n-1)$, $V_{1,2}^{(n)} \deff Z(a_n-2, a_n+1)$, and $V_{1,3}^{(n)} \deff
Z(a_n, l_n)$. Cover $V_1$ by 2 or, respectively, 3 successive cylinders $V_{1,
i}$ in an obvious way. Find diffeomorphisms $\psi_n : V_1 \to V_1$ identical in the neighborhood of the boundary of $V_1$ and such that $\psi_n(V_{1,i}) =
\sigma_n ^{-1} V^{(n)} _{1,i}$. Define the new parameterizations $\sigma'_n
\deff \sigma_n \scirc \psi_n$. Note that we may additionally assume that if the conformal radius of $\sigma'_n(V_{1,i})$ is independent of $n$ then the structure $\sigma'_n{}^*j_n \ogran_{V_{1,i}}$ is also independent of $n$.
By this construction we obtain the following property of the covering $\{ V_{1,i}\}$ and new parameterizations $\sigma'_n$. For any $V_i$ we have either the estimate $$\area(u_n(\sigma'_n(V_{1,i}))) \le (N-1)\,\eps.$$ or the structures $\sigma'_n{}^*j_n|_{V_{1,i}}$ do not depend on $n$. Thus we reduce our case to the situation which is covered either by the inductive assumption or by [*Case 1)*]{}.
The proof of the theorem can be finished by induction. The fact that the limit curve $(C_\infty, u_\infty)$ remains stable over $X$ is proved in [*Lemma 5.3.5*]{} below.
Remark. Here we explain the meaning of the constructions used in the proof. We start with [*Case 1)*]{}, where $J_n$-holomorphic maps from a fixed domain $D\subset \cc$ are treated. Bubbling points $y^*_i$ appear in this case as those where the strong convergence of maps $u_n: D \to (X, J_n)$ fails. The patching construction of Sacks and Uhlenbeck insures that the “convergence failure” set is finite and insures an effective estimate on the number of bubbling points by the upper bound of the area, $l \le 3N$ in our situation. The characterization property (5.3.2) of bubbling points is essentially due to Sacks and Uhlenbeck; the only difference is that we use the area of the map $u$ (which is equivalent to the energy of $u$, i.e., $L^2$-norm of $du$) , whereas in \[S-U\] the $\norm{du}_{L^\infty}$ is used. The next step, the construction of maps $v_n$ as the rescaling of the $u_n$ and the existence of the limit $v_\infty$ is also due to Sacks and Uhlenbeck.
Due to the explicit construction of the map $v_\infty$, it is useful to imagine the curve $(S^2, v_\infty)$ as a “bubbled sphere” and $y^*_1$ as the point where the “bubbling” occurs. Moreover, one obtains natural partitions (one for each $n>\!>0$) of $D$ into three pieces: $D$ minus a fixed small neighborhood of $y^*_1$; disks $(\Delta(x_n, br_n), u_n)$ representing pieces $(\Delta(0, b), v_\infty)$ and approximating a sufficiently big part $(\Delta(0, b), v_\infty)$ of the bubbled sphere and the “part between”.
These latter “parts between” appear to be the annuli of infinitely growing conformal radii, the situation considered in [*Case 4)*]{}. Since neither outer nor inner boundary circle should be preferred in some way, we consider them as long cylinders $C_n = Z(0,l_n)$ with $l_n \lrar \infty$, in the spirit of [*Definition 3.2*]{}. [*Lemma 5.2.2*]{} provides a “good” convergence model for long cylinders, stated above as convergence type [*A)*]{}. If such a convergence for a sequence $(C_n, u_n)$ fails, then there must exist subannuli $A_n \subset C_n$ of a constant conformal radius for which $\area(u_n(A_n)) \ge \eps$.
In both cases — a constant domain $D$ or long cylinders — we proceed by cutting the curves into smaller pieces. We arrive at a situation that is simpler in the following way. The obtained curves either converge or have the upper bound for the area smaller in the fixed constant $\eps$. Thus, the use of induction leads finally to a decomposition of the curves into pieces for which one of the convergence types [*A)–C)*]{} holds. The possibility of gluing these final pieces together is insured by the fact that the partitions above are represented by appropriate coverings satisfying the conditions of [*Theorem 5.3.2*]{}.
Considering curves with nodes, additional attention should be paid to the case where bubbling appears at a nodal point. This situation is considered in [*Case 3)*]{}. The constructed points $x_n$ and radii $r_n$ describe the “center” and the “size” of energy localization of the bubbling, represented by the sequence of the maps $v_n \lrar v_\infty$. So the convergence picture depends on whether the energy localization occurs near the nodal point ([*Subcase 3$'$)*]{}) or away from it ([*Subcase 3$''$)*]{}). As a result, the nodal point can either remain on the “bubbled” sphere $(S^2, v_\infty)$ or move into the “part between”, which is represented by long cylinders.
In [*Subcase 3$'$)*]{} we remove neighborhoods of the nodal point from the disks $(\Delta(0, b), v_n)$ and thus get four pieces of converging instead of three as in [*Case 1)*]{}. In [*Subcase 3$''$)*]{} the situation is more complicated, because we must take into account the position of the nodal point in the long cylinders — the “parts between”. Thus, we must consider now the sequence of cylinders with one marked point, i.e., the sequence of pants. The vanishing $R_n\to 0$ and $r_n / R_n \to$ mean that conformal structure of those pants is not constant and converges to one of the spheres with three punctures.
In order to have the covering pieces with the convergence types [*A)–C)*]{}, we choose an appropriate refinement of the covering. After that, we obtain two sequences of long cylinders, describing the appearance of two new nodal points. The first one corresponds to the part between the “original” nodal point and bubbled sphere and is represented by $V_5$, whereas the other one, represented by $V_2$, lies on the other side of the “original” nodal point. In addition, we fix a neighborhood of the “original” nodal point which has a constant complex structure and is topologically an annulus with the disc $\Delta''$ attached to the nodal point. To satisfy the requirements of [*Theorem 5.3.2*]{}, we cover the neighborhood by two pieces, the pants $V_3$ and the piece $V_4$ parameterizing the nodes $W^{(n)}_4$. This explains the appearance of six pieces of covering in [*Subcase 3$''$)*]{}.
Lemma 5.3.5. *The limit curve $(C_\infty, u_\infty)$ constructed in the proof of [*Theorem 1.1*]{} is stable over $X$.*
Proof. If $(C_\infty, u_\infty)$ is unstable over $X$, then either $C_\infty$ is a torus with $u_\infty$ constant, or $C_\infty$ should have a component $C' \subset C_\infty$ such that $\ti C'$ is a sphere with at most two marked points, and $u_\infty(C')$ is a point.
The case of a constant map from a torus is easy to exclude. In fact, in this case all $C_n$ must also be tori with $\area(u_n(C_n))$ sufficiently small for $n>\!>1$. Cover every $C_n$ by an infinite cylinder $Z(-\infty, +\infty)$ and consider compositions $\ti u_n: Z(-\infty, +\infty) \to X$ of $u_n$ with the covering maps. Since $\area(u_n(C_n))\approx 0$, [*Corollary 5.2.1*]{} can be applied to show that every $\ti u_n$ extends to a $J_n$-holomorphic map from $S^2$ to $X$. Consequently, $\area(\ti u_n(Z(-\infty, +\infty)))$ must be finite. On the other hand, $\area(u_n(C_n))>0$ due to the stability condition; hence $\area(\ti u_n(Z(-\infty, +\infty)))$ must be infinite. This contradiction excludes the case of a torus.
The same argumentations are valid in the case, where $C_\infty$ is the sphere with no marked points. Then the curves $C_n$ are also parameterized by the sphere $S^2$. The condition of instability means that $\area(u_\infty(C_\infty))=0$. Due to [*Corollary 2.5.1*]{}, $\area(u_n(C_n))$ must be sufficiently small for $n>\!>1$. Now [*Lemma 5.2.3*]{} and the stability of $(C_n,u_n)$ exclude this possibility.
Now consider the cases where the limit curve $C_\infty$ has a “bubbled” component $C'$, which is the sphere with one or two marked points. If $C'$ has one marked point, then $C'$ must appear as a “bubbled” sphere $(S^2, v_\infty)$ in the constructions of [*Cases 1)–3)*]{} in the proof of [*Theorem 1.1*]{}. But these constructions yield only non-trivial “bubbled” spheres, for which $v_\infty \not= const$. Thus, such a component $C'$ must be stable.
In the remaining case, a component $C'$ with two marked points, we consider a domain $U\subset C_\infty$, which is the union of the component $C'$ and neighborhoods of the marked point on $C'$. If $C'$ is an unstable component, then $\area(u_\infty(C'))=0$, and we can achieve the estimate $\area( u_\infty(
U))< \eps$ taking $U$ sufficiently small. Set $\Omega \deff \sigma_\infty
\inv(U)$, where $\sigma_\infty: \Sigma \to C_\infty$ is the parameterization of $C_\infty$. Let $\gamma_1$ and $\gamma_2$ be the pre-images of marked points on $C'$. Then $\Omega$ must be a topological annulus, and $\gamma_i$, $i=1,2$, disjoint circles generating the group $\pi_1 (\Omega) = \zz$. Further, $C'$ must be a “bubbling” component of $C_\infty$, i.e., at least for one of the circles $\gamma_i$, $i=1,2$, the images $\sigma_n(\gamma_i)$ are not nodal points of $C_n$ but smooth circles.
If these are both circles $\gamma_1$ and $\gamma_2$, then $U_n \deff
\sigma_n (\Omega)$ would satisfy the conditions of [*Lemma 5.2.2*]{}. In this case we should have the convergence type [*A)*]{}, and hence the limit piece $\sigma_\infty(\Omega)$ should be isomorphic to the node $\cala_0$.
In the case where only one circle, say $\gamma_1$, corresponds to nodal points on $C_n$, and for the other one the images $\sigma_n(\gamma_2)$ are smooth circles, then the domains $\sigma_n(\Omega)$ must be isomorphic to the node $\cala_0$. Furthermore, due to the condition $\area(u_\infty(\sigma_\infty(
\Omega))) < \eps$, we would have $\area(u_n(\sigma_n (\Omega))) < \eps$. Consequently, we would have the convergence type [*B)*]{}, and the unstable component $C'$ could not appear.
[Appendix 3]{}
[Compactness with Totally Real Boundary Conditions.]{}
[A3.1. Curves with Boundary on Totally Real Submanifolds.]{}
In this section we consider the behavior of complex curves over an almost complex manifold $(X,J)$ with a boundary on totally real submanifold(s). As in the “interior" case, we need to allow some type of boundary singularity.
Definition A3.1.1.
*The set $\cala^+ \deff \{ (z_1,z_2) \in
\Delta^2\;:\; z_1 \cdot z_2=0, \im z_1\ge 0, \im z_2\ge 0\}$ is called the [*standard boundary node*]{}. A curve $\barr C$ with boundary $\d C$ is called a [*nodal curve with boundary*]{} if*
$C$ is a nodal curve, possibly not connected;
$\barr C=C\cup \d C$ is connected and compact;
every boundary point $a\in \d C$ has a neighborhood homeomorphic either to the half-disk $\Delta^+ \deff \{ z \in \Delta
\;:\; \im z\ge0 \}$, or to the standard boundary node $\cala^+$.
In the last case $a\in \d C$ is called a [*boundary nodal point*]{}, whereas nodal points of $C$ are called [*interior nodal points*]{}. Both boundary and interior nodal points are simply called nodal points.
Definition A3.1.2. [*Let $(X,J)$ be an almost complex manifold. A pair $(\barr C, u)$ is called a [*curve with boundary over $(X,J)$*]{} if $\barr
C=C\cup \d C$ is a nodal curve with boundary, and $u: \barr C \to (X,J)$ is a continuous $L^{1,2}$-smooth map, which is holomorphic on $C$.* ]{}
A curve $(C,u)$ with boundary is stable if the same condition as in [*Definition 1.5*]{} on the automorphism groups of compact irreducible components is satisfied.
Remark. One can see that $\barr C$ has a uniquely defined real analytic structure such that the normalization $\barr C{}^{\sf nr}$ is a real analytic manifold with a boundary. More precisely, the pre-image of every boundary nodal point $a_i$ consists of two points $a'_i$ and $a''_i$. The normalization map $s: \barr C{}^{\sf nr} \to \barr C$ glues pairs $(a'_i, a''_i)$ of points on $\barr C{}^{\sf nr}$ into nodal points $a_i= s(a'_i)= s(a''_i)$ on $\barr
C$. This implies that the notion of a $L^{1,p}$-smooth map, $p>2$, as well as a continuous $L^{1,p}$-smooth map $u: \barr C \to X$ is well- defined.
Definition A3.1.3.
*We say that a real oriented surface with boundary $(\Sigma, \d\Sigma)$ [*parameterizes*]{} a nodal curve with boundary $C$ if there is a continuous map $\sigma :\barr\Sigma \to \barr C$ such that*
if $a\in C$ is an interior nodal point, then $\gamma_a \deff \sigma\inv
(a)$ is a smooth imbedded circle in $\Sigma$;
if $a\in\d C$ is a boundary nodal point, then $\gamma_a\deff \sigma
\inv (a)$ is a smooth imbedded arc in $\Sigma$ with end points on $\d\Sigma$, transversal to $\d\Sigma$ at this point;
if $a,b \in \barr C$ are distinct (interior or boundary) nodal points, then $\gamma_a\cap \gamma_b= \emptyset$;
$\sigma :\barr\Sigma \bs \bigcup_{i=1}^N\gamma_{a_i}\to \barr C \bs \{
a_1,\ldots ,a_N\} $ is a diffeomorphism, where $a_1,\ldots ,a_N$ are all (interior and boundary) nodal points of $\barr C$.
Recall that a real subspace $W$ of a complex vector space is called [*totally real*]{} if $W \cap \isl W =0$. Similarly, a $C^1$-immersion $f:W \to X$ is called [*totally real*]{} if for any $w\in W$ the image $d f(T_wW)$ is a totally real subspace of $T_{f(w)}X$.
Let $(\barr C, u)$ be a stable curve with boundary over an almost complex manifold $(X,J)$ of a complex dimension $n$.
Definition A3.1.4.
*We say that $(\barr C, u)$ satisfies the [*totally real boundary condition $\mib W$ of type $\bfbeta$*]{} if*
$\bfbeta= \{\beta_i\}$ is a collection of arcs with disjoint interiors, which defines a decomposition of the boundary $\d C= \cup_i \beta_i$; moreover, we assume that every boundary nodal point is an endpoint for four arcs $\beta_i$;
${\mib W} = \{(W_i, f_i)\}$ is a collection of totally real immersions $f_i: W_i \to X$, one for every $\beta_i$;
there are given continuous maps $u^{(b)}_i :\beta_i \to W_i$ [*realizing conditions $\mib W$*]{}, i.e., $f_i \scirc u^{(b)}_i = u|_{\beta_i}$.
Remarks. 1. We shall consider (immersed) totally real submanifolds only of [*maximal real dimension*]{} $n= \dimc X$.
2\. If $\bfbeta$ is a collection of arcs as above, a parameterization $\sigma: \barr \Sigma \to \barr C$ induces a collection of arcs $\sigma\inv(
\bfbeta) \deff \{ \sigma\inv(\beta_i)\;:\; \beta_i \in \bfbeta \} $ with the properties similar to of [*Definition A3.1.4*]{}. Thus, $\sigma\inv(
\beta_i)$ have disjoint interiors, $\cup_i\sigma\inv(\beta_i) = \d\Sigma$, and for any boundary node $a\in \barr C$ every endpoint of the arc $\beta_a=
\sigma\inv(a)$ is an endpoint of two arcs $\sigma\inv(\beta_i)$. Since $\bfbeta$ is completely determined by $\sigma\inv(\bfbeta)$, we shall denote both collections simply by $\bfbeta$ and shall not distinguish them when considering boundary conditions.
[A3.2. A priori Estimates near a Totally Real Boundary.]{}
A totally real boundary condition is a suitable elliptic boundary condition for an elliptic differential operator $\dbar$ of the Cauchy-Riemann type. In particular, all statements about “inner" regularity and convergence for complex curves remain valid near “totally real" boundary. As in the “inner” case, to get some “uniform” estimate at boundary one needs $W$ to be “uniformly totally real”.
Definition A3.2.1.
*Let $X$ be a manifold with a Riemannian metric $h$, $J$ a continuous almost complex structure, $W$ a manifold, and $A_W
\subset W$ a subset. We say that an immersion $f: W \to X$ is [*$h$-uniformly totally real along $A_W$ with a lower angle $\alpha = \alpha(W,
A_W, f) > 0$*]{},*
$df: TW \to TX$ is $h$-uniformly continuous along $A_W$;
for any $w \in A_W$ and any $\xi \not =0 \in T_wW$ the angle $\angle_h\bigl(Jdf(\xi), df(T_wW) \bigr) \ge \alpha$.
We start with an analog for the First A priori Estimate. Define the half-disks $\Delta^+(r) \deff \{ z\in \Delta(r) \;:\; \im z \ge 0 \}$ with $\Delta^+ = \Delta^+(1)$ and $\Delta^- \deff \{ z \in \Delta \; :\; \im z \le
0 \}$. Set $\beta_0 \deff (-1, 1) \subset \d\Delta^+$. Let $X$ be a manifold with a Riemannian metric $h$, $A\subset X$ a subset, $J^*$ a continuous almost complex structure, $f: W \to X$ a totally real immersion and $A_W
\subset W$ a subset.
Lemma A3.2.1.
*Suppose that $J^*$ is $h$-uniformly continuous on $A$ with the uniform continuity modulus $\mu_{J^*}$ and that $f : W \to X$ is $h$-uniformly totally real along $A_W$ with a lower angle $\alpha_f>0$ and the uniform continuity modulus $\mu_f$. Then for every $2< p<\infty $ there exists an $\eps^b_1 =\eps^b_1(\mu_{J^*}, \alpha_f, \mu_f )$ (independent of $p$) and $C_p= C(p, \mu_{J^*}, \alpha_f, \mu_f )$ such that the following holds.*
If $J$ is a continuous almost complex structure on $X$ with $\norm{ J- J^*}
_{L^\infty(A)} <\eps^b_1$, and if $u\in C^0 \cap L^{1,2}(\barr\Delta^+ ,X)$ is a $J$-holomorphic map with $u(\Delta) \subset A$ and with the boundary condition $u|_{\beta_0} =f \scirc u^b$ for some continuous $u^b: \beta_0
\to A_W \subset W$, then the condition $\norm{du}_{ L^2( \Delta^+ )}
<\eps^b_1$ implies the estimate $$\norm{du}_{L^p(\Delta^+({1\over 2}) )}\le C_p\cdot
\norm{du}_{L^2(\Delta^+ )}.\eqno(A3.2.1)$$
Proof. Suppose additionally that $\diam(u(\Delta^+))$ is sufficiently small. Then we may assume that $u(\Delta^+)$ is contained in some chart $U
\subset \cc^n$ such that $\norm{ J- J\st} _{L^\infty(U)}$ is also small enough. Let $z=(z_1, \ldots, z_n)$ be $J\st$-holomorphic coordinates in $U$ such that $u(0)= \{z_i= 0\}$. Making an appropriate diffeomorphism of $U$, we may additionally assume that $W_0 \deff f(W) \cap U$ lies in $\rr^n$ and that $J = J\st$ along $W_0$.
Consider a trivial bundle $E \deff \Delta \times \cc^n$ over $\Delta$ with complex structures $J\st$ and $J_u \deff J \scirc u$. We can consider $u$ as a section of $E$ over $\Delta^+$ satisfying the equation $\dbar_{J_u} u\deff \d_x u+ J_u \d_y u=0$. Over $\beta_0$ we obtain a $J_u$-totally real subbundle $F \deff \beta_0 \times \rr^n$ such that $u(\beta_0) \subset F$. Let $\tau$ denote a complex conjugation in $\Delta$ as well as a complex conjugation in $E$ with respect to $J\st$. Extend $J_u$ on $E|_{\Delta^-}$ as the composition $- \tau\scirc J_u \scirc \tau$. This means that for $z\in Delta^-$ we obtain $$J_u(z): v \mapsto \tau v \in E_{\tau z}
\mapsto J_u(\tau z)(\tau v) \in E_{\tau z}
\mapsto -\tau J_u(\tau z)(\tau v) \in E_z .
\eqno(A3.2.2)$$
Since $J_u=J \scirc u$ coincides with $J\st$ along $\beta_0$, this extension is also continuous. Further, for $v\in L^1( \Delta^+, E)$ define the extension $\ex(v)$ by setting $v(z)\deff \tau v(\tau z)$. This gives a continuous linear operator $\ex: L^p( \Delta^+, E) \to L^p( \Delta^+, E)$ for any $1\le p\le \infty$. An important property of operator $\ex$ is that if $v \in L^{1,p} (\Delta^+, E)$ with $1\le p\le \infty$ (resp. $v\in C^0(
\barr \Delta^+ )$) satisfies the boundary condition $v|_{\beta_0} \subset
F$, then $\ex v \in L^{1,p}(\Delta, \cc^n)$ (resp. $\ex v \in C^0 (\barr
\Delta)$). Let us denote by $L^{1,p} (\Delta^+, E,F)$ (resp. by $C^0
(\Delta^+, E,F) $) the corresponding spaces of $v$ with the boundary condition $v|_{\beta_0} \subset F$.
Since for $v\in L^{1,1}(\Delta^+, E)$ holds $\d_x(\tau v) =\tau (\d_x v)$ and $\d_y(\tau v) = - \tau (\d_y v)$, we obtain $\dbar_{J_u} (\ex v) =\ex (\dbar_{J_u} v)$ for any $v \in L^{1,p} (\Delta^+, E,F)$. In particular, for $\ti u \deff \ex u \in C^0 \cap L^{1,2} (\Delta, E)$ we have $\dbar_{J_u} \ti u=0$.
From this point we can repeat the steps of the proof of [*Lemma 2.4.1*]{}.
Remark. We shall refer to the construction of a complex structure $J_u$ in $E$ over $\Delta^-$ and (resp. of a section $\ti u$ of $E$ over $\Delta^-$) as an [*extension of $J \scirc u$ (resp. of $u$) by the reflection principle*]{}.
Let $X$ be a manifold with a Riemannian metric $h$, $J^*$ a continuous almost complex structure on $X$, $A\subset X$ a closed $h$-complete subset such that $J^*$ is $h$-uniformly continuous on $A$, and $f_0 :W \to X$ an immersion, which is $h$-uniformly totally real on some closed $f^*h$-complete subset $A_W\subset W$.
Corollary A3.2.2.
*Let $\{ J_n \}$ be a sequence of continuous almost complex structures on $X$ such that $J_n$ converge $h$-uniformly on $A$ to $J$, $f_n: W \to X$ a sequence of totally real immersion such that $df_n$ converge $h$-uniformly on $A_W$ to $df_0$, and $u_n\in C^0\cap L^{1,2}(
\Delta^+ ,X)$ a sequence of $J_n$-holomorphic maps such that $u_n(\Delta^+)
\subset A$, $\norm{du_n}_{L^2 (\Delta^+ )} \le \eps^b_1$, $u_n(0)$ is bounded in $X$, and $u_n| _{\beta_0} = f_n \scirc u^b_n$ for some continuous $u^b_n
:\beta_0 \to A_W$.*
Then there exists a subsequence $u_{n_k}$ where $L^{1,p}_\loc (\Delta^+
)$-converge to a $J$-holomorphic map $u_\infty$ for all $p < \infty$.
Here $\beta_0 = (-1,1) \subset \d \Delta^+$ and $L^{1,p}_\loc (\Delta^+
)$-convergence means $L^{1,p} (\Delta^+(r) )$-convergence for all $r<1$, i.e., convergence up to boundary component $\beta_0$.
Proof. As a statement itself, the proof of [*Corollary A3.2.2*]{} copies the one of [*Corollary 2.5.1*]{} with appropriate modifications and using the reflection principle.
[A3.3. Long Strips and the Second A priori Estimate.]{}
Consider now an analog of the Second A priori Estimate. An analog of “long cylinders” is now “long strips" satisfying appropriate boundary conditions.
Definition A3.3.1. [*Define a [*strip $\Theta(a,b) \deff (a,b)
\times [0,1]$*]{} with the complex coordinate $\zeta \deff t - i\theta$, $t \in (a,b)$, $ \theta \in [0,1]$. Define also $\Theta_n \deff \Theta(n-1,n)$, $\d_0 \Theta(a,b) \deff (a,b) \times \{0\}$, and $\d_1 \Theta(a,b) \deff (a,b)
\times \{1\}$.* ]{}
We are interested in maps $u: \Theta(a,b) \to X$, which are holomorphic with respect to the complex coordinate $\zeta$ on $\Theta(a,b)$ and a continuous almost complex structure $J$ on $X$, which satisfy the boundary conditions $$u\ogran_{\d_0\Theta(a,b)} = f_0 \scirc u^b_0,
\hskip6em
u\ogran_{\d_1\Theta(a,b)} = f_1 \scirc u^b_1,$$ with some $J$-totally real immersions $f_{0,1} :W_{0,1} \to X$ and continuous maps $u_{0,1} :\d_{0,1}\Theta(a,b) \to W_{0,1}$. First we consider the special linear case.
Lemma A3.3.1. [*Let $W_0$ and $W_1$ be $n$-dimensional totally real subspaces in $\cc^n=(\rr^n,J\st )$. Then there exist a constant $\gamma_W=
\gamma(n, W_0, W_1)$ with $0< \gamma_W <1$ such that for any holomorphic map $u: \Theta(0,3) \to \cc^n$ with the boundary conditions $$u(\d_0 \Theta(0,3)) \subset W_0 \qquad
u(\d_1 \Theta(0,3)) \subset W_1
\eqno(A3.3.1)$$ we have the following estimate: $$\int_{\Theta_2} |du|^2 dt\,d\theta \le \msmall{\gamma_W\over2}
\left( \int_{\Theta_1} |du|^2 dt\,d\theta +
\int_{\Theta_3} |du|^2 dt\,d\theta \right).
\eqno(A3.3.2)$$* ]{}
Proof. Let $L^{1,2}_W([0,1], \cc^n)$ be a Banach manifold of those $v(\theta) \in L^{1,2}([0,1], \cc^n)$, $v(0) \in W_0$ and $v(1) \in W_1$. Consider a nonnegative quadratic form $Q(v) \deff \int_0^1 |\d_\theta v(
\theta)|^2 d\theta$. Since $Q(v) + \norm{v} ^2_{L^2} =\norm{v}^2 _{L^{1,2}}$ and the imbedding $L^{1,2}_W([0,1], \cc^n) \hook L^2([0,1],\cc^n)$ is compact, we can decompose $L^{1,2}_W([0,1], \cc^n)$ into a direct Hilbert sum of eigenspaces $\ee_\lambda$ of $Q$ $\norm{v}^2_{L^2}$. This means that $v_\lambda$ belongs to $\ee_\lambda$ iff for any $w\in L^{1,2}_W([0,1],
\cc^n)$ $$\int_0^1 \<
\d_\theta v_\lambda(\theta), \d_\theta w(\theta) \> d\theta = \int_0^1
\lambda\< v_\lambda(\theta), w(\theta) \> d\theta ,\eqno(A3.3.3)$$ where $\<\cdot,\cdot \>$ denotes a standard [*$\rr$-valued*]{} scalar product in $\cc^n$. Integrating by parts yields $$\int_0^1 \< \d^2_{\theta\theta} v_\lambda(\theta)
+ \lambda v_\lambda(\theta), w(\theta) \> d\theta + \<\d_\theta v_\lambda(
\theta), w(\theta) \>|_{\theta=1} - \<\d_\theta v_\lambda(
\theta), w(\theta) \>|_{\theta=0} =0.$$ This implies that $v_\lambda$ belongs to $\ee_\lambda$ $\d^2_{\theta\theta} v_\lambda(\theta) + \lambda v_\lambda( \theta)=0$, $\d_\theta v_\lambda(1) \perp W_1$, and $\d_\theta v_\lambda(0) \perp W_0$. Since $J\st$ is $\<\cdot, \cdot\>$-orthogonal, we can conclude that $J\d_\theta v_\lambda(\theta) \in \ee_\lambda$.
Positivity and compactness of $Q$ $\norm\cdot_{L^2}$ imply that all $\ee_\lambda$ are finite dimensional and empty for $\lambda<0$. Further, since $\d_\theta v=0$ for any $v\in \ee_0$, $\ee_0$ consists of constant functions with values in $W_0 \cap W_1$.
Now let $u: \Theta(0,3) \to \cc^n$ be a holomorphic map with the boundary condition (5.2). We can represent $u$ in the form $u(t, \theta)= \sum_\lambda
u_\lambda(t,\theta)$ with $u_\lambda(t, \cdot) \deff \pr_\lambda (u(t,\cdot))
\in L^{1,2}([0,3],\ee_\lambda)$. Since $J\d_\theta$ is an endomorphism of every $\ee_\lambda$, every $u_\lambda(t,\theta)$ is also holomorphic. In particular, $u_0$ is also holomorphic and constant in $\theta$. Thus $u_0$ is constant.
Since $u$ is harmonic, i.e., $(\d^2_{tt} + \d^2_{\theta \theta})u =0$, it follows that $\d^2_{tt} u_\lambda(t,\theta)\allowbreak = \lambda u_\lambda(t,\theta)$. For $\lambda > 0$ this yields $u_\lambda(t,\theta) = e^{+\sqrt \lambda t}
v^+_\lambda(\theta) + e^{-\sqrt\lambda t} v^-_\lambda(\theta)$ with $v^\pm
_\lambda(\theta) \in \ee_\lambda$. Fixing an orthogonal $\rr$-basis of $\ee_\lambda $ $v^i_\lambda$, we write every $u_\lambda$ in the form $$u_\lambda(t,\theta)= \sum_i (a^i_\lambda e^{+\sqrt \lambda t} +
b^i_\lambda e^{-\sqrt \lambda t}) v^i_\lambda(\theta)$$ with real constants $a^i_\lambda$, $b^i_\lambda$. Since $u_0(t,\theta)$ is constant, $\norm{du}^2_{L^2(\Theta_k)} = 2\norm{\d_\theta u}^2_{L^2(\Theta_k} =
\sum_{\lambda,i} 2\lambda \int_{k-1}^k (a^i_\lambda e^{+\sqrt\lambda t}
+b^i_\lambda e^{-\sqrt\lambda t})^2 d\theta $. Here we used $(5.4)$ and $L^2$-orthonormality of $v^i_\lambda $. This leads us to the problem of determining the smallest possible constant $\gamma$ for the inequality $$\int_1^2 (ae^{\alpha t} + be^{-\alpha t})^2 dt
\le \msmall{\gamma\over2} \left(
\int_0^1 (ae^{\alpha t} + be^{-\alpha t})^2 dt +
\int_2^3 (ae^{\alpha t} + be^{-\alpha t})^2 dt \right)
\eqno(A3.3.4)$$ with $a,\,b\in\rr$ for given $\alpha>0$. Integration gives $$a^2 e^{3\alpha}\msmall{ e^\alpha - e^{-\alpha} \over 2\alpha}
+ b^2 e^{-3\alpha}\msmall{ e^\alpha - e^{-\alpha} \over 2\alpha}
+2ab \le$$ $$\le \msmall{\gamma\over2} \left(
a^2 e^{3\alpha}\msmall{ (e^\alpha - e^{-\alpha}) (e^{2\alpha} + e^{-2\alpha})
\over 2\alpha}+
b^2 e^{-3\alpha}\msmall{ (e^\alpha
- e^{-\alpha}) (e^{2\alpha} + e^{-2\alpha})
\over 2\alpha}
+ 4ab \right)$$ or equivalently $$a^2 e^{3\alpha}\msmall{ (e^\alpha - e^{-\alpha})
(e^{2\alpha} + e^{-2\alpha} -2/\gamma ) \over 2\alpha}
+
b^2 e^{-3\alpha}\msmall{ (e^\alpha - e^{-\alpha})
(e^{2\alpha} + e^{-2\alpha} -2/\gamma) \over 2\alpha}$$ $$+ 4ab(1-1/\gamma)
\ge 0$$ The determinant of the last quadratic form in $a,b$ is $$\left(\!\!\msmall{ (e^\alpha - e^{-\alpha})
(e^{2\alpha} + e^{-2\alpha} -2/\gamma )
\over 2\alpha}\!\!\right)^{\!\!2} - 4(1-1/\gamma)^2
=
4\left(\!\!\msmall{ \sh \alpha (\ch\,2\alpha -1/\gamma )
\over \alpha} \!\!\right)^{\!\!2} - 4(1-1/\gamma)^2$$ $$\ge 4(\ch\,2\alpha -1/\gamma )^2 -4(1-1/\gamma)^2 =
4(\ch\,2\alpha -1)(\ch\,2\alpha +1 -2/\gamma).$$ Thus, the inequality (5.5) holds for every $a,b \in\rr$ provided $\gamma \ge
{2 \over 1 + \ch\,2\alpha}<1$.
Note that there exists a minimal positive eigenvalue $\lambda_1>0$ of $Q$. Thus, the estimate (5.3) holds for $\gamma_W \deff {2 \over 1 + \ch(2
\sqrt{\lambda_1} )}<1$.
Remark. A dependence of $\gamma_W$ as a function of $\lambda_1=
\lambda_1(W_0, W_1)$ shows that $\gamma_W<1$ can be chosen the same for all pairs $(\wt W_0,\wt W_1)$ sufficiently close to $(W_0, W_1)$, provided $\dim
(\wt W_0 \cap \wt W_1) = \dim (W_0 \cap W_1)$. Vice versa, if we perform a sufficiently small deformation of $(W_0, W_1)$ into $(\wt W_0, \wt W_1)$ with $\dim (\wt W_0 \cap \wt W_1) < \dim (W_0 \cap W_1)$, then some $v \in
\ee_0(W_0, W_1)$ will wander into an eigenvector $\ti v \not\in \ee_0 (\wt
W_0, \wt W_1) = \wt W_0 \cap \wt W_1$, but with a sufficiently small eigenvalue $\lambda_1(\wt W_0,\wt W_1) >0$, so that the best possible $\gamma_{\wt W}$ will be arbitrarily close to $1$. Thus the uniform separation of $\gamma_W$ from 1 under a small perturbation of $(W_0, W_1)$ is equivalent to a uniform separation of $\lambda_1(W_0, W_1)$ from 0, and is equivalent to stability of $\dim(W_0 \cap W_1)$.
Another phenomenon, also connected with spectral behavior, is that for [*harmonic*]{} $u(t, \theta)$ from $\ee_0= \{\,const\, \}$ does not follow $u_0 = const$, but merely $u_0(t, \theta) = v_0 + v_1 t$ with $v_0,
v_1 \in \ee_0$, so that the inequality (5.3) is not true. This leads to a much more complicated bubbling phenomenon with energy loss for harmonic and harmonic-type maps, compare \[S-U\], \[P-W\], \[Pa\].
Note also that if $W_0$ and $W_1$ are [*affine*]{} totally real subspaces of $\cc^n$, then the inequality (5.3) for holomorphic $u: \Theta(0,3)\to \cc^n$ with boundary condition (5.2) is in general not true. An easy example is a natural imbedding $u:\Theta(0,3) \hook \cc$, with the nonconstant component $u_0 \equiv
u$ and with $\int_{\Theta(0,1)} |du|^2 = \int_{\Theta (1,2)} |du|^2 = \int_{
\Theta(2,3)} |du|^2 \not=0$. In general, those are $n$-dimensional totally real affine planes $W_0$ and $W_1$ in $\cc^n$ with an empty intersection. This can happen if $W_0$ and $W_1$ are parallel or [*skew*]{}. The latter means that the corresponding vector spaces $V_0$ and $V_1$ ($W_i = V_i + w_i$ for some $w_i \in \cc^n$) are different. In both cases the intersection $V_0 \cap
V_1$ is not zero, because otherwise $W_0 \cap W_1 \not= \emptyset$ by dimensional argumentation.
The considerations above show which properties should be controlled in order to obtain a reasonable statement in the nonlinear case.
Definition A3.3.2.
*Let $X$ be a manifold with a Riemannian metric $h$, $f_0 : W_0 \to X$ and $f_1 : W_1 \to X$ immersions and $A_0 \subset
W_0$, $A_1 \subset W_1$ subsets. We say that [*$f_0 : W_0 \to X$ and $f_1
:W_1 \to X$ are $h$-uniformly transversal along $A_0$ and $A_1$ with parameters $\delta >0$ and $M$*]{} if for any $x_0 \in A_0$ and $x_1 \in A_1$ the following holds:*
either $\dist_h(f_0(x_0), f_1(x_1)) > \delta$;
or there exists $x'_i \in A_i$ with $f_0( x'_0) =f_1(x'_1)$ such that $$\dist_h(x_0,x'_0) + \dist_h(x_1,x'_1) \le M\,\dist_h(f_0(x_0), f_1(x_1)).$$
Remark. Roughly speaking, the condition excludes the appearance of points, where $W_0$ and $W_1$ are “asymptotically parallel or skew" and ensures a uniform lower bound for the angle between $W_0$ and $W_1$.
Now let $X$ be a manifold with a Riemannian metric $h$, $J^*$ a continuous almost complex structure on $X$, $f_0 : W_0 \to X$ and $f_1 : W_1 \to X$ immersions, and $A\subset X$, $A_0 \subset W_0$, $A_1 \subset W_1$ subsets. Suppose that $J^*$ is $h$-uniformly continuous on $A$, $df_i: TW_i \to TX$ are $h$-uniformly continuous on $A_i$ and that $f_i$ are $h$-uniformly transversal along $A_i$ with parameters $\delta= \delta(f_0,f_1)>0$ and $M=
M(f_0,f_1)$.
Lemma A3.3.2.
*There exist constants $\eps^b_2 =\eps^b_2 (\mu_{J^*},
\, f_0,\, f_1,\, \delta,\, M) >0$ and $\gamma^b = \gamma^b(\mu_{J^*},\, f_0,\,
f_1,\, \delta,\, M) <1$ such that for any continuous almost complex $\ti J$ with $\norm{\ti J-J^*}_{L^\infty(A)}<\eps_2^b$, any immersions $\ti f_i :W_i
\to X$ with $\dist(\ti f_i, f_i)_{C^1(A_i)} <\eps_2^b$, and any $\ti J
$-holomorphic map $u \in C^0 \cap L^{1,2}( \Theta(0,5), X)$ with $u(
\Theta(0,5)) \subset A$, $u|_{\d_i\Theta(0,5)} = f_i \scirc u^b_i$ for some continuous $u^b_i : \d_i\Theta(0,5) \to A_i \subset W_i$ the conditions*
$\norm{ du}_{ L^2(\Theta_i)} <\eps_2^b$;
$\ti f_i: W_i \to X$ are $h$-uniformly transversal along $A_i$ with the same parameters $\delta$ and $M$
imply the estimate $$\norm{du}^2_{L^2(\Theta_3)}\le \msmall{\gamma^b \over 2} \cdot
\left( \norm{du}^2_{L^2(\Theta_2)} + \norm{du}^2_{L^2(\Theta_4)} \right)
\eqno(A3.3.5).$$
Proof. Suppose the statement of the lemma is false. Then there should exist a sequence of continuous almost complex structures $J_k$ with $\norm{
J_k- J^*} _{L^\infty (A)} \lrar 0$, a sequence of immersions $f_{k,i}: W_i
\to X$ with $f_{k,i} \lrar f_i$ in $C^1(A_i)$ such that $f_{k,i} :W_i \to X$ are $h$-uniformly transversal with the same parameters $\delta$ and $M$, and a sequence of $J_k$-holomorphic maps $u_k \in C^0\cap L^{1,2}(\Theta(0,5),
X)$ with $u_k(\Theta(0,5)) \subset A$ and $u_k |_{\d_i\Theta(0,5)} = f_{k,i}
\scirc u^b_{n,i}$ for some continuous $u^b_{n,i} : \d_i\Theta(0,5) \to A_i
\subset W_i$ such that $\norm{du_k}^2_{L^2(\Theta(0,5))} \lrar 0$ and $$\norm{du_k}^2_{L^2(\Theta_3)}\ge \msmall{\gamma_k \over 2} \cdot
\left( \norm{du_k}^2_{L^2(\Theta_2)} + \norm{du_k}^2_{L^2(\Theta_4)} \right)
\eqno(A3.3.6)$$ with $\gamma_k = 1- 1/k$. [*Lemmas 5.3.1*]{} and [*A3.2.1*]{} provide that in this case
$\diam_h(u_k( \Theta(1,4))) \allowbreak \lrar 0$.
Since $f_{k,i} :W_i \to X$ are $h$-uniformly transversal with the same parameters $\delta$ and $M$, there should exist sequences $x_k \in A$, $x_{k,
0} \in A_0$, and $x_{k,1} \in A_1$ such that $x_k = u_0(x_{k,0}) =u_1(x_{k,
1})$ and $u_k(\Theta(1,4)) \subset B(x_k, r_k)$ with $r_k \lrar 0$. The $h
$-uniform continuity of $J^*$ implies that there exist $C^1$-diffeomorphisms $\phi_k: B(x_k, r_k) \to B(0, r_k) \subset \cc^n$ with $\norm{J_k - \phi^*_k
J\st}_{L^\infty(B(x_k, r_k))} + \norm{h -\phi^*_k h\st} _{L^\infty(B(x_k,
r_k))} \lrar 0$.
Using $\phi_k$, we transfer our situation into $B(0, r_k) \subset \cc^n$ and rescale it. Namely, we set $\alpha_k \deff \norm{du_k}_{L^2(\Theta_3)}$ and define diffeomorphisms $\psi_k \deff {1\over \alpha_k} \scirc \phi_k: B(x_k,
r_k) \to B(0, R_k) \subset \cc^n$ with $R_k \deff \alpha_k\inv \cdot r_k$. Note that by [*Lemmas 5.3.1*]{} and [*A3.2.1*]{} we have $\alpha_k =
\norm{du_k} _{L^2(\Theta(2,3))} \le C \diam_h(u_k( \Theta(1,4))) \le C' r_k$, so that $R_k$ are uniformly bounded from below.
In $B(0, R_k)$ we consider Riemannian metrics $h_k \deff \alpha^{-2}_k \cdot
\psi_{k\,*}h_k$ (pushed forward and $\alpha^{-2}_k$-rescaled $h_k$), almost complex structures $J^*_k \deff \psi_{k\,*} J_k$ and $J^*_k
$-holomorphic maps $u^*_k \deff \psi_k \scirc u_k : \Theta(1,4) \to
B(0,R_k)$. Note that here we consider $h$ as a metric tensor; thus, multiplying $h$ by $\alpha^{-2}$, we increase $h$-norms and $h$-distances in $\alpha^{-1}$ and not in $\alpha^{-2}$ times.
Then $\norm{du^*_k}_{L^2(\Theta_3, h_k)}=1$, $\norm{du^*_k}^2_{L^2(\Theta_2,
h_k)} + \norm{du^*_k}^2_{L^2(\Theta_4, h_k)} \le {2k\over k-1}$, and $$\norm{J^*_k - J\st}_{L^\infty(B(0, R_k), h_k)} =
\norm{J_k - \phi^*_kJ\st}_{L^\infty(B(x_k, r_k), h)} \lrar 0.$$ The last equality uses the obvious relation $$\msmall{|F(\xi)|_{\alpha^{-2} \cdot h} \over |\xi|_{\alpha^{-2} \cdot h} }
= \msmall{\alpha\inv \cdot |F(\xi)|_h \over \alpha\inv \cdot |\xi|_\cdot h }
= \msmall{|F(\xi)|_h \over |\xi|_h }$$ for any linear $F: T_xX \to T_xX$ and $\xi\not = 0\in T_xX$. In a similar way we also obtain $\norm{h_k - h\st}_{L^\infty(B(0, R_k), h_k)}\lrar 0$.
Going to a subsequence, we may additionally assume that the tangent spaces $d\psi_k \scirc df_i (T_{x_{k,i}} W_i)$, $i=1,2$, converge at some spaces $W_i^* \subset \cc^n$. Since $W_i$ are uniformly totally real, $W_i^*$ are also totally real linear subspaces in $\cc^n$. Since the maps $df_i : TW_0
\to TX$ are uniformly continuous on $A_i \subset W_i$, $f_{k,i} \lrar f_i$ in $C^1(A_i)$, and since $r_k \lrar 0$, the images $W^*_{k,i} \deff \psi_k
\scirc f_{k,i}(B_{W_i}(x_{k,i}, r_k))$ of the balls $B_{W_i}(x_{k,i}, r_k)
\subset W_i$ are imbedded submanifolds of $\cc^n$ with $0 \in W^*_{k,i}$ which converge to $W^*_i$ in Hausdorff topology. Moreover, we can consider $W^*_{k,i}$ as graphs of maps $g_{k,i}$ from subdomains $U_{k,i} \subset
W^*_i \cap B(0,R_k)$ to $W^*_i {}^\perp$ and for any fixed $R \le \inf
\{R_k\}$ the restrictions $g_{k,i} |_{W^*_i \cap B(0,R)}$ converge to zero map from $W^*_i \cap B(0,R)$ to $W^*_i {}^\perp$.
The a priori estimates for the maps $u^*_k : \Theta(1,4) \to \cc^n$ provide that for any $p<\infty$ the maps $u^*_k$ converge in weak- $L^{1,p}$-topology to some $J\st$-holomorphic map $u^*: \Theta(1,4) \to \cc^n$. Furthermore, since $u^*_k$ satisfy totally real boundary conditions $u^*_k|_{\d_i\Theta(1,4)}
\subset W^*_{k,i}$, the same is true for $u^*$, i.e., $u^*|_{\d_i\Theta(1,4)}
\subset W^*_i$. Nice behavior of $W^*_{k,i}$ provides that on $\Theta_3$ we also have a strong convergence; hence $\norm{du^*}_{L^2(\Theta_3)} = \lim
\norm{du^*_k}_{L^2(\Theta_3)} =1$. In particular, $u^*$ is not constant. On the other hand, $\norm{du^*}^2_{L^2(\Theta_2)} + \norm{du^*}^2_{L^2(\Theta_4)}
\le \lim \norm{du^*_k}^2_{L^2(\Theta_2)} + \norm{du^*_k}^2_{L^2(\Theta_4)} \le
2$. The obtained contradiction to [*Lemma A3.3.1*]{} shows that [*Lemma A3.3.2*]{} is true.
Let $X$, $h$, $J$, $A$, $f_i: W_i \to X$, $A_i$, and the constant $\eps^b_2$ and $\gamma^b$ be as in [*Lemma A3.3.2*]{}. Suppose that $\ti J$ is a continuous almost complex structure on $X$ with $\norm{\ti J-J}_{L^\infty(A)} <\eps^b_2$, and $\ti f_i: W_i \to X$ are totally real immersions with $\dist(\ti f_i, f_i)
_{C^1(A_i)} \le \eps^b_2$ such that $f_i$ are $h$-uniformly transversal along $A_i$ with the same parameters $\delta$ and $M$ as $f_i: W_i \to X$.
Corollary A3.3.3.
*Let $u \in C^0\cap L^{1,2}(\Theta(0,l),X)$ be a $\ti J$-holomorphic map such that $u(\Theta(0,l))\subset A$, $u|_{\d_i
\Theta(0,l)} = \ti f_i \scirc u^b_i$ for some continuous $u^b_i: \d_i\Theta(
0,l) \to A_i \subset W_i$ and $\norm{du}_{ L^2(Z_k)}<\eps_2$ for any $k=1,\ldots,l$.*
Let $\lambda_b>1$ be the (uniquely defined) real number with $\lambda_b
= {\gamma^b \over 2}(\lambda_b^2+ 1)$. Then for $2\le k\le l-1$ the following holds: $$\norm{du}^2_{L^2(\Theta_k)}
\le \lambda_b^{-(k-2)} \cdot \norm{du}^2_{L^2(\Theta_2)} +
\lambda_b^{-(l-1-k)} \cdot \norm{du}^2_{L^2(\Theta_{n-1})}.
\eqno(A3.3.7)$$
Proof. It is the same as in [*Lemma 5.3.5*]{}.
The immediate corollary of this estimate is a lower bound of energy on a nonconstant “infinite strip”.
Lemma A3.3.4. [*Let $X$, $h$, $J$, $A$, $f_i: W_i \to X$, $A_i$, the constant $\eps^b_2$ and $\gamma^b$, also a structure $\ti J$, and immersions $\ti f_i : W_i \to X$ be the same as in [*Corollary A3.3.3*]{}. Let $u \in C^0\cap L^{1,2}(\Theta(-\infty, +\infty),X)$ be a nonconstant $\ti
J$-holomorphic map such that $u(\Theta(-\infty, +\infty))\subset A$ and $u|_{\d_i\Theta(0,l)} = \ti f_i \scirc u^b_i$ for some continuous $u^b_i :
\d_i\Theta(-\infty, +\infty) \to A_i \subset W_i$. Then $\norm{du}_{ L^2(
\Theta_k)} > \eps^b_2$ for some $k$. In particular, $\norm{du}_{ L^2(
\Theta(-\infty, +\infty))} > \eps^b_2$.* ]{}
Proof. [*Corollary A3.3.3*]{} provides that if $\norm{du}_{L^2(
\Theta_k)} \le \eps^b_2$ for all $k$, then $\norm{du}_{ L^2(\Theta_k)}
\allowbreak =0$, i.e., $u$ is constant.
Another consequence of [*Corollary A3.3.3*]{} is a generalization of Gromov’s result about removability of boundary point singularity, see \[G\]. An important improvement is the fact that the statement remains valid also when one has [*different*]{} boundary conditions on the left and on the right of a singular point. One can see such a point $x$ as a [*corner point*]{} for the corresponding complex curve. A typical example appears in symplectic geometry where one takes Lagrangian submanifolds as boundary conditions.
Define the punctured half-disk by setting $\check\Delta^+ \deff \Delta^+\bs
\{0\}$. Define $I_- \deff (-1,0) \subset \d\check\Delta^+$ and $I_+ \deff (0,+1) \subset \d\check\Delta^+$.
Corollary A3.3.5.
*(Removal of boundary point singularities). *Let $X$ be a manifold with a Riemannian metric $h$, $J$ a continuous almost complex structure, $f_i : W_i \to X$, $i=1,2$, totally real immersions and $A_i \subset W_i$ subsets. Let $u:(\check\Delta^+, J\st)
\to (X,J)$ be a holomorphic map. Suppose that**
$J$ is uniformly continuous on $A \deff u(\check\Delta)$ $h$, and closure of $A$ is $h$-complete;
$u$ satisfies boundary conditions of the form $u|_{I_+} = f_0
\scirc u^b_+$ and $u|_{I_-} = f_1 \scirc u^b_-$ with some continuous $u^b_+ :
I_+ \to A_0 \subset W_0$ and $u^b_- : I_- \to A_1 \subset W_1$;
$f_i$ are $h$-uniformly totally real on $A_i$ and $h$-uniformly transversal along $A_i$;
there exists $k_0$ such that for all half-annuli $R^+_k\deff \{
z\in \Delta^+ :{1\over e^{\pi(k+1)}}\le | z| \le {1\over e^{\pi k}}\}$ with $k\ge k_0$ one has $\norm{du}^2_{L^2(R^+_k)}\le \eps^b_2$, $\eps^b_2$ as in [*Lemma A3.3.2*]{}.
Then $u$ extends to origin $0\in \Delta^+$ as an $L^{1,p}$-map for some $p>2$.
Proof. Using the holomorphic map $\exp : \Theta(0,\infty) \to
\check\Delta^+$, $\exp(\theta + \isl t) \deff e^{\pi(-t + \isl\theta)}$, we can reduce our situation to the case of the holomorphic map $u^* \deff u
\scirc \exp$ form “infinite strip” $\Theta(0,\infty)$. By [*Corollary A3.3.3*]{}, for $k\ge k_0$ we obtain the estimate $\norm{ du^*} _{L^2(\Theta_k)} \le
\lambda_b^{-(k-k_0)/2} \norm{ du^*}_{L^2(\Theta_{k_0})}$ with some $\lambda_b
>1$. This is equivalent to the estimate $\norm{ du} _{L^2(R^+_k)} \le
\lambda_b^{-(k-k_0)/2} \norm{ du}_{L^2(R^+_{k_0})}$. [*Lemmas 5.3.1*]{} and [*A3.2.1*]{} and the scaling property of $L^p$-norms provide the estimate $$\norm{ du}_{L^p(R^+_k)} \le C e^{-k(\log\lambda_b/2 + \pi(2/p-1))}.\eqno(A3.3.8)$$ Thus, $du \in L^p(\Delta^+)$ for any $p$ with $\log\lambda_b/2 > \pi(1-2/p)$, which means $p < {4\pi \over 2\pi - \log \lambda_b}\cdot$
Remark. Unlike the “inner" and smooth boundary cases, it is possible that the map $u$, as in [*Corollary A3.3.5*]{}, is not $L^{1,p}$-regular in the neighborhood of a “corner point” $0\in \Delta^+$ for some $p>2$. For example, the map $u(z) = z^\alpha$ with $0<\alpha <1$ satisfies totally real boundary conditions $u(I_+) \subset \rr$, $u(I_-) \subset e^{\alpha\pi \isl}\rr$ and is $L^{1,p}$-regular only for $p< p^* \deff {2\over1-\alpha}\cdot$
As in the “inner” case, for the proof of the boundary compactness theorem we need a description of a convergence of a “infinitely long strip”. Let $X$ be a manifold with a Riemannian metric $h$, $J$ a continuous almost complex structure, $A\subset X$ a closed $h$-complete subset such that $J$ is $h$-uniformly continuous on $A$, and let $\{J_n\}$ be a sequence of almost complex structures converging $h$-uniformly on $A$ to $J$. Also let $f_0: W_0
\to X$ and $f_1: W_1\to X$ be immersions, $A_i \subset W_i$ subsets such that $df_i$ are uniformly $h$-uniformly totally real on $A_i$ and $f_i$ are $h$-uniformly transversal along $A_i$. Let $f_{n,i} : W_i \to X$ be totally real immersions, which $C^1$-converge to $f_i$ on $A_i$ such that $f_{n,0}$ and $f_{n,1}$ are $h$-uniformly transversal along $A_i$ with uniform in $n$ parameters $\delta$ and $C^*$. Finally, let $\{l_n\}$ be a sequence of integers with $l_n\to \infty$, and $u_n: \Theta(0,l_n) \to X$ a sequence of $J_n$-holomorphic maps, satisfying boundary conditions $u_n|_{\d_i\Theta(0,
l_n)} = f_{n,i}\scirc u^b_{n,i}$ with some continuous $u^b_{n,i}:
\d_i\Theta(0,l_n) \to A_i \subset W_i$.
Lemma A3.3.6.
*In the described situation, suppose additionally that $u_n(\Theta(0,l_n)) \allowbreak \subset A$ and $\norm{du_n} _{ L^2(
\Theta_k) }\le \eps^b_2$ for all $n$ and $k\le l_n$. Take a sequence $k_n\to
\infty$ such that $k_n<l_n-k_n\to \infty$. Then*
[*1)*]{} $\norm{du_n}_{L^2(\Theta(k_n,l_n-k_n))}\to 0$ and $\diam\bigl(u_n (\Theta(k_n,l_n-k_n))\bigr)\to 0$.
[*2)*]{} There is a subsequence $\{ u_n \}$, still denoted $\{ u_n \}$ such that both $u_n |_{\Theta(0,k_n)}$ and $u_n |_{\Theta(k_n,l_n)}$ converge in $L^{1,p}
$-topology on compact subsets in $\check\Delta^+ \cong \Theta(0, +\infty)$ to $J^*$-holomorphic maps $u^-_\infty$ and $u^+_\infty$. Moreover, both $u^+_\infty$ and $u^-_\infty$ extend to the origin and $u^+_\infty(0)=
u^-_\infty(0)$.
Proof. It follows from the above considerations.
[A3.4. Gromov Compactness for Curves with Totally Real Boundary Conditions.]{}
Let us turn to the Gromov Compactness Theorem for curves with boundary on totally real submanifolds. To give a precise statement we need to modify the definition of the Gromov convergence ([*Definition 4.1.6*]{}). The reason to do this is the following. Considering [*open*]{} curves $C_n$ with changing complex structures, we want to fix some kind of a common “neighborhood of infinity" $i_n: C^*\hookrightarrow C_n$ of every $C_n$. Thus, we can imagine that all changes of complex structure take place “outside of infinity”, i.e., in a relatively compact part $C_n \bs i_n(C^*) \Subset C_n$. This is done to insure that $C_n$ do not approach infinity in an appropriate moduli space.
On the other hand, it is more natural to consider curves $(\barr C_n, u_n)$ with totally real boundary conditions as compact objects without “infinity”. In fact, in this case the behavior of $u_n$ near the boundary $\d C_n$ can be controlled. The obtained a priori estimates near a “totally real boundary” can be viewed as a part of such a “control”. So for curves with totally real boundary conditions we can hope to extend the Gromov convergence up to the boundary.
Further, as in the “inner case”, an appropriate modification of the Gromov convergence in this case should allow boundary bubbling and the appearance of boundary nodes. This means, however, that the structure of the boundary can change during the approach to the limit curve and cannot be considered as fixed. Instead, one should fix a type of boundary conditions. We shall consider the following general situation.
Let $u_n: \barr C_n \to X$ be a sequence of stable $J_n$-complex curves over $X$ with parameterizations $\delta_n: \barr\Sigma \to \barr C_n$. Also let $\bfbeta =\{ \beta_i \}_{i=1}^m$ be a collection of arcs $\beta_i$ in $\d
\Sigma$ such that $\cup_{i=1}^m \beta_i= \d\Sigma$ and that the interiors of $\beta_i$ are mutually disjoint and do not intersect pre-images of boundary nodal points of $C_n$. Let further $\{ W_i\} _{i=1}^m$ be a collection of real $n$-dimensional manifolds, $f_{n,i} : W_i \to X$ a sequence of totally real immersions and $u^b_{n,i}:
\beta_{n,i} \to W_i$ a sequence of continuous maps from $\beta_{n,i} \deff
\delta_n( \beta_i)$. Then ${\mib W}_n \deff \{ (W_i,
f_{n,i} )\} _{i=1}^m$ are totally real boundary conditions on $(\barr C_n,
u_n)$ of the same type $\bfbeta$.
Definition A3.4.1.
*In the situation above we say that the sequence of boundary conditions ${\mib W}_n$ of the same type $\bfbeta$ [*converges $h$-uniformly transversally*]{} to $J^*$-totally real boundary conditions ${\mib W}$ on subsets $A_i \subset W_i$ if*
${\mib W}= \{ (W_i, f_i) \}_{i=1}^m$, where $f_i: W_i \to X$ are $J^*$-totally real immersions;
$f_{n,i}$ converge to $f_i$ in $C^1$-topology and this convergence is $h$-uniform on $A_i$;
for any $n$ immersions $\{f_{n,i}\}_{i=1}^m$ are mutually $h$-uniformly transversal along $A_i$ with parameters $\delta>0$ and $M$, and these parameters are independent of $n$.
Note that condition implies that the limit immersions $f_i$ are also mutually $h$-uniformly transversal along $A_i$ with the same parameters $\delta>0$ and $M$.
Definition A3.4.2.
*We say that the sequence $(\barr C_n, u_n)$ [*converges up to the boundary*]{} to a stable $J^*$-holomorphic curve $(\barr
C_\infty, u_\infty)$ over $X$ if the parameterizations $\sigma_n: \barr\Sigma
\to \barr C_n$ and $\sigma_\infty: \barr\Sigma \to \barr C_\infty$ can be chosen in such a way that the following holds:*
$u_n\scirc \sigma_n$ converges to $u_\infty\scirc \sigma_\infty$ in $C^0( \barr\Sigma, X)$-topology;
if $\{ a_k \}$ is the set of the nodes of $C_\infty$ and $\{ \gamma_k
\}$, $\gamma_k \deff \sigma_\infty\inv(a_k)$ are the corresponding circles and arcs in $\barr\Sigma$, then on any compact subset $K\comp \barr\Sigma \bs
\cup_k\gamma_k$ the convergence $u_n\scirc \sigma_n\to u_\infty\scirc
\sigma_\infty$ is $L^{1,p}(K, X)$ for all $p< \infty$;
for any compact subset $K\comp \barr\Sigma \bs \cup_k\gamma_k$ there exists $n_0=n_0(K)$ such that $ \sigma_n^{-1}(\{ a_k \}) \cap K= \emptyset$ for all $n\ge n_0$ and complex structures $\sigma_n^*j_{C_n}$ smoothly converge to $\sigma_\infty^*j_{C_\infty}$ on $K$.
Theorem A3.4.1.
*Fix a metric $h$ on $X$, and an $h$-complete subset $A\subset X$, and subsets $A_i \subset W_i$. Suppose that*
[*a)*]{} $J_n$ are continuous almost complex structures on $X$, converging $h$-uniformly on $A$ to a continuous almost complex structure $J^*$;
[*b)*]{} $u_n(C_n) \subset A$ and $\area [u_n (C_n)]\le M$ with a constant $M$ independent of $n$;
[*c)*]{} ${\mib W}_n \deff \{(W_i, f_{n,i}) \}_{i=1}^m$ are totally real boundary conditions of the same type $\bfbeta = \{ \beta_i \}_{i=1}^m$ such that ${\mib W}_n$ converge $h$-uniformly transversally to a boundary condition ${\mib W}= \{(W_i, f_i) \}_{i=1}^m$ on subsets $A_i \subset W_i$;
[*d)*]{} immersions $f_i: W_i \to (X,J^*)$ are $h$-uniformly totally real along $A_i$;
[*e)*]{} there exist maps $u^b_{i,n}: \beta_i \to A_i\subset W_i$, realizing boundary conditions ${\mib W}_n$.
Then there exits a subsequence of $\{( \barr C_n, u_n )\}$, still denoted $\{( \barr C_n, u_n )\}$, and parameterizations $\sigma_n: \barr \Sigma
\to \barr C_n$ such that $(C_n, u_n, \sigma_n)$ converges up to the boundary to a stable $J^*$-holomorphic curve $(\barr C_\infty, u_\infty, \sigma_\infty)$ over $X$.
If, in addition, $A_i \subset W_i$ are $f_i^*h$-complete, then the limit curve $(\barr C_\infty, u_\infty)$ satisfies real boundary conditions $\mib
W$ with maps $u^b_i: \beta_i \to A_i \subset W_i$.
Our main idea of the proof is to apply arguments used in the demonstration of [*Theorem 1.1*]{}. To realize this, we use the following trick. We replace every pair $(C_n, u_n)$ by a triple $(C^d_n, \tau_n, u^d_n)$, where $C^d_n$ is the [*Schottky double*]{} of $C_n$ with an antiholomorphic involution $\tau_n$ and $u^d_n: C_n^d \to X$ a $\tau_n$-invariant map. Then we shall show the changes of all the constructions in the proof to make them $\tau_n$-invariant in an appropriate sense. In particular, the convergence $(C^d_n, \tau_n, u^d_n) \lrar (C^d_\infty, \tau_\infty, u^d_\infty)$ will be equivalent to the convergence $(C_n, u_n) \lrar (C_\infty, u_\infty)$.
We start with a construction of the Schottky double of a nodal curve $\barr C$ with boundary. Take two copies $\barr C^+ \equiv \barr C$ and $\barr C^-$ of $\barr C$. Equip $\barr C^-$ with the opposite complex structure, so that the identity map $\tau: \barr C^+ \to \barr C^-$ now becomes antiholomorphic. Glue $\barr C^+$ and $\barr C^-$ together along their boundaries, identifying $\d C^+$ and $\d C^-$ by means of the identity map $\tau: \d C^+ \buildrel
\cong \over \lrar \d C^-$. The union $C^d \deff \barr C^+ \cup_{\d C} \barr
C^-$ obeys the unique structure of a closed nodal curve compatible with imbeddings $\barr C^\pm \hook C^d$. The boundary $\d C$ becomes the fixed point set of $\tau$.
The map $\tau$ induces an antiholomorphic involution of $C^d$ which we also denote by $\tau$. We call the obtained curve $C^d$ the [*Schottky double*]{} of $\barr C$. Note that every boundary nodal point $a_i\in \d C$ defines a $\tau
$-invariant nodal point $a_i$ on $C^d$, whereas an inner nodal point $b_i \in
C$ defines a pair of nodal points $b_i^\pm$ on $C^d$ interchanged by $\tau$. If $\sigma: \barr\Sigma \to \barr C$ is a parameterization of $\barr C$, then we obtain in an obvious way the double $\Sigma^d$ with the involution $\tau:
\Sigma^d \to \Sigma^d$ and the parameterization $\sigma^d: \Sigma^d \to C^d$ compatible with the involutions.
Remark. The introduced notation $C^d$ for the [*Schottky double*]{} of a nodal curve $\barr C$ with boundary coincides with that for the [*holomorphic double*]{}, used in [*Section 2*]{}. Since in the present section only the Schottky double is considered, this should not lead to confusion.
Suppose, additionally, that an almost complex structure $J$ on $X$ and a $J$-holomorphic map $u: \barr C \to X$ are given. Suppose, also, that the curve $(\barr C,u)$ satisfies the totally real boundary conditions $\mib W$ of type $\bfbeta$. In particular, $\bfbeta$ defines a certain system of arcs $\{\beta_i\}$ on $\d C$. In order to take into account the type of boundary conditions, we fix the ends of $\beta_i$ which are not boundary nodal points of $\barr C$ and declare these points as marked points of $C^d$. Note that these and the nodal points are the only “corner” points of $(\barr C,
u)$. The latter means that in a neighborhood of these points the map $u$ cannot be $L^{1,p}$-smooth for all $p< \infty$. The example in the [*Remark*]{} following [*Corollary A3.3.5*]{} explains the notion of a “corner point”. Considering the Schottky double, we shall always equip $C^d$ with this set of marking points. Note also that every boundary circle of $\barr C$ contains at least one nodal or marked point as above.
For $(\barr C,u)$ as above, we extend the $J$-holomorphic map $u: \barr C \to
X$ to a map $u^d: C^d \to X$ by setting $u^d(x) \deff u(\tau(x))$ for $x\in
C^-$. By the construction, $u^d$ is $\tau$-invariant, $u^d \scirc \tau= u^d$, but $u^d$ [*is not $J$-holomorphic*]{} (with the only trivial exception $u
\equiv const$). However, the analysis already described in this section provides necessary $L^{1,p} $-estimates for $u^d$, at least for some $p^*>2$.
In the situation of [*Theorem A3.4.1*]{}, such an exponent $p^*>2$ can be chosen to be the same for all curves $(\barr C_n, u_n)$. This depends only on the topology of $\barr C_n$ and the geometry of immersions $f_n: W_n \to X$. In particular, every $u_n^d$ is continuous.
The next step of the proof is to find a $\tau_n$-invariant decomposition of $C^d_n$ into pants. This implies that the corresponding graph $\Gamma_n$ becomes $\tau_n$-invariant. In the construction which follows we shall use the fact that $\tau_n$ is an isometry on the union of the non-exceptional components of $C^d_n$. This is provided by the uniqueness of the intrinsic metric.
Lemma A3.4.2. *Let $C$ be a nodal curve with boundary, $\sigma: \barr
\Sigma \to \barr C$ a parameterization, and $\{ x_i \}_{i=1}^m$ a set of marked points on the boundary $\d C$. Let $C^d$ be the Schottky double of $C$ with the antiholomorphic involution $\tau$.*
Then there exists a $\tau$-invariant decomposition of $C^d\bs \mapo$ into pants such that the intrinsic length of corresponding boundary circles is bounded by a constant $l^+$ depending only on genus $g$ of $\Sigma^d$ and the number of marked points $m$.
Moreover, every [*short*]{} geodesic appears as a boundary circle of some pants of the decomposition.
Remark. Recall (see [*Remark*]{} on page 25) that a closed geodesic $\gamma$ is called [*short*]{} if $\ell(\gamma) < l^*$, where $l^*$ is the universal constant $l^*$ with the following property. For any simple closed geodesics $\gamma'$ and $\gamma''$ on the conditions $\ell(\gamma') <l^*$ and $\ell( \gamma'') <l^*$ imply $\gamma' \cap \gamma'' = \emptyset$.
Proof. Since the genus of the parameterizing real surface $\Sigma^d$ and the number of marked points is fixed, we obtain a uniform upper bound on the possible genera and the number of marked points of non-exceptional components of $C^d$, as well as on the number of exceptional components. This implies that there exists a decomposition of every non-exceptional component $C_i$ of $C^d$ into pants $S_\alpha$ such that the intrinsic length of boundary circles of $S_\alpha$ is bounded by the constant $l^+$ depending only on $g$ and $m$. The idea of the proof of our lemma is to show that the construction of such a decomposition, given in \[Ab\], Ch.II, §3.3, can be modified to produce a $\tau$-invariant decomposition.
Let us first describe the construction itself, say, for a given smooth curve $C^*$ with marked points $\{x_i\}$ of non-exceptional type. The procedure is done inductively by choosing at every step a non-trivial simple closed geodesic $\gamma_{J^*} \subset C^* \bs\mapo$, disjoint from an already chosen geodesic $\gamma_j$, $j<J^*$. Moreover, at every step there exists a geodesic $\gamma_{J^*}$ as above whose intrinsic length is bounded by a constant $l^+_{J^*}$ depending only on the genus of $C^*$, the number of marked points, and the maximum of the lengths of the already chosen geodesics $\gamma_j$, $j<J^*$.
Take any non-exceptional component $C^d_i$ of $C^d$. Two cases can occur: either $C^d_i$ is $\tau$-invariant, or $\tau \bigl( C^d_i \bigr)$ is another component $C^d_{i'}$. Two separate cases are distinguished: $C^d_i$ intersects the boundary $\d C$ (first case) or not (second one).
The existence of $\tau$-invariant decomposition into pants for every pair of non-exceptional components $C^d_i$ and $\tau(C^d_i) \not= C^d_i$ is obvious. We choose an appropriate decomposition of $C^d_i$ and transfer it on $\tau(
C^d_i)$ by means of $\tau$.
It remains to consider the case of a $\tau$-invariant non-exceptional component $C^d_i$.
Suppose that at some step we have already chosen a $\tau$-invariant set $\{\gamma_1, \ldots, \allowbreak \gamma_{J^*-1}\}$ of simple disjoint geodesics on $C^d_i \bs \mapo$. Take a simple geodesic $\gamma$ of the length $\ell(\gamma) \le l^+_{J^*}$, where $l^+_{J^*}$ is the upper bound introduced above. By the construction of the double $C^d$, the fixed point set of $\tau$ on $C^d_i$ is $C^d_i \cap \d C$ and is non-empty. Denote $C_i \deff C \cap
C^d_i$, so that $C^d_i \cap \d C= \d C_i$. Note that any boundary circle of $C_i$ contains at least one marked point of $C^d_i$. Consequently, it has an infinite length the intrinsic metric on $C^d_i \bs\mapo$. Thus the chosen geodesic $\gamma$ cannot lie on $\d C_i$. Only three cases can happen.
[*Case 1.*]{} $\gamma$ is disjoint from $\d C_i$. Then $\gamma$ lies either in $C_i$ or in $\tau (C_i)$. In any case, $\gamma\cap \tau(\gamma) = \emptyset$. Thus, we can set $\gamma_{J^*} = \gamma$ and $\gamma_{J^*+1}= \tau(\gamma)$, obtaining the $\tau$-invariant set $\{\gamma_1, \ldots, \gamma _{J^*+1}\}$ of simple disjoint geodesics. This will be discussed in the next two steps of our construction.
[*Case 2.*]{} $\gamma \cap \d C_i \not=0$ and $\gamma$ is $\tau$-invariant. We set $\gamma_{J^*}=\gamma$ and proceed inductively. Note that in this case $\gamma \cap \d C_i$ consists of 2 points, in which $\gamma$ is orthogonal to $\d C_i$.
The last expression means that we move along corresponding arcs in the prescribed order, as it is shown on Fig. 11. Note that only one part of $C^d_i$ is drawn, namely $C_i$. The rest of the picture is symmetric the involution $\tau$. Thus we can see only half of the geodesics in classes $[\ti\gamma_i]$, $i=2,3,4$.
Each of the classes $[\ti\gamma_k]$ is either represented by a closed geodesic or corresponds to a curve which winds around some marked point of $C^d_i$. To shorten notation, we say in the last case that the class $[\ti\gamma_i]$ corresponds to a marked point of $C^d_i$.
If one of the classes $[\ti\gamma_k]$, $k=1,2,3$, is represented by the geodesic $\ti\gamma_k$, which is different and disjoint from the already chosen geodesics $\gamma_j$, $j<J^*$, then we can set $\gamma_{J^*} =
\ti\gamma_k$. If $k=1$ we also set $\gamma_{J^*} = \ti\gamma_1$ and $\gamma_{J^*+1} = \tau(\ti\gamma_1)$. Then we proceed inductively.
To finish the proof it remains to consider the following situation. Under the conditions of [*Case 3*]{}, each of the classes $[\ti\gamma_k]$, $k=1,2,3$, either corresponds to a marked point or is represented by a closed geodesic $\ti\gamma_k$, which intersects or coincides with one from the already chosen geodesics $\gamma_j$, $j<J^*$.
We state that, in fact, a proper intersection cannot happen, i.e., each class $[\ti\gamma_k]$, $k=1,2,3$, either corresponds to a marked point or is represented by an already chosen geodesic $\gamma_j$, $j<J^*$. To show this we note that $\gamma_j \cap \tau(\gamma) =\emptyset$ for all $j<J^*$. Otherwise, we could have a contradiction with the conditions $\gamma_j \cap
\gamma =\emptyset$ and $\tau$-invariance of the set of the geodesics $\gamma_j$, $j<J^*$. Consequently, each class $[\ti\gamma_k]$ is represented by a circle $\alpha_k \subset C^d_i \bs\mapo$, $k=1,2,3,4,$ with $\alpha_k
\cap \gamma_j = \emptyset$.
Now assume that the proper intersection of $\ti\gamma_k$ and some $\gamma_j$, $j<J^*$ occurs. Let $\ell_k\deff \ell(\ti\gamma_k)$ be the intrinsic metric of $\ti\gamma_k$. As in the proof of [*Lemma 4.3.2*]{} construct the annulus $A= \{ (\rho, \theta) : |\rho| < {\pi^2 \over \ell} \} \times \{0\le
\theta \le 2\pi\}$ with the metric $({\ell_k \over 2\pi} / \cos {\ell_k \rho
\over 2\pi})^2 (d\rho^2 + d\theta^2)$ and an isometric covering of $C^d_i
\bs\mapo$ by $A$, which sends the geodesic $\beta_k \deff \{ \rho=0\} \subset
A$ onto $\ti\gamma_k \subset C^d_i$. Find a lift of $\gamma_j$ to a geodesic line $L_j \subset A$ with $L_j\cap \beta_k \not= \emptyset$ and a lift of a circle $\alpha_k$ to a circle $\ti\alpha_k \subset A$ homotopic to $\beta_k$. Then the intersection $L_j \cap \beta_k$ must consist of exactly one point, and, consequently, the homology intersection index $[L_j] \cdot [\beta_k]$ is equal to $\pm1$. This would imply that $[L_j]\cdot [\ti\alpha_k] =[L_j]
\cdot [\beta_k] \not =0$ and consequently $L_j \cap \ti\alpha_k \not=
\emptyset$. But this would contradict $\gamma_j \cap \alpha_k =\emptyset$.
Summing up, we see that in the situation we are considering we must have the picture of [*Fig. 11*]{}. Namely, both geodesics $\gamma$ and $\tau(\gamma)$ lie in a $\tau $-invariant domain $\Omega$ on $C^d_i$ with four components of the boundary; these components of $\d\Omega$ are either marked points or geodesics corresponding to the classes $[\ti\gamma_1]$, $[\tau(\ti\gamma_1)]$, $[\ti
\gamma_2]$, $[\ti\gamma_3]$; finally, every boundary circle of $\Omega$ is one of the geodesic $\gamma_j$. We conclude that the class $[\ti\gamma_4]$ is represented by a $\tau$-invariant geodesic $\ti\gamma_4$, which can be chosen at this step of the construction of $\tau$-invariant decomposition of $C^d_i$ into pants.
Note that by construction for the intrinsic length of $\gamma_{J^*}$ we obtain $\ell(\gamma_{J^*}) \le 2\ell(\gamma) \le 2l^+_{J^*}$. This means that in our construction we do not lose control of the intrinsic length of the chosen geodesics. This shows the existence of a constant $l^+$ stated in the lemma.
Finally, the definition of a [*short*]{} geodesic provides that the geodesic $\gamma$ in [*Case 3*]{} above cannot be short. This implies that the set of short geodesics on $C^d$ is disjoint. Since the involution $\tau$ is an isometry, the set of short geodesics on $C^d$ is also $\tau$-invariant. Thus, in our construction of decomposition into pants we can start with this set of geodesics. This shows the last statement of the lemma.
Remark. To explain the meaning of [*Lemma A3.4.2*]{}, let us consider pants $S$ with a complex structure $J_S$ and an antiholomorphic involution $\tau$ acting on $S$. It is easy to see that only two types of such an action, illustrated by Figs. 8a) and 8b), are possible.
In the first case, Fig. 12a), the involution $\tau$ interchanges two boundary components $\gamma_1$ and $\gamma_2$ of $S$ and leaves the third one $\gamma_3$ invariant. The fixed point set $\beta$ of $\tau$ is a geodesic arc with both ends on the $\tau$-invariant boundary component $\gamma_3$. This case includes subcases where some boundary components of $S$ are not geodesics but marked points (punctures). In particular, if $\gamma_3$ is a marked point, then the set $\beta$ is an (infinite) geodesic line with both ends approaching $\gamma_3$. The set $\beta$ divides $S$ into two parts, $S^+$ and $S^-$ (see Fig. 12a)), which are interchanged by $\tau$. Topologically, each part $S^\pm$ is an annulus.
In the second case, Fig. 12b), all three boundary components $\gamma_1$, $\gamma_2$, and $\gamma_3$ are invariant. The fixed point set of $\tau$ consists of geodesic arcs $\beta_1$, $\beta_2$, and $\beta_3$. These are the shortest simple geodesics between $\gamma_2$ and $\gamma_3$, resp., $\gamma_3$ and $\gamma_1$ and resp., $\gamma_2$ and $\gamma_3$. If some boundary component of $S$ is not a geodesic but a marked point, then corresponding arcs have ends of infinite length approaching this boundary component. The arcs $\beta_k$, $k=1,2,3$, divide $S$ into two parts, $S^+$ and $S^-$ (see Fig. 12b)), which are interchanged by $\tau$. In this case, each part $S^\pm$ is topologically a disc.
We call pieces $S^\pm$ [*half-pants of the first or second type*]{}, respectively. Note that in both cases $\tau$-invariant arcs $\beta$ or $\beta_i$ are orthogonal to corresponding boundary circles $\gamma_j$.
Return to the situation of a nodal curve $\barr C$ with boundary and marked points. Let $C^d$ be the Schottky double and $\tau$ the antiholomorphic involution. Suppose that $C^d \bs\mapo$ is non-exceptional. Use [*Lemma A3.4.2*]{} and find a $\tau$-invariant decomposition into pants $C^d = \cup_j
S_j$. Set $S_j^+ \deff S_j \cap \barr C$. Then we obtain a decomposition $\barr C= \cup_j S^+_j$ such that the pieces $S^+_j$ are either pants (which means $S^+_j = S_j$), or half-pants of the first or second type. This decomposition is a suitable one for the situation of Gromov convergence up to the boundary of curves with totally real boundary conditions. In particular, we obtain arcs $\beta_{j,k}$ as $\tau$-fixed point sets of $S^+_j$, which define a decomposition $\d C= \cup_{j,k} \beta_{j,k}$ of the boundary of $\barr C$. The collection $\bfbeta' \deff \{ \beta_{j,k} \}$ of these arcs satisfies condition of [*Definition 5.4*]{}, but it can be different from the collection $\bfbeta =\{ \beta_i \}$ which was given. The reason is that in the construction of the pants-decomposition $C^d = \cup_j S_j$ we can subdivide original arcs $\beta_i \in \bfbeta$ into smaller pieces, so that every arc $\beta_i \in \bfbeta$ is a union of arcs $\beta_{j,k}$ from $\bfbeta'$. This means compatibility of $\bfbeta$ and $\bfbeta'$.
The next step is to establish an analog of [*Theorem 5.3.2*]{}. Assume that the hypothesis of [*Theorem A3.4.1*]{} is fulfilled. For each curve $C_n$ denote by $C^d_n$ its Schottky double and by $\tau_n$ the corresponding involution.
Lemma A3.4.3. *In the situation above, after passing to a subsequence, there exist parameterizations $\sigma^d_n: \Sigma^d \to C^d_n$, a finite covering $\calv$ of $\Sigma^d$ by open sets $\{ V_\alpha \}$, and a set $\{x^*_1, \ldots, x^*_m\}$ of marked points on $\Sigma$ such that the conditions [*(a), (c)–(f)*]{} of [*Theorem 4.2*]{} and the following additional conditions [*(b’)*]{} and [*(h)*]{} are satisfied:*
[*(b’)*]{} $\sigma_n\{x^*_1, \ldots, x^*_m\}$ is the set of marked points on $C^d_n$ corresponding to the decomposition of the boundary $\d C_n$ into arcs $\beta_{n,i}$; moreover, each such point $x^*_j$ lies in a single piece of covering $V_\alpha$ which is a disc;
[*(h)*]{} there exists an involution $\tau : \Sigma^d \to \Sigma^d$ which is compatible with the covering $\calv$ and with parameterizations $\sigma^d_n$, i.e.,$\calv$ is $\tau$-invariant and $\tau_n \scirc \sigma^d_n = \sigma^d_n
\scirc \tau$. In particular, each marked point $x^*_i$ of $\Sigma^d$ is fixed by $\tau$.
Remark. The condition [*(g)*]{} of [*Theorem 5.3.2*]{} is trivial in this case, because $C^d_n$ and $\Sigma^d$ are closed.
Proof. One can use the proof of [*Theorem 5.3.2*]{} with minor modifications. Note that the starting points of that proof were the intrinsic metric on non-exceptional components of nodal curves $C_n$ and the decomposition of $C_n$ into pants. Now the existence of a $\tau$-invariant decomposition of the curves $C^d_n$ into pants is provided by [*Lemma A3.4.2*]{}, whereas the $\tau$-invariance of the intrinsic metrics follows from the fact that any (anti)holomorphic isomorphism of curves with marked points is an isometry the intrinsic metric. Thus the constructions of the proof of [*Theorem 5.3.2*]{} yield $\tau$-invariant objects. Condition [*(b’)*]{} does not cause much difficulty.
Now we are ready to finish the
Proof of Theorem A3.4.1. As was mentioned, our main idea is to modify the construction used in the proof of [*Theorem 4.1.1*]{} to make it $\tau$-invariant. The main work has already been done. We have the necessary a priori estimates, the construction of a $\tau$-invariant pants-decomposition of the double $C^d_n$ of the curve $\barr C_n$ and the appropriate covering $\calv$ of the real surface $\Sigma^d$ parameterizing the doubles $C^d_n$.
As in the proof of [*Theorem 4.1.1*]{}, we consider the curves $C_{ \alpha ,n}
\deff \sigma^d_n(V_\alpha)$. Due to the presence of the involutions $\tau_n$, the geometric situation in now different. This involves new phenomena and needs additional considerations and constructions. In particular, the pieces $C_{
\alpha ,n}$ are divided into two groups depending on whether they are disjoint from the boundary $\d C_n$ or intersect it. In the last case $C_{
\alpha ,n}$ is $\tau_n$-invariant. In this case we shall use the notation $C^+_{ \alpha ,n} \deff C_{ \alpha,n} \cap C_n$ for the part of $C_{ \alpha
,n}$ lying in $C_n$. In addition, we denote $V^+_\alpha \deff V_\alpha \cap
\Sigma$. Then $V_\alpha$ appear as the union of domains $V^+_\alpha$ and $\tau(V^+_\alpha)$, interchanged by $\tau$. Similarly, it is true for $C_{ \alpha,n}$.
To prove the theorem, we want to construct a refined covering $\wt\calv$ of $\Sigma$ and refined parameterizations $\ti\sigma_n: \Sigma \to C_n$ such that for every $V_\alpha \in \wt\calv$ the sequence $(C_{ \alpha ,n},
u_{ \alpha ,n})$ with $C_{ \alpha ,n} \deff \ti\sigma_n(V_\alpha)$ one the following convergence types holds:
In the case [*B$''$)*]{} the strong convergence of maps $\ti u^+_n: \cala^+
_0 \to X$ is the one in the $L^{1,p^*}$-topology for some $p^*>2$ [*up to the boundary intervals containing the nodal point*]{}. An equivalent requirement is the usual $L^{1,p^*}$-convergence of the doubles $\ti u^d_n: C_{\alpha, n}
\to X$ on compact subsets of $C_{\alpha, n} \cong \cala_0$.
To obtain a desired refinement, we use the same inductive procedure as in the proof of [*Theorem 1.1*]{}. To insure convergence near the boundary $\d C_n$, we take a new value for the constant determining the inductive step. We choose a positive $\eps^b$ such that $\eps^b \le \eps$ and all a priori estimates of this section are valid for maps with area.$\le 3\eps^b$. This will yield the convergence of type [*A)–C)*]{} for sequences of curves with totally real boundary conditions and with the upper bound $\eps^b$ on the area.
In fact, essential modifications of the constructions of [*Theorem 1.1*]{} are needed only if the covering piece $V_\alpha$ is $\tau$-invariant. Indeed, if $V_\alpha$ is not $\tau$-invariant, then we can apply all the argumentations and constructions used in [*Cases 1)–4)*]{} in the proof of [*Theorem 1.1*]{}, and then “transfer” them onto $\tau(V_\alpha)$ by means of $\tau$. This gives the inductive step preserving $\tau$-invariance.
Hence, it remains to consider the situation when the covering piece $V_\alpha$ is $\tau$-invariant. As in [*Theorem 1.1*]{}, we must consider four cases.
*Case 1$_b$): $C_{\alpha, n}$ have constant complex structure different from the one of the standard node.*
Case 2$_b$): $C_{\alpha, n}$ are annuli of changing conformal radii $R_n$ such that $R_n \to R <\infty$.
Case 3$_b$): $C_{\alpha, n}$ are isomorphic to the standard node, so that $C^+_{\alpha, n}$ are isomorphic to the standard boundary node $\cala_0^+$.
Case 4$_b$): $C_{\alpha, n}$ are annuli of infinitely growing conformal radii $R_n$.
The subindex $(\cdot)_b$ indicates that we will consider the cases where $V_\alpha$ intersects the [*b*]{}oundary of $\Sigma$. As was mentioned, the last property is equivalent to the fact that $V_\alpha$ is $\tau$-invariant. References to [*Cases 1)–4) *without**]{} the subindex will mean the corresponding parts of the proof of [*Theorem 1.1*]{}.
[*Case 1$_b$)*]{}. Without loss of generality we may assume that $V_\alpha$ is a domain with a fixed complex structure and a fixed antiholomorphic involution $\tau$, and that $u_{\alpha, n}: V_\alpha \to X$ is a sequence of $\tau$-invariant maps which are (anti)holomorphic outside the set of $\tau
$-invariant points of $V_\alpha$. If we have the convergence of type [*C)*]{}, there is nothing to do. Otherwise, we fix a $\tau$-invariant metric on $V_\alpha$ compatible with the complex structure. Repeating the constructions from [*Case 1)*]{}, we distinguish the “bubbling” points $y^*_1, \ldots, y^*_l$ where the strong convergence fails.
Take the first point $y^*_1$. Suppose $y^*_1$ is disjoint from $\d\Sigma$. Then we may assume that $y^*_1\in V^+_\alpha$. Thus, we can repeat the rest of the constructions from [*Case 1)*]{}. The only correction needed at this point is that the neighborhood $\Delta( y^*_1, \varrho)$ of $y^*_1$ must be small enough and lie in $V^+_\alpha$. Transferring all these constructions into $\tau(V_\alpha)$, we realize the inductive step preserving the $\tau$-invariance.
It remains to consider the case, where $y^*_1 \in \d\Sigma$. This means that $y^*_1$ is $\tau$-invariant. Let $z$ be a holomorphic coordinate in a neigborhood of $y^*_1$ on $V_\alpha$ such that $z=0$ in $y^*_1$, the involution $\tau$ corresponds to the conjugation $z \mapsto \bar z$ and $\im
z >0$ in $\Sigma$. Find the sequences $r_n\lrar0$ of the radii and $x_n \to
y^*_1$, using the constructions from [*Case 1)*]{}. Note that the sequence $\tau(x_n)$ has the same property. Thus, replacing some points $x_n$ by $\tau(x_n)$, we may additionally assume that all $x_n$ lie in $\barr
V^+_\alpha$. Let $v_n:\Delta (0,{\varrho \over 2r_n}) \to (X,J_n)$ be the rescalings of maps $u_n$ defined by $v_n(z) \deff u_n(x_n + {z\over r_n})$. Argumentations of [*Case 1*]{} show that there exists the limit $v_\infty:
\cc \to X$ of (a subsequence of) $\{ v_n \}$ which extends to a map $v_\infty
: S^2 \to X$.
Denote by $\rho_n$ the distance from $x_n$ to $\d\Sigma$ and by $\ti x_n$ the point on $\d\Sigma$ closest to $x_n$. Then $x_n = \ti x_n + \isl \rho_n$ in the coordinate $z$ introduced above. In addition, $\lim \ti x_n = y^*_1$. We consider two subcases according to the possible behavior of $\rho_n$ and $r_n$.
[*Subcase 1$'_b$): $\bigl\{{ \rho_n \over r_n }\bigr\}$ is bounded*]{}. Passing to a subsequence, we may assume that ${\rho_n \over r_n}$ converges. Fix an upper bound $b$ for the sequence ${\rho_n \over r_n}$. In particular, $b \ge \lim {\rho_n \over r_n}$.
For $n>\!>1$ define maps $v_n:\Delta (0,{\varrho \over 2r_n}-b) \to (X,J_n)$ and $\ti v_n:\Delta (0,{\varrho \over 2r_n}-b) \to (X,J_n)$ setting $v_n(z)
\deff u_n (x_n +r_n z)$ and $\ti v_n(z)\deff u_n(\ti x_n +r_n z)$, respectively. Then every $\ti v_n$ is the shift of the map $v_n$ by $\isl{\rho_n\over r_n}$, i.e., $\ti v_n(z)= v_n\bigl(z + \isl {\rho_n\over
r_n}\bigr)$. The arguments of [*Case 1)*]{} show that $v_n$ converge on compact subsets of $\cc$ to a non-constant map. Consequently, $\ti v_n$ also converge on compact subsets of $\cc$ to a non-constant map $\ti v_\infty: \cc
\to X$. Moreover, since $\area( \ti v_\infty( \cc))$ is finite, $\ti v_\infty$ extends to a map $\ti v_\infty: S^2 \to X$. By the choice of $\eps^b$, $\area(\ti v_\infty (S^2)) \ge 3\eps^b$. Changing the choice of the constant $b$, we can additionally assume that $\area(\ti v_\infty( \Delta(0,b)) \ge
2\eps^b$. Then for all sufficiently big $n$ we obtain $$\area(\ti v_n(\Delta(0,b)) \ge \eps^b.
\eqno(A3.4.1)$$
For $n>\!>1$ we define the coverings of $V_\alpha$ by three sets $$V^{(n)}_{\alpha,1} \deff V_\alpha
\bs \barr\Delta (0, {\textstyle{\varrho\over2}}),
\qquad
V^{(n)}_{\alpha,2} \deff \Delta (0, \varrho)
\bs \barr \Delta (\ti x_n, br_n),
\qquad
V^{(n)}_{\alpha,3} \deff \Delta (\ti x_n, 2br_n).$$ Fix $n_0$ sufficiently big. Denote $V_{\alpha,1} \deff V^{(n_0)}_{\alpha,1}$, $V_{\alpha,2} \deff V^{(n_0)}_{\alpha,2}$ and $V_{\alpha,3} \deff V^{(n_0)}
_{\alpha,3}$. There exist diffeomorphisms $\psi_n: V_1 \to V_1$ such that $\psi_n: V_{\alpha,1} \to V^{(n)}_{\alpha,1}$ is an identity, $\psi_n:
V_{\alpha, 2} \to V^{(n)}_{\alpha,2}$ is a diffeomorphism, and $\psi_n:
V_{\alpha,3} \to V^{(n)}_{\alpha,3}$ is biholomorphic the complex structures, induced from $C_n$ by means of $\sigma^d_n$. Note that the sets $V^{(n)}_{\alpha,i}$ are $\tau$-invariant. Moreover, we can choose the maps $\psi_n$ in such a way that $\psi_n$ are also $\tau$-invariant.
The covering $\{ V_{\alpha,1}, V_{\alpha,2}, V_{\alpha,3} \}$ of $V_1$ and parameterizations $\ti\sigma_n \deff \sigma_{\alpha,n} \scirc \psi_n: V_1 \to
C_{\alpha, n}$ satisfy the conditions of [*Lemma A3.4.3*]{}. Moreover, inequality (A3.2.1) implies $\area(u_n(\ti\sigma_n(V_{\alpha,i}))) \le
%%%%%%%%%%%%%%%%%%%%%%%question%%%%%%%%%%%%%%%%%%%%%
(N-1)\eps^b$. Consequently, we can apply the inductive assumptions for the sequence of curves $\ti\sigma_n(V_{\alpha,i})$ and finish the proof by induction.
[*Subcase 1$''_b$): $\bigl\{{ \rho_n \over r_n}\bigr\}$ is unbounded*]{}. Passing to a subsequence, we may assume that ${\rho_n \over r_n}$ increases infinitely. However, $\rho_n \lrar 0$ since $x_n \lrar y^*_1 \in \d\Sigma$.
Define maps $v_n:\Delta (0,{\varrho \over 2r_n}) \to (X,J_n)$, setting $v_n(z)
\deff u_n(x_n +r_n z)$. As in [*Case 1)*]{}, $v_n$ converge on compact subsets of $\cc$ to a non-constant map $v_\infty: \cc \to X$, which extends to a map from the whole sphere $S^2$. Choose $b>0$ satisfying (A3.2.1). For $n>\!>1$ we define the coverings of $V_\alpha$ by five sets $$\mathsurround=0pt
\matrix\format\l\ \ &\l\\
V^{(n)}_{\alpha,1} \deff V_\alpha
\bs \barr\Delta (0, {\textstyle{\varrho\over2}}),
&
V^{(n)}_{\alpha,2} \deff \Delta(0, \varrho) \bs
\barr \Delta(0, 2\rho_n)
\cr
\noalign{\vskip5pt}
\rlap{$
V^{(n)}_{\alpha,3} \deff \Delta(0, 4\rho_n) \bs
\bigl(\barr \Delta (x_n, br_n) \cap \barr \Delta (\tau(x_n), br_n) \bigr)
$}&
\cr
\noalign{\vskip5pt}
V^{(n)}_{\alpha,4} \deff \Delta (x_n, 2br_n),
&
V^{(n)}_{\alpha,5} \deff \Delta(\tau(x_n), 2br_n).
\endmatrix$$ Fix $n_0$ sufficiently big. Denote $V_{\alpha,i} \deff V^{(n_0)}_{\alpha,i}$, $i=1,\ldots,5$. Then for every $n>\nobreak\!>1$ there exists a diffeomorphism $\psi_n: V_1 \to V_1$ with the following properties:
$\psi_n$ maps $V_{\alpha,i}$ onto $V^{(n)}_{\alpha,i}$ diffeomorphically;
$\psi_n: V^{(n)}_{\alpha,1} \to V^{(n)}_{\alpha,1}$ is the identity;
$\psi_n: V_{\alpha, 2} \to V^{(n)}_{\alpha,2}$ and $\psi_n:
V_{\alpha,3} \to V^{(n)}_{\alpha,3}$ are diffeomorphisms;
$\psi_n: V_{\alpha,3} \to V^{(n)}_{\alpha,3}$ is biholomorphic the complex structures, induced from $C_n$ by means of $\sigma^d_n$; and, finally
$\psi_n$ are $\tau$-invariant: $\tau \scirc\psi_n =\psi_n \scirc\tau$. Note that the last property is obtained due to the fact that the sets $V^{(n)} _{\alpha, i}$ are $\tau$-invariant. The remaining constructions are the same as in [*Subcase 1$'_b$)*]{}.
[*Case 2$_b$)*]{}. Consider the parameterizations $\sigma_n: V_\alpha \to
C_{\alpha,n}$. Without loss of generality we may assume that the complex structures $\sigma_n^* j_n\ogran_{V_\alpha}$ are constant near the boundary $dV_\alpha$ and converge to some complex structure. If we have the convergence of type [*C)*]{}, i.e., the strong convergence, there is nothing to do. Otherwise, there exists only a finite set of points $\{y^*_1,\ldots, y^*_l\}$ where the strong convergence fails. Changing the parameterizations $\sigma_n$, we may additionally assume that the structures $\sigma_n^* j_n \ogran
_{V_\alpha}$ are constant in the neighborhood of these points. Then we repeat the argumentations of [*Case 1$_b$)*]{}.
[*Case 3$_b$)*]{}. Fix identifications $C_{\alpha,n} \cong \cala_0$ such that every $C_{\alpha,n}^+$ is mapped onto $\cala_0^+$ and the induced parameterization maps $\sigma_{\alpha, n} : V_\alpha \to \cala_0$ are the same for all $n$ and $\tau$-invariant. Fix the standard representation of $\cala_0$ as the union of two discs $\Delta'$ and $\Delta''$ with identification of the centers $0\in \Delta'$ and $0\in \Delta''$ into the nodal point of $\cala_0$, still denoted by $0$. Let $\Delta' (x, r)$ denote the subdisc of $\Delta'$ with the center $x$ and the radius $r$.
Denote by $u'_n :\Delta' \to X$ and $u''_n :\Delta'' \to X$ the corresponding “components” of the maps $u_{\alpha,n} : C_{\alpha,n} \to X$. Find the common collection of bubbling points $y^*_i$ for both sequences of maps $u'_n
:\Delta' \to X$ and $u''_n :\Delta'' \to X$. If there are no bubbling points, then we obtain a convergence of type [*B)*]{}, and the proof can be finished by induction. Otherwise, consider the first such point $y^*_1$, which lies, say, on $\Delta'$. If $y^*_1$ is distinct from the nodal point $0 \in \Delta'$, then we simply repeat all the constructions in [*Case 1$_b$)*]{}.
It remains to consider the case $y^*_1=0 \in \Delta'$. The following modifications of the argumentations are needed. Repeat the construction of the radii $r_n\lrar 0$ and the points $x_n \lrar y^*_1=0$ from [*Case 1$_b$*]{}. Then $\{x_n \}$ is a sequence in the half-disk $\delta^{\prime+}
\deff \{ z\in \Delta' : \im z \ge0 \}$. Set $\ti x_n \deff \re(x_n)$, $\rho_n
\deff \im(x_n)$ and $R_n \deff |x_n|$. Thus, $x_n= \ti x_n +\isl\rho_n$, $R_n$ is the distance from $x_n$ to the point $0=y^*_1 \in \Delta'$, whereas $\rho_n$ is the distance from $x_n$ to the interval $]-1,1[ \subset \Delta'$, the set $\tau$-invariant points of $\Delta'$. Thus $\rho_n \le R_n$. Fix $\varrho>0$ such that the disc $\Delta'(0,\varrho)$ contains no bubble points $y^*_i\not=0 \in \Delta'$.
Depending on the behavior of the sequences $r_n$, $\rho_n$ and $R_n$, we consider four subcases. [*Subcase 3$'_b$): The sequence $\bigl\{{ R_n \over r_n }\bigr\}$ is bounded.*]{} Then the sequences $\bigl\{{ {\rho_n\over r_n} }\bigr\}$ and $\bigl
\{ {\ti x_n\over r_n} \bigr\}$ are also bounded. Passing to a subsequence, we may assume that the corresponding limits exist. Let $b$ be some upper bound for the sequence $\bigl\{{ R_n \over r_n }\bigr\}$. Consider the maps $\ti
v_n: \Delta(0, {\varrho \over 2r_n}-b) \to X$ defined by $\ti v_n(z) \deff
u_n(\ti x_n+ {z\over r_n})$. Then $\ti v_n$ are $\tau$-invariant, $\ti v_n
\scirc \tau= \ti v_n$, $\ti v_n$ converge to a nonconstant map $\ti v_\infty:
\cc \to X$ on compact subsets of $\cc$, and $\ti v_\infty$ extends to a map $\ti v_\infty: S^2 \to X$.
Since $\ti v_\infty$ is nonconstant, $\norm{ d\ti v _\infty} ^2 _{L^2( S^2)}
= \area( \ti v_\infty( S^2) )\ge 3\eps_b$. Choose $b>0$ in such a way that $$\norm{d\ti v_\infty}_{L^2(\Delta (0,b))}^2 \ge 2\eps_b
\eqno(A3.4.2)$$ and $b \ge 2\lim {R_n \over r_n}+2$. Due to [*Corollary A3.2.2*]{}, for $n>\!>1$ we obtain the estimate $$\norm{du'_n}_{L^2(\Delta' (\ti x_n, br_n))}^2 =
\norm{ d\ti v_n}_{L^2(\Delta (0,b))}^2 \ge \eps_b.
\eqno(A3.4.3)$$ Note that $0\in \Delta' (\ti x_n, (b-1)r_n))$ for $n>\!>1$ by the choice of $b$.
Define the coverings of $\cala_0$ by four sets $$\mathsurround=0pt
\matrix\format\l\ \ &\l\\
W^{(n)}_1 \deff \Delta'
\bs \barr\Delta' (0, {\textstyle{\varrho\over2}}),
&
W^{(n)}_2 \deff \Delta' (0, \varrho)
\bs \barr \Delta' (\ti x_n, br_n),
\cr
\noalign{\vskip4.5pt}
W^{(n)}_3 \deff \Delta' (\ti x_n, 2br_n)
\bs \barr\Delta'(0,{\textstyle {r_n \over 2}}),
&
W^{(n)}_4 \deff \Delta' (0, r_n) \cup
\Delta'',
\endmatrix$$ and lift them to $V_\alpha$ by putting $V^{(n)}_{\alpha,i} \deff \sigma\inv
_{\alpha,n}(W^{(n)}_i)$. Choose $n_0 >\!> 0$ such that $|x_n| < (b-1)r_n$ and the relation (A3.4.3) holds for all $n \ge n_0$. Set $V_{\alpha,i} \deff
%%%%%%%%%%%%%%%%%%%question%%%%%%%%%%%%%%%%%%%%%%%%%
V^{( n_0)} _{\alpha, i}$. Fix diffeomorphisms $\psi_n: V_\alpha \to V_\alpha$ such that $\psi_n: V_{\alpha,1} \to V^{(n)} _{\alpha,1}$ is the identity map, $\psi_n: V_{\alpha,2} \to V^{(n)}_{\alpha,2}$ and $\psi_n: V_{\alpha,3} \to
V^{(n)} _{\alpha,3}$ are diffeomorphisms, and $\psi_n: V_{\alpha,4} \to
V^{(n)} _{\alpha,4}$ correspond to isomorphisms of nodes $W^{(n)}_4 \cong
\cala_0$. Set $\sigma'_n \deff \sigma_n \scirc \psi_n$. The choice above can be made in such a way that the refined covering $\{ V_{\alpha,i} \}$ of $V_\alpha$ and parameterization maps $\sigma'_n: V_\alpha \to C_{\alpha,n}$ have the properties of [*Lemma A3.4.3*]{}. Relation (A3.4.3) implies the estimate $\area(u_n (\sigma'_n(V_{\alpha,i})) \le (N-1)\,\eps$. This provides the inductive conclusion for [*Subcase 3$'_b$)*]{}.
[*Subcase 3$''_b$): The sequence $\bigl\{{R_n \over r_n }\bigr\}$ increases infinitely but $\bigl\{{\rho_n \over r_n }\bigr\}$ remains bounded*]{}. Note that in this subcase we still have the relation $R_n \lrar
0$, or equivalently, $x_n \lrar 0$. On the other hand, $\lim {\rho_n \over
R_n} =0$. This implies that for $\ti R_n \deff |\ti x_n|$ we have $\lim
{\ti R_n \over R_n} =1$ since $R_n^2 = \ti R_n^2 + \rho_n^2$.
We proceed as follows. Define the maps $\ti v_n: \Delta(0, {\varrho \over
2r_n}-b) \to X$ setting $\ti v_n(z) \deff u'_n(\ti x_n + {z\over r_n})$. Then the $\ti v_n$ have the same properties as in [*Subcase 3$'_b$)*]{}. Choose $b>0$ obeying the relation (A3.4.2). Then for $n>\!>0$ the property (A3.4.3) follows.
For $n>\!>0$ define the coverings of $\cala_0$ by six sets $$\mathsurround=0pt
\matrix\format\l\ \ &\l\\
W^{(n)}_1 \deff \Delta'
\bs \barr\Delta' (0, {\textstyle{\varrho\over2}}),
&
W^{(n)}_2 \deff \Delta' (0, \varrho)
\bs \barr \Delta' (\ti x_n, 2\ti R_n),
\cr\noalign{\vskip5pt}
W^{(n)}_3 \deff \Delta' (\ti x_n, 4\ti R_n)
\bs \bigl(\barr \Delta' (\ti x_n, {\ti R_n \over 6})
\cup \barr \Delta' (0, {\ti R_n \over 6}) \bigr)
&
W^{(n)}_4 \deff \Delta' (0, {\ti R_n \over 3}) \cup \Delta'',
\cr\noalign{\vskip5pt}
W^{(n)}_5 \deff \Delta' (\ti x_n, {\textstyle{\ti R_n \over 3}})
\bs \barr\Delta'(\ti x_n, br_n),
&
W^{(n)}_6 \deff \Delta' (0, 2br_n ),
\endmatrix$$ and lift them to $V_\alpha$ by putting $V^{(n)}_{\alpha,i} \deff \sigma\inv
_{\alpha,n}(W^{(n)}_i)$. Choose $n_0 >\!> 0$ such that $R_{n_0} >\!> b
r_{n_0} $, and set $V_{\alpha,i} \deff V^{(n_0)} _{\alpha, i}$. Choose diffeomorphisms $\psi_n: V_\alpha \to V_\alpha$ such that $\psi_n: V
_{\alpha, 1} \to V^{(n)} _{\alpha,1}$ is the identity map, $\psi_n: V
_{\alpha, 2} \to V^{(n)}_{\alpha,2}$, $\psi_n: V_{\alpha,4} \to V^{(n)}
_{\alpha,4}$ and $\psi_n: V_{\alpha,5} \to V^{(n)}_{\alpha,5}$ are diffeomorphisms, and finally, $\psi_n: V_{\alpha,6} \to V^{(n)}_{\alpha,6}$ corresponds to isomorphisms of nodes $W^{(n)}_6 \cong \cala_0$. Set $\sigma'_n \deff \sigma_n \scirc \psi_n$. Note that the choices can be made in such a way that $\{ V_{\alpha,i} \}$ and parameterization maps $\sigma'_n:
V_\alpha \to C_{\alpha,n}$ have the properties of [*Lemma A3.4.3*]{}. As above, we get the estimate $\area(u_n (\sigma'_n(V_{\alpha,i} )) \le (N-1)\, \eps$ due to (A3.4.3). Thus we obtain the inductive conclusion for [*Subcase 3$''_b$)*]{} and can proceed further.
[*Subcase 3$'''_b$): The sequence $\bigl\{{\rho_n \over r_n} \bigr\}$ increases infinitely, but $\bigl\{{ R_n \over \rho_n} \bigr\}$ remains bounded*]{}. Then $\bigl\{{R_n \over r_n} \bigr\}$ also increases infinitely, but both sequences $\{R_n \}$ and $\bigl\{{ \rho_n }\bigr\}$ converge to 0. We may also assume that $\bigl\{{ \rho_n \over R_n }\bigr\}$ and $\bigl\{{ \ti x_n \over R_n }
\bigr\}$ also converge. Set $a_1 \deff \lim {\ti x_n \over R_n}$, $a_2 \deff
\lim{ \rho_n \over R_n}$, $a\deff a_1 + \isl a_2$ and $\barr a \deff a_1
- \isl a_2$. Note that $0 < a_2 \le 1$ and the involutions $\tau_n$ in $C_{\alpha, n}$ correspond to the complex conjugation $z\to \barr z$ in $\Delta'$. In particular, $\barr x_n = \tau_n(x_n)$.
Consider maps $v_n: \Delta(0, {\varrho \over 2r_n}) \to X$ defined by $v_n(z)
\deff u'_n(x_n + {z\over r_n})$. Then the sequence $\{ v_n \}$ converges on compact subsets to a nonconstant map which extends to the map $v_\infty :S^2
\to X$. Moreover, we can fix sufficiently big $b>0$ such that for $n>\!>0$ we get the property (A3.4.3).
For $n>\!>0$ define the coverings of $\cala_0$ by eight sets $$\mathsurround=0pt
\matrix\format\l\ \ \ \ \ &\l\\
W^{(n)}_1 \deff \Delta'
\bs \barr\Delta' (0, {\textstyle{\varrho\over2}}),
&
W^{(n)}_2 \deff \Delta' (0, \varrho)
\bs \barr \Delta' (0, 2R_n),
\cr\noalign{\vskip6pt}
\rlap{$W^{(n)}_3 \deff \Delta' (0, 4R_n)
\bs \bigl(\, \barr \Delta' ( a\, R_n , {a_2 R_n \over 4})
\cup \barr \Delta' ( \barr a\, R_n , { a_2 R_n \over 4})
\cup \barr \Delta' (0, {a_2 R_n \over 4})
\,\bigr)\qquad$}
&
\cr\noalign{\vskip6pt}
W^{(n)}_4 \deff \Delta' (0, {a_2 R_n \over 3}) \cup \Delta''.
\cr\noalign{\vskip5pt}
W^{(n)}_5 \deff \Delta' (a R_n, {a_2 R_n \over 3})
\bs \barr\Delta'(x_n, br_n),
&
W^{(n)}_6 \deff \Delta' (x_n, 2br_n ),
\cr\noalign{\vskip5pt}
W^{(n)}_7 \deff \Delta' (\barr a R_n, {a_2 R_n \over 3})
\bs \barr\Delta'(\barr x_n, br_n),
&
W^{(n)}_8 \deff \Delta' (\barr x_n, 2br_n ),
\endmatrix$$ and lift them to $V_\alpha$ by putting $V^{(n)}_{\alpha,i} \deff \sigma\inv
_{\alpha,n}(W^{(n)}_i)$. Fix sufficiently big $n_0 >\!> 0$,
and set $V_{\alpha, i} \deff V^{(n_0)} _{\alpha, i}$. Choose diffeomorphisms $\psi_n:
V_\alpha \to V_\alpha$ mapping $V _{\alpha, i}$ diffeomorphically onto $V^{(n)} _{\alpha, i}$ such that the assertions of [*Lemma A3.4.3*]{} are fulfilled. As above, we obtain the estimate $\area(u_n (\sigma'_n(V_{\alpha,i} )) \le (N-1)\, \eps$. This gives the inductive conclusion for [*Subcase 3$'''_b$)*]{}.
[*Subcase 3$''''_b$): The sequences $\bigl\{{\rho_n \over r_n} \bigr\}$ and $\bigl\{{ R_n \over \rho_n} \bigr\}$ increase infinitely*]{}. Thus $\lim {\ti R_n \over R_n} =1$. We consider the sequence of maps $\{ v_n \}$. It is defined in the same way as in the previous subcase and has the same properties. In particular, $\{ v_n \}$ converges to the map $v_\infty :S^2 \to X$ and there exists a sufficiently big $b>0$ such that for $n>\!>0$ we get the property (A3.4.3).
For $n>\!>0$ define the coverings of $\cala_0$ by ten sets $$\mathsurround=0pt
\matrix\format\l\ \ \ \ \ &\l\\
W^{(n)}_1 \deff \Delta'
\bs \barr\Delta' (0, {\textstyle{\varrho\over2}}),
&
W^{(n)}_2 \deff \Delta' (0, \varrho)
\bs \barr \Delta' (0, 2R_n),
\cr\noalign{\vskip6pt}
\rlap{$W^{(n)}_3 \deff \Delta' (0, 4R_n)
\bs \bigl(\, \barr \Delta' ( 0, {\ti R_n \over 4})
\cup \barr \Delta' ( \ti x_n, {\ti R_n \over 4})
\bigr),$}
&
\cr\noalign{\vskip6pt}
W^{(n)}_4 \deff \Delta' (0, {\ti R_n \over 3}) \cup \Delta'',
&
W^{(n)}_5 \deff \Delta' (\ti x_n, {\ti R_n \over 3})
\bs \barr\Delta'(\ti x_n, 2\rho_n),
\cr\noalign{\vskip6pt}
\rlap{$W^{(n)}_6 \deff \Delta' (\ti x_n, 4\rho_n )
\bs \bigl(\, \barr \Delta' ( x_n, {\rho_n \over 4})
\cup \barr \Delta' ( \barr x_n, {\rho_n \over 4})
\bigr), $}
\cr\noalign{\vskip5pt}
W^{(n)}_7 \deff \Delta' (x_n, {\rho_n \over 3})
\bs \barr\Delta'(x_n, br_n),
&
W^{(n)}_8 \deff \Delta' (x_n, 2br_n ),
\cr\noalign{\vskip5pt}
W^{(n)}_9 \deff \Delta' (\barr x_n, {\rho_n \over 3})
\bs \barr\Delta'(\barr x_n, br_n),
&
W^{(n)}_{10} \deff \Delta' (\barr x_n, 2br_n ).
\endmatrix$$ The remaining “manipulations” with $W^{(n)}_i$ are the same as in the previous subcases. As a result, we obtain the covering of $V_\alpha$ by sets $V_{\alpha,
i} \deff \sigma\inv _{\alpha, n_0}(W^{(n_0)}_i)$ with an appropriate $n_0 >\!>
0$ and refined parameterizations $\sigma'_n: V_\alpha \to C_{\alpha, n}$, for which the assertions of [*Lemma A3.4.3*]{} are fulfilled. As above, we obtain the estimate $\area(u_n (\sigma'_n(V_{\alpha,i} ))
\le (N-1)\, \eps$. This gives the inductive conclusion for [*Subcase 3$''''_b$)*]{}.
[*Case 4$_b$): $V_\alpha$ is a cylinder such that conformal radii of $(V_\alpha, \sigma_n^*j_n)$ increase infinitely*]{}. We can simply repeat the contructions made in [*Case 4)*]{} from the proof of [*Theorem 1.1*]{}. Additional attention is needed to preserve the $\tau$-invariantness.
The proof of the theorem can now be finished by induction.
Remark. Here we give some explanation of the geometric meaning of the constructions of the proof of [*Theorem 5.9*]{} and describe the picture of the bubbling. We restrict ourselves to [*Case 3$_b$)*]{} as the most complicated one; the constructions of the other cases can be treated similarly. The reflection principle allows us to reduce [*Case 3$_b$)*]{} to the consideration of $\tau$-invariant maps $u^d_n: \cala_0 \to X$ from the standard node which are $J_n$-holomorphic on $\cala^+_0$. The situation is different from those in [*Theorem 1.1*]{}, where the bubbling appears in the nodal point. In this case we must take into consideration not only the parameters $r_n$ describing the size of energy localization of the bubbled sphere, but also the additional parameters $R_n$ and $\rho_n$. These describe the position of the localization centers $x_n$ the nodal point and the set of $\tau$-invariant points of $\cala_0$. Depending on the behavior of $r_n$ $\rho_n$ and $R_n$, we can have four different types of bubbling and corrsponding [*Subcases 3$'_b$)–3$''''_b$)*]{}.
=12.3pt plus 1.5pt Turning back from a “doubled” description by $\tau$-invariant objects to the original maps $u_n: \cala^+_0 \to X$ with a totally real boundary condition, we obtain the following picture. Since every covering piece $W^{(n)}_i$ is $\tau$-invariant, for $\cala_0^+$ we obtain the covering piece $W^{(n)\,+}_i
\deff W^{(n)}_i \cap \cala_0^+$. Thus we obtain a bubbled disk represented by $W^{(n)\,+}_3$ instead of the bubbled sphere represented by $W^{(n)}_3$, the sequence of long strips $W^{(n)\,+}_2$ instead of the sequence of long cylinder $W^{(n)\,+}_2$ and so on.
However, because of the presence of the nodal point (piece $W^{(n)}_4$ on the figure), this “part inbetween” is topologically not a cylinder (an annulus) but pants. Furthermore, the complex structures on the pants are not constant. To get pants with a constant structure (piece $W^{(n)}_3$), we cut off the annuli $W^{(n)}_2$ and $W^{(n)}_5$. Since $\lim R_n =0 = \lim {r_n\over
R_n}$, the conformal radii of these annuli increase infinitely. This shows that $W^{(n)}_2$ and $W^{(n)}_5$ are sequences of long cylinders and that the sequence $W^{(n)}_3$ defines in the limit a sphere with three nodal points.
As in [*Subcase 3$'_b$)*]{} every covering piece $W^{(n)}_i$ is $\tau$-invariant, whereas $W^{(n)\,+}_i \deff W^{(n)}_i \cap \cala_0^+$ is the “half” of $W^{(n)}_n$. Thus, for a sequence of undoubled maps $u_n:
\cala^+_0 \to X$ we obtain the following bubbling picture. The limit contains two bubbled disks, represented by $W^{(n)\,+}_6$ and $W^{(n)\,+}_3$, a boundary node $W^{(n)\,+}_4$ and possibly further bubbled pieces which can appear in the limit of long strips $W^{(n)\,+}_2$ and $W^{(n)\,+}_5$. Note also that the action of the involution $\tau$ on the pants $W^{(n)}_3$ is described in Fig. 12b).
In [*Subcase 3$'''_b$)*]{} the bubbling takes place near but not at the boundary. Indeed, since ${\rho_n \over r_n}\lrar \infty$, the bubbled sphere which appears as the limit of the sequence $\{v_n \}$ is not $\tau$-invariant. To see this phenomenon, we note that for any fixed $b>0$ the covering pieces $W^{(n)}_5= \Delta' (x_n,2br_n)$ representing a sufficient big part of this sphere lie in $\cala^+_0$ for $n>\!>0$. This implies that the sequence $v_n
\scirc \tau$ converges to another bubbled sphere, which is $\tau$-symmetric to the first and represented by $W^{(n)}_7$.
Another bubbled sphere, represented by $W^{(n)}_3$, appears from pants between the first two spheres and the disk $\Delta'$. Since $\{ {R_n
\over \rho_n}\}$ remains bounded, the original nodal point remains on this latter sphere.
[A3.5. Attaching an Analytic Disk to a Lagrangian Submanifold of $\cc^n$]{}.
As in the first lecture we consider $\cc^n \approx \rr^{2n}$ together with some symplectic form $\omega $ taming the standart complex structure $J\st $. An $n$-dimensional submanifold $W\subset \cc^n$ is called $\omega $-Lagrangian if $\omega\mid_W\equiv 0$.
**Exercise 1. Prove that every Lagrangian submanifold of $\cc^n$ is totally real.**
[**2.**]{} Prove that for a Lagrangian manifold $W\subset \cc^n$ and the unit circle $\ss^1\subset \cc $ the manifold $\ss^1\times W$ is Lagrangian in $\cc^{n+1}$ with respect to $\hat\omega = {i\over 2}dz_1\wedge d\bar z_1 +
\omega $.
A holomorphic map $u:\Delta \to \cc^n$ “sufficiently smooth” up to the boundary and such that $u(\d \Delta )\subset W$ we shall call *an analytic disk attached to $W$.*
Our goal in this paragraph is to prove the following theorem of Gromov:
Theorem A3.5.1. [*Let $W$ be a compact Lagrangian submanifold of $\cc^n$. Then there exists a non-constant analytic disk attached to $W$.* ]{}
We shall closely follow the exposition of H. Alexander, \[Al\]. Fix some point $w_0\in W$ and denote by $u_0(z)\equiv w_0$ the constant holomorphic map. Fix $p>2$ and consider the Banach manifold $L^{2,p}(\Delta , \d \Delta , 1; \cc^n , W, w_0)$ of $L^{2,p}$-maps from $\bar\Delta $ to $\cc^n$ which map $\d \Delta $ to $W$ and $1$ to $w_0$, and which are homotopic to the constant map $u_0\equiv w_0$ as $(\Delta , \d \Delta , 1)\to (\cc^n , W, w_0)$ mappings. Note that due to the Sobolev imbedding $L^{2,p}\subset C^{1,1-{2\over p}}$ our mappings are smooth up to the boundary.
Take $u\in L^{2,p}(\Delta , \d \Delta , 1; \cc^n , W, w_0)$. Denote by $E$ as usually the pull-back by $u$ of the tangent bundle of $\cc^n$. In fact $E=\Delta\times \cc^n \to \Delta $, the trivial bundle over $\Delta $. Denote by $F$ the pull-back $u^*TW$ of the tangent to $W$ bundle. $F$ is a totally real subbundle of $E$ of real dimension $n$. The tangent space to $L^{2,p}(\Delta , \d \Delta , 1; \cc^n , W, w_0)$ at $u$ is $T_uL^{2,p}(\Delta , \d \Delta , 1; \cc^n , W, w_0) =
\{ h\in L^{2,p}(\Delta ,\cc^n): h\mid_{\d \Delta }\in F, h(1)=0\} $. In the cartesian product $L^{2,p}(\Delta , \d \Delta , 1; \cc^n , W, w_0)
\times L^{1,p}(\Delta ,\cc^n)$ consider the submanifold $\cale =\{ (u,v):
\dbar u=v\} $ with the natural projection $\pi :\cale \to L^{1,p}(\Delta ,\cc^n)$.
**Exercise 1. In $\rr^{2n+2}$ consider the operator given by the matrix**
$$J_v = \left(
\matrix
0 & -1 & ...... & v_2 & -v_1 \cr
1 & 0 & ...... & -v_1 & -v_2 \cr
. & . & ....... & ...... & ... \cr
0 & 0 & ....... & v_{2n} & -v_{2n-1} \cr
0 & 0 & ....... & -v_{2n-1} & -v_{2n} \cr
0 & 0 & ....... & 0 & -1 \cr
0 & 0 & ....... & 1 & 0 \cr
\endmatrix
\right)$$ Prove that $J_v$ defines an almost complex structure in $\rr^{2n+2}$, which for every $z\in \rr^2$ on the vertical slice $\{ z\} \times \rr^{2n}$ coincides with the standart structure $J\st$ of $\cc^n$.
[**2.**]{} Prove that the equation $\dbar u=v$ for a $C^1$-map $u:\cc \to \cc^n$ is equivalent to the $J_v$-holomorphicity of the section $\hat u :z\to (z,u(z))$ of the fibration $(\rr^{2n+2},J_v)\to (\rr^2, J\st)$, i.e. to the equation
$${\d (z,u)\over \d x} + J_v(z,u)[{\d (z,u)\over \d y}] = 0.$$
Now we shall prove the following alternative:
Lemma A3.5.2. [*If there is no nonconstant analytic disk $u\in L^{2,p}(\Delta , \d \Delta , 1;$ $ \cc^n , W, w_0)$, then the projection $\pi :\cale \to L^{1,p}(\Delta ,\cc^n)$ is surjective.* ]{} Proof.
Suppose that a nonconstant analytic disk $u\in L^{2,p}(\Delta , \d \Delta , 1; \cc^n , W, w_0)$ doesn’t exists. We are going to prove that in this case $\pi $ is surjective.
[*Step 1. *$\pi $ is a proper mapping, i.e. for the converging sequence $v_k\to v_0$ in $L^{1,p}(\Delta ,\cc^n)$ and for the sequence $u_k$ with $(u_k,v_k)\in \cale $, there is a converging subsequence $u_{k_n}$.** ]{}
Note that $\dbar u_k=v_k$. According to the Exercise above this means that the sections $\hat u_k:=(z,u_k)$ are $J_{v_k}$-holomorphic with $J_{v_k}$ converging to $J_{v_0}$ in $C^0$-sence. Note also that boundaries of our disks are on the Lagrangian (and thus totally real) submanifold $\hat W:=\ss^1\times W$ and they are homotopic to each other. From here we see that
$$\area(\hat u_k(\Delta )) = \int_{\hat u_k(\Delta )}\hat\omega =
\int_{\hat u_k(\d \Delta )}\d \lambda ,$$ where $\lambda $ is some primitive of $\hat\omega $. The second integral does not depend on the homology class of $\hat u_k(\Delta )$ in $\sfh_1(\hat W,
\rr)$, because $\hat\omega\mid_W = \d \lambda\mid_{\hat W} \equiv 0$ ($\hat W$ is $\hat\omega $-Lagrangian!).
So by the Theorem A3.4.1 either the limit of some subequence, still denoted as $u_k(\Delta )$, contains a nonconstant complex sphere (this is impossible in $\cc^n$), or the limit of $u_k(\Delta )$ contains some nonconstant analytic disk with boundary on $W$ (this is prohibited by our assumption), or $u_k$ $C^1$-converge.
[*Step 2. *$d\pi_{(u,v)} : T_{(u,v)}\cale \to L^{1,p}(\Delta ,\cc^n)$ is Fredholm of index zero for every $u\in \cale_v :=\pi^{-1}(v)$.** ]{} The fact that the boundary value problem $\dbar h=v, h\mid_{\d \Delta }\in F$ is Fredholm is classical, see \[Ga\]. Our manifold $\cale $ is connected, so the index of $d\pi_{(u,v)}$ doesn’t depend on $(u,v)$ and can be calculated at $(u_0,0)\in \cale $ where $d\pi $ is a bijection. Therefore $ind(d\pi )$ is everywhere zero.
One says that for the smooth mapping $\pi :\cale \to L^{1,p}(\Delta ,\cc^n)$ the point $(u,v)$ is regular if $d\pi_{(u,v)}$ is surjective. $v$ is a regular value is it is not an image of a nonregular point.
[*Step 3. (Smale’s theorem). *Let $\pi :\cale \to \calm $ be a proper Fredholm map (i.e. $d\pi_x $ is Fredholm for all points $x\in \cale $. Then the set of regular values is dence in $\calm $. Moreover, for every regular value $v\in \calm $ the set $\cale_v:=\pi^{-1}(v)$ is a manifold of dimension equal to the $\ind(d\pi_x) $ at $x\in \cale_v$. Moreover, for any two regular values $v_1$ and $v_2$ the manifolds $\cale_{v_1}$ and $\cale_{v_2}$ are cobordant.** ]{} See \[Sm\].
In our case $0$ is a regular value, so for a dence subset of $v$’s $\cale_{v}$ is cobordant to a point, therefore is a point intself. Properness of $\pi $ implies now that $\cale_v$ is allways a point.
It is not difficult to show that $\pi $ cannot be surjective. This will imply the existence of non constant analytic disk attached to $W$.
Lemma A3.5.3. [*The projection $\pi :\cale \to L^{1,p}(\Delta ,\cc^n)$ is not surjective.* ]{} Proof. [Othervice, for $v^C=(C,0,....,0)$ find $u^C=(u_1^C,...,u_n^C)$ with $\dbar u^C = v^C$. Therefore $\bar \d u_1^C=C$. This implies that $u_1^C=C\bar z - h^C$, where $h^C$ is a holomorphic function on $\Delta $. Since $u^C(\d \Delta )\subset W$ the family $u^C$ is uniformly bounded on $\d \Delta $ by a constant $k$ independent of $C$. Therefore $\vert \bar z -
h^C(z)/C\vert \le k/C$ for $z\in \d \Delta $. Since $\bar z - h^C(z)/C$ is harmonic on $\Delta $, the bound holds for all $z\in \Delta $. This implies that $\bar z$ can be uniformly approximated on $\Delta $ by holomorphic functions. Contradiction. ]{}
[Chapter III. Global Properties and Moduli Spaces.]{}
Let $(X, J)$ be an almost-complex manifold of complex dimension $n$. An almost complex structure is always assumed to have smoothness of class $C^1$. Further, let $(S, J_S)$ denote a Riemann surface with complex structure $J_S$.
In Lecture 7 we show that for a $J$-complex curve $u:(S,J_S) \to (X,J_0)$ the pulled-back bundle $E\deff u^*TX$ possesses a natural [*holomorphic*]{} structure (the corresponding sheaf of holomorphic sections will be denoted as $\calo(E)$) such that the differential $du:TS\to E$ is a [*holomorphic*]{} homomorphism. This allows us to define the order of vanishing of the differential $du$ at point $s\in S$. We denote this number by ${\sl ord}_sdu$.
This also gives the following short, exact sequence: $$0\longrightarrow \calo(TS) \buildrel du \over\longrightarrow
\calo(E) \buildrel \pr \over\longrightarrow
\calo(N_0)\oplus \caln_1 \longrightarrow 0,\eqno(3.1)$$ where $\calo(N_0)$ denotes a free part of the quotient $\calo(E)/du(\calo(TS))$, and $\caln_1$ is supported on a finite set of cusps of $u$ ([**]{} points of vanishing of $du$).
On the Sobolev space $L^{1,p}(S,N_0)$ of $L^{1,p}$-smooth sections of the bundle $N_0$ the natural Gromov operator $D_N: L^{1,p}(S,N_0) \to
L^p(S,\Lambda^{0,1}S \otimes N_0)$ is defined. Put $\sfh^0_D(S, N_0):=
\ker D_N$ and $\sfh^1_D(S, N_0):=\coker D_N$.
We prove the following Theorem 3.1.
*Let $u: (S, J_S) \to (X, J_0)$ be a nonconstant irreducible and non- multiply covered holomorphic map such that $\sfh^1_D(S, N_0)=0$. Then*
in a neighborhood of $M\deff u(S)$ the Moduli space of nonparameterized $J_0$-holomorphic curves $\calm_{[\gamma], g, J_0}$ is a manifold whose tangent space is $T_M\calm_{[\gamma], g, J_0} =
\sfh^0_D(S, N_0) \oplus \sfh^0(S, \caln_1)$;
further, there is a neighborhood $V\ni J$ in the Banach manifold $\calj$ of $C^1$-smooth almost-complex structures on $X$ and a neighborhood $W$ of $M$ in $\calm_{[\gamma], g, V}:=\bigcup_{J\in V}
\calm_{[\gamma], g, J}$ such that the natural projection $\pr_\calj:
W\to V$ is a trivial Banach bundle;
if $\dim_{\rr }X=4$ and $c_1(E)[M] > \sum_{s\in S}{\sl ord}_sdu$, then $\sfh^1_D(S, N_0)=0$; thus, the conclusions and hold.
Let us explain our second result in this chapter, which will be strongly relied upon in Chapter IV and also in the proof of the Gromov non-squeezing theorem. Denote by $\calj $ some Banach manifold of $C^k$-smooth almost-complex structures on a manifold $X$, $k\ge 1$. In our applications $\calj $ can be
1\) the manifold $\calj_{\omega }$ of all a.-c. structures $J$ on $X$, which are tamed by some fixed symplectic form $\omega $; i.e., $\omega (u,Ju)>0$ for all $u\in TX\setminus \{ 0\} $.
2\) the manifold $\calj_U$ of all a.-c. structures $J$ on a complex Kähler manifold $(X,J\st ,\omega )$ tamed by $\omega $ and which are different from $J\st $ only on some relatively compact subset of $U$, i.e., $\{ x\in X:J_x\not= J\st \} \comp U$.
Denote by $\calm =\calm_{\calj ,[\gamma ],\Sigma }$ the (topological) space of all $J$-complex curves in $X$ ($J$ runs over all structures in $\calj $) representing some fixed homology class $[\gamma ]\in \sfh_2(X,\zz )$, and parameterized by some fixed compact surface $\Sigma $, see Chapter II. For $\calj $, as in 1) and 2), above the space $\calm $ is a manifold in all its non- multiply covered points, see [*Lemma 8.2.2*]{}.
Also let $h:[0,1]\to \calj $ be some smooth map (more generally one can consider a smooth map $h:Y\to \calj $, where $Y$ is a compact real manifold with boundary). By $\calm_h$ we denote the subset of $\{ (u,J_S,t)
\in \cals \times \ttt_{\Sigma }\times [0,1]\} $ which consists of $h(t)$-complex curves $u:(S,J_S)\to (X,h(t))$ modulo the natural action of the group $G$ of reparameterizations, see §7.4 and Definition 8.3.1. Roughly speaking, $\calm_h$ is a moduli space of $h(t)$-complex curves in $X$, $t\in [0,1]$. This is a closed subset of $\calm $. For a point $(M_0,t_0)$ in $\calm_h$ (for a $h(t_0)$-complex curve $M_0$) one can define a component $\calm_h(M_0,t_0)$ of $\calm_h$ containing $(M_0,t_0)$ as was shown in [*Definition 8.3.2*]{}.
Theorem 3.2. [*If the component $\calm_h(M_0,t_0)$ is not compact, then there exists a continuous path $\beta :[0,1]\to \calm_h(M_0,t_0)$, $\beta (t) =(M_t,J_t)$ starting at $(M_0,t_0)$ such that $\beta (t_n)$ is not bounded in $\calm_h(M_0,t_0)$ for some sequence $t_n\to 1$, but $J_{t_n}$ converge to some $J^*\in \calj $.* ]{}
This means, in fact, that in the limit the sequence $M_{t_n}$ breaks into several irreducible $J^*$-complex components.
Let us consider an example where such behavior is not possible.
Example.
Suppose that $X=S^2\times Y$ and $\omega =\omega_1
+ \omega_2$, where $\omega_1$ is some positive form on $S^2$ and $\omega_2$ is some symplectic form on a compact manifold $Y$. As a homology class $[\gamma ]$ take $S^2\times \{ pt\} $ and as a parameterizing surface for our curves again $\Sigma :=S^2$.
Observe that for any structure $J$ on $X$ tamed by $\omega $ no $J$-complex curve $C\in [\gamma ]$ can be decomposed as $C=C_1\cup C_2$ for some $J_k$-complex curves $C_k$, where $J_k\in \calj_{\omega }$!
Exercise. [Prove this.]{}
In this example $\calm_h(M_0,t_0)$ will be always be compact.
Theorem 3.3.
*Under the above conditions suppose that $S$ is a sphere $S^2$ and that $\calm_h(M_0,t_0)$ is compact. Then the path $h$ can be $C^1$-approximated by the paths $h_n$, all starting at $(M_0,t_0)$ such that for all points $(M,t)\in \calm^0_{h_n}$ the associated $D_N$-operator is surjective, i.e., $\sfh^1_{D_N}(S^2,N_{M_t})=0$.*
Moreover, $\calm_{h_n}$ is a trivial bordism: $\calm_{h_n(0)}\times [0,1]$. In particular for every $h_n(0)$-holomorphic sphere $M_0\in \calm_{h_n(0)}$ there exists a continuous family of $h_n(t)$-holomorphic spheres $M_{n,t}=
u_{n,t}(S^2)$ with $M_{n,0}=M_0$.
[Lecture 6]{}
[First Variation of the $\dbar_J$-equation.]{}
In this paragraph we want to introduce, following Gromov, a $D_{u,J}$-operator associated to the $J$ - complex curve $u : S \to X$ in an almost complex manifold $(X,J)$. For the convenience of the reader we now recall the notion of a connection.
*6.1. Symmetric Connections.*
Let $M$ be a real manifold, $E$ a real vector bundle over $M$. By $\Gamma (E)$ we denote the space of smooth sections of $E$ and by $\Lambda^p$, a space of smooth $p$-forms on $M$.
Definition 6.1.1.
*A connection on the real vector bundle $E\to M$ is a mapping $\nabla :\Gamma (E) \to \Lambda^1\otimes \gamma (E)$ such that*
$\nabla 1$. $\nabla $ is $\rr $-linear;
$\nabla 2$. for any smooth function $f$ on $M$ and any section $e\in \Gamma (E)$
$$\nabla (f\cdot e) = df\otimes e + f\cdot \nabla e.\eqno(6.1.1)$$
If $X\in \Gamma (TM)$ is a vector field on $M$, one denotes $\nabla_Xe :=
\nabla (e)(X)$, a covariant derivative of $e$ along $X$.
Remarks. 1. Choose a local frame $e_1,...,e_n$ of $E$ and write $\nabla (e_i) =
\Gamma_i^j\otimes e_j$, where $\Gamma_i^j$ are some $1$-forms. Then $\nabla $ is uniquely determined by $\{ \Gamma_i^j\}_{i,j=1}^n$.
2\. The dual connection $\nabla^*$ is defined on the dual bundle $E^*\to M$ by $$d\< h,e\> = \<\nabla^*h,e\> + \<h, \nabla e\>, \eqno(6.1.2)$$ where $h\in \Gamma (E^*)$.
3\. One defines an extension of $\nabla $ onto $E^{\otimes p}\otimes
E^{\otimes q}\to M$ as $$\nabla (e\otimes h) = \nabla e\otimes h + e\otimes \nabla h.
\eqno(6.1.3)$$ In particular, since $\hom (E,E) \equiv E\otimes E^*$, $\nabla $ extends onto $\hom (E,E)$.
We shall usually denote all these extensions by the same symbol $\nabla$. One remarks, of course, that (6.1.1), (6.1.2) and (6.1.3) are nothing but the Leibnitz formula.
4\. Mapping $\nabla^2 = \nabla\scirc \nabla :\Gamma (E) \to \Lambda^2\otimes
\Gamma (E)$ is called the curvature of $\Gamma $ and is, in fact, linear over $\cala (M)$: $\nabla^2(fe)=\nabla (df\otimes e+f\nabla e)=
d^2f\otimes e-df\wedge \nabla e+df\wedge \nabla e+f\nabla^2e= f\nabla^2e$.
Definition 6.1.2. [*Let $E=TM$. A connection $\nabla $ is called symmetric if $\nabla_XY-\nabla_YX = [X,Y]$.* ]{}
Exercise. [Take $e_i={\d \over \d x_i}$ for some local coordinates $\{ x_i\} $. Write $\Gamma_i^j=\gamma_{ik}^jdx^k$, i.e., $\nabla_{e_k}e_i =
\Gamma_{ik}^je_j$. What does symmetricity of $\nabla $ mean in terms of $\{
\Gamma_{ik}^j\} $? ]{}
Exercise.
Choose a symmetric connection $\nabla $. Then the Nijenhuis tensor of an almost-complex structure $J$ is defined as
$$4N_J(X,Y) = (\nabla_XJ\scirc J)Y - (\nabla_YJ\scirc J)X + (\nabla_{JX}J)Y -
(\nabla_{JY}J)X.$$ Prove that this defintion doesn’t depend on the choice of a symmetric connection, i.e. express $N_J(X,Y)$ in the terms of Lee brackets.
*6.2. Definition of the $D_{u, J}$-operator.*
Recall that a $C^1$ map $u:S\to X$ from a Riemann surface $S$ with a complex structure $J_S$ is called holomorphic with respect to $J_S$ and $J$ if it satisfies the equation
$$du + J\scirc du\scirc J_S = 0. \eqno(6.2.1)$$
This simply means that $du\scirc J_S = J\scirc du$, [*i.e.,*]{} $du:T_pS \to
T_{u(p)}X$ is a [*complex*]{} linear map for every $p\in S$. Equation (6.2.1) is an elliptic quasi-linear PDE of order one. We are interested in the behavior of the solutions of (6.2.1), in particular, when the structures $J$ and $J_S$ change. So we need to choose appropriate functional spaces both for solutions and for the coefficient of (6.2.1). Our choice is based on the following facts:
[*a)*]{} The minimal reasonable smoothness of an almost complex structure $J$ on $X$, for which the Gromov operator $D_{u, J}$ can be defined, is $C^1$, see the explicit formula (6.2.5).
[*b)*]{} It is more convenient to operate with Banach spaces and manifolds and thus with finite regularity like $C^k$, $C^{k,\alpha}$ or $L^{k,p}$ than with $C^\infty$-smoothness defining only a Frechet-type topology.
[*c)*]{} Equation (6.2.1) is defined also for $u$ lying in Sobolev-type spaces $L^{k,p}(S, X)$ with $k\ge1$, $1\le p\le\infty$ and $kp>2$; such solutions are $C^1$-smooth and the topology on the space of a solution is, in fact, independent of the particular choice of such a Sobolev space.
[*d)*]{} For $J\in C^k$ the coefficients of the Gromov operator $D_{u, J}$ are, in general, only $C^{k-1}$-continuous, see (6.2.5), and hence solutions of the “tangential equation” $D_{u, J}v=0$ are only $L^{k,p}$-smooth, $1\le
p<\infty$; thus for obtaining a [*smooth*]{} structure on a space of (parameterized) $J$-complex curves, one should use Sobolev spaces $L^{k', p}(S, X)$ with $k'\le k$.
Fix a compact Riemann surface $S$, [*i.e.,*]{} a compact, connected, oriented, smooth manifold of real dimensions 2. Recall that the Sobolev space $L^{k, p}(S, X)$, $kp>2$, consists of those continuous maps $u: S \to X$, which are represented by $L^{k, p}$-functions in local coordinates on $X$ and $S$. This is a Banach manifold, and the tangent space $T_uL^{k, p}(S, X)$ to $L^{k, p}(S, X)$ in $u$ is the space $L^{k,p }(S, u^*(TX))$ of all $L^{k,p}$-sections of the pull-back under $u$ of the tangent bundle $TX$. One has the Sobolev imbeddings $$\matrix
L^{k,p}(S, X)& \hookrightarrow & L^{k-1, q}(S, X),&
\quad & \hbox{for $1\le p<2$} & \hbox{and} & 1\le q\le {2p\over2-p},
\cr
L^{k, p}(S, X)& \hookrightarrow & C^{k-1, \alpha}(S, X),&
\quad & \hbox{for $2<p\le\infty$}& \hbox{and} & 0\le\alpha\le 1-{2\over p}.
\cr
\endmatrix$$
Let $[\gamma ]$ be some homology class in $\sfh_2(X, \zz)$. Fix $p$ with $2<p<\infty$ and consider the Banach manifold $$\cals = \{u\in L^{1,p}(S, X):u(S)\in [\gamma]\}$$ of all $L^{1,p }$-smooth mappings from $S$ to $X$, representing the class $[\gamma]$. This makes sence, because $L^{1,p}\subset C^{0,1-{2\over p}}$, and therefore all $u$ from $\cals $ are continuous. Remark that the tangent space to $\cals $ at $u$ is
$$T_u\cals = L^{1,p}(S,u^*(TX))$$ the space of $L^{1,p}$-sections of the pulled-back by $u$ the tangent bundle of $X$.
Denote by $\calj$ the Banach manifold of $C^1$-smooth almost complex structures on $X$. In other words, $\calj = \{ J\in C^1
(X, \endo (TX)):J^2=-\id \} $. The tangent space to $\calj$ at $J$ consists of $C^1$-smooth $J$-antilinear endomorphisms of $TX$, $$T_J\calj = \{ I\in C^1(X, \endo (TX)) : JI + IJ = 0\}
\equiv C^1(X, \Lambda^{0,1}X \otimes TX),$$ where $\Lambda^{0,1}X$ denote the complex bundle of $(0,1)$-form on $X$.
Denote by $\calj_S$ the Banach manifold of $C^1$-smooth complex structures on $S$. Thus, $\calj_S = \{ J_S\in C^1(S, \endo (TS)):J_S^2=-\id \}$ and the tangent space to $\calj_S$ at $J_S$ is $$T_{J_S}\calj = \{ I\in C^1(S, \endo (TS)) : J_S I + IJ_S = 0\}
\equiv C^1(S, \Lambda^{0,1}S\otimes TS).$$
Consider also the subset $\calp\subset \cals\times \calj_S \times
\calj$ consisting of all triples $(u, J_S, J)$ with $u$ being $(J_S,J)$-holomorphic, [*i.e.,*]{}
$$\calp = \{(u, J_S, J)\in \cals\times \calj_S\times \calj: du +
J\scirc du\scirc J_S = 0 \}. \eqno(6.2.2)$$
Lemma 6.2.1. *Let $J$ and $J_S$ be continuous almost-complex structures on $X$ and $S$, respectively, and let $u\in L^{1,p}(S, X)$. Then $$du + J\scirc du\scirc J_S \in L^p(S,\Lambda^{0, 1}S \otimes u^*(TX)).$$*
Proof. One can easily see that $du \in L^p(S,\hom_\rr(TS, u^*(TX))$. On the other hand, $$(du + J\scirc du\scirc J_S ) \scirc J_S= -J \scirc
(du + J\scirc du\scirc J_S ),$$ which means that $du + J\scirc du\scirc J_S$ is $L^p$-integrable $u^*(TX)$-valued (0,1)-form.
Consider a Banach bundle $\calt\to \cals\times \calj_S\times \calj$ with a fiber $$\calt_{(u, J_S, J)} = L^p(S,\Lambda^{0, 1}S\otimes
u^*(TX)),$$ where $TS$ and $TX$ are equipped with complex structures $J_S$ and $J$, respectively. $\calt$ has two distinguished sections:
1\) $\sigma_0 \equiv 0$, the zero section of $\calt$;
2\) $\sigma_{\dbar}(u, J_S, J) = du + J\scirc du\scirc J_S$.
By definition $\calp$ is the zero-set of $\sigma_{\dbar}$.
Let us compute the tangent space to $\calp$ at the point $(u, J_S, J)$. Let $(u_t, J_S(t),J(t))$ be a curve in $\calp$ such that $(u_0, J_S(0), J(0)) =
(u, J_S, J)$. Let $$(v,\dot J_S,\dot J) \deff
\left({du_t\over dt}|_{t=0},
{dJ_S(t)\over dt}|_{t=0},
{dJ(t)\over dt}|_{t=0}\right)$$ be the tangent vector to this curve and hence to $\calp$ at $t=0$. The condition $(u_t, J_S(t), J(t)) \in \calp$ means that $$du_t + J(u_t, t)\scirc du_t\scirc J_S(t) = 0 \eqno(6.2.3)$$ in $ L^p(S,\Lambda^{0, 1}(S)\otimes u_t^*(TX))$. Let $\nabla $ be some symmetric connection on $TX$, [*i.e.,*]{} $\nabla_YZ - \nabla_ ZY = [Y,
Z]$. The co-variant differentiation of (6.2.3), with respect to $t$, gives $$\nabla_{\partial \over \partial t}(du_t) + (\nabla_vJ)(du_t\scirc J_S)
+ J(u_t, t)\scirc \nabla_{\partial \over \partial t}(du_t)\scirc J_S +$$ $$+ J\scirc du_t\scirc \dot J_S +
\dot J\scirc du_t\scirc J_S = 0.$$ Let us show that $\nabla_{\partial \over \partial t}(du_t) = \nabla v$. Indeed, for $\xi \in TS$ one has $$(\nabla_{\partial \over \partial t}
du_t)(\xi ) = \nabla_{\partial \over \partial t}[du_t(\xi )] =
\nabla_{\partial \over \partial t}({\partial u_t\over \partial \xi }) =
\nabla_{\xi }({du_t\over dt}) = \nabla_{\xi }v.$$ So every vector $(v,\dot J_S,\dot J)$ which is tangent to $\calp$ satisfies the equation
$$\nabla v + J\scirc \nabla v\scirc J_S + (\nabla_vJ)\scirc (du\scirc
J_S) + J\scirc du\scirc \dot J_S +
\dot J\scirc du\scirc J_S = 0. \eqno(6.2.4)$$
[**Definition 6.2.1.**]{} Let $u$ be a $J$ - complex curve in $X$. Define the operator $D_{u, J}$ on $L^{1,p}$-sections $v$ of $u^*(TX)$ as $$D_{u, J}(v) = \msmall{1\over2}\bigl(\nabla v + J\scirc\nabla v\scirc J_S
+ (\nabla_vJ) \scirc (du\scirc J_S). \bigr)
\eqno(6.2.5)$$
[**Remark.**]{} This operator plays a crucial role in studying properties of complex curves. In its definition we use the symmetric connections instead of those compatible with $J$, as is shown in \[G\]. The matter is that one can use the same connection $\nabla$ for changing almost complex structures $J$. The lemmas below justify our choice.
[**Lemma 6.2.2.** ]{}*$D_{u, J}$ does not depend on the choice of a symmetric connection $\nabla $ and is an $\rr$-linear operator from $L^{1,p}(S,
u^*(TX))$ to $L^p(S, \Lambda^{0, 1}S \otimes u^*(TX))$.*
[**Proof.**]{} Let $\widetilde \nabla $ be another symmetric connection on $TX$. Consider the bilinear tensor on $TX$, given by formula $Q(Z, Y) {:=} \nabla_ZY
- \widetilde \nabla_ZY$. It is easy to see that $Q$ is symmetric on $Z$ and $Y$. Also note that $\nabla_{\xi }v - \widetilde\nabla_{\xi }v =
\nabla_{du(\xi )}v
- \widetilde \nabla_{du(\xi )}v = Q(du(\xi), v)$ and, in addition, that $(\nabla_ZJ)(Y) - (\widetilde \nabla_ZJ)(Y) = Q(Z, JY) - JQ(Z, Y)$. From here one obtains $$2(D_{u, J}v)(\xi ) - 2(\widetilde D_{u, J}v)(\xi) =$$ $$= \nabla_{\xi }v - \widetilde \nabla_
{\xi }v + J(\nabla_{J_S\xi }v - \widetilde \nabla_{J_S\xi }v) + (\nabla_vJ -
\widetilde \nabla_vJ)du(J_S\xi ) =$$ $$= Q(du(\xi ), v) + JQ(du(J_S\xi ), v) + Q(v, JduJ_S
\xi ) - JQ(v, du(J_S\xi)) =$$ $$=Q(du(\xi ), v) + Q(JduJ_S\xi , v) = Q((du + JduJ_S)(\xi ), v) = 0.$$
Now let us show that $D_{u, J}(v)$ is $J_S$-antilinear: $$2D_{u, J}(v)[J_S\xi] = \nabla_{J_S\xi }v + J\scirc \nabla_{J^2
_S\xi }v + (\nabla_vJ)\scirc (du\scirc J^2_S)(\xi ) =$$ $$= \nabla_{J_S\xi }v -
J(\nabla_{\xi }v) - (\nabla_vJ)(du(\xi )) = -J[\nabla_{\xi }v + J(\nabla
_{J_S\xi }v) - J\nabla_vJdu(\xi )] =$$ $$= -J[\nabla_{\xi }v + J\scirc \nabla_{J_S\xi}v + (\nabla_vJ)
(du\scirc J_S(\xi )] = -2J\, D_{u, J}(v)[\xi].$$ Here we use the fact that $J\scirc \nabla_vJ + \nabla_vJ\scirc J = 0$ and $du\scirc J_S = J\scirc du $.
This lemma allows us to obtain expression (6.2.4) also by computation in local coordinates $x_1, x_2$ on $S$ and $u_1,\ldots, u_{2n}$ on $X$, choosing as connection $\nabla$ the de Rham differential $d$. Really, wright $u_t(x) =
u(x,t) = (u_1(x_1,x_2,t),...,u_{2n}(x_1,x_2,t))$ for the $J_t$-holomorphic map $u:S\to X$ (note that $u_t$ and structure $J_t$ are both time dependent). In the local basis ${\d \over \d u_1},...,{\d \over \d u_{2n}}$ of the tangent bundle $TX$ $J_t$ is represented by a time dependent $2n\times 2n$ matrix, which we will also denote as $J(\cdot ,t)$. Differential of $u_t$ will be denoted as $du_t(x)$ or as ${du_t\over dx}$ and is $2\times 2n$ matrix $$du_t(x) =
\left( \matrix
{\d u_1\over \d x_1}(x,t)&{\d u_1\over \d x_2}(x,t)\cr
\cdot \cdot \cdot \cdot &\cdot \cdot \cdot \cdot \cr
{\d u_{2n}\over \d x_1}(x,t)&{\d u_{2n}\over \d x_2}(x,t)\cr
\endmatrix \right).
\eqno(6.2.6)$$
Complex structure $J_S(t)$ on the surface is also time dependent and in the local frame ${\d \over \d x_1},{\d \over \d x_2}$ of $TS$ is represented by a $2\times 2$ matrix, which we shall denote also as $J_S(t)$. Equation of holomorphicity (6.2.3) now reeds in matrix form as $${du_t\over dx} + J(u_t,t)\cdot {du_t\over dx}\cdot J_S(t)\equiv 0.
\eqno(6.2.7)$$ Differentiating this with respect to $t$ we get
$${d\over dx}\left[{du_t\over dt}\right] +
{\d J\over \d t}(u_t,t)\cdot {du_t\over dx}\cdot J_S(t) +
\left\<{\d J(u_t,t)\over \d u},{du_t\over dt}\right\>
\cdot {du_t\over dx} \cdot J_S(t)
+$$ $$+ J(u_t,t)\cdot {d\over dx}
\left[{du_t\over dt}\right]\cdot J_S(t) + J(u_t,t)\cdot
{du_t\over dx}\cdot {dJ_S(t)\over dt} \equiv 0.
\eqno(6.2.8)$$ Here we put $\<{\d J(u_t,t)\over \d u},{du_t\over dt}\>:=\sum_{j=1}^{2n}{\d
J\over \d u},{du_j(t)\over dt}$. Therefore the vector $$(v,\dot J_S,\dot J) \deff
\left({du_t\over dt}|_{t=0},
{dJ_S(t)\over dt}|_{t=0},
{dJ(t)\over dt}|_{t=0}\right)$$ is tangent to $\calp $ at $(u_0,J_S(0),J_0)$ if and only if $${dv\over dx} + \left\<{\d J(u_t,t)\over \d u},v \right\>
\cdot {du_t\over dx}\cdot J_S +
J(u_0,0)\cdot {dv\over dx}\cdot J_S(0) +$$ $$+ \dot J\cdot {du\over dx}\cdot J_S +
J\cdot {du\over dx}\cdot \dot J_S = O.\eqno(6.2.9)$$ Note that $\<{\d J(u_t,t)\over \d u},v\>={dJ\over dv}$ is the derivative of $J$ along the vector $v$. Therefore (6.2.9) with $\nabla =d$ gives the same expression as (6.2.4).
*6.3. $\dbar $-type Operators.*
Now we need to understand the structure of the operator $D_{u,J}$ in more detail. The problem arising here is that $D\deff D_{u,J}$ is only $\rr$-linear. So we decompose it into $J$-linear and $J$-antilinear parts. For the rest of this text we shall denote by $E$ the pulled-back $u^*TX$ by $u$ tangent bundle of $X$. For $\xi \in C^1 (S, TS)$ and $v\in L^{1,p}(S, E)$ write $D_\xi v= {1\over2}[
D_{\xi }v - JD_{\xi }(Jv)] + {1\over2} [D_{\xi }v + JD_\xi(Jv)] =
\dbar_{u, J}[v](\xi) + R(v,\xi )$.
Definition 6.3.1. The operator $\dbar_{u, J}$, introduced above as the $J$-linear part of $D_{u, J}$, we shall call the [*$\dbar$-operator for a $J$-complex curve $u$*]{}.
Lemma 6.3.1. [*$\dbar_{u,J}: L^{1,p}(S,E) L^p(S,\Lambda^{0,1}S
\otimes E)$ is a first order differential operator satisfying $$\dbar_{u,J}(fv) = \dbar_{J_S}f\otimes v + f\otimes \dbar_{u,J}v .\eqno(6.3.1)$$ where $\dbar_{J_S}f={1\over 2}[df + i\cdot df\circ J_S]$ is $\dbar $-operator on $S$.* ]{}
Proof. Really, $$2D_{u,J}(fv)=\nabla (fv) + J\circ \nabla (fv)\circ
J_S + (\nabla_{fv}J)\circ (du\circ J_S) =$$
$$= f\nabla v + fJ\circ \nabla v\circ J_S + f\nabla_vJ\circ (du\circ J_S) +$$ $$+ df\otimes v + Jdf\otimes vJ_S = f2D_{u,J}v + 2\dbar_{J_S}f\otimes v.$$ From here we see that $2JD_{u,J}(Jfv) = 2fJD_{u,J}(Jv) + 2J\dbar f\otimes Jv$. So $$4\dbar_{u,J}(fv) = 2[D_{u,J}(fv) - JD_{u,J}(Jfv)] = 2f[D_{u,J}v-JD_{u,J}(Jv)]
+ 4\dbar f\otimes v =$$ $$= 4f\dbar_{u,J}v + 4\dbar_{J_S}f\otimes v .$$
The following statement is well known in the smooth case. The line bundles case can be found in \[Hf-L-Sk\].
[**Lemma 6.3.2.**]{} [*Let $S$ be a Riemann surface with a complex structure $J_S$ and $E$ a $L^{1,p}$-smooth complex vector bundle of rank $r$ over $S$. Let also $\dbar_E : L^{1,p}(S, E) \to
L^p(S,\Lambda^{(0, 1)}S\otimes E)$ be a differential operator, satisfying the condition $$\dbar_E(f\xi) = \dbar_S f \otimes \xi
+ f\cdot \dbar_E\xi, \eqno(6.3.2)$$ where $\dbar_S$ is the Cauchy-Riemann operator, associated to $J_S$. Then the sheaf $$U\subset S \mapsto {\cal O}(E)(U) := \{\,\xi \in L^{1,p}(U, E) \, :\,
\dbar_E\xi=0 \,\} \eqno(6.3.3)$$ is analytic and locally free of rank $r$. This defines the holomorphic structure on $E$, for which $\dbar_E$ is an associated Cauchy-Riemann operator.* ]{}
[**Remark.**]{} The condition $(6.3.2)$ means that $\dbar_E$ of order 1 and has the Cauchy-Riemann symbol.
[**Proof.**]{} It is easy to see that the problem is essentially local. So we may assume that $S$ is a unit disk $\Delta$ with the standard complex structure. Let $\xi=(\xi_1,\ldots,\xi_r)$ be a $L^{1,p}$-frame of $E$ over $\Delta$. Let also $\Gamma \in L^p(\Delta,\Lambda^{0, 1}\Delta
\otimes Mat(r,\cc))$ be defined by relation $$\dbar_E \xi_i = \sum_j \Gamma_i^j \xi_j,
\qquad \hbox{or in matrix form, } \qquad
\dbar_E \xi = \xi \cdot \Gamma.
\eqno(6.3.4)$$ Then for any section $\eta=\sum g^i \xi_i$ the equation $\dbar_E
\eta =0$ is equivalent to $$\dbar g^i + \sum_j \Gamma^i_j g^j=0,
\qquad \hbox{or in matrix form, } \qquad
\dbar g + \Gamma \cdot g =0.
\eqno(6.3.5)$$
Let the map $\tau_t : \Delta \to \Delta$ be defined by formula $\tau_t(z) = t\cdot z$, $0<t<1$. One can easily check that $$\Vert \tau_t^*\Gamma \Vert_{L^p(\Delta)}
\le t^{1-2/p} \Vert \Gamma \Vert_{L^p(\Delta)}.$$ So taking pull-backs $\tau_t^*E$ with $t$ sufficiently small we may assume that $\Vert \Gamma \Vert_{L^p(\Delta)}$ is small enough. Now consider the mapping $F$ from $L^{1,p}(\Delta,
Gl(r,\cc))\subset $
$ \subset L^{1,p}(\Delta,
Mat(r,\cc))$ to $L^p(\Delta,\Lambda^{0, 1}\Delta
\otimes Mat(r,\cc))$ defined by formula $F(g):=
\dbar g \cdot g^{-1}$. It is easy to see that the derivation of $F$ in $g\equiv \id$ equals to $\dbar$. Due to the [*Lemma 3.2.1*]{} and the implicit function theorem for any $\Gamma$ with $\Vert \Gamma \Vert_{L^p(\Delta)}$ small enough there exists $g$ in $L^{1,p}(\Delta, Gl(r,\cc))$ with $g(0)=\id$, satisfying the equality $F(g)=-\Gamma$, which is equivalent to (6.3.4). Consequently, in the neighbourhood of every point $p\in S$ there exists a frame $\eta =(\eta_1,\ldots,\eta_r)$ of $E$ consisting of sections of ${\cal O}(E)$. This implies that ${\cal O}(E)$ is analytic and locally free of rank $r$. This lemma has an interesting corollary, which will be not used in this notes. Let $E\to S$ be real vector bundle over a Riemann surface $(S,J_S)$, and let $J$ be a complex structure on $E$, i.e. $J\in
End(E)$ with $J^2=-\id $. Take some connection $\nabla $ on $E$ and consider a $J$-linear first order differential operator $D^{\nabla }v:={1\over 2}[\nabla v - J\nabla (Jv)]:\Gamma^{1,p}(S,E)\to
\Gamma^p(S,\Lambda^1_S\otimes E)$. Define
$$\dbar^{\nabla }v = {1\over 2}(D^{\nabla }v - JD^{\nabla }_{J_S\cdot }v).
\eqno(6.3.6)$$
Lemma 6.3.3. [*Operator $\dbar^{\nabla }$ acts from $\Gamma^{1,p}(S,E)$ to $\Gamma^p(S,\Lambda^{0,1}_S\otimes E)$ and satisfies condition (6.3.2) i. e. $$\dbar^{\nabla }(fv) = \dbar_S f \otimes v
+ f\cdot \dbar^{\nabla }v , \eqno(6.3.7)$$ and therefore defines by [*Lemma 6.3.2*]{} a holomorphic structure on $E$ compatible with a given complex structure $J$.* ]{} Proof.
We need to check only the antilinearity of the form $\dbar^{\nabla }v$and relation (6.3.7). Take a $\xi \in TS$, then $$(\dbar^{\nabla }v)(J_S\xi ) = {1\over 2}[D^{\nabla }_{J_S\xi }v -
JD^{\nabla }_{\xi }v] = -J{1\over 2}[D^{\nabla }_{J\xi }v +
JD^{\nabla }_{J_S\xi }v] = -J\dbar^{\nabla }v(\xi ).$$ This shows antilinearity of $\dbar^{\nabla }v$ with respect to $TS$-variable $\xi $, and therefore $\dbar^{\nabla }$ acts from $\Gamma^{1,p}(S,E)$ to $\Gamma^p(S,\Lambda^{0,1}_S\otimes E)$. Note that $\dbar^{\nabla }$ is $J$-linear on the variable $v$ as $D^{\nabla }$ is. Further,
$$D^{\nabla }(fv) = {1\over 2}[df\otimes v + f\nabla v - Jf\nabla (Jv) -
Jdf\otimes Jv] =$$
$$= fD{\nabla }v + {1\over 2}(df\otimes v + df\otimes v) = fD^{\nabla }v +
df\otimes v.\eqno(6.3.8)$$ Therefore $$\dbar^{\nabla }(fv) = {1\over 2}[D^{\nabla }(fv) - JD^{\nabla }_{J_S\cdot }
(fv)] = {1\over }[fD^{\nabla }v + df\otimes v - fJD^{\nabla }_{J_S\cdot }v
- Jdf\otimes v] =$$
$$= f\dbar^{\nabla }v + {1\over 2}[df\otimes v -Jdf\otimes v] =
f\dbar^{\nabla }v + \dbar _{J_S}f\otimes v .$$ This is because $$D^{\nabla }_{J_S\cdot }(fv) = {1\over 2}[\nabla_{J_S\cdot } (fv) -
J\nabla_{J_S\cdot }(fv)] = fD^{\nabla }_{J_S\cdot }v + {1\over 2}[df\circ J_S
\otimes v - Jdf\circ J_S\otimes v] =$$ $$= fD^{\nabla }_{J_S\cdot }v + Jdf\circ J_S\otimes v =
fD^{\nabla }_{J_S\cdot }v + \dbar_Sf\otimes v.$$ We used here the definition $if\otimes v = f\otimes Jv$ of multiplication by $i$ in the bundle $E$.
*6.4. Holomorphic Structure on the Induced Bundle.*
Relation (6.3.1) tells us that the $J$-linear operator $\dbar_{u, J} :
L^{1,p}(S, E) \to L^p(S, \allowbreak
\Lambda^{0, 1}S\otimes E)$ satisfies the condition (6.3.2) and therefore defines in way explained in the previous paragraph a holomorphic structure on the bundle $E$. We shall denote by ${\cal
O}(E)$ the sheaf of holomorphic sections of $E$.
Now let us turn to the antilinear part $R$ of Gromov’s operator.
**Lemma 6.4.1. *$R$ is a continuous $J$-antilinear operator from $E$ to $\Lambda^{0,1}\otimes E$ of the order zero, satisfying $$R(v,\xi ) = N(v, du(\xi )).\eqno(6.4.1)$$ and consequently $$R\scirc du \equiv 0.\eqno(6.4.2)$$ Here the second relations means that for all $\xi , \eta \in TS$ we have $R(du(\eta ),\xi )=0$.***
[**Proof.**]{} $J$ anti-linearity of $R$ is given by its definition. Compute $R(v,\xi )$ for $v\in L^{1,p}(S, E)$ and $\xi \in C^1(S, TS)$, setting $w\deff du(\xi)$ and $D\deff D_{u, J}$ to simplify the notations
$$4R(v,\xi ) = 2D[v](\xi ) + 2JD[Jv](\xi ) =$$
$$=\nabla_{\xi }v + J\nabla_{J_S\xi }v + \nabla_vJ\scirc du\scirc J_S(\xi )$$ $$+J\nabla_{\xi }(Jv) + J^2\nabla_{J_S(\xi )}(Jv) +
J \nabla_{Jv}J\scirc du\scirc J_S(\xi ) =$$
$$= \nabla_\xi v + J\nabla_{J_S\xi }v + \nabla_vJ \scirc du\scirc J_S(\xi ) +$$ $$+ J^2\nabla_{\xi }v + J(\nabla_wJ)v + J^3\nabla_{J_S\xi }v +
J^2(\nabla_{Jw}J)v + J\scirc \nabla_{Jv}J\scirc Jw.$$
Here we used the relations $\nabla_{du(\xi )}J = \nabla_wJ$, $du(J_S\xi) =Jw$ and $\nabla_{du(J_S\xi )}J = \nabla_{Jw}J$. Contracting terms, we obtain $$4R(v,\xi ) = \nabla_vJ(Jw) + J(\nabla_wJ)v - (\nabla_{Jw}J)v +
J(\nabla_{Jv}J(Jw)) =$$ $$= (\nabla_vJ\scirc J)w - (\nabla_wJ\scirc J)v - (\nabla_{Jw}J)v +
(\nabla_{Jv}J)w = 4N(v, w),$$ where $N(v, w)$ denotes the torsion tensor of the almost-complex structure $J$, see \[Li\], p.183, or \[Ko-No\], vol.II., p.123, where another normalization constant for the almost complex torsion is used. Finally we obtain $$R(v,\xi ) = N(v, du(\xi )).$$
$N$ is antisymmetric and $J$-antilinear on both arguments, so $$-JR\bigl(du(\eta ),\xi \bigr) = R\bigl(du(\eta ), J_S\xi \bigr) =
N\bigl(du(\eta ), du(J_S\xi ) \bigr)
= N\bigl(du(\eta ), du(\eta )\bigr) =0$$ if $\xi$ and $\eta$ were chosen in such a way that $J_S(\xi )=\eta $. The relation $R(du(\xi ),\xi )=0$ obviously follows from (6.4.1).
Remarks. 1. [If our structure $J$ is integrable, i.e. $N_J\equiv 0$ then by (6.4.2) $R\equiv 0$ and therefore $D_{u,J}=\dbar_{u,J}$ is $J$-linear. In fact we know that $J$ is integrable iff there exists a symmetric connection $\nabla $ on $TX$ compatible with $J$, i.e. $\nabla J=0$, see \[Ko-No\]. From (6.2.5) we see that in this case $D_{u,J}(v)={1\over 2}(\nabla v + J\circ
\nabla v\circ J_S)$. So $D_{u,J}$ is onviously $J$-linear. ]{}
**2. The tangent bundle $TS$ to the Riemann surface $S$ carries a natural holomorphic structure. We shall denote by $\calo(TS)$ the corresponding analytic sheaf. If the structure $J$ on $X$ is integrable then the differential $du:(TS,\calo(TS))\to (S,E)$ of a holomorhpic map $u:S\to X$ is an analytic morphism of sheaves. We shall see know that this fact is stil true in nonintegrable case.**
**Lemma 6.4.2. *Let $u : (S, J_S) \to (X, J)$ be a non-constant complex curve in almost complex manifold $X$. Then $du$ defines an injective analytic morphism of analytic sheaves $$0\longrightarrow \calo(TS)
\buildrel du \over \longrightarrow \calo(E), \eqno(6.4.3)$$***
where $E = u^*(TX)$ is equipped with a holomorphic structure defined as above by the operator $\dbar_{u, J}$.
[**Proof.**]{} Injectivity of a sheaf homomorphism is equivalent to its nondegeneracy, which is our case.
To prove holomorphicity of $du$ it is sufficient to show that for $\xi,\eta \in
C^1(S, TS)$ one has $$\bigl(\dbar_{u, J}(du(\xi ))\bigr)(\eta ) =
du\bigl((\dbar_S\xi ) (\eta )\bigr), \eqno(6.4.4)$$
where $\dbar_S$ is the usual $\dbar $-operator on $TS$. We shall use the relation which is, in fact, the definition for $\dbar_S$:
$$(\dbar_S\xi )(\eta ) =
\msmall{1\over2}\left( \nabla_{\eta }\xi + J_S\nabla_{J_S\eta}\xi \right).
\eqno(6.4.5)$$ Here $\nabla$ is a symmetric connection on $S$, compatible with $J_S$. One has $$2\cdot (\dbar_{u, J}du(\xi ))(\eta ) = \nabla_{\eta }(du(\xi) ) +
J\nabla_{J_S\eta }(du(\xi)) + (\nabla_{du(\xi )}J)(du(J_S\eta )) =$$ $$= (\nabla_{\eta }du)(\xi ) + du(\nabla_{\eta }\xi ) +
J(\nabla_{J_S\eta}du)(\xi ) + J(du(\nabla_{J_S\eta }\xi ))
+ (\nabla_{du(\xi )}J)(du(J_S\eta )) =$$ $$= du\bigl[ \nabla_{\eta }\xi + J_S \nabla_{J_S\eta }\xi \bigr] +
\bigl[ (\nabla_{\eta }du)(\xi ) + J(\nabla_{J_S\eta}du)(\xi )
+ (\nabla_{du(\xi )}J)(du(J_S\eta )) \bigr]. \eqno(6.4.6)$$ The first term of $(6.4.6)$ is $2\cdot du(\dbar_S\xi)(\eta)$. To cancel the second one we use the identities $(\nabla_\xi du)[\eta]
= (\nabla_\eta du)[\xi]$, $\nabla_wJ\scirc J = - J\scirc \nabla_w J$, and $(\nabla_\xi du)\scirc J_S= J\scirc(\nabla_\xi du) + \nabla_{du(\xi)}J\scirc
du$. The last identity is obtained via co-variant differentiation of $du\scirc J_S=
J\scirc du$. Consequently, we obtain $$(\nabla_{\eta }du)(\xi ) + J(\nabla_{J_S\eta}du)(\xi ) +
(\nabla_{du(\xi )}J)(du(J_S\eta ))=$$ $$(\nabla_{\xi }du)(\eta ) + J(\nabla_\xi du)(J_S\eta ) +
(\nabla_{du(\xi )}J)(du(J_S\eta ))=$$ $$(\nabla_{\xi }du)(\eta ) +
J^2(\nabla_\xi du)(\eta ) +
J (\nabla_{du(\xi)}J)(du(\eta))
+ (\nabla_{du(\xi )}J)(du(J_S\eta) )=$$ $$=(J \scirc\nabla_{du\xi}J)(du\eta)
+ (\nabla_{du(\xi )}J\scirc J)(du(\eta ))=0.$$
Remark. We can give an alternative proof to both [*Lemmas 6.3.1*]{} and [*6.3.2*]{}, which does not use direct calculation. Fix a complex structure $J_S$ on $S$ and let $\phi_t$ be the one parameter group of diffeomorphisms of $S$, generated by a vector field $\xi$. Then ${d\over dt}|_{t=0}(\dbar\phi_t)=D_{J_S, \id }\xi=\dbar_S \xi$. Let a $J$-holomorphic map $u:S\to X$ also be fixed. Then $
{d\over dt}|_{t=0}(u\scirc \phi_t)=du(\xi)$ and consequently $$D_{J, u}(du(\xi))={d\over dt}\bigm|_{t=0}\dbar_J(u\scirc \phi_t)=
{d\over dt}\bigm|_{t=0}(du \scirc \dbar \phi_t)=
du\bigl({d\over dt}\bigm|_{t=0}(\dbar_{J_S} \phi_t)=$$ $$= du(\dbar_S\xi)$$ or equivalently $D_{J, u}\scirc du=du \scirc \dbar$. Taking the $J$-antilinear part of the last equality we obtain $R\scirc du=0$. Nevertheless we shall use the explicit form of the Gromov operator:
$$D_{u,J}(v)[\xi ] = \dbar_{u,J}(v)[\xi ] + N_J(v,du(\xi )).\eqno(6.4.6)$$
The zeroes of analytic morphism $du : \calo(TS) \to \calo
(E)$ are isolated. So we have the following
**Corollary 6.4.3. (\[Sk\]). *The set of critical points of a complex curve in almost the complex manifold $(X, J)$ is discrete, provided $J$ is of class $C^1$.***
For $C^\infty$-structures the result is due to McDuff, see \[McD-1\].
[Lecture 7]{}
[Fredholm Properties of the Gromov Operator]{}
*7.1. Generalized Normal Bundle.*
[*Lemma 6.3.2*]{} makes it possible to define the order of vanishing of the differential of $(J_S,J)$-holomorphic map $u:(S,J_S)\to (X,J)$, provided $J\in C^1$.
**Definition 7.1.1. *By the order of zero ${\sf ord}_p du$ of the differential $du$ at a point $p\in S$, we understand the order of vanishing at $p$ of the holomorphic morphism $du : \calo(TS)\to \calo(E)$.***
From (6.3.2) we obtain the following short exact sequence: $$0\longrightarrow \calo(TS) \buildrel du \over\longrightarrow \calo(E)
\longrightarrow \caln\longrightarrow 0. \eqno(7.1.1)$$
Here $\caln$ is a quotient-sheaf $\calo(E)/du(TS)$. We can decompose $\caln = \calo(N_0)\oplus \caln_1$, where $N_0$ is a holomorphic vector bundle and $\caln_1 =
\bigoplus_{i=1}^P \cc_{a_i}^{n_i}$. Here $\cc_{a_i}^{n_i}$ denotes the sheaf, supported at the critical points $a_i\in S$ of $du$ and having a stalk $\cc^{n_i}$ with $n_i = \ord_{a_i}du$, the order of zero of $du$ at $a_i$. We shall call $\caln$ a normal sheaf and $N_0$ a generalized normal bundle.
Denote by $[A]$ the divisor $\sum_{i=1}^Pn_i[a_i]$, and by $\calo([A])$ a sheaf of meromorphic functions on $S$ having poles in $a_i$ of order at most $n_i$. Then (7.1.1) gives rise to the exact sequence
$$0\longrightarrow \calo(TS)\otimes \calo([A])\buildrel{du}\over
{\longrightarrow} \calo(E)
\longrightarrow \calo(N_0)\longrightarrow 0. \eqno(7.1.2)$$
Denote by $L^p_{(0, 1)}(S, E)$ the space of $L^p$-integrable (0, 1)-forms with coefficients in $E$. Then (7.1.2) together with [*Lemma 6.3.1*]{} implies that the following diagram is commutative
$$\def\normalbaselines{\baselineskip20pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\smash{\mathop{\longrightarrow}\limits^{#1}}}
\def\mapdown#1{\Big\downarrow\rlap{$\vcenter{\hbox{$\scriptstyle#1$}}$}}
%
\matrix
0&\mapright{}&L^{1,p}(S, TS\otimes [A])&\mapright{du}&
L^{1,p}(S, E)&\mapright{\pr}&L^{1,p}(S, N_0)&\mapright{}&0\cr
& &\mapdown{\dbar_S}& &\mapdown{D_{u, J}}& &\mapdown{}& & \cr
0&\mapright{}&L^p_{(0,1)}(S, TS\otimes [A])&\mapright{du}&
L^p_{(0,1)}(S, E)&\mapright{}&
L^p_{(0,1)}(S, N_0)&\mapright{}&0. \cr
\endmatrix
\eqno(7.1.3)$$
This defines an operator $D_{u, J}^N : L^{1,p}(S, N_0) \longrightarrow
L^p_{(0,1)}(S, N_0)$ which has the form $D_{u, J}^N =
\dbar_N + R$. Here $\dbar_N$ is a usual $\dbar $-operator on $N_0$ and $R\in C^0(S, \hom_\rr(N_0,\Lambda^{0,1} \otimes N_0))$. This follows from the fact that $D_{u, J}$ has the same form.
Definition 7.1.2. Let $E$ be a holomorphic vector bundle over a compact Riemann surface $S$ and let $D:L^{1,p}(S, E)\to L^p(S,
\Lambda^{0,1}S \otimes E)$ be an operator of the form $D=\dbar + R$, where $R\in L^p\bigl(S,\,\hom_\rr(E,\,\Lambda^{0, 1}S\otimes E) \bigr)$ with $2<p<\infty$. Define $\sfh^0_D(S, E)\deff \ker D$ and $\sfh^1_D(S, E)
\deff \coker D$.
Remark. It is shown in [*Lemma 7.2.2*]{} below that given $S$, $E$ and $R\in L^p$, $2<p<\infty$, one can define $\sfh^i_D(S, E)$ as a (co)kernel of the operator $\dbar +R: L^{1,q}(S, E) \to
L^q(S,\,\Lambda^{0, 1}S\otimes E)$ for any $1<q\le p$. Thus, the definition is independent of the choice of a functional space.
By the standard lemma of homological algebra we obtain from (7.1.2) the following long exact sequence of $D$-cohomologies.
$$\def\normalbaselines{\baselineskip20pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\smash{\mathop{\longrightarrow}\limits^{#1}}}
\def\mapdown{\Big\downarrow}
%
\matrix
0& \mapright{}& \sfh^0(S, TS\otimes [A]) &\mapright{} & \sfh^0_D(S, E)
& \mapright{}& \sfh^0_D(S, N_0) &\mapright{\delta} &\cr
%
\vphantom{\mapdown}&&&&&&&&\cr
%
& \mapright{}& \sfh^1(S, TS\otimes [A]) &\mapright{} & \sfh^1_D(S, E)
& \mapright{}& \sfh^1_D(S, N_0) &\mapright{} &0.
\endmatrix
\eqno(7.1.4)$$
*7.2. Surjectivity of $D_{u, J}$.*
We shall use a result of Gromov (\[G\]) and Hofer-Lizan-Sikorav (\[Hf-L-Sk\]) about surjectivity of $D^N_{u, J}$, namely a vanishing theorem for $D$-cohomologies. First we prove some technical statements.
Lemma 7.2.1.
*Let $X$ and $Y$ be Banach spaces and $T:X\to Y$ a closed dense defined unbounded operator with the graph $\Gamma=\Gamma_T$ endowed with the graph norm $\Vert x \Vert_\Gamma= \Vert x \Vert_X + \Vert Tx
\Vert_Y$. Suppose that the natural map $\Gamma\to X$ is compact. Then*
$\ker(T)$ is finite-dimensional;
$\im(T)$ is closed;
the dual space $\bigl(Y/\im(T)\bigr)^*$ is naturally isomorphic to $\ker(T^*:Y^*\to X^*)$.
Proof. Obviously, for $x\in \ker(T)$ one has $\Vert x \Vert_X =
\Vert x \Vert_\Gamma$. Let $\{x_n\}$ be a sequence in $\ker(T)$ which is bounded in the $\Vert \cdot \Vert_X$- norm. Then it is bounded in the $\Vert \cdot
\Vert_\Gamma$- norm and hence relatively compact with respect to the $\Vert \cdot \Vert_X$- norm. Thus, the unit ball in $\ker(T)$ is compact which implies the statement of the lemma.
Due to finite-dimensionality, there exists a closed complement $X_0$ to $\ker(T)$ in $X$. Let $\{x_n\}$ be a sequence in $X$ such that $Tx_n \lrar y
\in Y$. Without losing generality we may assume that $x_n$ belong to $X_0$. Suppose that $\Vert x_n \Vert_X \lrar \infty$. Denote $\tilde x_n \deff {x_n
\over \Vert x_n \Vert_X }$. Then $\Vert \tilde x_n \Vert_\Gamma$ is bounded and hence some subsequence of $\{\tilde x_n\}$, still denoted by $\{\tilde
x_n\}$, converges in $X_0$ to some $\tilde x$. Note that $\tilde x\not=0$, because $\Vert \tilde x \Vert_X = \lim \Vert \tilde x_n \Vert_ X =1$. On the other hand, one can see that $T\tilde x_n \lrar 0\in Y$. Since $\Gamma$ is closed, $(\tilde x, 0)\in \Gamma$ and hence $\tilde x\in \ker(T) \cap X_0
=\{0\}$. The contradiction shows that the sequence $\{x_n\}$ must be bounded in $X$. Since $\{Tx_n \}$ is also bounded in $Y$, some subsequence of $\{x_n\}$, still denoted by $\{x_n\}$, converges in $X_0$ to some $x$. Due to the closeness of $\Gamma$, $Tx=y$. Thus $\im (T)$ is closed in $Y$.
Denote $Z\deff\ker(T^*:Y^*\to X^*)$ and let $h\in \bigl(Y/\im(T)
\bigr)^*$. Then $h$ defines a linear functional on $Y$,i.e., some element $h'\in Y^*$, which is identically zero on $\im(T)$. Thus, for any $x$ from the domain of the definition of $T$ one has $\langle h', Tx \rangle =0$, which implies $T^*(h')=0$. Consequently, $h'$ belongs to $Z$. Conversely, every $h'\in Z$ is a linear functional on $Y$ with $h'(Tx)= \langle T^*h', x\rangle
=0$ for every $x$ from the domain of the definition of $T$. Thus, $h'$ is identically zero on $\im(T)$ and is defined by some unique $h\in
\bigl(Y/\im(T)\bigr)^*$.
Lemma 7.2.2. [*(Serre Duality for $D$-cohomologies.)*]{} [*Let $E$ be a holomorphic vector bundle over a compact Riemann surface $S$, and let $D:L^{1,p}(S, E)\to L^p(S, \Lambda^{0, 1}S\otimes E)$ be an operator of the form $D=\dbar + R$, where $R\in L^p\bigl(S,\,
\hom_\rr(E,\Lambda^{0, 1}S\otimes E) \bigr)$ with $2<p <\infty$. Also let $K\deff \Lambda^{1, 0}S$ be the canonical holomorphic line bundle of $S$. Then there exists the naturally defined operator $$D^*=\dbar- R^* : L^{1,p}(S, E^* \otimes K)
\to L^p(S,\Lambda^{0, 1} \otimes E^* \otimes K)$$ with $R^* \in L^p\bigl(S,\,\homr(E^*\otimes K,\,
\Lambda^{0, 1}S \otimes E^*\otimes K) \bigr)$ and the natural isomorphisms $$\sfh^0_D(S,\, E)^*\cong \sfh^1_{D^*}(S,\, E^*\otimes K),$$ $$\sfh^1_D(S,\, E)^*\cong \sfh^0_{D^*}(S,\, E^*\otimes K).$$ If, in addition, $R$ is $\cc$-antilinear, then $R^*$ is also $\cc$-anti-linear.* ]{}
Proof. For any $1<q\le p$ we associate with $D$ an unbounded dense defined operator $T_q$ from $X_q\deff L^q(S, E)$ into $Y_q\deff L^q(S,\Lambda^{0, 1}S\otimes E)$ with the domain of definition $L^{1, q}(S, E)$. The elliptic regularity of $D$ (see [*Lemma 3.2.1*]{} above ) implies that $$\Vert \xi \Vert_{L^{1,q}(S, E)} \le C(q)
\left(
\Vert \dbar\xi +R\xi\Vert_{L^q(S, E)} +
\Vert \xi \Vert_{L^q(S, E)} \right).$$ Consequently, $T_q$ are closed and satisfy the hypothesis of [*Lemma 7.2.1*]{}. For $q>q_1$ we also have the natural imbedding $X_q
\hookrightarrow X_{q_1}$ and $Y_q \hookrightarrow Y_{q_1}$ which commutes with the operator $D$. Moreover, due to the regularity of $D$ this imbedding maps $\ker T_q$ [*identically*]{} onto $\ker T_{q_1}$. Thus we can identify $\sfh^0_D(S, E)$ with any $\ker T_q$.
Now note that for $q'\deff q/(q-1)$ we have the natural isomorphisms $$\eqalign{
X_q^*\equiv &(L^q(S, E))^*
\cong L^{q'}(S,\Lambda^{0, 1}S \otimes E^* \otimes K),\cr
Y_q^*\equiv &(L^q(S,\Lambda^{0, 1}S\otimes E))^*
\cong L^{q'}(S, E^* \otimes K),\cr
}$$ induced by the pairing of $E$ with $E^*$ and by integration over $S$. One can easily check that the dual operator $T_q^*$ is induced by the differential operator $-D^*: L^{1,q'}(S, E^* \otimes K)
\to L^{q'}(S,\Lambda^{0, 1} \otimes E^* \otimes K)$ of the form $D^* =\dbar - R^*$. In fact, for $\xi \in L^{1, q}(S, E)$ and $\eta\in L^{1, q'}(S, E^*\otimes K)$ one has $$\langle T_q\xi,\,\eta \rangle=
\int_S \langle \dbar\xi + R\xi,\,\eta \rangle =
\int_S \dbar\langle \xi ,\,\eta \rangle +
\int_S \langle \xi,\, -(\dbar - R^*)\eta \rangle =
\int_S \langle \xi ,\, -D^*\eta \rangle,$$ since the integral of any $\dbar$-exact $(1, 1)$-form vanishes. From [*Lemma 7.2.1*]{} we obtain the natural isomorphisms $\sfh^1_D(S, E)^*\equiv (\coker
T_p)^*\cong \ker T_p^*$ and $\sfh^0_D(S, E)^*\equiv (\ker
T_p)^*\cong \coker T_p^*$, which yields the statement of the lemma.
Corollary 7.2.3. (\[G\], \[H-L-Sk\].) [*(Vanishing Theorem for $D$-cohomologies.)*]{} [*Let $S$ be a Riemann surface $S$ of the genus $g$. Also let $L$ be a holomorphic [*line*]{} bundle over $S$, equipped with a differential operator $D=\dbar + R$ with $R\in L^p\bigr(S,
\homr(L,\, \Lambda^{0, 1}S\otimes L) \bigl)$, $p>2$. If $c_1(L)<0$, then $\sfh^0_D (S,\, L)=0$. If $c_1(L)>2g-2$, then $\sfh^1_D (S,\, L)=0$.* ]{}
Proof. Suppose $\xi$ is a nontrivial $L^{1,p}$-section of $L$ satisfying $D\xi=0$. Then due to [*Lemma 3.1.1*]{}, $\xi$ has only finitely many zeros $p_i\in S$ with [*positive*]{} multiplicities $\mu_i$. One can easily see that $c_1(L)=\sum \mu_i \ge 0$. Consequently $\sfh^0_D (S,\, L)$ vanishes if $c_1(L)<0$. The vanishing result for $\sfh^1_D$ is obtained via the Serre duality of [*Lemma 7.2.2*]{}.
*7.3. Tangent Space to the Moduli Space.*
Recall that in (7.1.4) we obtained the following long exact sequence
$$\def\normalbaselines{\baselineskip20pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\smash{\mathop{\longrightarrow}\limits^{#1}}}
\def\mapdown{\Big\downarrow}
%
\matrix
0& \mapright{}& \sfh^0(S, TS\otimes [A]) &\mapright{} & \sfh^0_D(S, E)
& \mapright{}& \sfh^0_D(S, N_0) &\mapright{\delta} &\cr
%
\vphantom{.}&&&&&&&&\cr
%
& \mapright{}& \sfh^1(S, TS\otimes [A]) &\mapright{} & \sfh^1_D(S, E)
& \mapright{}& \sfh^1_D(S, N_0) &\mapright{} &0.
\endmatrix$$ It is most important for us to associate a similar long exact sequence of $D$-cohomologies to the short exact sequence (7.1.1). Note that, due to [*Lemmas 6.3.1*]{} and 6.3.2, we obtain the short exact sequence of complexes $$\def\normalbaselines{\baselineskip20pt\lineskip3pt \lineskiplimit3pt }
\setbox1=\hbox{$\lrar$}
\def\mapright#1{\,\,\smash{\mathop{{\hbox to
\wd1{\hss\hbox{$\displaystyle\longrightarrow$}\hss}}}\limits^{#1}}\,\,}
%
\def\mapdown#1{\Big\downarrow\rlap{$\vcenter{\hbox{$\scriptstyle#1$}}$}}
%
\matrix\format\c&\c&\c&\c&\c&\c&\c&\c&\l\\
0&\mapright{}&L^{1,p}(S, TS)&\mapright{du}&
L^{1,p}(S, E)&\mapright{\barr\pr}&
L^{1,p}(S, E){\bigm/}du(L^{1,p}(S, TS))
&\mapright{}&0\cr
& &\mapdown{\dbar_S}& &\mapdown{D}& &\mapdown{\barr D}& &
(7.3.1)%\eqno(7.3.1)
\cr
%
0&\mapright{}&L^p_{(0,1)}(S, TS)&\mapright{du}&
L^p_{(0,1)}(S, E)&\mapright{}&
L^p_{(0,1)}(S, E){\bigm/}
du(L^p_{(0,1)}(S, TS))&\mapright{}&0\cr\vphantom{.}
\endmatrix
%\eqno(7.3.1)$$
where $\overline D$ is induced by $D\equiv D_{u, J}$.
[**Theorem 7.3.1.**]{} [*For $\overline D$ just defined, $\ker \overline D=\sfh^0_D(S, N_0) \oplus \sfh^0(S, \caln_1)$ and $\coker \overline D=\sfh^1_D(S, N_0)$.* ]{}
**Proof. Let $\pr_0: \calo(E) \to \calo(N_0)$ and $\pr_1: \calo(E) \to \caln_1$ denote the natural projections induced by $\pr: \calo(E) \to \caln\equiv \calo(N_0) \oplus \caln_1$. Also let $A$, as in (2.2.7), denote the support of $\caln_1$, [*i.e.,*]{} the finite set of the vanishing points of $du$. Then $\pr_0$ defines maps $$\matrix
\pr_0:L^{1,p}(S , E) & \longrightarrow
& L^{1,p}(S , N_0)
\cr
\vphantom{.}\cr
\pr_0:L^p_{(0, 1)}(S , E) & \longrightarrow
& L^p_{(0, 1)}(S, N_0).
\cr
\endmatrix
\eqno(7.3.2)$$ Furthermore, in a neighborhood of every point $p\in A$ the sequence (7.1.1) can be represented in the form $$0\longrightarrow \calo
\buildrel \alpha_p \over\longrightarrow \calo^{n-1}
\oplus \calo
\buildrel \beta_p \over\longrightarrow \calo^{n-1}
\oplus\caln_1\vert_p
\longrightarrow 0 \eqno(7.3.4)$$ with $\alpha_p(\xi)= (0, z^{\nu_p} \xi)$ and $\beta(\xi,\eta)=
(\xi, j^{(\nu-1)}_p\eta)$. Here $z$ denotes local a holomorphic coordinate on $S$ with $z(p)=0$, $\nu_p$ is the multiplicity of $du$ in $p$, $j^{(\nu-1)}_p\eta$ is a $(\nu-1)$-jet of $\eta$ in $p$ and $\caln_1
\vert_p$ is a stalk of $\caln_1$ in $p$.**
Now let $\overline\xi\in L^{1,p}(S, E)/du(L^{1,p}(S, TS))$ which satisfies $\overline D(\overline\xi)=0$. This means that $\overline\xi$ is represented by some $\xi \in L^{1,p}(S, E)$ with $D\xi= du(\eta)$ for some $\eta \in L^p_{(0, 1)}(S, TS)$. It is obvious that there exists $\zeta
\in L^{1,p}(S, TS)$ such that $\overline\partial \zeta = \eta$ in a neighborhood of $A$. Consequently, $D(\xi- du(\zeta))=0$ in a neighborhood of $A$. Denote $\xi_1 := \xi -du(\zeta)$. Due to (6.3.1), in a neighborhood of $p\in A$ the equation $D\xi_1=0$ is equivalent to $$\overline\partial \xi_1+ N(\xi_1, du)=0.$$ Due to [*Lemma 3.1.2*]{} above, $\xi_1= P(z)+o(|z|^{\nu_p} )$ with some (holomorphic) polynomial $P(z)$. This gives the possibility of defining $\varphi_0 := \pr_0(\xi) \in L^{1,p}(S, N_0)$ and $\varphi_1 :=
\pr_1(\xi- du(\zeta)) \in \sfh^0(S, \caln_1)$. Due to (7.1.2) and (7.1.3), $D_N\varphi_0=0$. If $\overline\partial \zeta' = \eta$ in a neighborhood of $A$ for some other $\zeta'\in L^{1,p}(S, TS)$, then $\zeta - \zeta'$ is holomorphic in a neighborhood of $A$, and consequently $\pr_1(du(\zeta - \zeta'))=0$. Thus, the map $\iota^0: \ker \overline D \to \sfh^0_D(S, N_0) \oplus \sfh^0(S, \caln_1)$, $\iota^0(\overline\xi)=(\varphi_0,\varphi_1)$ is well-defined.
Assume that $\iota^0(\overline\xi)=0$ for some $\overline\xi\in
\ker\overline D$ and that $\overline\xi$ is represented by $\xi\in
L^{1,p}(S, E)$ with $D\xi= du(\eta)$ for some $\eta \in
L^p_{(0, 1)}(S, TS)$. Let $\zeta \in L^{1,p}(S, TS)$ satisfy $\overline\partial \zeta = \eta$ in a neighborhood of $A$. The assumption $\iota^0(\overline\xi)=0$ implies that $\pr_1(\xi- du(\zeta))=0$ and that $\pr_0(\xi- du(\zeta))=0$ in a neighborhood of $A$. Consequently, $\xi-
du(\zeta)=du(\psi)$ for some $\psi \in L^{1,p}(S, TS)$ and $\xi \in
du(L^{1,p}(S, TS))$. This means that $\overline\xi=0\in \ker\overline D$ and $\iota^0$ is injective.
Let $\varphi_0 \in \sfh^0_D(S, N_0)$ and $\varphi_1 \in \sfh^0(S, {\cal
N}_1)$. For every $p\in A$ fix a neighborhood $U_p$ and representation of (7.1.1) in the form (7.3.4) over $U_p$. In every $U_p$ we find $\xi=(\xi_0,
\xi_1) \in L^{1,p}(U_p,\cc^{n-1} \times\cc)$ satisfying the following properties:
$D\xi=0$;
$\xi_0$ coincide with $\varphi_0\vert_{U_p}$ under the identification $\calo(N_0)\vert_{U_p} \cong \calo^{n-1}$;
$j^{(\nu-1)}_p \xi = \varphi_1\vert_p \in
\caln_1 \vert_p$.
The corresponding $\xi_1 \in L^{1,p}(U_p,\cc)$ can be constructed as follows. Let $D(\xi_0,\xi_1) = (\eta_0,\eta_1)$. From $\pr_0(D\xi)= D_N(\pr_0\xi)=0$ one has $\eta_0=0$. In the representation (7.3.4) the identity $R\scirc du =0$ of [*Lemma 6.3.1*]{} means that $D(0,\xi_1) = (0,\dbar\xi_1)$. Thus, one can find $\xi_1$ with $\dbar\xi_1
=-\eta_1$, which gives $D(\xi_0,\xi_1) =0$. Consequently, $j^{(\nu-1)}_p
\xi$ is well-defined. Adding an appropriate [*holomorphic*]{} term to $\xi_1$ one can satisfy condition [*c)*]{}. Using an appropriate partition of unity, we can construct $\xi'
\in L^{1,p}(S, E)$ such that $\xi'$ coincides with $\xi$ in a (possibly smaller) neighborhood of every $p\in A$ and such that $\pr_0 \xi'
= \varphi_0$ in $S$. Thus from $D_N\varphi_0=0$ and (2.2.8) one obtains $D\xi' \in du(L^p_{(0, 1)}(S, TS \otimes [A]))$. But $D\xi'
=0$ in a neighborhood of every $p\in A$, which means that $D\xi' \in du(
L^p_{(0, 1)}(S, TS))$. This shows surjectivity of $\iota^0: \ker
\overline D \to \sfh^0_D(S, N_0) \oplus \sfh^0(S, \caln_1)$.
Now consider the case of $\coker \overline D$. Since the operator $\pr_0$ satisfies the identities $\pr_0\scirc D_E=D_N\scirc \pr_0$ and $\pr_0\scirc du=0$, the induced map
$\displaystyle
\iota^1 : \coker \overline D\equiv L^p_{(0,1)}(S, E)\bigm/
\bigl(D_E(L^{1,p}(S, E) \oplus du(L^p_{(0,1)}(S, TS)\bigr)
\longrightarrow $
$\displaystyle
\longrightarrow\sfh^1_D(S, N_0)\equiv
L^p_{(0,1)}(S, N_0)/ D_N(L^{1,p}(S, N_0))
\qquad$
is well-defined. Moreover, the surjectivity of $\pr_0:
L^p_{(0,1)}(S, E) \to L^p_{(0,1)}(S, N_0)$ easily yields the surjectivity of $\iota^1$.
Now assume that $\iota^1 \overline\xi =0$ for some $\overline\xi \in
\coker \overline D$ and $\overline \xi$ is represented by $\xi
\in L^p_{(0,1)}(S, E)$. Then the condition $\iota^1
\overline\xi=0$ means that $\pr_0 \xi= D_N \eta$ for some $\eta
\in L^{1,p}(S, N_0)$. Find $\zeta \in L^{1,p}(S, E)$ such that $\pr_0 \zeta =\eta$. Then $\pr_0 (\xi - D\zeta)=0$, and, due to (7.1.3), $\xi- D\zeta=du(\varphi)$ for some $\varphi \in
L^p_{(0,1)}(S, TS\otimes [A])$. Find $\psi \in
L^{1,p}(S, TS\otimes [A])$ such that $\dbar \psi =\varphi$ in some neighborhood of $A$. Then $$\xi- D\zeta - D(du(\psi))= du(\varphi-\dbar\psi) \in
du(L^p_{(0,1)}(S, TS)).$$ Consequently, $\overline \xi \in \im\overline D$. This shows the injectivity of $\iota^1$.
Corollary 7.3.2. *The short exact sequence $(7.1.1)$ induces the long exact sequence of $D$-cohomologies $$\def\normalbaselines{\baselineskip20pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\,\,\smash{\mathop{\longrightarrow}\limits^{#1}}\,\,}
\def\mapdown{\Big\downarrow}
%
\matrix\format\c&\c&\c&\c&\c&\c&\c&\c&\c\\
0& \mapright{}& \sfh^0(S, TS) &\mapright{} & \sfh^0_D(S, E)
& \mapright{}& \sfh^0_D(S, N_0)\oplus \sfh^0(S, \caln_1)
&\mapright{\delta} &\cr
%
%\vphantom{.}&&&&&&&&\cr
%
& \mapright{}& \sfh^1(S, TS) &\mapright{} & \sfh^1_D(S, E)
& \mapright{}& \sfh^1_D(S, N_0) &\mapright{} &0.
\endmatrix$$*
*7.4. Reparameterizations.*
Let $S$ be an oriented compact real surface without boundary. Consider the Teichmüller space $\ttt_g$ of marked complex structures on $S$. This is a complex manifold of a dimension $$\dimc \ttt_g = \cases 0 &\text{ if $g=0$;} \\
1 &\text{ if $g=1$;} \\
3g-3 &\text{ if $g\ge2$}
\endcases$$ which can be completely characterized in the following way. The product $S\times \ttt_g$ possesses a complex (holomorphic) structure $J_{S\times
\ttt}$ such that
the natural projection $\pi\mid_{\ttt }: S\times \ttt_g \to \ttt_g$ is holomorphic, so that for any $\tau\in \ttt_g$ the identification $S \cong S
\times \{\tau\}$ induces the complex structure $J_S(\tau)\deff J_{S\times
\ttt}\ogran_{S\times \{\tau\}}$ on $S$;
for any complex structure $J_S$ on $S$ there exists a uniquely defined $\tau\in\ttt_g$ and a diffeomorphism $f: S\to S$ such that $J_S =
f^*J_S(\tau)$ ($f: (S, J_S) \to (S, J_S(\tau))$ is holomorphic) and $f$ is homotopic to the identity map $\id_S: S\to S$.
Denote the automorphism group of $S\times \ttt_g$ by $\bfg$. It is known that $$\bfg=\cases {\bold PGl}(2,\cc) &\text{ for $g=0$,} \\
{\bold Sp}(2,\;\;\zz) \ltimes T^2 &\text{ for $g=1$,} \\
{\bold Sp}(2g,\zz) &\text{ for $g\ge2$.}
\endcases$$ We shall use the following information about $\ttt_g$.
For $g=0$ the surface $S$ is a Riemann sphere $S^2$ and all complex structures on $S^2$ are equivalent to the standard one when $S^2\cong
\cc\pp^1$. Thus, $\ttt_0$ consists of one point, and the group $\bfg={\bold
PGl}(2,\cc)$ is the group of biholomorphisms of $\cc\pp^1$.
For $g=1$ the surface $S$ is a torus $T^2$ and $\ttt_1$ is an upper half-plane $\cc_+=\{\,\tau\in\cc\;:\;\im(\tau)>0\,\}$. The product $S\times \ttt_1$ can be identified by the quotient $(\cc\times \cc_+)/\zz^2$ with respect to the holomorphic action $$\bigl((n,m),\, (z,\tau)\bigr) \in\zz^2 \times \cc\times \cc_+ \longmapsto
(m,n)\cdot(z,\tau)\deff (z+m +n\,\tau,\tau) \in \cc\times \cc_+.$$ In this case the subgroup $T^2\equiv \rr^2/\zz^2$ is a connected component of the identity ${\rom e}\in\bfg$, in particular $T^2$ is normal. The group $\bfg= {\bold Sp}(2,\;\; \zz) \ltimes T^2$ is a semi-direct product. The holomorphic action of $T^2$ on the quotient $(\cc\times \cc_+)/\zz^2$ is given by $$\eqalign{
\bigl([t_1,t_2],\, ([z],\tau)\bigr) &\in T^2 \times (\cc\times \cc_+)/\zz^2
\longmapsto
\cr
[t_1,t_2]\cdot([z],\tau)&\deff ([z+t_1 +t_2\tau],\tau )
\in (\cc\times \cc_+)/\zz^2.
}$$
For any $g\ge0$ the action of $\bfg$ on $S\times \ttt_g$ is effective and preserves fibers of the projection $\pi\mid_{\ttt }: S\times \ttt_g \to
\ttt_g$. This induces the action of $\bfg$ on $\ttt_g$. Furthermore, given $\tau \in \ttt_g$ and $f\in\bfg$, there exists a unique diffeomorphism of $\hat f_\tau:S \to S$ such that $$f\cdot(x,\tau)=(\hat f_\tau(x), f\cdot\tau).
\eqno(7.4.1)$$
For any $\tau\in\ttt_g$ we have natural isomorphisms $T_\tau\ttt_g \cong
\sfh^1(S,\, \calo_\tau(TS))$ and $T_{\rom e}\bfg \cong \sfh^0(S,\,
\calo_\tau(TS))$, where $\calo_\tau(TS)$ denote the sheaf of a section of $TS$ which are holomorphic with respect to the complex structure $J_S(\tau)$.
Later on, we shall denote elements of $\ttt_g$ by $J_S$ and consider them as corresponding complex structures on $S$.
[Lecture 8 ]{}
[Transversality.]{}
*8.1. Moduli Space of Nonparameterized Curves.*
Let $(X,\omega, J\st)$ be a symplectic manifold with some distinguished $w$-tamed almost-complex structure $J\st$. In our applications $J\st $ will be integrable, providing, together with $w$, the Kähler structure on $X$. Let $U\Subset X$ be an open relatively compact subset which can coincide with $X$ if $X$ is compact. Also let $S$ be a (fixed) compact oriented surface of genus $g\ge0$, $u_0: S \to X$ a non-constant $C^1$-smooth map such that $u_0(S)\cap
U \not = \emptyset$.
Fix $2<p<\infty$. The Banach manifold $L^{1,p}(S,X)$ of all $L^{1,p}
$-smooth maps $u:S \to X$ is smooth with a tangent space $T_uL^{1,p}(S,X)$ equal to the space $L^{1,p}(S,u^*TX)$ of $L^{1,p}$-smooth sections of a pulled-back tangent bundle to $X$. Denote by $\cals_U$ a Banach manifold of those $u\in L^{1,p}(S,X) $, for which $u$ is homotopic to $u_0$ and $u(S) \cap U \not=\emptyset$. Fix an integer $k\ge1$ and denote by $\calj^k_U$ the set of $C^k$-smooth almost complex structures $J$ in $X$ such that $\{ x\in X:J(x)\not=J\st(x) \} \comp U$, and which are tamed by $\omega$. The latter means that $J$ is $\omega$-positive, i.e., $\omega(\xi, J\xi)>0$ for every $x\in X$ and every nonzero $\xi\in T_xX$.
The evaluation map $\ev: S\times \cals_U \times \ttt \times \calj^k_U \to X$, given by formula $\ev(x, u, J_S,J)\deff u(x)$, defines a bundle $E\deff \ev^*(TX)$ over $S\times \cals_U \times \ttt \times \calj^k_U$. We equip $E$ with the natural complex structure, which is equal to $J(u(x))$ in the fiber $E_{(x,
u, J_S,J)}\cong T_{u(x)}X$. We shall denote by $(E_u,J)$ the restriction of $E$ onto $S\times \{(u,J_S,J)\}$ which is isomorphic to $u^*TX$.
The bundle $E$ with the complex structure $J$ induces complex Banach bundles $\wh\cale$ and $\wh\cale'$ over a product $\cals_U \times \ttt \times \calj^k_U$ with fibers $$\eqalign{
\wh\cale_{(u,J_S,J)} & \deff L^{1,p}(S, E_u),\cr
\wh\cale'_{(u,J_S,J)} & \deff L^p(S, E_u\otimes\Lambda^{(0,1)}S).
}$$ Here $S$ is equipped with the complex structure $J_S$, $\Lambda^{(0,1)}S$ is a complex bundle of (0,1)-forms on $S$ and $\otimes$ means the tensor product over $\cc$ of bundles with the corresponding complex structures. The bundle $\wh\cale$ is simply a pull-back of a tangent bundle $TL^{1,p}(S,X)$ with respect to the projection $(u,J_S,J)\in \cals_U \times \ttt \times \calj^k_U
\mapsto
u \in L^{1,p}(S,X)$.
On the other hand, the bundle $\wh\cale'$ is the target manifold of the $\dbar$-operator for the map $u\in L^{1,p}(S,X)$. Namely, we have a distinguished section $\sigma_\dbar$ of $\wh\cale'$, $$\sigma_\dbar(u,J_S,J)\deff \dbar_{J_S,J}u \deff
\half\bigl(du + J\scirc du \scirc J_S\bigr).
\eqno(8.1.1)$$ If $f$ is another section of $\wh\cale'$ (given by some explicit geometric construction), then one can consider a $\dbar$-equation $$\dbar_{J_S,J}u=f(u,J_S,J).$$ We shall consider only a homogeneous case where $f(u,J_S,J)\equiv
0$. The corresponding set of solutions is closed in $\cals_U \times \calj^k_U$ and will be denoted by $$\calp \deff \{\, (u,J_S, J) \in \cals_U \times \ttt \times \calj^k_U \;:\;
\dbar_{J_S,J}u=0 \,\}. \eqno(8.1.2).$$ A map $u\in L^{1,p}(S,X)$ such that
$$\dbar_{J_S,J}u=0\eqno(8.1.3)$$ for some $J_S\in\ttt_g$ and $J\in \calj^k_U$ is called $J$-holomorphic, or $(J_S,J)$-holomorphic, and its image $M\deff u(S)$ is called a $J$-complex curve.
The regularity theory for $\dbar$-equation says that $\calp$ is a closed subset of $$\calx \deff \{\,(u,J_S,J)\in\cals_U\times \ttt \times \calj^k_U\;:\;
u\in C^1(S,X)\,\}.$$ Set $
\calx^* \deff \{\,(u,J_S,J)\in \calx \,:\,
\hbox{$u$ is an imbedding in a neighborhood of some $y\in S$ }\},
$ and let $\calp^*\deff \calp\cap \calx^*$. Then $\calx^*$ is open in $\calx$ and $\calp^*$ is open in $\calp$. A group $\bfg$ acts on $\cals_U\times \ttt \times \calj^k_U$ and on $\calx$ in a natural way by composition, $$(f,u,J_S,J) \in \bfg\times \cals_U\times \calj^k_U
\mapsto f\cdot(u,J_S,J) \deff (u\scirc \hat f\inv, f\cdot J_S,J),$$ where $\hat f:S\to S$ is a diffeomorphism induced by $f\in\bfg$ and $J_S$, see (7.4.1). As usual, the inverse $(\hat f)\inv$ is introduced to preserve the associative law $(f\cdot g)\cdot(u,J_S,J)=f\cdot(g\cdot(u,J_S,J))$. The sets $\calx^*$, $\calp$ and $\calp^*$ are invariant with respect to this action.
Let $\calm\deff\calp^*/\bfg$ be a quotient with respect to this action and $\pi_\calp: \calp^* \to \calm$ the corresponding projection. Also let $\pi_\calj: \calm\to\calj^k_U$ be a natural projection.
**Lemma 8.1.1. *The projections $\calx^* \longto \calx^*/\bfg$ and $\pi_\calp:\calp^* \longto \calm$ are principal $\bfg$-bundles.***
**Proof. We consider the case $g\ge2$ first. It is known that in this case the group $\bfg$ acts proper discontinuously on $\ttt_g$. This implies that the same is true for the action of $\bfg$ on $\calx$. Moreover, the definition of $\calx^*$ provides that $\bfg$ acts freely on this set. Consequently, the map $\calx^*\longto \calx^*/\bfg$ is simply an (unbranched) covering.**
Now we consider the case $g=0$. Note that in this case $S=S^2$ and $\ttt_0=\{J\st\}$. Fix some $(u^0,J\st,J^0)\in \calx^*$. Let $y_1$, $y_2$, $y_3$ be distinct points on $S^2$ such that $u^0$ is an imbedding in a neighborhood of every $y_i$, in particular $du^0$ is non-vanishing in $y_i$. Also let $Z_i$ be smooth submanifolds of codimension 2 in $X$, intersecting $u^0(S^2)$ transversally in each $u^0(x_i)$, respectively.
Let $V\ni(u^0,J\st,J^0)$ be an open set in $\calx^*$, $W$ its projection on $\calx^*/\bfg$, and $\bfg\cdot V\deff \{\,f\cdot(u,J\st,J):
f\in \bfg, (u,J\st,J)\in V\,\}$ its $\bfg$-saturation. Consider a set $$\calz \deff\{\, (u,J\st,J)\in \bfg\cdot V\;:\; u(y_i)\in Z_i\,\}.$$ One can easily show that if $V$ is chosen sufficiently small, then $\calz$ is a smooth closed Banach submanifold of $\bfg\cdot V$, intersecting every orbit $\bfg \cdot (u,J\st,J)$ transversally at exactly one point. Moreover, we have a $\bfg$-invariant diffeomorphism $\bfg\cdot V \cong \bfg \times \calz$, so that $\calz$ is a local slice of $\bfg$-action at $(u^0,J\st,J^0)$. This equips a quotient $\calx^*/\bfg$ with a structure of a smooth Banach manifold such that the projection $\calx^*\to\calx^*/\bfg$ is a smooth principle $\bfg$-bundle.
The case $g=1$ is a combination of the two above cases. We fix some $(u^0,J_S^0, J^0)\in \calx^*$ and a point $y$ on $S$ such that $u^0$ is an imbedding in a neighborhood of $y$; in particular $du^0$ is non-vanishing in $y$. Also let $Z$ be a smooth submanifold of codimension 2 in $X$, intersecting $u^0(S)$ transversally in $u^0(y)$. As in the case $g=0$, we fix a small open set $V\ni(u^0,J^0,J^0)$ in $X$ and consider a set $$\calz \deff\{\, (u,J_S,J)\in \bfg\cdot V\;:\; u(y)\in Z\,\},$$ where $\bfg\cdot V \deff \{\,f\cdot(u,J\st,J): f\in \bfg, (u,J\st,J)\in V\,\}$ is $\bfg$-saturation of $V$. If $V$ is chosen sufficiently small, then $\calz$ is a slice to the action of $T^2\subset \bfg$. Thus, the projection $\calx^*\longto \calx^*/T^2$ is a principle $T^2$-bundle.
Now we consider the action of ${\bold Sp}(2,\;\; \zz) =\bfg/T^2$ on $\calx^*/T^2$. The same arguments as in the case $g\ge2$ show that this action is free and proper discontinuous. This implies that the projection $\calx^*/T^2\longto \calx^*/\bfg$ is a covering. Consequently, $\calx^*\longto \calx^*/\bfg$ is a principle $\bfg$-bundle.
The natural inclusion $\calm \hookrightarrow \calx^*/\bfg$ is continuous and closed. Further, we have a natural $\bfg$-invariant homeomorphism $\calp^* \cong \calm \,\times_{\calx^*/\bfg} \calx^*$, which gives a desired structure of a principle $\bfg$-bundle on $\calp^*$ with a base $\calm$.
**Remark. One can show that if $(u,J_S,J)\in \calp^*$, then $u$ is an imbedding in all but finitely many points of $S$, in particular $J_S$ is completely defined by $u$ and $J$. Similarly, every class $\bfg\cdot
(u,J_S,J) \in \calm$ is completely defined by $J\in \calj^k_U$ and a $J$-complex curve $M\deff u(S)$. Therefore we shall denote elements of $\calm$ by $(M,J)$. Our motivation is that the object we really exploit is a complex curve $M=u(S)$ rather than its concrete parameterization $u$. We hope that the reader will not be confused by such a formal incorrectness.**
Using [*Lemma 8.1.1*]{}, one can obtain from $\bfg$-invariant objects on $\calp^*$ corresponding objects on $\calm$. For example, on $\calp$ we have a (trivial) $S$-bundle with a total space $\calp\times S$, natural projection on $\calp$, a complex structure $J_S$ on a fiber over $(u,J_S,J)$, and a map $\ev:\calp\times S \to X$ such that $\ev(u,J_S,J;y)\deff u(y)$ which is invariant with respect to the $\bfg$-action. It gives a $\bfg$-bundle $\pi_\calc:\calc\to \calm$ with the total space $\calc\deff\calp^* \times_\bfg S$ and fiber $S$. Moreover, $\calc$ is equipped with a map $\ev: \calc\to X$. We shall imagine $\calm$ as a moduli space of all complex curves in $X$ of appropriate topological properties, $\pi_\calc:\calc \to\calm$ as a corresponding universal bundle and $\ev:\calc \to X$ as an evaluation map. In particular, every fiber $\calc_{(M,J)} \deff \pi_\calc\inv(M,J)$ over $(M,J)\in \calm$ possesses a canonical complex structure $J_{\calc_{(M,J)}}
=J_S$.
In a similar way we define complex Banach bundles $\cale$ and $\cale'$ over $\calx^*/\bfg\supset \calm$. Note that we have a natural continuous $\bfg$-equivariant inclusion $\calx^*\hook \cals_U\times \calj^k_U$. Let $(u,J_S,J)\in \cals_U\times \calj^k_U$, $f\in\bfg$, and let $\hat f:S\to S$ be a diffeomorphism defined as in (1.1), so that $f\cdot(u,J_S,J)= (u\scirc\hat f\inv,f\cdot J_S,J)$. Define the operator $f_*$ by setting $$f_*:s\in\wh\cale_{(u,J_S,J)} \mapsto (\hat f\inv)^*s\in
\wh\cale_{f\cdot(u,J_S,J)}.$$ This defines a natural lift of a $\bfg$-action on $\cals_U\times \calj^k_U$ to a $\bfg$-actions on $\wh\cale$. In the same way we define the action of $\bfg$ on $\wh\cale'$. Since the projection $\calx^*\to \calx^*/\bfg$ admits a local $\bfg$-slice, there exist uniquely defined bundles $\cale$ and $\cale'$ over $\calx^*/\bfg$, whose lifts onto $\calx^*$ is $\bfg$-equivariant isomorphic to $\wh\cale$ and $\wh\cale'$, respectively. In particular, for any $(u,J_S,J)\in \calx^*$ representing $\bfg\cdot(u,J_S,J)\in \calx^*/\bfg$ we have the natural isomorphisms $$\eqalign{
\cale_{\bfg\cdot(u,J_S,J)} \cong \wh\cale_{(u,J_S,J)}
& = L^{1,p}(S, E_u),\cr
\cale'_{\bfg\cdot(u,J_S,J)} \cong \wh\cale'_{(u,J_S,J)}
& = L^p(S, E_u\otimes\Lambda^{(0,1)}S).
}$$
The question we are interested in is whether we can deform a given (compact) $J_0$-holomorphic curve $M_0=u_0(S)$ into a compact complex curve $M_1$ which is holomorphic with respect to the given (integrable, for example) complex structure $J\st$ on $X$. The idea is to use the continuity method: one finds an appropriate homotopy $h(t)=J_t$, $t\in[0,1]$, of almost complex structures connecting $J_0$ with $J\st=J_1$, and shows that there exists a continuous deformation $u_t:S\to X$ of a map $u_0$ into $u_1$ such that $u_t$ is $J_t$-holomorphic for all $t\in[0,1]$. This can be successfully done by studying the linearization of the equation $\dbar u=0$.
**Lemma 8.1.2. *Let $\calx$ be a Banach manifold, $\cale\to \calx$ and $\cale'\to\calx$ $C^1$-smooth Banach bundles over $\calx$, $\nabla$ and $\nabla'$ linear connections in $\cale$ and $\cale'$, respectively, $\sigma$ a (local) $C^1$-section of $\cale$ and $D: \cale \to \cale'$ a $C^1$-smooth bundle homomorphism.***
If $\sigma(x)=0$ for some $x\in \calx$, then the map $\nabla\sigma_x:
T_x\calx \to \cale_x$ is independent of the choice of a connection $\nabla$ in $\cale$;
Set $K_x\deff \ker(D_x:\cale_x\to \cale'_x)$ and $Q_x\deff
\coker(D_x:\cale_x\to \cale'_x)$ with the corresponding imbedding $i_x:K_x \to \cale_x$ and projection $p_x:\cale'_x \to Q_x$. Let $\nabla^\hom$ be a connection in $\hom(\cale,\cale')$ induced by connections $\nabla$ and $\nabla'$. Then the map $$p_x \scirc(\nabla^\hom\!\! D_x)\scirc i_x:
T_x\calx\to \hom(K_x,Q_x)$$ is independent of the choice of connections $\nabla$ and $\nabla'$.
**Remark. Taking into account this lemma, we shall use the following notation. For $\sigma\in \Gamma(\calx, \cale)$, $D\in \Gamma(\calx,
\hom(\cale, \cale'))$ and $x\in \calx$ as in the hypothesis of the lemma, we shall denote by $\nabla\sigma_x:T_x\calx \to \cale_x$ and $\barr{\nabla\!\!D}:
T_x\calx\times \ker D_x \to \coker D_x$ the corresponding operators without pointing out which connections were used to define them.**
**Proof. Let $\wt\nabla$ be another connection in $\cale$. Then $\wt\nabla$ has a form $\wt\nabla=\nabla +A$ for some $A\in\Gamma(\calx,\hom(T\calx,\endo(\cale)))$. Thus, for $\xi\in T_x\calx$ we get $\wt\nabla_\xi\sigma -\nabla_\xi\sigma =
A(\xi,\sigma(x))=0$.**
Similarly, let $\wt\nabla'$ be another connection in $\cale'$, and let $\wt\nabla^\hom$ be a connection in $\hom(\cale,\cale')$ induced by $\wt\nabla$ and $\wt\nabla'$. Then $\wt\nabla'$ also has the form $\wt\nabla=\nabla +A'$ for some $A'\in\Gamma(\calx,\hom(T\calx,\endo(\cale')))$. Thus, for $\xi\in T_x\calx$ we obtain $\wt\nabla_\xi^\hom D -\nabla_\xi^\hom D=
A'(\xi)\scirc D_x - D_x \scirc A(\xi)$. The statement of the lemma now follows from the identities $p_x\scirc D_x=0$ and $D_x\scirc i_x=0$.
[8.2. Transversal Mappings.]{}
To have the possibility of deforming a complex curve along a given path of almost-complex structures, it is useful to know in which points $(u,J_S,J)$ the set $\calp$ of holomorphic maps is a Banach manifold. Note that by definition the set $\calp$ is essentially an intersection of the zero section and the $\sigma_\dbar$-section in the total space of $\cale' $ over $S\times \ttt \times \calj_U^k$. Thus, we are interested in which points these sections meet transversally.
Definition 8.2.1. [*Let $\calx$, $\caly$, and $\calz$ be Banach manifolds with $C^k$-smooth maps $f:\caly \to \calx$ and $g:\calz\to\calx$, $k\ge1$. Define the [*fiber product*]{} $\caly\times_\calx \calz $ by setting $\caly\times_\calx \calz \deff \{\,(y,z)\in \caly\times\calz \;:\;
f(y)=g(z)\,\}$. The map $f$ is called [*transversal to $g$*]{} at point $(y,z)\in \caly\times_\calx \calz$ with $x\deff f(y)=g(z)$, and $(y,z)$ is called a [*transversality point*]{}, a map $df_y\oplus -dg_z: T_y\caly \oplus T_z\calz \to T_x\calx$ is [*surjective*]{} and its kernel admits a closed complement.* ]{}
The set of transversality points $(y,z)\in \caly\times_\calx \calz$ we shall denote by $\caly\times^\trans_\calx \calz$, with $\trans$ symbolizing the transversality condition.
In the special case, when the map $g:\calz \to \calx$ is a closed imbedding, the fiber product $\caly\times_\calx \calz$ is simply the pre-image $f\inv\calz$ of $\calz\subset\calx$. In particular, every point $(y,z)\in\caly\times_\calx \calz$ is completely defined by the point $y\in\caly$, $z=f(y)\in\calz\subset \calx$. In this case we simply say that $f:\caly \to \calx$ is [*transversal to $\calz$ in $y\in\caly$*]{}, $(y,f(y))$ is a transversal point of $\caly\times_\calx \calz\cong
f\inv\calz$.
Lemma 8.2.1. [*The set $\caly\times^\trans_\calx \calz$ is open in $\caly\times_\calx \calz$ and is a $C^k$-smooth Banach manifold with a tangent space $$T_{(y,z)}\caly\times^\trans_\calx \calz =
\ker\bigl( df_y\oplus d(-g_z): T_y\caly \oplus T_z\calz \to T_x\calx \bigr).$$*]{}
Proof. Fix $w_0\deff(y_0,z_0)\in \caly\times^\trans_\calx \calz$ and set $K_0 \deff \ker( df_{y_0}\oplus dg_{z_0}:
T_{y_0}\caly \oplus T_{z_0}\calz \to T_x\calx \bigr)$. Let $Q_0$ be a closed complement to $K_0$. Then the map $df_{y_0}\oplus dg_{z_0}:Q_0\to T_x\calx$ is an isomorphism.
Due to the choice of $Q_0$, there exists a neighborhood $V\subset
\caly\times\calz$ of $(y_0,z_0)$ and $C^k$-smooth maps $w':V\to K_0$ and $w'':V\to Q_0$ such that $dw'_{w_0}$ (resp. $dw''_{w_0}$) is the projection from $T_{y_0}\caly \oplus T_{z_0}\calz$ onto $K_0$ (resp. onto $Q_0$), so that $(w',w'')$ are coordinates in some smaller neighborhood $V_1\subset
\caly\times\calz$ of $w_0=(y_0,z_0)$. It remains to consider the equation $f(y)=g(z)$ in new coordinates $(w',w'')$ and apply the implicit function theorem.
We have defined $\calp$ as a pre-image of a zero section $\sigma_0$ of $\cale'$ with respect to the map $\sigma_\dbar:\cals_U\times \calj_S\times \calj^k_U
\to \cale'$. Due to [*Lemma 8.2.1*]{}, the set $\calp$ is a Banach manifold in those points $(u,J_S,J)\in\calp$, where $\sigma_\dbar$ is transversal to $\sigma_0$. However, in any point $(u,J_S,J;0)$ on the zero section $\sigma_0$ of $\cale'$ we have the natural decomposition $$T_{(u,J_S,J;0)}\cale' =
d\sigma_0\bigl( T_{(u,J_S,J)}(\cals_U\times \calj_S\times \calj_U^k) \bigr)
\oplus \cale'_{(u,J_S,J)},$$ where the first component is the tangent space to the zero section of $\cale'$ and the second is the tangent space to the fiber $\cale'_{(u,J_S,J)}$. Let $p_2$ denote the projection on the second component. Then the transversality $\sigma_\dbar$ and $\sigma_0$ is equivalent to the surjectivity of the map $p_2\scirc d\sigma_\dbar: T_{(u,J_S,J)}(\cals_U\times \calj_U^k)\to
\cale'_{(u,J_S,J)}$. But, due to [*Lemma 8.1.2* ]{}, this map is a linearization of $\sigma_\dbar$ at $(u,J_S,J)$ and has a form (8.1.2).
Thus, the transversality of $\sigma_\dbar$ to $\sigma_0$ at $(u,J_S,J)\in \calp$ is equivalent to the surjectivity of the operator $$\eqalign{
\nabla\sigma_\dbar:\;&T_uL^{1,p}(S,X)\oplus T_{J_S}\ttt_g \oplus T_J\calj^k_U
\longto \wh\cale'_{(u,J_S,J)}
\cr
\nabla\sigma_\dbar:\;&(v,\dot J_S,\dot J) \longmapsto
D_{(u,J)}v + J\scirc du \scirc \dot J_S
+ \dot J\scirc du \scirc J_S.
}\eqno(8.2.1)$$ [*Definition 7.1.2*]{} provides that the quotient of $\wh\cale'_{(u,J_S,J)}=
L^p_{(0,1)}(S,E_u)$ by the image of $D_{u,J}$ is $\sfh^1_D(S,E_u)$. The induced map $\dot J_S \in T_{J_S}\ttt_g \mapsto J\scirc du \scirc \dot J_S \in
\sfh^1_D(S,E_u)$ is also easy to describe. From equality (8.2.1) and [*Corollary 7.3.2*]{} it follows that its image is equal to the image of the homomorphism $du\scirc J_S:\sfh^1(S,TS) \to \sfh^1_D(S,E_u)$ and its cokernel is $\sfh^1_D(S,N_u)$.
It remains to study the image of $T_J\calj^k_U$ in $\sfh^1_D(S,N_u)$. For $(u,J_S,J)\in \calp$ we define $\Psi=\Psi_{(u,J)}:T_J\calj^k_U \to
\wh\cale'_{(u,J_S,J)}$ by setting $\Psi_{(u,J)}(\dot J)\deff \dot J\scirc du
\scirc J_S$. Let $\barr\Psi=\barr\Psi_{(u,J)}:T_J\calj^k_U \to
\wh\cale'_{(u,J_S,J)}$ be induced by $\Psi$. Recall that if $(u,J_S,J)\in
\calp$, then $J_S$ is determined by $u$ and $J$.
Lemma 8.2.2. *(Infinitesimal Transversality). *Let $(u,J_S,J)\in
\calp^*$. Then the operator $\barr\Psi:T_J\calj^k_U \to \sfh^1_D(S, N_u)$ is surjective.**
Proof. In [*Lecture 3*]{} we proved that for $(u,J_S,J)\in
\calp^*$ mapping $u$ is an imbedding in the neighborhood of all but a finite number of points $x\in S$. Thus, there exists such a nonempty open set $V\subset S$ that $u(V)\subset U$ and $u\mid_V$ is an imbedding.
By [*Lemma 7.2.2*]{} we have an isomorphism $\sfh^0_D(S,N_u^*\otimes K_S)
\cong \sfh^1_D(S,N_u)^*$. The fact that operators of the type $D=\dbar + R: L^{1,p}(S,E)\to L^p_{(0,1)}(S,E)$ on the compact Riemann surface $(S,J_S)$ are Fredholm ensures the existence of the finite basis $\xi_1,...,\xi_l$ of the space $\sfh^0_D(S,N_u^*\otimes K_S)$. By [*Lemma 3.1.1*]{} every $\xi \in \sfh^0_D(S,N_u^*\otimes K_S)$ vanishes not more than in $c_1(N_u^*\otimes K_S)[S]$ points on $S$ (compare to the proof of [*Corollary 7.2.3*]{}). From here it follows that there exists $\psi_i\in C_c^k(V, N\otimes \Lambda^{0,1}), i=1,...,l$, generating an $\rr $-basis of the space $\sfh^1_D(S,N)$.
Take some $\psi_i\in C^k_c(V, N\otimes \Lambda^{0,1})$. It is a $\cc$-antilinear $C^k$-smooth homomorphism from $TS\ogran_V$ into $N\ogran_V$ vanishing outside some compact in $V$. Since $u\ogran_V$ is a $C^k$-imbedding and $u(V)\subset U$, $\psi_i$ can be represented in the form $\psi_i =\pr_N \scirc \dot J \scirc du \scirc J_{S}$ for some $J$-anti-linear $C^k$-smooth endomorphism $\dot J$ of $TX$ vanishing outside some compact in $U$. Thus, $\dot J\in
T_J\calj^k_U$ and $\Psi\dot J=\psi_i^{(\nu)}$.
Corollary 8.2.3. *$\calm$ and $\calp^*$ are $C^k$-smooth Banach manifolds and $\pi_\calj:\calm \to \calj^k_U$ is a Fredholm map. For $(M,J)\in\calm$ with $M=u(S)$ there exist natural isomorphisms $$\eqalign{
\ker(d\pi_\calj:T_{(M,J)}\calm \to T_J\calj^k_U)&\;\cong\; \sfh^0_D(S,
\caln_M),
\cr
\coker(d\pi_\calj:T_{(M,J)}\calm \to T_J\calj^k_U)&\;\cong\; \sfh^1_D(S,
\caln_M),
}$$ where $\caln_M=\calo(N_u)\oplus \caln\sing_u$ is a normal sheaf to $M$, $\sfh^0_D(S, \caln_M)$ denotes $\sfh^0_D(S,N_u)$ $\oplus\sfh^0(S,\caln_u\sing)$ and $\sfh^1_D(S, \caln_M)$ denotes $\sfh^1_D(S, N_u)$. The index of $\pi_\calj$ $$\ind_\rr(\pi_\calj)=\ind_\rr(\caln_M)\deff \dimr\sfh^0_D(S, \caln_M)
-\sfh^1_D(S, \caln_M)$$ is equal to $2(c_1(X)[M]+ (n-3)(1-g))$, where $n\deff\dimc X$.*
**Proof. One easily sees that the section $\sigma_{\dbar }$ is $C^k$-smooth if $J\in \calj^k_U$. For $\calp^*$ the statement follows from [*Lemmas 8.2.1*]{} and [*8.2.2*]{}. Moreover, $\calp^*$ is a $C^k$-smooth submanifold of $\calx^*$. Further, local slices of the projection $\pi_\calp:\calp^*\to \calm$ are also $C^k$-smooth. This induces the structure of $C^k$-smooth Banach manifold on $\calm$.**
Consider the natural projection $\pi: \calp^* \to \calj^k_U$. The tangent space $T_{(u,J_S,J)}\calp^*$ consists of $(v,\dot J_S,\dot J)$ with $D_{u,J}v + \half J\scirc du \scirc \dot J_S + \half\dot J\scirc du
\scirc J_S =0$, and the differential $d\pi: T_{(u,J_S,J)}\calp^* \to
T_J\calj^k_U$ has a form $(v,\dot J_S,\dot J)\in T_{(u,J_S,J)}\calp^*\mapsto
\dot J\in T_J\calj^k_U$.
The kernel $\ker(d\pi) T_{(u,J_S,J)}\calp^*$ consists of solutions of the equation $$D_{u,J}v + \half J\scirc du \scirc \dot J_S =0
%\eqno(2.22)$$ with $v\in\cale_{(u,J_S,J)}$ and $\dot J_S\in T_{J_S}\ttt_g$. Since the map $\pi_\calp:\calp^*\to\calm$ is a principal $\bfg$-bundle, the kernel $\ker(d\pi_\calj:T_{(M,J)}\calm \to T_J\calj^k_U)$ is obtained from $\ker(d\pi)$ by taking a quotient with respect to the tangent space to a fiber $\bfg\cdot(u,J_S,J)$, which is equal to $du(\sfh^0(S,TS))$. Using relations $\sfh^0(S,TS)=
\ker(\dbar_{TS}:L^{1,p}(S,TS)\to L^p(S,TS\otimes\Lambda^{(0,1)}S)$, $T_{J_S}\ttt_g\cong\sfh^1(S,TS)=\coker(\dbar_{TS})$ and $du\scirc \dbar_{TS}= D_{(u,J)}\scirc du$, we conclude that the space $\ker(d\pi_\calj)$ is isomorphic to the quotient $$\{v\in L^{1,p}(S,E_u)\;:\; Dv=du(\phi) \hbox{ for some }\phi\in
L^p(S,TS\otimes\Lambda^{(0,1)}S)\}
\!\!\bigm/\!\!
du\bigl(L^{1,p}(S,TS)\bigr).$$ Hence, $\ker(d\pi_\calj:T_{(M,J)}\calm\to T_J\calj^k_U)\cong
\sfh^0_D(M,\caln_M)$ by [*Theorem 7.3.1*]{}. In particular, $\ker(d\pi_\calj)$ is finite dimensional.
Similarly, the image of $d\pi_\calj$ consists of those $\dot J$ for which the equation $$D_{u,J}v + \half J\scirc du \scirc \dot J_S + \half\dot J\scirc
du \scirc J_S =0$$ has a solution $(v,\dot J_S)$. Hence, $\im(d\pi_\calj)=\ker\barr\Psi$, and $\coker(d\pi) \cong \sfh^1_D(S,N_u)$. Thus, $d\pi_\calj$ is a Fredholm map. This implies the Fredholm property for the projection $\pi:\calp^* \to
\calj^k_U$.
Due to [*Corollary 7.3.2*]{}, $\ind_\rr(\caln_M)=\ind_\rr(E_u)-\ind_\rr(TS)$. Using the index and Riemann-Roch theorems, we get from $c_1(E)=c_1(X)[M]$ and $c_1(TS)=2-2g$ the needed formula $\ind_\rr(\caln)=2\bigl(c_1(X)[M] + n(1-g) - (3-3g)\bigr)=
2(c_1(X)[M] + (n-3)(1-g))$.
A straightforward application of this statement and the Sard Lemma for Fredholm maps gives the following
Corollary 8.2.4.
*If smoothness $k$ of the structures in $\calj^k$ is large enough and with $\ind_\rr (\pi_{\calj })=2(c_1(X)[\gamma ]
+(n-3)(1-g))<0$ for the homology class $\gamma $, then the following holds:*
1\) For any $(M_0,J_0)\in \calm_{\calj^k}$ with $[M]=[\gamma ]$ there is a neighborhood $W$ fo $J_0$ in $\calj^k_{\omega }$ and a closed subset $S
\subset W$ of Hausdorff codimension close to $2$ such that for $J\in W
\setminus S$ $\calm_J$ is empty;
2\) In the manifold $C^k_{[\gamma ]}(I,\calm_{\calj^k})$ of $k$-smooth paths in $\calm_{[\gamma ]} $ there is a closed subset $R$ of Hausdorff codimension close to one such that for any path $h\in C^k(I,\calm_{\calj^k}\setminus R$ the moduli space $\calm_h$ is empty.
[8.3. Components of the Moduli Space.]{}
Before stating the next results, we introduce some new notations.
Definition 8.3.1. Let $Y$ be a $C^k$-smooth finite-dimensional manifold, possibly with $C^k$-smooth boundary $\d Y$, and $h:Y \to \calj^k_U$ a $C^k$-smooth map. Define the [*relative moduli space*]{} $$\calm_h \deff Y\times_{\calj^k_U}\calm \cong
\{\,(u,J_S,y)\in \cals_U\times \ttt_g\times Y\,:\, (u,J_S, h(y))\in \calp^*
\,\}/\bfg$$ with the natural projection $\pi_h:\calm_h \to Y$. In the special case $Y=\{J\}\hook \calj^k_U$, we obtain the moduli space of $J$-complex curves $\calm_J\deff \pi_\calj\inv(J)$. The projection $\pi_h:\calm_h \to Y$ is a fibration with a fiber $\pi_h\inv(y)=\calm_{h(y)}$. We shall denote elements of $\calm_h$ by $(M,y)$, where $M=u(S)$ with $h(y)$-holomorphic map $u:S\to X$.
**Lemma 8.3.1. *Let $Y$ be a $C^k$-smooth finite-dimensional manifold, and $h:Y \to \calj^k_U$ a $C^k$-smooth map. Suppose that for some $y_0\in Y$ and $u_0\in \calm_{h(y_0)}$ the map $\barr\Psi \scirc dh: T_{y_0}Y \to \sfh^1_D(S, N_{u_0})$ is surjective. Then $\calm_h$ is a $C^k$-smooth manifold in some neighborhood of $(M_0,y_0)\in \calm_h$ with the tangent space $$T_{(M,y)}\calm_h=\ker\bigl(D\;\oplus\; \Psi\scirc dh:
\cale_{u,h(y)}\oplus T_yY \longto \cale'_{u,h(y)} \bigr)
\bigm/ du(\sfh^0(S, TS)).
\eqno(8.3.1)$$***
**Proof. Let $y\in Y$, $(u,J_S,h(y)\in\calp^*$ and $M=u(S)$, so that $(M,y)\in \calm_h$. From the proof of [*Corollary 8.2.3*]{} it follows that the image of the map $d\pi_\calj:T_{(M,h(y))}\calm\to T_{h(y)}\calj^k_U$ is equal to $\ker\bigl(\barr\Psi_{(u,h(y))}\bigr)$, and that the cokernel of $d\pi_\calj$ is mapped by $\barr\Psi$ isomorphically onto $\sfh^1_D(S, N_u)$. The statement of the lemma now follows from [*Lemma 8.2.1*]{}.**
**Definition 8.3.2. Let $Y$ be a compact manifold, $h:Y\to \calj^k_U$ a $C^k$-smooth map, $\calm_h\subset \calm \times Y$ a corresponding moduli space and $(M_0,y_0)\in \calm_h$ a point. A [*component $\calm_h(M_0,y_0)$ of $\calm_h$ through $(M_0,y_0)$*]{} is the set of those $(M,y)\in\calm_h$, that for every open neighborhood $W$ of $h(Y)\subset
\calj^k_U$ there exists a continuous path $\gamma:[0,1]\to \calm$, with the following properties:**
[*a)*]{} $\gamma(0)=(M_0,h(y_0))$ and $\gamma(1)=(M,h(y))$, i.e., $\gamma$ connects $(M_0,y_0)$ and $(M,y)$ in $\calm$;
[*b)*]{} $J_t\deff\pi_\calj(\gamma(t))\in W\subset\calj^k_U$ for any $t\in[0,1]$, i.e., the corresponding path of almost complex structures $J_t$ lies in the given neighborhood $W$ of $h(Y)$.
**Lemma 8.3.2 *Let $h:Y\to \calj^k_U$, $(M_0,y_0)\in \calm_h$ and $\calm_h(M_0,y_0)$ be as in Definition 8.3.2. Then***
$\calm_h(M_0,y_0)$ is a closed subset of $\calm_h$,
if $\calm_h(M_0,y_0)$ is compact, then there exists a set $\calm_h^0$ containing $\calm_h(M_0,y_0)$, which is compact and open in $\calm_h$;
if $\calm_h(M_0,y_0)$ is not compact, then there exists a continuous path $\gamma:[0,1)\to \calm_h$ with the following properties:
[*a)*]{} $\gamma(0)=(M_0,h(y_0))$, i.e., $\gamma$ begins at $(M_0,y_0)$;
[*b)*]{} there exists a sequence $t_n\nearrow 1$ such that the sequence $(M_n,J_n)\deff\gamma(t_n)$ lies in $\calm_h$ and is discrete there, but the sequence $\{J_n\}$ converges to some $J^*\in \calj_U^k$.
**Proof. 1.2em Let $(M',y')\in \barr{\calm_h(M_0,y_0)}\subset \calm_h$ and $J'=h(y')$. Let $W$ be any open neighborhood of $h(Y)\subset \calj_U^k$, and let $\{(M_n,y_n)\}$ be a sequence in $\calm_h$ converging to $(M',y')$. Then there exists some ball $B\ni(M',J')$ in $\calm$ whose projection on $\calj^k_U$ lies in $W$. Since some $(M_n,J_n)$ with $J_n=h(y_n)$ lies in $B$, there exists a path $(M_t,J_t)$ in $\calm$ connecting $(M_0,h(y_0))$ with $(M',J')$ such that $J_t\in W$ for any $t\in [0,1]$. This proves the closedness of $\calm_h(M_0,y_0)$.**
1.2em Let $(M',y')\in \calm_h$ and $J'=h(y')$. Fix a finite dimensional subspace $F\subset T_{J'}\calj^k_U$ such that the map $D_{u',J'} \oplus \Psi:
\cale_{u',J'} \oplus F \to \cale'_{u',J'}$ is surjective. Let $B\ni0$ be a ball in $F$. Find a $C^k$-smooth map $H:Y\times B$ such that $H(y,0)\equiv h(y)$ and $dH_{(y',0)}: T_{(y',0)}(Y\times B) \to T_{J'}\calj^k_U$ induces isomorphism $T_0B\buildrel\cong\over \longto F\subset T_{J'}\calj^k_U$. Then $\calm_H$ contains a neighborhood $W\ni (M',y',0)$ which is $C^k$-smooth manifold such that $W\cap \calm_h$ is closed in $W$. This implies that $\calm_h$ is a locally compact topological space.
Since $\calm_h(M_0,y_0)$ is a compact subset of $\calm_h$, it has an open neighborhood $V$ whose closure $\barr V \subset\calm_h$ is also compact. Let $W_i\subset \calj^k_U$ be a fundamental system of neighborhoods of $h(Y)$, so that $\cap_i W_i =h(Y)$. Let $V_i$ denote the set of those $(M,y)\in \calm_h$ such that $(M,h(y))$ and $(M_0,h(y_0))$ can be connected by a continuous path $(M_t, J_t)$ in $\calm$ with $J_t$ lying in $W_i$. Then $\cap V_i =\calm_h(M_0,y_0)$. The same arguments as in the part of the proof show that every $V_i$ is both open and closed in $\calm_h$.
We state that there exists a number $N\in \nn$ such that $\bigl(\cap_{i=1}^N V_i\bigr) \cap\barr V \subset V$. If it is false, then for any $n\in \nn$ there would exist $(M_n,y_n)\in \bigl(\cap_{i=1}^n V_i\bigr)
\cap\barr V \bes V$. But, in this case, some subsequence of $\{(M_n,y_n)\}$ would converge to some $(M^*,y^*)\in\bigl(\cap_{i=1}^\infty V_i\bigr)
\cap\barr V \bes V$, which is impossible. For such $N\in \nn$ the set $$\def\capl{\mathop{\cap}}
\calm^0_h\deff \left(\capl_{i=1}^N V_i\right) \cap\barr V
=\left(\capl_{i=1}^N V_i\right) \cap V$$ satisfies the conditions of the lemma.
1.2em Suppose that $\calm_h(M_0,y_0)$ is not compact. Then there exists a discrete sequence $\{(M_n,y_n)\}$ in $\calm_h(M_0,y_0)$. Since $Y$ is compact, we may assume that $y_n$ converges to some $y^*$. For any $n\in \nn$ fix a path $\gamma_n:[0,1] \to \calm_h$ connecting $(M_{n-1}, h(y_{n-1}))$ with $(M_n,h(y_n))$. Set $t_n\deff
1-2^{-n}$. For $t\in [t_{n-1},t_n]$ define $\gamma(t)\deff
\gamma_n(2^n(t-t_{n-1}))$. Then $\gamma:[0,1)\to \calm_h$ and $t_n\nearrow1$ satisfy the conditions of the lemma.
Theorem 8.3.3.
*Let $(M_0,J_0)\in \calm$ and let $h:[0,1]\to \calj^k_U$ be $C^k$-smooth and with $h(0)=J_0$. Suppose that there exists a subset $\calm_h^0$ of $\calm_h$ which is compact, open and contains $(M_0,J_0)$. Suppose also that $\ind(\pi_\calj)=2(c_1(X)[M_0]+ (n-3)(1-g))$ is non-negative. Then $h$ can be $C^k$-approximated by smooth maps $h_n:[0,1] \to \calj^k_U$ such that*
every $\calm_{h_n}$ contains a component $\calm_{h_n}^0$ which is a smooth connected compact manifold of the expected dimension $\dimr(\calm_{h_n}^0)=\ind(\pi_\calj)+1$;
$\calm_{h_n(0)}^0\deff \pi_{h_n}\inv(0)\cap \calm_{h_n}^0$ and $\calm_{h_n(1)}^0\deff \pi_{h_n}\inv(1)\cap \calm_{h_n}^0$ are also smooth manifolds of the expected dimension $\ind(\pi_\calj)$ ($\calm^0_{h_n(1)}$ may be empty), and $\calm_{h_n}$ is a smooth bordism between $\calm_{h_n(0)}^0$ and $\calm_{h_n(1)}^0$;
every $\calm_{h_n(0)}^0$ is connected with $(M_0,J_0)$ by a path in $\calm$, i.e., there exists a $C^k$-path $\gamma_n: [0,1] \to \calm$ such that $\gamma_n(0)= (M_0,J_0)$ and $\gamma_n(1)\in \calm_{h_n(0)}^0$; in particular, every $\calm_{h_n(0)}^0$ is nonempty;
for every $(M,J)\in \calm_{h_n}$ one has $\dimr\sfh^1_D(S,N_M)\le1$.
Proof. Denote $\calm_h^0$ by $K$. Let $\cale_K$ and $\cale'_K$ be the pull-backs onto $K$ of the bundles $\cale\to \calm$ and $\cale'\to
\calm$, respectively. Further, let $\calt\deff h^*T\calj^k_U$ be the pull-back of the tangent bundle $T\calj^k_U$ onto $[0,1]$.
Due to [*Lemma 8.2.2*]{}, for every $(M,J)\in K$ with $M=u(S)$ we find that $n_{(M,J)}\in \nn$ and a $C^k$-smooth homomorphism $P_{(M,J)}: F_{(M,J)} \to \calt$ form a trivial vector bundle $F_{(M,J)}$ over $[0,1]$ of rank $\rank F_{(M,J)}=n_{(M,J)}$ such that the operator $$D_{(u,J)}\oplus \Psi_{(u,J)} \scirc P_{(M,J)} :
\cale_{(M,J)}\oplus F_{(M,J)} \to \cale'_{(M,J)}
\eqno(8.3.2)$$ is surjective.
Since $K$ is compact, we can take an appropriate finite collection of $(F_{\!(M,J)},$ $P_{\!(M,J)})$ and construct a homomorphism $P: F\to \calt$ from a trivial vector bundle $F\cong [0,1]\times \rr^n$ such that $D_{(u,J)}\oplus \Psi_{(u,J)} \scirc P : \cale_{(M,J)}
\oplus F \to \cale'_{(M,J)}$ is surjective for all $(M,J)\in K$.
For $\dot J$ lying in some small ball in $T_J\calj^k_U$ we set $\exp_J(\dot J)\deff J{(1-J\dot J/2)\over(1+J\dot J/2)}$. Differentiating the identity $J^2=-1$ gives the relation $J\dot J =- \dot JJ$, and consequently the equality $$\left(J{(1-J\dot J/2)\over(1+J\dot J/2)} \right)^2=-1,$$ which means that $\exp_J$ takes values in $\calj^k_U$. Further, it is easy to see that the differential of $\exp_J$ in $0\in T_J\calj^k_U$ is the identity map of $T_J\calj^k_U$. Thus, $\exp_J$ is a natural exponential map for $\calj^k_U$.
Take a sufficiently small ball $B=B(0,r)\subset \rr^n$ and define a map $H^*:[0,1]\times B \to \calj^k_U$ by setting $H^*(t, y)\deff \exp_{h(t)}(P(t,y))$. The construction of $H^*$ provides that for all $(M,J)\in K$ with $J=h(t)$ the map $D_{(u,J)}\oplus \Psi_{(u,J)} \scirc dH^*(t,0) : \cale_{(M,J)}
\oplus T_{(t,0)}([0,1]\times B) \to \cale'_{(M,J)}$ is surjective. Making an appropriate small perturbation of $H^*$, we obtain a $C^k$-smooth function $H: [0,1]\times B \to \calj^k_U$ with the following properties:
for $H(t,0)=h(t)$, i.e., $H$ is a deformation of $h$ with a parameter space $B$;
$
%\displaystyle
D_{(u,J)}\oplus \Psi_{(u,J)} \scirc dH(t,0) : \cale_{(M,J)}
\oplus T_0B \longto \cale'_{(M,J)}
$ is surjective for all $(M,t)\in K$ with $J=h(t)$ and $M=u(S)$;
in a neighborhood of every $(t,y)\in [0,1]\times B$ with $y\not=0$ the map $H$ is $\calj^\infty_U$-valued and $C^\infty$-smooth.
Let us identify $K$ and $\calm_h$ with the subset of $\calm_H$ using natural continuous imbedding $(M,t)\in \calm_h \mapsto (M,t,0)\in \calm_H$. Due to [*Lemma 8.2.2*]{}, there exists a neighborhood $V$ of $K$ in $\calm_H$ which is $C^k$-smooth manifold. Taking a smaller neighborhood of $K$ if it is needed, we may assume that the closure $\barr V$ of $V$ is compact and does not meet connected components of $\calm_h$ different from $\calm_h^0=K$. Let $p:V \to B$ be the natural projection such that $(u,t,y)\mapsto y$.
We state that there exists a smaller ball $B_1=B(0,r_1)\subset B$ such that for every $y\in B_1$ the set $p\inv(y)\subset V$ is compact. Suppose the contrary is true. Then there would exist a sequence $y_n\in B$ converging to $0\in B$ such that $p\inv(y_n)$ are not compact. Since $\barr V$ is compact, there would also exist such $u_n$ and $t_n$ that $(u_n,t_n,y_n)$ lies in the closure $\barr{p\inv(y_n)}\subset \barr V$ but not in $V$. Taking an appropriate subsequence we may assume that $(u_n,t_n,y_n)$ converges to $(u^*,t^*,0)$. However, in this case $(u^*,t^*,0)\in \calm_h \cap \barr V$ and hence $(u^*,t^*,0)\in V$. On the other hand $\barr V\bes V$ is compact, and thus $(u^*,t^*,0)$ must belong to $\barr V\bes V$. The obtained contradiction shows that the statement is true.
Set $V_1\deff p\inv(B_1)$. Due to the choice of $H$ the resticted projection $p:V_1\bes K\to \check B_1 \deff B_1\bes\{0\}$ is $C^\infty$-smooth. Due to the Sard lemma (see, e.g., \[Fed\], §[**3.4**]{}), there exists a dense subset $B_1^* \subset B_1$ such that for any $y\in B_1^*$ the set $p\inv(y)$ is a $C^\infty$, a smooth compact manifold. Fix a sequence $y_n\in B_1^*$, converging to $0\in B$ and set $h_n(t)\deff
H(t,y_n)$. Also set $\calm_{h_n}^0 \deff \calm_{h_n} \cap V$, so that $\calm_{h_n}^0 = p\inv(y_n)$. Then every $\calm_{h_n}^0$ is a $C^\infty$-smooth nonempty manifold, which is connected with $(M_0,J_0)$ by a path in $\calm$.
Due to [*Lemma 8.2.3*]{}, the tangent space to $V_1\subset \calm_H$ at $(u,t,y)$ is canonically isomorphic to $$\ker\!\Bigl(\!D_{u,H(t,y)}\,\oplus\, \Psi\scirc dH:
\cale_{u,H(t,y)}\,\oplus\, T_{(t,y)}\bigl([0,1]{\times} B_1\!\bigr)
\to \cale'_{u,H(t,y)} \!\Bigr)
\!\!\!\Bigm/ \!\!\!du(\sfh^0(S\!\!,\! TS)).$$ Since $p: V_1 \to B_1$ is a projection of the form $(u,t,y)\in V_1
\mapsto y\in B_1$, the differential $dp_{(u,t,y)}$ maps the tangent vector of the form $(\dot u,\dot t,\dot y)\in T_{(u,t,y)}V_1$ into $\dot y
\in T_yB_1$. This means that $dp_{(u,t,y)}$ is a restiction on $\ker(D_{u,H(t,y)}\,\oplus\, \Psi\scirc dH)$ of the linear projection $p_B:\cale_{u,H(t,y)} \oplus\, T_{(t,y)}\bigl([0,1]{\times} B_1\!\bigr)
\to T_y B_1$ such that $p_B(\dot u,\dot t,\dot y)=\dot y$. In particular, for $y=y_n$ the map $dp_{(u,t,y)}$ is surjective, which means the surjectivity of the map $$p_B: \ker \bigl(D_{u,H(t,y)}\,\oplus\, \Psi\scirc dH \bigr)
\longto T_{y_n}B_1.$$ This is equivalent to the surjectivity of $$D_{u,H(t,y_n)}\,\oplus\, \Psi\scirc dH\,\oplus\,p_B:
\cale_{u,H(t,y_n)}\,\oplus\, T_{(t,y_n)}\bigl([0,1]{\times} B_1\!\bigr)
\to \cale'_{u,H(t,y_n)} \oplus T_{y_n}B_1$$ and hence to the the surjectivity of $$D_{u,h_n(t)}\,\oplus\, \Psi\scirc dh_n:
\cale_{u,h_n(t)}\,\oplus\, T_t[0,1]
\to \cale'_{u,h_n(t)}.
\eqno(8.3.3)$$ Consequently, $\dimr\sfh^1_D(S, N_M)\le1$ for any $(M,t)\in \calm_{h_n}\cap
V_1$.
Corollary 8.3.4. *Under the conditions of Theorem 8.3.3 suppose additionally that $S$ is a sphere $S^2$. Then for all points $(M,t)\in \calm^0_{h_n}$ the associated $D_N$-operator is surjective, i.e., $\sfh^1_{D_N}(S^2,N_{M_t})=0$.*
Moreover, $\calm_{h_n}$ is a trivial bordism: $\calm_{h_n(0)}\times [0,1]$. In particular, for every $h_n(0)$-holomorphic sphere $M_0\in \calm_{h_n(0)}$ there exists a continuous family of $h_n(t)$, holomorphic spheres $M_{n,t}=
u_{n,t}(S^2)$ with $M_{n,0}=M_0$.
*Proof. Suppose that for some $(M,t)\in \calm^0_{h_n} $ we have $\sfh^1_{D_N}(M,N_M)\not= 0$. Then by (iv) of [*Theorem 8.3.3*]{} $\sfh^1_{D_N}(M,N_M)= 1$. But this contradicts [*Theorem 7.3.1*]{} for the case $S=S^2$ and $L=N_M$.*
Let for $(M, t)\in \calm^0_{h_n}$ we have $M =u(S)$ and $J=h_n(t)$. Let also $\dot J \not= 0 \in dh_n(T_t[0,1])$. Then by [*Lemma 8.2.1*]{} and [*Corollary 8.2.3*]{} the tangent space $’_{(M,t)} \calm^0_{h_n}$ is canonically isomorphic to
$$\ker\!\Bigl(\!D_{u,J}\,\oplus\, \Psi:
\cale_{u,J}\,\oplus\, \rr\,\dot J\!\bigr)
\to \cale'_{u,J} \!\Bigr)
\!\!\!\Bigm/ \!\!\!du(\sfh^0(S\!\!,\! TS)),$$ and the differential of the projection $d\pi_{h_n}: ’_{(M,t)} \calm^0_{h_n}
\to T_t[0,1] \cong \rr$ is of the form $d\pi_{h_n}[v,a\dot J] = a$. While in the case $S=S^2$ the space $\sfh^1(S, TS)$ is trivial, [*Corollary 8.3.4*]{} insures the surjectivity of the operator $D_{u,J}: \cale_{u,J} \to \cale'_{u,J}$. Thus for $a\not=0 \in \rr$ there exists such a $v\in \cale_{u,J}$ that $[v,a\dot J]\in ’_{(M,t)} \calm^0_{h_n}$. This means that for every $(M, t)\in \calm^0_{h_n}$ the projection $d\pi_{h_n}: ’_{(M,t)}
\calm^0_{h_n} \to T_t[0,1]$ is surjective. While the manifold $\calm^0_{h_n}$ is compact, there is a diffeomorphism $\calm_{h_n} \cong
\calm_{h_n(0)}\times [0,1]$.
*8.4. Moduli of Parameterized Curves.*
Sometimes it may be useful to consider the moduli spaces of parameterised complex curves. Therefore, in this paragraph $\calp $ will denote the space of parameterized complex curves on $X$. Suppose that $(u, J_S, J) \in \calp$, [*i.e.,*]{} $u$ belongs to $C^1(S, X)$ and satisfies the equation $du\scirc J_S = J \scirc
du$. Set $$T_{(u, J_S, J)}\calp\deff \bigl\{\, (v,\dot J_S,\dot J) \in T_u\cals
\times T_{J_S} \calj_S \times T_J\calj \, :\,
2D_{u, J}v + \dot J \scirc du \scirc J_S +
\dot J_S \scirc du \scirc J =0 \,\bigr\}.$$ Let $\pr_\calj: \cals \times \calj_S \times \calj \to \calj$ and $\pr_{(u, J_S, J)} : T_{(u, J_S, J)} \calp \to T_J\calj$ denote the natural projections.
Theorem 8.4.1.
*The map $\pr_{(u, J_S, J)} : T_{(u, J_S, J)}\calp
\to T_J\calj$ is surjective $\sfh^1_D(S, N_0)$ $=0$, and then the following hold:*
the kernel $\ker(\pr_{(u, J_S, J)})$ admits a closed complementing space;
for $(\tilde u,\tilde J_S,\tilde J)\in \calp$ close enough to $(u, J_S, J)$ the projection $\pr_{(\tilde u,\tilde J_S,\tilde J)}$ is also surjective;
for some neighborhood $U\subset \cals \times \calj_S
\times \calj$ of $(u, J_S, J)$ the set $\calp \cap U$ is a Banach submanifold of $U$ with the tangent space $T_{(u, J_S, J)}\calp$ at $(u, J_S, J)$;
there exists a $C^1$-map $f$ from some neighborhood $V$ of $J\in \calj$ into $U$ with $f(V)\subset \calp$, $f(J) = (u,J_S,J)$ and $\pr_\calj
\scirc f = \id_V$ such that $\im (df: T_J\calj \to
T_{(u, J_S, J)}\calp)$ is complementing to $\ker(\pr_{(u, J_S, J)})$.
Proof. Denote by $\tilde A$ the set of a singular point of $M=u(S)$, [*i.e.*]{}, the set of cuspidal and self-intersection points of $M$. Let $\xi \in L^p_{(0, 1)} (S, E)$. Using the local solvability of the $D$-equation (proved essentially in [*Lemma 3.2.1*]{}) and an appropriate partition of unity, we can find $\eta\in L^{1,p}(S, E)$ such that $\xi -2D\eta \in C^1_{(0, 1)}(S, E)$ and $\xi -2D\eta =0$ in a neighborhood of $\tilde A$. Then we find $\dot J\in T_J\calj \equiv
C^1_{(0, 1)}(X, TX)$ such that $\dot J \scirc du \scirc J_S =\xi-
2D\eta$. The surjectivity of $\pr_{(u, J_S, J)}$ and [*Theorem 7.3.1*]{} easily yield the identity $\sfh^1_D(S, N_0)=0$.
*From now on and until the end of the proof we suppose that $\sfh^1_D(S, N_0)=0$.*
Let $ds^2$ be some Hermitian metric on $S$ and let $\calh_{(0, 1)} \subset
C^1_{(0, 1)}(S, TS)$ be a space of $ds^2$-harmonic $TS$-valued $(0, 1)$-forms on $S$. Then the natural map $\calh_{(0, 1)} \to
\sfh^1(S, TS)$ is an isomorphism. Furthermore, due to [*Corollary 7.3.2*]{}, the map $$g\deff(du, 2D): \calh_{(0, 1)} \oplus L^{1,p}(S, E) \to
L^p_{(0, 1)}(S, E)$$ is surjective and has a finite-dimensional kernel. Consequently, there exists a closed subspace $Y\subset \calh_{(0, 1)} \oplus L^{1,p}(S, E)$ such that $g : Y \to L^p_{(0, 1)}(S, E)$ is an isomorphism. Let $$h=(h_{TS},h_E): L^p_{(0, 1)}(S, E) \to Y\subset \calh_{(0, 1)}
\oplus L^{1,p}(S, E)$$ be the inversion of $g\vert_Y$.
Take $\dot J\in T_J\calj= \{\, I\in C^1(X, {\sl End}(TX))\, :\,
JI+IJ=0\,\} \equiv C^1_{(0, 1)}(X, TX)$. Then $\dot J \scirc du
\scirc J_S$ lies in $C^0_{(0, 1)}(S, E)$. Let $h(\dot J \scirc du
\scirc J_S) =(\xi,\eta)$ with $\xi\in \calh_{(0, 1)}
\subset C^1_{(0, 1)}(S, TS)$ and $\eta\in L^{1,p}(S, E)$. Then we obtain $$2D(-\eta) + \dot J \scirc du \scirc J_S +
J \scirc du ( J_S \xi)=0,\eqno(8.4.1)$$ where we use the identity $J\scirc du \scirc J_S =-du$. Using that $T_{J_S}\calj_S =$ $ C^1_{(0, 1)}(S, TS)$, we conclude that the formula $F(\dot J)= \bigl(J_S h_{TS}(\dot J \scirc du
\scirc J_S), -h_E(\dot J \scirc du \scirc J_S),\dot J \bigr)$ defines a [*bounded*]{} linear operator $F: T_J\calj \to T_{(u, J_S, J)}\calp$ such that $\pr_{(u, J_S, J)}\scirc F = \id_{T_J\calj}$. In particular, $\sfh^1_D(S,
N_0)=0$ implies the surjectivity of $\pr_{(u, J_S, J)}$.
The image of the just defined operator $F$ is closed, because the convergence of $F(\dot J_n)$, $J_n\in T_J\calj$ obviously yields the convergence of $J_n=\pr_{(u, J_S, J)}\scirc F(\dot J_n)$. One can easily see that $\im(F)$ is a closed complementing space to $\ker(\pr_{(u, J_S, J)})$.
Further, for $(\tilde u, \tilde J_S,\tilde J) \in \calp$ close enough to $(u, J_S, J)$ the map $\tilde g \deff (d\tilde u, 2D_{\tilde u,\tilde J}):
Y \to L^p_{(0, 1)}(S,\tilde E)$ is also an isomorphism. This implies the surjectivity $\pr_{(\tilde u,\tilde J_S,\tilde J)}$.
The statements and can easily be obtained from [*ii)*]{} and the implicit function theorem.
*8.5. Gromov Non-squeezing Theorem.*
We shall finish this chapter with the proof of the Gromov non-squeezing theorem.
Consider an infinite cylinder $\zz (\lambda ):=\Delta (\lambda )
\times \rr^{2n-2}$ in $\rr^{2n}$. Here $\Delta (\lambda )$ is a disk of radius $\lambda $ in the $x_1,y_1$-plane, and on $\rr^{2n-2}$ we fix coordinates $x_2,y_2,...,x_n,y_n$. By $\bb (R)$ we denote the ball of radius $R$ in $\rr^{2n}$. We consider on $\rr^{2n}$ the standard symplectic form $\omega =\Sigma_{i=1}^ndx_i\wedge dy_i$.
Let $D\subset \rr^{2n}$ be a domain. Recall that a smooth map $f:D\to \rr^{2n}$ is called a symplectomorphism if $f^*\omega =\omega $. In particular, symplectomorphisms preserve the Euclidean volume.
Exercise. [Find a volume preserving a linear map from $\bb (R)$ to $\zz (\lambda )$ for arbitrary $R$ and $\lambda $. ]{}
The following theorem shows how far symplectomorphisms are from volume preserving maps.
Theorem 8.5.1. [*(Gromov Non-squeezing Theorem). If there exists a symplectomorphism $f:\bb (R) \to \zz (\lambda )$, then $R\le \lambda$*]{}.
Proof.
Taking a smaller ball, if nessessary, and translating the image, we can suppose that $f(\bb (R))$ is contained in some compact $\bar\Delta(0,\lambda')\times [0,a]^{2n-2}$ of $\Delta (0,\lambda )\times
\rr^{2n-2}$. Factorizing $\rr^{2n-2}$ by the lattice $a\cdot \zz^{2n-2}$, we observe that $\overline{f(\bb (R))}\comp \Delta (0,\lambda )\times
T^{2n-2}$. Moreover, our syplectic form $\omega $ descends onto this factor and still can be decomposed as $\omega = \omega_1 + \omega_2$. Here $\omega_1 = dx_1\wedge dy_1$ and $\omega_2=\Sigma_{i=2}^ndx_i\wedge dy_i$.
Further, we compactify the disk $\Delta (0,\lambda )$ to a two-dimensional sphere $S^2$ by an $\eps $-disk $\Delta (0,\eps )$ and extend $\omega_1$ to a smooth strictly positive $(1,1)$-form, still denoted by $\omega_1$ on $S^2$, and having $$\int_{S^2}\omega_1 = \pi \lambda^2 + \eps . \eqno(8.5.1)$$ Note that now we have a symplectic imbedding $f:\overline{\bb (R)}
\to S^2\times T^{2n-2}$. Also note that our symplectic form $\omega $ on $S^2\times T^{2n-2}$ tames the standard complex structure $J\st$ on this manifold and satisfies $$\int_{[\gamma ]}\omega = \pi \lambda^2 + \eps , \eqno(8.5.2)$$ where $[\gamma ]=[S^2\times \{ pt\} ]$. We further observe that $\calm_{[\gamma ],J\st}$ consists of $\{ \cc\pp^1
\times \{ a\} :a\in T^{2n-2}\} $, i.e., this moduli space is diffeomorphic to a torus $T^{2n-2}$.
Lemma 8.5.2. *(a) For a generic almost-complex structure $J$ on $S^2\times T^{2n-2}$ tamed by $\omega $ the moduli space $\calm_{[\gamma ],J}$ is diffeomorphic to $T^{2n-2}$.*
\(b) Moreover, for any (not nessessarily generic!) $J$ and for any point $p\in S^2\times T^{2n-2}$ there exists a $J$-complex curve from $\calm_{[\gamma ],J}$ passing through $p$.
Proof.
\(a) For any $J\in \calj_{\omega }$, any $J$-complex curve $C$ representing $[\gamma ]$ obviously cannot be decomposed as $C=C_1\cup C_2$, where $C_k$ are $J_k$-complex for some $J_k\in \calj_{\omega }$. Therefore, we can apply Theorem 3.2 from the introduction to this chapter. Thus (a) is proved.
\(b) In fact, for a generic $J$ we have a little bit more. Namely, for a generic curve $h:[0,1]\to \calj_{\omega }$ with $h(0)=J\st $ consider an evaluation map $\ev :\calc_h\to S^2\times T^{2n-2}$, where $\calc_h$ is the universal $\cc\pp^1$-bundle over $\calm_{h}$. Recall that $\calm_h$ is a manifold diffeomorphic to $\calm_{h(0)}\times [0,1]$. This yields that $\calc_h$ is diffeomorphic to $\calc_{h(0)}\times [0,1]$.
Now, from the uniqueness properties of complex curves it easily follows that for any $t\in [0,1]$ the map $\ev :\calc_{h(t)}\to
S^2\times T^{2n-2}$ is surjective. Thus, we prove $\ev :\calc_{J}\to
S^2\times T^{2n-2}$ is surjective for generic $J\in \calj_{\omega }$.
Since generic structures are dense in $\calj_{\omega }$, we immediately obtain that the evaluation map $\ev :\calc_{h(t)}\to
S^2\times T^{2n-2}$ is surjective for all $J\in \calj_{\omega }$.
This completes the proof of the lemma.
Let us return to the proof of our theorem. Using [*Proposition 1.2.1*]{} from [*Lecture 1*]{}, we can extend the structure $f_*J\st $ from $f(\bb (R))$ onto $S^2\times T^{2n-2}$ to a $\omega $-tamed structure $J$. By [*Lemma 8.5.2*]{} there is a $J$-complex rational curve $C\ni f(0)$. Remark that $$\int_C\omega = \pi \lambda^2 + \eps , \eqno(8.5.3)$$ because $[C]=[S^2\times \{ pt\} ]$. While $f$ is a symplectomorphism, $$\int_{f^{-1}(C\cap f(\bb (R))}\omega\st \le \pi \lambda^2 + \eps .
\eqno(8.5.4)$$ But $f^{-1}(C\cap f(\bb (R))$ is a complex curve in $\bb (R)\subset \cc^n$ passing through zero. By a well-known estimate due to Alexander-Taylor-Ullman, see \[A-T-U\], the area of such a curve is at least $\pi R^2$. So $\pi R^2\le
\pi \lambda^2 +\eps $.
[8.6. Exceptional Spheres in Symplectic $4$-Manifolds.]{}
[*Corollary 8.3.4*]{} enables us to prove an interesting result about exceptional spheres in symplectic $4$-manifolds. Recall that a smooth rational curve $C$ in a smooth complex surface $X$ is called exceptional if $[C]^2=-1$. Such curve can be contructed to a point, i.e., there exists a smooth complex surface $Y$ and a holomorphic map $h:X\to Y$ such that $h(C)=y$ is a point and $h\mid_{X\setminus C}:X\setminus C\to Y\setminus
\{ y\} $ is a biholomorphism.
Consider now a symplectic $4$-manifold $(X,\omega )$ and let $M$ be a symplectic $2$-sphere imbedded to $X$, i.e., there is an imbedding $u:\ss^2\to M \hookrightarrow X$ with $u^*\omega$ nowhere zero.
Definition 8.6.1. [*Call $M$ exceptional if $[M]^2=-1$.* ]{}
Using [*Lemmas 1.4.2 and 1.4.3*]{} we can construct an almost-complex structure $J\in \calj_{\omega }$, which is moreover integrable in the neighborhood of $M$ and such that $M$ becomes $J$-complex. Now the complex analytic statement mentioned above allows us to construct this $M$ to a point.
Corollary 8.6.1. [*Take a maximal, linearly independant in $\sfh_2(X,\zz )$ system of exceptional symplectic spheres $M_1,...,M_k$ in $X$. Then there exists $J\in \calj_{\omega }$ making some symplectic spheres $\tilde
M_1,...,\tilde M_k$, with $\tilde M_j$ being isotopic to $M_j$ for $1\le j\le k$, all $J$-complex.* ]{} Proof.
For $k=1$ this is the statement of [*Lemma 1.4.2*]{}. Suppose we prove this statement for $k-1$ exceptional spheres. Take our spheres $M_j,j=1,...,k$ and find a structure $J$ such that $\tilde M_j$ are isotopic to $M_j$ and $J$- complex, $j=1,...,k-1.$ The normal Gromov operator $D_{J,\tilde M_j}$ is surjective on all $\tilde M_j,j=1,...1k-1$ by [*Corollary 7.2.3*]{} and thus by [*Theorem 7.3.1*]{} and [*Corollary 8.2.3*]{} the projection $\pi_{\calj }:\calm \to \calj $ is a diffeomorphism in the neighborhood of $\tilde M_j$ and $J$ for all $j=1,...,k-1$.
Take a structure $J_1$ such that $M_k$ becomes $J_1$-complex. Again the projection $\pi_{\calj }$ is regular in the neighborhood of $M_k$. Using [*Corollary 8.2.4*]{}, we take a generic path $h$ starting close to $J$, say at $J´$ and ending close to $J_1$ such that all $\calm_h(\tilde M_j´,J´)$ are compact and open in corresponding moduli spaces. Here $\tilde M_j´$ are $J´$-complex spheres isotopic to $\tilde M_j$. Applying [*Corollary 8.3.4*]{}, we obtain an isotopy of $\tilde M_j´$ to a $J''$-complex spheres $\tilde M_j''$ with $J''$ close to $J_1$. It remains to take a $J''$-complex sphere $\tilde M_k''$ isotopic to $M_k$.
[Chapter IV. Envelopes of Meromorphy of Two-spheres.]{}
This chapter is devoted to the study of envelopes of meromorphy of neighborhoods of two-spheres in complex algebraic surfaces. The original question which motivated our studies was asked by A. Vitushkin.
*Let $M$ be a “small” perturbation of the complex line in $\cc\pp^2$. Does there exist a nonconstant holomorphic function in the neighborhood of such an $M$?*
It was asked as a test question on the way to searching for the solution to the Jakobian conjecture, and the answer, as expected, was negative.
Let us briefly test the possible approaches to the answer to this question. It is more or less clear that one should try to extend holomorphic (or meromorphic via expected nonexistence of holomorphic) functions onto the whole $\cc\pp^2$ and then conclude
[*1.*]{} If $M$ is a complex line itself, then nonconstant holomorphic functions do not exist in any neighborhood of $M$ for the following reason. First note that in this case $\cc\pp^2\setminus M=\cc^2$. Let $B_N$ denote the closed ball of radii $N$ in $\cc^2$. Then $V_N:=\cc\pp^2\setminus B_N$ is a fundumental system of strictly pseudoconcave neighborhoods of $M$. Any function holomorphic in $V_N$ holomorphically extends onto the ball $B_N$ by Hartogs’ theorem, and thus becomes holomorphic on the whole $\cc\pp^2$, i.e., constant.
One can try to construct such an exhaustion for any $M$. But the generic $M$ is totally real outside of three positive elliptic points (this follows from the Lai formulae and cancellation theorem of Kharlamov-Eliashberg, see Appendix IV) and one can check that a totally real disk has no small concave tubular neighborhoods.
[*2.*]{} As was just pointed out, a generic perturbation $M$ will have exactly three elliptic points with complex tangents. This makes it problematic to attempt to construct a family of complex disks (or any other Riemann surfaces) with boundary on $M$ starting from our elliptic points. Such an approach to the construction of the holomorphic envelope of $M$ seems to be perfect only for specially imbedded spheres, see \[B-G\],\[B-K\],\[E\], \[Sch\] and \[F-M\].
[*3.*]{} The method used here to study envelopes of meromorphy of spheres is based on Gromov’s theory of pseudoholomorphic curves. We remark first of all that a small perturbation of a complex sphere is symplectic:
Definition. *A $C^1$-smooth immersion $u\:S\to(X,\omega)$ of a real surface $S$ into a symplectic manifold $(X,\omega )$ is called [*symplectic*]{} if $u^*\omega$ does not vanish anywhere on $S$.*
Thus, we shall study envelopes of meromorphy of neighborhoods of two-spheres symplectically immersed in complex surfaces. Here a [*complex surface*]{} means a (Hausdorff) connected complex two-dimensional manifold $X$ countable at infinity.
The idea is to perturb the complex structure in the given neighborhood of $M$ in such a way that $M$ becomes complex; see [*Lemma 1.4.2*]{}, where an appropriate family $J_t$ of tamed almost complex structures is constructed. Using results from Chapter III, we then construct a family $M_t$ of $J_t$-complex spheres and try to extend functions along this family. In more formal language to extend “along a family” means “onto the envelope of meromorphy”.
Let $U$ be a domain in $X$. Its envelope of meromorphy $(\widehat U,\pi)$ is the maximal domain over $X$ satisfying the following conditions:
\(i) there exists a holomorphic embedding $i\:U\to\widehat U$ with $\pi\circ i=\id_U$;
\(ii) each meromorphic function $f$ on $U$ extends to a meromorphic function $\widehat f$ on $\widehat U$, that is, $\widehat f\circ i=f$.
The envelope of meromorphy exists for each domain $U$. This can be proved, for example, by applying the Cartan–Thullen method to the sheaf of meromorphic functions on $X$, see \[Iv-1\].
In the sequel we shall restrict ourselves to Kähler complex surfaces, that is, we assume that $X$ carries a strictly positive closed $(1,1)$-form $\omega$.
The aim of the present chapter is to prove the following result.
Theorem 4.1. *Let $u\:S^2\to X$ be a symplectic immersion of the two-sphere $S^2$ in a disk-convex Kähler surface $X$ such that $M:=u(S)$ has only positive double points. Assume that $c_1(X)[M]>0$. Then the envelope of meromorphy $(\widehat U,\pi)$ of an arbitrary neighborhood $U$ of $M$ contains a rational curve $C$ with $\pi^*c_1(X)[C]>0$.*
The definition of disk-convexity is given in Lecture 10. At this point, we only observe that all compact manifolds are disk-convex. As usual, $c_1(X)$ is the first Chern class of $X$. In Appendix IV, Corollary A4.3.2 we explain why the condition $c_1(X)[M]>0$ is nessessary.
The exposition includes a construction of a complete family of holomorphic deformations of a non-compact complex curve in a complex manifold, parameterized by a finite codimension analytic subset of a Banach ball. This will be carried out in Lecture 9.
The existence of this family is used to prove a generalization of Levi’s continuity principle, which is applied to describe envelopes of meromorphy.
[*4.*]{} Another natural approach is due to S. Nemirovski. Suppose that for some neighborhood $V\supset M$ there is a nonconstant holomorphic function in $V$. Then from the result of Fujita, \[Fu\], it follows that there exists a Stein domain over $\cc\pp^2$ which contains $M$. One can imbed by the Stout theorem, \[St\], some neighborhood of $M$ into a compact algebraic surface $X$ and observe that $b_+^2(X)>1$. The Seiberg-Witten theory imposes an adjunction inequality on $M$: $$\vert c_1(X)[M]\vert + [M]^2\le 0,$$ see \[K-M\]. But $c_1(X)[M]=c_1(\cc\pp^2)[M]=3$ and $[M]^2=1$, a contradiction.
This method seems to work only in some special algebraic surfaces like $\cc\pp^2$ or $\cc\pp^1\times \cc\pp^1$ but does not require that $M$ be symplectic! We shall give more details in Appendix V.
[Lecture 9]{}
[Deformation of Noncompact Curves.]{}
[9.1. Banach Analytic Sets.]{}
Now we shall study a process of local deformation of noncompact complex curves in complex manifolds. We shall see that the moduli have nice analytic structure. Let start from the following
Definition 9.1.1.
*By a [*Banach ball* ]{} we mean a ball in some complex linear Banach space. A subset $\calm$ of the Banach ball $B$ is said to be a Banach analytic set of the finite codimension, b.a.s.f.c., if there exists a holomorphic map $F:B\to \cc^N$ with $N<\infty$ such that $\calm =
\{ x\in B: F(x)=0\}$.*
The importance of this notion consists in the fact that contrary to the general Banach analytic sets b.a.s.f.c., their properties are similar to the finite dimensional analytic sets. Namely, the following is true.
Theorem 9.1.1. (\[Ra\]) *Let $B$ be a ball in a Banach space $\cal F$, $\calm \subset B$ — a b.a.s.f.c. and $x_0$ a point in $\calm $. Then there exists a neighborhood $U\ni x_0$ in $B$ such that $\calm \cap U$ is a union of finite number of irreducible b.a.s.f.c. $\calm_j$.*
Moreover, each $\calm_j$ can be represented as a finite ramified covering over a domain in some closed linear subspace ${\cal F}_j
\subset {\cal F}$ of finite codimension.
The aim of this paragraph is to prove an existence of a complete family of holomorphic deformations of a stable curve over a complex manifold $X$, which is parameterized by a b.a.s.f.c. Before stating the result, let us introduce the following definition.
Let $C$ be a nodal curve, $E$ a holomorphic vector bundle over $C$, and $C= \cup_{i=1}^l C_i$ a decomposition of $C$ into irreducible components. Suppose that $E$ extends sufficiently smoothly to the boundary $\d C$.
Definition 9.1.2. [*Let us define an [*$L^{1,p}$-section $v$*]{} of the bundle $E$ over $C$ as a couple $(v_i)_{i=1}^l$ of $v_i \in
L^{1,p}(C_i,E)$ such that at any nodal point $z \in C_i \cap C_j$ $v_i(z) = v_j(z)$. Let us define an [*$E$-valued $L^p$-integrable $(0,1)$-form $\xi$ on $C$*]{} as a couple $(\xi_i)_{i=1}^l$ of $(0,1)$-forms $\xi_i \in L^p(C_i,
E\otimes \Lambda^{(0,1)} )$.* ]{}
Let $L^{1,p}(C,E)$ denote a Banach space of $L^{1,p}$-sections of $E$ over $C$, and $L^p(C, E\otimes \Lambda ^{(0,1)} )$ a Banach space of $L^p$-integrable $(0,1)$-forms $C$. Denote by $\calh^{1,p}(C,E)$ a Banach space of [*holomorphic*]{} $L^{1,p}$-sections of $E$ over $C$.
Analogously, for a complex manifold $X$ by $L^{1,p}(C,X)$ we denote the set of couples $u=(u_i)_{i=1}^l$ such that $u_i \in L^{1,p}(C_i,
X)$, with $u_i(z) = u_j(z)$ at any nodal point $z \in C_i \cap C_j$. One can see that $L^{1,p}(C,X)$ is a Banach manifold with a tangent space $T_uL^{1,p}(C,X) = L^{1,p}(C,u^*TX)$. Further by $\calh^{1,p}(C,X)$ we shall denote the manifold of [*holomorphic*]{} $L^{1,p}$-maps from $C$ to $X$. Note that for $u\in \calh^{1,p}(C,X)$ one has $u(C) \subset u(\barr C)
\Subset X$, because $L^{1,p}\subset C^{0,1-{2\over p}}$ and therefore $u$ is continuous up to the boundary.
*9.2. Solution of a Cousin-type Problem.*
An important role plays the following result on the possibility to solve the following Cousin-type problem.
**Lemma 9.2.1. *Let $C$ be a nodal curve and $E$ a holomorhpic vector bundle over $C$, $C^1$-smooth up to the boundary. Let $\{ V_i\}_{i=1}^l$ be a finite covering of $C$ by Stein domains with piecewise smooth boundaries. Put $V_{ij} \deff V_i \cap V_j$ and suppose that all triple intersections $V_i \cap V_j \cap V_k$ á $i \not= j \not= k \not= i$ are empty.***
Then for $2 \le p< \infty$ the Čech-differential $$\matrix
\delta :& \sum_{i=1}^l \calh ^{1,p}(V_i, E)& \lrar&
\sum_{i<j} \calh ^{1,p}(V_{ij}, E)
\cr
\delta :& (v_i)_{i=1}^l & \longmapsto & (v_i -v_j)
\endmatrix
\eqno(9.1.1)$$ possesses the following properties:
the image $\im(\delta)$ is closed and has finite codimension; moreover, $\coker(\delta) = \sfh^1(C, E) = \sfh^1(C_{\sf comp}, E)$, where $C_{\sf comp}$ denotes the union of compact irreducible components of $C$;
the kernel $\ker(\delta)$ is isomorphic to $\calh ^{1,p}(C, E)$ and admits a closed direct complement.
Proof. Before considering the Čech complex, let us look at the corresponding $\dbar$-problem. Consider the following operator $$\dbar : L^{1,p}(C,E) \lrar L^p(C, E\otimes \Lambda^{(0,1)}_C).
\eqno(9.1.2)$$ First we shall prove that this operator possesses the same properties as $(9.1.1)$, i.e., that $\ker(\dbar)$ admits a closed direct complement, $\im(\dbar)$ has finite codemension and is closed, and $\coker(\dbar) =
\sfh^1(C, E) = \sfh^1(C_{\sf comp}, E)$. Moreover, we shall construct a natural isomorphism between (co)kernels of $(9.1.1)$ and $(9.1.2)$.
Since the boundary $C$ is smooth, there exist nodal curves $C^+$ and $C^{++}$ such that $C \Subset C^+ \Subset C^{++}$ and differences $C^+ \bs \barr C$ (resp. $C^{++} \bs \barr C^+$) consist of an annulai $A^+_\alpha$ (resp. $A^{++}_\alpha$) adjacent to the corresponding components $\gamma_\alpha$ (resp. $\gamma^+ _\alpha$) of the boundary $\d C$ (resp. $\d C^+$, see. Fig. 17.). Now $E$ extends to a holomorphic vector bundle over $C^{++}$, again denoted as $E$.
Consider the following sheaves on $C^{++}$ $$\matrix \format \r\;\; &\c\; &\; \c\; &\; \l \\
L^{1,p}_\loc(\cdot ,E): &
V & \mapsto & L^{1,p}_\loc(V,E) \vphantom{\int_3};
\cr
L^p_\loc(\cdot ,E\otimes \Lambda^{(0,1)}_{C^{++}}): &
V & \mapsto & L^p_\loc(V,E\otimes \Lambda^{(0,1)}_{C^{++}}),
\endmatrix$$ and a sheaf homomorphism defined by the operator $$\dbar: L^{1,p}_\loc(V,E) \lrar L^p_\loc(V,E\otimes \Lambda^{(0,1)}_{C^{++}}).$$ Then sheaves $L^{1,p}_\loc(\cdot ,E)$ and $L^p_\loc(\cdot ,E\otimes \barr
\Omega_{C^{++}})$ together with a $\dbar$-homomorphism form a fine resolution of the (coherent) sheaf $\calo^E$ of holomorphic sections of $E$ over $C^{++}$. In smooth points of $C^{++}$ this fact follows from $L^p$-regularity of the (elliptic) operator $\dbar$, and in nodal points this can be seen from the following considerations.
Let $z\in C$ be a nodal point which lies on the intersection of irreducible components $C_i$ and $C_j$ of the curve $C$. Let $\xi_i$ (resp. $\xi_j$) be an $E$-valued $L^p_\loc$-integrable (0,1)-form, defined in the neighborhood of the point $z$ in $C_i$ (resp. $C_j$). And let $v_i$ (resp $v_j$) be $L^{1,p}_\loc$-solutions of the equation $\dbar v_i = \xi_i$ (resp. $\dbar v_j= \xi_j$). Adding a local [*holomorphic*]{} section of $E$ over $C_i$, we can achieve an equality $v_i(z)= v_j(z)$. Then the pair $(v_i, v_j)$ defines a section of the sheaf $L^{1,p}_\loc(\cdot ,E)$ in the neighborhood of $z$. This shows that the Dolbeault lemma is valid for the holomorphic bundle $E$ also in the neighborhood of the nodal points.
This gives the natural isomorphisms $$\matrix \format \r&\c & \c & \l &\c\\
\ker{\Bigm(}\dbar: L^{1,p}_\loc(C^{++},E) &\lrar &
L^p_\loc(C^{++},E\otimes \Lambda^{(0,1)}_{C^{++}} )
{\Bigm)} &=& \sfh^0(C^{++}, E),
\vphantom{\int_3}
\cr
%\noalign{\rlap{and}}
\coker{\Bigm(}\dbar: L^{1,p}_\loc(C^{++},E) & \lrar&
L^p_\loc(C^{++},E\otimes \Lambda^{(0,1)}_{C^{++}} )
{\Bigm)} &=& \sfh^1(C^{++}, E).
\endmatrix
\eqno(9.1.3)$$ Note that one also has the same isomorphisms for $C$ and $C^+$. Note also that there are natural isomorphisms $\sfh^1(C^{++}, E)= \sfh^1(C^+, E) =
\sfh^1(C, E)= \sfh^1(C_{\sf comp}, E)$, induced by the restrictions $L^p_\loc(C^{++},E\otimes \Lambda^{(0,1)}_{C^{++}}) \to L^p_\loc(C^+,E\otimes
\Lambda^{(0,1)}_{C^+} ) \to L^p_\loc(C,E\otimes \Lambda^{(0,1)}_{C} ) \to
L^p_\loc(C_{\sf comp},E\otimes \Lambda^{(0,1)}_C) $.
Now take some $\xi \in L^2(C^+,E \otimes \Lambda^{(0,1)}_{C^+})$, which defines a zero cohomology class in $\sfh^1(C^+, E)$. We can extend $\xi$ by zero to the element $\ti \xi\in L^2_\loc(C^{++}, E\otimes
\Lambda^{(0,1)}_{C^{++}})$. While $[\ti\xi]_\dbar= [\xi]_\dbar =0$, there exists $\ti v \in L^{1,2}_\loc(C^{++},E)$ such that $\dbar\ti v =
\ti\xi$. The restriction $v \deff \ti v\ogran_{C^+}$ satisfies $\dbar v = \xi$ and $v \in L^{1,2}(C^+,E)$. This shows that the image of a continuous (!) operator $$\dbar: L^{1,2}(C^+,E) \lrar L^2(C^+,E\otimes
\Lambda^{(0,1)}_{C^+}) \eqno(9.1.4)$$ is of finite codimension. From here and from the Banach open mapping theorem it follows that this image is closed. Furthermore, because $L^{1,2}(C^+,E)$ is a Hilbert space, the kernel of the operator $(9.1.4)$ admits a direct complement $Q \subset L^{1,2}(C^+,E)$. Moreover, operator $(9.1.4)$ maps $Q$ isomorphically onto its image. Thus, operator $(9.1.4)$ splits, i.e., there exists a continuous operator $$\def\dash{\raise.2pt\hbox{-}} T^+:L^2(C^+,E\otimes
\Lambda^{(0,1)}_{C^+}) \lrar L^{1,2}(C^+,E) \eqno(9.1.5)$$ such that $\im(T^+)=Q$ and for any $\xi \in L^2(C^+,E\otimes
\Lambda^{(0,1)}_{C^+})$ with $[\xi]_\dbar=0 \in \sfh^1(C^+, E)$ one has $\dbar(T^+\xi)=\xi$.
Define an operator $T:L^2(C, E\otimes \Lambda^{(0,1)}_C) \lrar
L^{1,2}(C,E)$ in the following way. For $\xi \in
L^2(C, E\otimes \Lambda^{(0,1)}_C)$ extend $\xi$ by zero to $\ti \xi \in
L^2(C^+,E\otimes \Lambda^{(0,1)}_{C^+})$ and put $T(\xi) =
T^+(\ti\xi)\ogran_{C}$. $T$ is obviously continuous; moreover $$\norm{T^+(\ti\xi)}_{L^{1,2}(C^+)} \le c\cdot \norm{\xi}_{L^2(C)}$$ with constant $c$ independent of $\xi$. By the $L^p$-regularity for the elliptic $\dbar$-operator (see ex., \[Mo\]) for $2\le p<\infty$ and
$v \in L^{1,p}_\loc(C^+,E)$ there is an interior estimate $$\norm{v}_{L^{1,p}(C)} \le c'\cdot
\bigl( \norm{v}_{L^{1,2}(C^+)} + \norm{\dbar v}_{L^p(C^+)} \bigr)
\eqno(9.1.6)$$ with the constant $c'$ independent on $v$. From this it follows that for $2\le p<\infty$ and $\xi \in L^p(C, E\otimes \Lambda^{(0,1)}_C)$ with $[\xi]_\dbar=0 \in \sfh^1(C, E)$ one has $$\norm{T(\xi)}_{L^{1,p}(C)} \le c''\cdot \norm{\xi}_{L^p(C)}$$ with a constant $c''$ independent on $\xi$. This means that the operator $T$ is a splitting off the operator $(9.1.2)$. This means that operator $(9.1.2)$ possesses the properties and of [*Lemma 9.2.1*]{}.
Let us return to the Čech operator $(9.1.1)$. Fix some partition of unity $\1=\sum_{i=1}^l\phi_i$, subordinate to the covering $\{V_i\}_{i=1}^l$ of the curve $C$. Take a cocycle $w=(w_{ij}) \in \sum_{i<j}\calh ^{1,p}(V_{ij}, E)$. For $i>j$ put $w_{ij} \deff -w_{ji}$. Define $f_i \deff \sum_j \phi_j w_{ij}$. Then $f_i \in L^{1,p}(V_i, E)$ and $f_i- f_j = w_{ij}$. Consequently, $\dbar f_i \in L^p(V_i, E\otimes
\Lambda^{(0,1)}_C)$ and $\dbar f_i = \dbar f_j$ ¢ $V_{ij}$. So, $\dbar f_i = \xi\ogran_{V_i}$ for the correctly defined $\xi \in L^p(C,
E\otimes \Lambda^{(0,1)}_C)$. Moreover, $(w_{ij})$ and $\xi$ define the same cohomology class $[w_{ij}]=[\xi]$ in $\sfh^1(C,E)$.
Suppose additionally that the induced cohomology class $[w_{ij}]$ is trivial. Put $f\deff T(\xi)$ and $v_i= f_i-f$. Then $v_i \in
L^{1,p}(V_i, E)$, $v_i- v_j = w_{ij}$, and $\dbar v_i= \dbar f_i -\dbar f=0$. Thus, $v\deff(v_i) \in \sum_{i=1}^l \calh ^{1,p}(V_i, E)$ and $\delta(v)= w$. From here it follows that the formula $T_\delta : w \mapsto v$ defines an operator $T_\delta$, which is a splitting of $\delta$. The explicit construction shows that $T_\delta$ is continuous. This completes the proof of [*Lemma 9.2.1*]{}.
*9.3. Case of a Stein Curve.*
Lemma 9.3.1. *Let $C$ be a Stein nodal curve with a piecewise smooth boundary, and $X$ a complex manifold. Then*
$\calh^{1,p}(C,X)$ possesses a natural structure of a complex Banach manifold with a tangent space $T_u\calh^{1,p}(C,X) =
\calh^{1,p}(C,u^*TX)$ at $u\in \calh^{1,p}(C,X)$.
If $C'\subset C$ is again a nodal curve, then the restriction map $\calh^{1,p}(C, X) \to \calh^{1,p}(C', X)$ is holomorphic, and its differential at $u\in \calh^{1,p}(C,X)$ is again the restriction map $\calh^{1,p}(C, u^*TX) \to \calh^{1,p}(C', u^*TX)$, $v \mapsto v|_{C'}$.
The **proof will be given in several steps.**
[*Step 1.*]{} Suppose first that the image $u(C)$ lies in the complex chart $U \subset X$ with complex coordinates $w= (w_1, \ldots, w_n) : U \buildrel \cong
\over \lrar U' \subset \cc^n$. Then the set $\calh^{1,p}(C,U)$ can be naturally identified with the set $\calh^{1,p}(C,U')$, which is an open set in the Banach space $\calh^{1,p}(C,\cc^n)$. This defines on $\calh^{1,p}(C,U)$ the structure of a complex Banach manifold with the tangent space $T_u\calh^{1,p}(C,U) \cong \calh^{1,p}(C,\cc^n) \cong \calh^{1,p}(C,u^*TU)$ at the point $u\in \calh^{1,p}(C,U)$.
Note that if $u_t$, $t\in [0,1]$, is a $C^1$-curve in $\calh^{1,p}(C,U)$, then the tangent vector $v\in \calh^{1,p}(C,u^*TU)$ to $u_t$ at $u_0$ is given by $v(z)= {\d u \over\d t}(z) \in T_{u(z)}U$. This last expression does not depend on the choice of the complex coordinates $w= (w_1,\ldots, w_n) : U \to \cc^n$ in $U$. This implies two things.
First, [*the complex structure on $\calh^{1,p}(C, U)$ does not depend on the choice of the complex coordinates $w= (w_1, \ldots, w_n) : U \to \cc^n$ in $U$*]{}. Second, for $C$ property of the lemma is fulfilled.
Thus, if $u(C)$ is contained in a coordinate chart, then [*Lemma 9.3.1*]{} is proved.
[*Step 2.*]{} Now fix $u_0 \in \calh^{1,p}(C,X)$ and suppose that a finite covering $\{V_i\}_{i=1}^l$ of the curve $C$ is chosen such that first, the conditions of [*Lemma 9.2.1*]{} are satisfied, and second, for every $V_i$ [*Lemma 9.3.1*]{} holds, e.g., every $u_0(V_i)$ is contained in some coordinate chart $U_i \subset X$.
Set $V_{ij}\deff V_i\cap V_j$. Choose balls $B_{ij} \subset\calh^{1,p}
(V_{ij}, u_0^*TX) \cong T_{u_0}\calh^{1,p}(V_{ij},X)$ such that there exists a biholomorhpism $\psi_{ij} : B_{ij} \buildrel \cong \over \lrar
B'_{ij}
\subset \calh^{1,p}(V_{ij},X)$ with $\psi_{ij}(0) = u_0|_{V_{ij}}$ and $d\psi_{ij}
(0) = \id: \calh^{1,p}(V_{ij}, u_0^*TX) \to\calh^{1,p}(V_{ij}, u_0^*TX)$. Then take an open sets $B_i \subset \calh^{1,p}(V_i, X)$ such that $u_0|_{V_i}
\in B_i$ and for every $u_i \in B_i$ one has $u_i|_{V_{ij}} \in B'_{ij}$.
Now, the holomorphic mappings $\phi_{ij} :B_i \to B_{ij}
\subset \calh^{1,p}(V_{ij}, \allowbreak u_0^*TX)$, $\phi_{ij} : u_i
\mapsto \psi_{ij}\inv (u_i|_{V_{ij}})$ are uniquely defined. They define a holomorphic map $$\matrix
\Phi: \prod_{i=1}^l B_i & \lrar & \sum_{i<j} \calh^{1,p}(V_{ij},u_0^*TX)
\cr
(u_i)_{i=1}^l & \mapsto & \phi_{ij}(u_i) - \phi_{ji}(u_j).
\endmatrix$$ One can easily see that the map $\Phi$ defines the condition of compatibility of local holomorphic mappings $u_i : V_i \to X$, namely $(u_i)_{i=1}^l \in \prod_{i=1}^lB_i$ define a holomorphic map $u: C \to X$ if and only if $\Phi(u_i)=0$. Moreover, the differential $d\Phi$ at point $(u_0|_{V_i})$ coincides with the co-boundary Ĉech operator $(9.1.1)$. Since $C$ is Stein, $\sfh^1(C, u_0^*TX) =0$. From [*Lemma 9.2.1*]{} and the implicit function theorem, we see that the statements and of [*Lemma 9.3.1*]{} are fulfilled in some neighborhood of the map $u_0\in \calh^{1,p} (C, X)$.
[*Step 3.*]{} One can show that for any Stein nodal curve $C$ and any $u \in \calh^{1,p} (C, X)$ the statements and of [*Lemma 9.3.1*]{} are fulfilled in some neighborhood of the mapping $u$, applying [*Step 2*]{} sufficiently many times. If, for example, $C$ is an annulus $A_{r,R}$, then one can cover it by the narrow annulai $A_{r_i, R_i}$, $0< R_i - r_i <\!\!< 1$, and every $A_{r_i, R_i}$ with small sectors $V_{ij} = \{ z= \rho
e^{\isl \theta} \in \cc \,:\, r_i<\rho < R_i, \alpha_j < \theta < \beta_j\}$ á $0< \beta_j -\alpha_j <\!\!< 1$. Details are left to the reader.
*9.4. Degeneration to a Node.*
One of the diffuculties in constructing [*holomorphic*]{} families of stable curves with boundaries comes from the fact that the moduli space of complex structures on the noncompact surface with boundary $\Sigma$ does not have a natural complex structure; moreover it can have an odd real dimension. If, for example, $\Sigma$ is an annulus, then it is biholomorhic to the standard annulus $A_{r,1}$ with uniquely defined $r\in (0,1)$; thus, the interval $(0,1)$ is a moduli space in this case. In the general case, if $\Sigma$ has genus $g$ and $k$ boundary components, then the real dimension of the moduli space is equal to $d= 6g -6 + 3k$, except in four cases, where $\Sigma$ is either a sphere ($g=0$, $k=0$), a torus ($g=1$, $k=0$), a disk ($g=0$, $k=1$) or an annulus ($g=0$, $k=2$), see e.g., \[Ab\]. Note that exactly in this cases the group of biholomorphic automorphisms of corresponding complex curves $(\Sigma, J)$ has positive dimension.
One can correct the situation by introducing additionally $k$ parameters, namely, fixing $k$ marked points, one on each boundary component. Let $A$ be an annulus with boundary circles $\gamma_0$ and $\gamma_1$ and $X$ be a complex manifold.
Theorem 9.4.1. *There exist complex Banach manifolds $\calm(A, X)$ and $\calc(A, X)$, a holomorphic projection $\pi_\calc: \calc(A, X)\to \calm(A,X)$ and holomorphic mappings $\ev: \calc(A, X) \to X$, $z_1 :\calc(A, X) \to \Delta$, $z_2 :\calc(A, X)
\to \Delta$ and $\lambda_\calm :\calm(A, X) \to \Delta$ with the following properties:*
for any $y\in \calm(A,X)$ the fiber of the projection $C_y\deff \pi_\calc\inv(y)$ is a nodal curve, parameterized by an annulus $A$; moreover, the mapping $(z_1, z_2) : C_y \to \Delta^2$ defines a biholomorphism onto the curve $\{ (z_1, z_2)
\in \Delta^2 \;:\; z_1 \cdot z_2 = \lambda_\calm(y) \}$; in particular, $C_y$ is either a standard node if $\lambda_\calm(y)=0$, or a holomorphic annulus $\{ |\lambda_\calm(y)| <|z_1| <1 \}$;
the diagram $$\def\normalbaselines{\baselineskip18pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\!\!\smash{\mathop{\hbox to50pt{\cleaders\lowminus\hfill}
\mkern-13mu\rightarrow}\limits^{#1}}\!\!}
\def\mapdown#1{\Big\downarrow\rlap{$\vcenter{\hbox{$\scriptstyle#1$}}$}}
%
\matrix
\calc(A, X) & \mapright{ (\ev, z_1, z_2) }& X \times \Delta^2
\cr
\mapdown{\pi_\calc} & &
\hphantom{\hbox{$X\times{.}$}} \mapdown{\lambda= z_1 \cdot z_2}
\cr
\calm(A,X)& \mapright{ \lambda_\calm } &
\hphantom{\hbox{$X\times{.}$}} \Delta
\endmatrix
\eqno(9.4.1)$$ is commutative; moreover, for every $y\in \calm(A,X)$ the restriction $\ev|_{C_y}$ is in
$\calh^{1,p} (A_a, X)$ with $a=\lambda_\calm(y)$, and mappings $\ev_1: y\in \calm(A,X) \mapsto \ev|_{C_y}(z_1^{-1}(1))$ and $\ev_2: y\in \calm(A,X) \mapsto \ev|_{C_y}(z_2^{-1}(1))$ are holomorphic;
let $C$ be an annulus or a nodal curve with smooth boundary $\d C =\gamma_1\coprod \gamma_2$, $p_i \in \gamma_i$ be marked points, and $u:C \to X$ be a holomorphic $L^{1,p}$-smooth mapping, then there exist [*a unique*]{} $y\in \calm(A,X)$ and a [*unique*]{} biholomorphism $\phi: C \to C_y$ such that $\ev \scirc \phi = u : C\to X$ ¨ $z_i \scirc
\phi (p_i) =1\in \barr \Delta$; in other words, $\calm(A,X)$ parameterizes holomorphic mappings to $X$ from annulai and nodes with marked points on the boundary;
if the commutative diagram with complex spaces $\calw$ and $\calz$ and holomorphic mappings $$\def\normalbaselines{\baselineskip18pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\!\!\smash{\mathop{\hbox to50pt{\cleaders\lowminus\hfill}
\mkern-13mu\rightarrow}\limits^{#1}}\!\!}
\def\mapdown#1{\Big\downarrow\rlap{$\vcenter{\hbox{$\scriptstyle#1$}}$}}
%
\matrix
\calz & \mapright{ (\ev^\calz, \ti z_1, \ti z_2) }& X \times \Delta^2
\cr
\mapdown{\pi_\calz} & &
\hphantom{\hbox{$X\times{.}$}} \mapdown{\lambda= \ti z_1 \cdot \ti z_2}
\cr
\calw & \mapright{ \lambda_\calw } &
\hphantom{\hbox{$X\times{.}$}} \Delta
\endmatrix
\eqno(9.4.2)$$ possesses the properties and , in particular, the fibers $\calz_w \deff\pi_\calz\inv(w)$ should be the nodal curves with induced maps $f_w\deff\ev^\calz|_{\calz_w} \in \calh^{1,p}(\calz_w, X)$, then diagrams $(9.4.1)$ and $(9.4.2)$ uniquely fit together to form the commutative diagram $$\def\normalbaselines{\baselineskip18pt\lineskip3pt \lineskiplimit3pt }
\def\mapright#1{\!\!\smash{\mathop{\hbox to50pt{\cleaders\lowminus\hfill}
\mkern-13mu\rightarrow}\limits^{#1}}\!\!}
\def\mapdown#1{\Big\downarrow\rlap{$\vcenter{\hbox{$\scriptstyle#1$}}$}}
%
\matrix
\calz & \buildrel \wt F \over \lrar &
\calc(A, X) & \mapright{ (\ev, z_1, z_2) }& X \times \Delta^2
\cr
\mapdown{\pi_\calz} & &
\mapdown{\pi_\calc} & &
\hphantom{\hbox{$X\times{.}$}} \mapdown{\lambda= z_1 \cdot z_2}
\cr
\calw & \buildrel F \over \lrar &
\calm(A,X)& \mapright{ \lambda_\calm } &
\hphantom{\hbox{$X\times{.}$}} \Delta
\endmatrix
\eqno(9.4.3)$$ here $\lambda_\calm \scirc F = \lambda_\calw$ and $(\ev, z_1, z_2) \scirc
\wt F= (\ev^\calz, \ti z_1, \ti z_2)$;
differential $d\lambda_\calm: T_y\calm(A,X) \to T_{\lambda_\calm(y)}
\Delta \cong \cc$ is not degenerate in any point $y \in \calm(A,X)$, and for any $a \in \Delta$ the fiber $\lambda_\calm\inv(a)$ is naturally isomorphic to the manifold $\calh^{1,p} (A_a, X)$, where $A_a$ is a curve $\{ (z_1, z_2)\in
\Delta^2 : z_1 \cdot z_2 = a \}$; in particular, for any $y \in \calm(A,X)$ there is a biholomorphism $C_y \cong A_{\lambda_\calm(y)}$ and a natural exact sequence $$0 \lrar
\calh^{1,p} (C_y, u^*TX) \buildrel \iota_y \over\lrar T_y \calm(A,X)
\buildrel d\lambda_\calm(y) \over\lrar \cc \lrar 0.$$
Proof. Let $(A, p_1, p_2)$ be a smooth annulus with marked points, one on each boundary component $\gamma_i \cong S^1$, and let $J$ be a complex structure on $A$. It is known that $(A, J)$ is biholomorphic to some annulus $A_{r,1}= \{ r < |z| <1 \}$. It is easy to see that there is only one isomorphism $\psi: (A, J) \to A_{r,1}$ which smoothly extends to the diffeomorphism $\psi: \barr A \to \barr A_{r,1}$ with $\phi(p_1) =1$. Put $a \deff \phi(p_2)$. Now it is obvious that there is a unique biholomorphism $\phi: (A, J) \to A_a \deff \{ (z_1, z_2)\in
\Delta^2 \;:\; z_1 \cdot z_2 = a \}$ such that $\phi(p_1) =1$ ¨ $\phi(p_2)
=a$.
Thus, the mapping $\lambda: \Delta^2 \to \Delta$, $\lambda(z_1, z_2)
= z_1 \cdot z_2$ with a fiber $A_a$ over $a\in \Delta$ represents a holomorphic moduli space of annulai with marked points on the boundaries, completed at $a=0$ by a standard node. If $a\not =0$ the coordinates $z_i$, $i=1, 2$, realize an imbedding of each $A_a$ as an annulus into $\cc$ in such a way that the circle $\gamma_i$ becomes an outer unit circle. When $a \lrar 0$, the annulus $A_a$ degenerates into a standard node, and each $z_i$ becomes a standard coordinate on the corresponding component of the node.
Remark. In the sequel we denote by $A_a$ an annulus (or corr. a node) [*with fixed points $a$ and $1$ on the boundary $\d A_a$*]{}, and also together with uniquely determined coordinates $z_1$ and $z_2$.
Fix $r$ with $0 <r< 1$. For $|a|<r$ define a mappings $\zeta^a_1,
\zeta^a_2: A_{r,1} \to A_a$ as $\zeta^a_1(z) \deff z$ and $\zeta^a_2(z)
\deff a/z$, so that $\zeta^a_i$ are the inverse of the coordinates $z_i$. Consider a mapping $$\matrix
\Psi_r: \coprod_{|a|<r} \calh^{1,p}(A_a, X) & \lrar &
\calh^{1,p}(A_{r,1}, X) \times \calh^{1,p}(A_{r,1}, X) \times \Delta(r)
\cr
u \in \calh^{1,p}(A_a, X) & \mapsto & (u \scirc \zeta^a_1, u \scirc
\zeta^a_2, a).
\endmatrix$$ One easily sees that $\Psi_r$ is holomorphic on each $\calh^{1,p}(A_a, X)$ and the image of $\Psi_r$ consists of such triples $(u_1,u_2,a)$ that each $u_i \in \calh^{1,p}(A_{r,1}, X)$ extends to $u_i \in \calh^{1,p}
(A_{|a|,1}, X)$ and $u_2(z) = u_1(a/z)$. Thus $\Psi_r$ is injective and has closed image. On the disjoint union $\calm(A, X) \deff \coprod_{a\in
\Delta}\calh^{1,p} (A_a, X)$ define the topology induced by the mappings $\Psi_r$. Clearly it agrees with the topologies on each slice $\calh^{1,p}(A_a, X)$.
Our task now is to construct an appropriate holomorphic structure on $\calm(A, X)$, which agrees with the above introduced holomorhpic structures on each slice
$\calh^{1,p}(A_a, X)$, and also with the topology on $\calm(A, X)$, defined above.
*Case 1. Let us start with the special case $X= \cc^n$. For $a\not =0$ every function $f\in \calh^{1,p} (A_a, \cc^n)$ in a unique way can be decomposed into a Laurent series $f(z_1) =\sum_{i= -\infty} ^\infty c_i
z_1^i$. Set $f^+(z_1) \deff \sum_{i=0}^\infty c_i z_1^i$ and $f^-(z_1)
\deff \sum_{i=-\infty} ^0 c_i z_1^i$. It is convenient to consider the function $f^-$ as a function of the variable $z_2 = a/z_1$; then $f^-(z_2)= \sum_{i=0}^\infty c_{-i} (z_2/a)^i$. One has $f^+ \in \calh^{1,p}(\{ |z_1| <1\}, \cc^n)$, $f^-\in \calh^{1,p}(\{ |z_2| <
1\}, \cc^n)$, $f^+(0) = f^-(z_2 =0) = c_0$, and $f(z_1) = f^+(z_1) +
f^-(a/z_1) - c_0$, so that a pair $(f^+, f^-)$ defines the holomorhpic function $\hat f \in \calh^{1,p}(A_0, \cc^n)$. The canonical isomorphisms $\calh^{1,p}(A_a, \cc^n)\cong \calh^{1,p}(A_0, \cc^n)$ obtained in this was define on $\coprod_{|a|<1} \calh^{1,p} (A_a, \cc^n)$ the structure of the trivial Banach bundle with base $\{ |a| <1\}$ and fiber $\calh^{1,p}(A_0, \cc^n)$, and thus a structure of a Banach manifold.*
One can see that the mapping $\Psi_r : \coprod_{|a|<r} \calh^{1,p}(A_a,
\cc^n) \to
\calh^{1,p}(A_{r,1}, \cc^n)^2 \times \Delta(r)$ is holomorphic.
*Case 2. Let $X = U \subset \cc^n$ be an open subset in $\cc^n$. Then $\calm(A, U)$ is also open in $\calm(A, \cc^n)$ and consequently inherits a holomorphic structure. The natural projection $\lambda_\calm : \calm(A, U) \to \Delta$ becomes holomorphic. One easily sees that the differential $d\lambda_\calm$ is not degenerate. Thus, we can define [*a universal family of curves*]{} $\calc(A, U)$ as a fiber product $\calm(A, U) {\times}_\Delta
\Delta^2$ with respect to the mappings $\lambda_\calm : \calm(A, U) \to
\Delta$ and $\lambda: \Delta^2 \to \Delta$, which is a complex Banach manifold due to the nondegeneracy of $d\lambda_\calm$.*
Denote by $\pi_\calc : \calc(A, U) \to \calm(A, U)$ the natural projection. Then the fiber $C_y$ over $y \in \calm(A, U)$ is biholomorphic to $A_a$ with $a=\lambda_\calm(y)$. The natural projection from $\calc(A, U)$ onto $\Delta^2$ induces holomorphic functions $z_1$ and $z_2$ on $\calc(A, U)$ taking values in $\Delta$ such that the property of [*Theorem 9.4.1*]{} is satisfied.
Now let $a\not=0 \in \Delta$ and $f\in \calh^{1,p} (A_a, \cc^n)$. Represent $f$ in the form $f(z_1) = f^+(z_1) + f^-(a/z_1) - f_0$, where $f^\pm \in \calh^{1,p} (\Delta, \cc^n)$ with $f^+(0) = f^-(0) = f_0$. Analogously, for $f\in \calh^{1,p} (A_0, \cc^n)$ we have $f= (f^+, f^-)$, where again $f^\pm \in \calh^{1,p} (\Delta, \cc^n)$ á $f^+(0) = f^-(0) = f_0$. Consider a holomorphic function $\ti f(z_1, z_2) \deff f^+(z_1) + f^-(z_2)-
f_0$, $\ti f\in \calh^{1,p} (\Delta^2, \cc^n)$. Define mappings $\ti\ev_a:\calh^{1,p} (A_a, \cc^n) \times \Delta^2 \lrar \cc^n$ by the formula $\ti\ev (f, z_1, z_2) \deff \ti f(z_1, z_2)$. One easily sees that $\ti\ev_a$ define a holomorphic map $\ti\ev :\calm(A, \cc^n) \times \Delta^2
\to \cc^n$. Let $\ev$ denote the restriction of this map onto $\calc(A, \cc^n) \subset \calm(A, \cc^n) \times \Delta^2$. It is straightforward to check that $\calm(A, \cc^n)$, $\calc(A, \cc^n)$, $\ev: \calc(A, \cc^n) \to \cc^n$ and $z_{1,2} : \calc(A, \cc^n) \to \Delta$ satisfy the statements of [*Theorem 9.4.1*]{}.
Thus, for $U \subset \cc^n$ we can define $\ev: \calc(A, U) \to U$ as a restriction $\ev: \calc(A, \cc^n) \to \cc^n$. Statements of [*Theorem 9.4.1*]{} again remain true. In particular, if $G: U \to U'\subset \cc^n$ is biholomorphic, then the natural bijections $\calm(A,U) \buildrel \cong \over\lrar \calm(A, U')$ and $
\calc(A, U) \buildrel\cong \over \lrar \calc(A, U')$ are also biholomorphisms. This means that the complex structure in $\calm(A, U)$ does not depend on the imbedding $U \subset \cc^n$.
Let $C= A_{r,1}$ be an annulus and $u: C \to X$ be a holomorphic imbedding. Then $du : TC \to u^*TX$ is an imbedding of holomorphic bundles over $C$, and this defines a holomorphic normal bundle as a factor-bundle $N_C \deff u^*TX/ TC$. Since $C$ is Stein, the bundle $N_C$ is holomorphically trivial. Fix a holomorphic frame $\sigma_1, \ldots, \sigma_{n-1} \in
\calh^{1,p} (C, N_C)$, $n= \dimc X$, and also its lift $\ti\sigma_1, \ldots,
\ti \sigma_{n-1} \in \calh^{1,p} (C, u^*TX)$. Denote by $B^{n-1}(r)$ the ball of radius $r$ in $\cc^{n-1}$ with coordinates $w= (w_1, \ldots, w_{n-1})$. As follows from [*Lemma 9.3.1*]{}, there is a holomorphic map $\Psi : C\times B^{n-1}(r) \to X$ such that $\d \Psi / \d w_i|_{z\in C,w=0}
=\sigma_i(z)$. Thus $\Psi$ is biholomorphic in the neighborhood of $C\equiv C\times \{0\}$. In particular, for $r$ sufficiently small an image $U\deff \Psi(C \times B^{n-1}(r))$ is a coordinate chart with coordinates $(z, w_1, \ldots, w_{n-1})$.
*Step 3. Consider a general case where $C\cong A_a$ is arbitrary, and $u: C \to X$ is a holomorphic map.*
If $a=0$, then $C$ is a node; thus, there exists a neigborhood $V_0$ of the nodal point such that its image $u(\barr V_0)$ lies in some coordinate chart of $X$. If $a\not=0$ and the image $u(\barr C)$ is not contained in any chart of $X$, then $u$ is not constant. Thus, for some $|a|< r <1$ mapping $u$ will be an imbedding in some neighborhood of the circle $S^1_r \deff \{|z_1| =r\} \subset C \cong \{|a|< |z_1|<1\}$.
In all cases there is a covering $\{V_0, V_1, V_2\}$ of $C=\{(z_1,
z_2) \in \Delta^2 : z_1 \cdot z_2 =a\}$ of the form $V_1 =\{ (z_1,z_2)\in C
: r_1 <|z_1| <1 \}$, $V_2 =\{ (z_1,z_2)\in C : r_2 <|z_2| <1 \}$, $V_0 =\{ (z_1, z_2) \in C : |z_1| < R_1, |z_2|< R_2 \}$, where $0<r_1 <R_1
<1$, $0< r_2 <R_2 <1$, $r_1 \cdot R_2 > |a| <r_2 \cdot R_1$, and $V_0$ is such that the image $u(\barr V_0)$ lies in some chart $U$ of $X$.
Fix coordinate $z_1$ on $V_1$ and coordinate $z_2$ on $V_2$. On $V_0$ define new coordinates $\ti z_1 \deff z_1 /R_1$ and $\ti z_2 \deff z_2 /R_2$, and also the marked points $\ti p_1 \deff R_1$ and $\ti p_2 \deff a/R_2$. Set $\ti a \deff a/(R_1R_2)$. Then $V_0 \cong A_{\ti a}= \{(\ti z_1, \ti z_2)
\in \Delta^2 : \ti z_1\cdot \ti z_2 = \ti a\}$ with $\ti z_i(\ti p_1)=1$. It is easy to see that the change of the complex parameter $a= z_1 z_2$ on $C$, which parameterizes the holomorphic structures of annulai with marked points on the boundary, can be reduced to the change of the analogous parameter in $V_0$. Namely, let $C' \cong A_{a'} = \{ (z_1,z_2) \in \Delta^2:
z_1 \cdot z_2= a'\} $ be obtained as a result of a small deformation of $a$. Put $\ti a{}'
\deff a'/(R_1R_2)$. Look on $C'$ as the result of patching of complex curves $V_0'$, $V_1$ ¨ $V_2$, where $V_0'$, $V_1$ and $V_2$ are defined in $\Delta^2$ in the same way $V_0'= \{ (z_1,z_2) \in C': |z_1| <r_1, |z_2| <r_2\}$, $V_i= \{r_i <|z_i|<1\}$ when $i=1,2$, but now the coordinates $(z_1,z_2)$ satisfy a new relation $z_1\cdot z_2 = a'$, and $V_0'\cong A_{\ti a{}'}$.
Thus, for a sufficiently small neighborhood $\calw \subset \calm(A, X)$ of the curve $(C,u)$ over $X$ a holomorphic map is defined as $$\matrix \format \c\;&\c\;&\c\;& \c& \c& \c& \c& \c \\
\Theta :& \calw &\lrar& \calm(A,U) &\llap{$\times$}&
\calh^{1,p} (V_1, X) &\times&
\calh^{1,p} (V_2, X);
\cr
\noalign{\vskip2pt}
\Theta :&(C',u') &\mapsto& ((A_{\ti a{}'}, u|_{V_0'}), &&
u|_{V_1},&& u|_{V_2}).
\endmatrix$$ On the other hand, a couple of maps $u_0 : V_0' \to X$, $u_1 : V_1 \to X$ and $u_2 : V_2 \to X$ define a map $u': C' \to X$, exactly where $u_1$ coincides with $u_0$ on $W_1 \deff V_1 \cap V_0'$, and $u_2$ coincides with $u_0$ on $W_2 \deff V_2 \cap V_0'$. Note that domains $W_i \subset V_i$ do not change in the process of deformation of the complex structure on $C$. A construction of gluing from [*Step 2*]{} of the proof of [*Lemma 9.3.1*]{} finishes the proof of the theorem.
*9.5. Banach Analytic Structure on the Stable Neighborhood.*
We are prepared now to prove the main result of this lecture. Theorem 9.5.1.
*Let $(X,J)$ be a complex manifold, and $(C_0,u_0)$ a stable complex curve over $X$, parameterized by a real surface $\Sigma$. There exists a b.a.s.f.c. $\calm $ and $\calc$ and holomorphic maps $F: \calc \lrar X$ and $\pi: \calc \lrar \calm$ such that*
a\) for any $\lambda \in \calm $ a fiber $C_\lambda =\pi^{-1}(\lambda )$ is a nodal curve parameterized by $\Sigma$, and $C_{\lambda_0} \cong C_0$ for some $\lambda_0$;
b\) for $F_\lambda \deff F|_{C_\lambda}$ a pair $(C_\lambda ,F_\lambda )$ is a stable curve over $X$; moreover $F_{\lambda_0}=u_0$;
c\) if $(C',u')$ is a stable curve over $X$, sufficiently close to $(C_0,u_0)$ in the Gromov topology, then there exists $\lambda' \in \calm$ such that $(C',u')=(C_{\lambda '},F_{\lambda'} )$;
d\) for an appropriate $N{\in}\nn$ and for a small ball $B$ in Banach space
$\calh^{1,p}(C_0, u_0^*TX) \allowbreak \oplus \cc^N$ b.a.s.f.c. $\calm$ is realized as a zero set of a holomorphic map $\Phi$ from $B$ to the finite dimensional space $\sfh^1(C,u_0^*TX)$.
**Proof. The proof is based on the construction of local deformations of stable curves and on the analysis of the gluing of local models.**
Let $(C_0, u_0)$ be a stable curve over a complex manifold $X$ with parameterization $\sigma_0: \Sigma
\to C_0$. Using [*Proposition 3.2*]{}, fix a covering $\{ V_\alpha,
V_{\alpha\beta}\}$ of the surface $\Sigma$, having properties [*i)*]{}–[*v$\!$i)*]{} of the above-mentioned proposition. In particular, there exist biholomorphisms $\phi^0 _{\alpha \beta}: \sigma_0 (V_{\alpha \beta}) \lrar
A_{\lambda^0 _{\alpha \beta}}$. From [*i)*]{}–it follows that for every couple $\bflamb \deff (\lambda _{\alpha \beta})$, sufficiently close to $\bflamb^0 \deff (\lambda^0 _{\alpha \beta})$, there is a nodal curve $C$ and parameterization $\sigma: \Sigma \to C$ such that properties [*i*]{}–[*v$\!$i*]{} remain true and there exist biholomorphisms $\phi _{\alpha \beta}: \sigma (V_{\alpha \beta}) \lrar A_{\lambda _{\alpha
\beta}}$. In particular, the complex structures on each $\sigma (W_{\alpha
\beta})$ do not change. Moreover, in $V_{ \alpha\beta}$ there exist holomorphic coordinates $z_1$ and $z_2$ such that $z_1 |_{W_{\alpha,\beta}}$ and $z_2|_{W_{\beta, \alpha}}$ do not change by changing $\lambda _{\alpha \beta}$, and $z_1 \cdot z_2 \equiv
\lambda _{\alpha \beta}$. This means that a disk $\Delta_{\alpha
\beta}\deff \{ \lambda _{\alpha \beta} \,:\, |\lambda^0 _{\alpha \beta}-
\lambda _{\alpha \beta} | \le \epsi \,\} $ parameterizes [*a holomorphic*]{} family of curves of the type $\sigma(V_{\alpha\beta })$.
One can illustrate the changing of the complex structure in $V_{\alpha\beta}$ by the following picture.
Let $N \in \nn$ be the number of the elements of our covering of the type $V_{\alpha \beta}=V_{\beta \alpha}$. Then for sufficiently small $\epsi>0$ the polydisk $$\Delta^N_\lambda \deff \{ \bflamb \deff (\lambda _{\alpha \beta}) \;:\;
|\lambda^0 _{\alpha \beta}- \lambda _{\alpha \beta} | \le \epsi \,\}$$ parameterizes a holomorphic family of nodal curves $\{C_\bflamb\}_{\bflamb
\in \Delta^N _\lambda }$. Every curve $C_\bflamb = \sigma(\Sigma)$ is obtained by gluing the pieces $\sigma(V_\alpha)$ and $\sigma(V_{\alpha \beta})$. Denote $\calm \deff $
$\cup_{\bflamb \in\Delta^N _\lambda} \calh^{1,p}
(C_\bflamb, X)$. Since $\sigma(V_\alpha)$ does not contain nodal points and the complex structures on $\sigma(V_\alpha)$ are constant and do not depend on $\bflamb =(\lambda _{\alpha \beta})$, we can suppose that $V_\alpha$ are complex curves. We have natural isomorphisms $\sigma(V_
{\alpha \beta}) \cong A_{\lambda _{\alpha \beta}}$. Thus, the following map is well- defined $$\matrix \format \c\;& \c \;& \c& \c& \c&\c \\
\Theta: & \calm &\lrar&
\prod_\alpha \calh^{1,p} (V_\alpha, X)&\times&
\prod_{\alpha \beta} \calm(V_{\alpha \beta}, X)
\cr
\noalign{\smallskip}
\Theta: & (C_\bflamb, u) & \mapsto & \bigl(u|_{V_\alpha},
(\lambda _{\alpha \beta}, u|_{V _{\alpha \beta}}) \bigr).
\endmatrix$$ It is easy to see that a couple $\bigl( u_\alpha, (\lambda_{\alpha \beta},
u_{\alpha
\beta}) \bigr)\in \prod_\alpha \calh^{1,p} (V_\alpha, X)\times \prod_{\alpha
\beta} \calm(V_{\alpha \beta}, X)$ belongs to the image of $\Theta$, exactly if for all pairs $(\alpha,\beta)$ the gluing conditions $u_\alpha
|_{W_{\alpha, \beta}} = u_{\alpha \beta}|_{W_{\alpha, \beta}} $ are fulfilled.
Let us repeat the gluing procedure from [*Step 2 of Lemma 3.6.*]{} For this take the balls $$\matrix \format \l\,& \l\,& \l\,& \l\,& \l \\
B_\alpha &\subset& \calh^{1,p} (V_\alpha, u_0^*TX)
&\cong& T_{u_0}\calh^{1,p} (V_\alpha,X)
\cr
B_{\alpha \beta} &\subset& \calh^{1,p} (V_{\alpha \beta},u_0^*TX) \oplus \cc
&\cong& T_{u_0}\calm(V_{\alpha \beta}, X)
\cr
B'_{\alpha, \beta} &\subset& \calh^{1,p} (W_{\alpha, \beta}, u_0^*TX)
&\cong& T_{u_0} \calh^{1,p}(W_{\alpha, \beta}, X)
\endmatrix$$ such that there exist biholomorphisms $$\matrix \format \l\,& \l\,& \l\,& \l\,& \l\,& \l \\
\psi_\alpha: & B_\alpha &\buildrel \cong \over \lrar &\psi_\alpha(B_\alpha)&
\subset& \calh^{1,p}(V_\alpha, X)
\cr
\psi_{\alpha \beta}:& B_{\alpha \beta}&
\buildrel \cong \over \lrar& \psi_{\alpha \beta}(B_{\alpha \beta})&
\subset& \calm(V_{\alpha \beta}, X)
\cr
\psi'_{\alpha,\beta}:& B'_{\alpha,\beta}&
\buildrel \cong \over \lrar& \psi'_{\alpha,\beta} (B'_{\alpha,\beta})&
\subset &\calh^{1,p}(W_{\alpha,\beta}, X)
\endmatrix$$ having the properties $$\matrix \format \l\,& \l\,& \l\, \hskip 7pt & \l\,& \l\,& \l \\
\psi_\alpha(0) &=& u_0|_{V_\alpha}, &
d\psi_\alpha(0) &=&
\id : T_{u_0} \calh^{1,p}(V_\alpha, X) \to T_{u_0} \calh^{1,p}(V_\alpha, X),
\cr
\psi_{\alpha\beta}(0) &=& u_0|_{V_{\alpha\beta}}, &
d\psi_{\alpha\beta}(0) &=&
\id : T_{(\lambda^0_{\alpha,\beta}, u_0)} \calm(V_\alpha, X) \to
T_{(\lambda^0_{\alpha,\beta}, u_0)} \calm(V_\alpha, X),
\cr
\psi'_{\alpha, \beta}(0) &=& u_0|_{W_{\alpha, \beta}}, &
d\psi'_{\alpha, \beta}(0) &=&
\id: T_{u_0} \calh^{1,p}(W_{\alpha, \beta}, X)
\to T_{u_0} \calh^{1,p}(W_{\alpha, \beta}, X).
\endmatrix$$ Shrinking the balls $B_\alpha$ and $B_{\alpha \beta}$, if nessessary, we can suppose that for all $\xi_\alpha\in B_\alpha$ and $\xi_{\alpha\beta} \in B_{\alpha \beta}$ the restrictions $\psi_\alpha(\xi_\alpha)|_{W_{\alpha,\beta}}$ and $\psi_{\alpha \beta}(\xi_{\alpha \beta})|_{ W_{\alpha, \beta}}$ belong to the image $\psi'_{\alpha, \beta} (B'_{\alpha, \beta })$. Consider a holomorphic map $$\matrix\format \c\,& \c& \c & \c\,& \c \,& \l\\
\Psi:& \prod_\alpha B_\alpha &\times&
\prod_{\alpha <\beta} B_{\alpha \beta}
& \lrar & \sum_{\alpha, \beta} \calh^{1,p}(W_{\alpha, \beta}, u_0^*TX)
\cr
\Psi:& (v_\alpha, && v_{\alpha \beta}) & \mapsto &
\psi^{\prime\;\;-1}_{\alpha, \beta}
\bigl(\psi_\alpha(v_\alpha)|_{W_{\alpha, \beta}}\bigr)
- \psi^{\prime\;\;-1}_{\alpha, \beta}
\bigl(\psi_{\alpha \beta}(v_{\alpha \beta})|_{W_{\alpha, \beta}}\bigr).
\endmatrix$$ As in analogous situations, which have already appeared in this paper, the map $\Psi$ gives the gluing condition of local holomorphic maps $\psi_\alpha (v_\alpha): V_\alpha \to X$ and $\psi_{\alpha \beta} (v_{\alpha \beta}): V_ {\alpha \beta}\to X$. Thus we can identify $\calm \cap \prod_\alpha B_\alpha \times \prod_{\alpha
<\beta} B_{\alpha \beta}$ with the set $\Psi\inv(0)$.
Let us study in detail the behavior of $\Psi$ in the point $y_0 \in \prod
B_\alpha \times \prod B_{\alpha \beta}$, $y_0 = \bigl(\psi_\alpha \inv (u_0
|_{V_\alpha}), \psi_{\alpha \beta} \inv (u_0| _{V_{\alpha \beta}})\bigr)$, $\Psi(y_0) =0 \in \sum_{\alpha, \beta} \calh^{1,p}(W_{\alpha, \beta},
u_0^*TX)$. One easily sees that the tangent space at $y_0$ is $$T_{y_0}\bigl( \prod B_\alpha \times \prod B_{\alpha \beta}\bigr)
=
\sum_\alpha \calh^{1,p} (V_\alpha, u_0^*TX) \oplus
\sum_{\alpha \beta} \calh^{1,p} (V_{\alpha \beta}, u_0^*TX)
\oplus \cc^N,$$ and the differential $d\Psi(y_0)$ on the summand $\sum\calh^{1,p} (V_\alpha,
u_0^*TX)\oplus \sum \calh^{1,p} (V_{\alpha \beta},$
$ u_0^*TX )$ coincides with the Čech codifferential $(9.1.1)$ with respect to the covering $\{V_\alpha, V_{\alpha\beta}\}$ of the curve $C_0$. By [*Lemma 9.2.1*]{} we can represent $\sum
\calh^{1,p} (W_{\alpha, \beta}, u_0^*TX)$ as a direct sum $\calw \oplus
\calq$, where $\calw= \im(d\Psi(y_0))$, and $\calq$ is isomorphic to $\sfh^1(C_0, u_0^*TX)$ and of finite dimension. Let $\Psi_\calw$ and $\Psi_\calq$ be the components of $\Psi= (\Psi_\calw, \Psi_\calq)$ with respect to this decomposition, and let $\wt\calm \deff \Psi_\calw\inv(0)$. By [*Lemma 9.2.1*]{} and the implicit function theorem, $\wt\calm$ is a complex submanifold in $\prod B_\alpha \times\prod B_{\alpha\beta}$ with a tangent space at $y_0\in \wt\calm$ equal to $\calh^{1,p} (C_0, u_0^*TX) \oplus
\cc^N$, and $\calm$ is defined in $\wt\calm$ as a zero set of a holomorphic mapping $\Phi\deff \Psi_\calq
|_{\wt\calm}: \wt\calm \to \calq \cong \sfh^1(C, u_0^* TX)$. This defines on $\calm$ a structure of the Banach analytic set of finite definition.
All that remains is to construct the corresponding holomorphic family of nodal curves $\pi: \calc\to \calm$ and the holomorphic mapping $F: \calc \to X$. Note that to each ball $B_\alpha$ we can naturally associate a trivial family $\pi_\alpha: \calc_\alpha
\deff B_\alpha \times V_\alpha \to B_\alpha$, and to each ball $B_{\alpha\beta}$ a holomorphic family $\pi_{\alpha\beta}:
\calc_{\alpha\beta} \to B_{\alpha \beta}$, with fiber $\pi_{\alpha\beta}
\inv( v_{\alpha\beta} )$ equal to $A_{\lambda_{\alpha\beta}}$, where $\lambda_{ \alpha\beta}$ is uniquely determined by the relation $\psi_{\alpha\beta} (v_{\alpha\beta}) = (\lambda_{\alpha\beta},
u_{\alpha\beta}) \in \calm( V_{\alpha\beta}, X)$.
Extend these families to the families $\ti\pi_\alpha: \wt\calc_\alpha \to\prod B_\alpha \times\prod B_{\alpha\beta}$ and $\ti\pi_{\alpha\beta}: \wt\calc_{ \alpha \beta} \to\prod B_\alpha
\times\prod B_{\alpha\beta}$. It is obvious that $\wt\calc_\alpha$ and $\wt\calc_{\alpha\beta}$ canonically patch into a globally defined family of nodal curves $\ti\pi: \wt\calc \to
\prod B_\alpha \times\prod B_{\alpha\beta}$. $\calc$ will be a Banach manifold. Moreover, we shall obtain correctly defined holomorphic maps $F_\alpha: \wt\calc_\alpha \to X$ and $F_{\alpha\beta}: \wt\calc_{
\alpha \beta} \to X$ such that for $z\in V_\alpha$ one has $F(v_\alpha,
v_{\alpha\beta}, z) \deff \psi_\alpha(v_\alpha)[z]$ plus analogous relations for $F_{\alpha\beta}$.
Let us define $\calc$ as a restriction $\calc \deff \wt\calc|_\calm$. Note that the restriction of the [*trivial*]{} holomorphic family $\wt\calc_\alpha= V_\alpha \times \prod B_\alpha \times \prod
B_{\alpha\beta}$ $\calm$ is also a trivial holomorphic family.
Thus, $\calc$ is a b.a.s.f.d. in the neighborhood of the points $y \in \calc \cap
\wt\calc_\alpha$. Analogously, every holomorphic family of curves $\wt\calc_{\alpha\beta}$ is defined in the trivial bundle $\Delta^2 \times \prod B_\alpha \times \prod B_{\alpha \beta} \buildrel\pr
\over\lrar \prod B_\alpha \times\prod B_{\alpha\beta}$ by the condition $z_1\cdot z_2- \lambda_{\alpha\beta} =0$, where $\lambda_{\alpha\beta}
:B_{\alpha \beta} \to \Delta$ is a holomorphic parameter of the deformation of the complex structure in $V_{\alpha \beta}$, and $z_1$ with $z_2$ are the standard coordinates in $\Delta^2$. So $\calc$ is a b.a.s.f.d. also in the neighborhood of points $y \in \calc \cap \wt\calc_{\alpha \beta}$. Since $\calm$ was defined by the condition of the coincidence of local mappings $F_\alpha$ and $F_{\alpha\beta}$; thus, on $\calc$ a global holomorphic map $F: \calc \to X$ is defined.
Properties [*a), b) [and]{} d) of *Theorem 9.5.1**]{} for $\calm$, $\calc$ and $F$ follow directly from the construction, and one obtains property [*c)*]{} by using [*Theorem 5.3.2*]{}.
*9.6. Drawing Families of Curves.*
In the proof of the Continuity Principle for meromorphic mappings, which we need in this paper (see [*Theorem 4.2*]{} of the next paragraph or [*Theorem 5.1.3*]{} from \[I-S\]), we use the following consequence of the main results of the present paragraph.
Let $(C_n, u_n)$ be a sequence of irreducible stable curves over $X$, converging to a stable curve $(C_\infty, u_\infty)$.
Lemma 9.6.1. *There is an index $N$ and a complex (maybe singular) surface $Z$ with holomorphic mappings $\pi_Z :Z \to \Delta$ and $F: Z \to
X$, which define together a holomorphic family of stable nodal curves over $X$, joining $(C_N, u_N)$ with $(C_\infty, u_\infty)$. More precisely, the following are true:*
$1)$ For every $\lambda\in \Delta$ the fiber $C_\lambda=\pi_Z^{-1}(\lambda)$ is a connected nodal curve with boundary $\d C_\lambda$; a pair $(C_\lambda, u_\lambda)$ with $u_\lambda \deff F|_{C_\lambda}$ is a stable curve over $X$.
$2)$ All curves $C_\lambda$, except for finite numbers, are connected and smooth.
$3)$ $(C_0,u_0)$ is equal to $(C_\infty, u_\infty)$, and there exists $\lambda_N\in \Delta$ such that $(C_{\lambda_N},u_{\lambda_N})=(C_N,u_N)$.
$4)$ There exist open sets $V_1,\ldots,V_m$ in $Z$ such that each $V_j$ is biholomorphic to $\Delta \times A_j$, where $A_j$ is an annulus. Moreover, the diagram $$\matrix
V_j&\buildrel \cong \over \lrar &\Delta\rlap{$\times A_j$}
\cr
\llap{$\pi$}\downarrow & &\downarrow\rlap{$\pi_{\Delta }$}
\cr
\Delta & = & \Delta
\endmatrix$$ is commutative; each annulus $C_\lambda \cap V_j \cong \{\lambda\}
\times A_j$ is adjacent exactly to one of the components of the boundary $\d C_\lambda$, and the number $m$ of domains $V_j$ is equal to the number of the boundary components of each curve $C_\lambda$.
Remark. This lemma was stated without proof in [*Proposition 5.1.1*]{} from \[I-S\].
Proof. Let $\pi_\calc: \calc \to \calm$ and $\ev: \calc \to
X$ define a complete family of holomorphic deformations of the stable nodal curve over $X$ $(C_\infty, u_\infty)$, constructed in [*Theorem 9.4.1*]{}. Let $\lambda^* \in \calm$ parameterize the curve $(C_\infty, u_\infty)$. Choose a sequence $\lambda_n \lrar \lambda^*$ ¢ $\calm$ such that $(C_n, u_n) \cong (\pi_\calc\inv(\lambda_n),
\ev|_{\pi_\calc\inv(\lambda_n)})$ for all sufficiently big $n$.
While $(C_\infty, u_\infty) = (\pi_\calc\inv (\lambda^*),
\ev|_{ \pi_\calc \inv(\lambda^*)})$ lifts in the neighborhood of the boundary of $\hat U$, thus by shrinking $\calm$, if nessessary, we can suppose that this is true for all $\lambda \in \calm$.
By Theorem 9.5.1 the space $\calm$ is a b.a.s.f.d. therefore, we can apply [*Theorem 9.1.1*]{}. In particular, there are only finitely many irreducible components at $\lambda^*$ of $\calm$. Let $\calm_1$ be a component of $\calm$, which contains infinitely many $\lambda_n$. Represent $\calm_1$ in the neighborhood of $\lambda^*$ as a proper ramified covering $\pi_1: \calm_1 \to B_1$ over a Banach ball $B_1$. If $\Delta$ is an imbedded disk in $B_1$, then $\pi_1\inv(\Delta)$ is a one-dimensional analytic set, irreducible components which can be also parameterized by the disks. Thus, there exists a holomorphic mapping $\phi: \Delta \to \calm_1$, passing through $\lambda^*$ and $\lambda_N$ for some $N>\!>1$. The pre-image of the family $\pi_\calc: \calc\to \calm$ with respect to $\phi$ defines a holomorphic family of stable nodal curves over $\Delta•$ with total space $\pi_Z: Z \to\Delta$ and mapping $F:Z \to X$, containing $(C_\infty,
u_\infty)$ and $(C_N, u_N)$.
While $C_N$ is smooth, the general fiber $C_\lambda =\pi_Z\inv(\lambda)$, and again is smooth. Taking if nessessary a smaller disk, we can suppose that $C_\lambda$ are singular only for a finite number of $\lambda \in \Delta$. All other properties 1)–5) follow from the construction of the family $\pi_Z: Z\to \Delta$ and a mapping $F:Z \to X$.
[Lecture 10]{}
[Envelopes of Meromorphy of Two-Spheres.]{}
*10.1. Continuity Principles Relative to Kähler Spaces.*
Our aim in this paragraph is to prove [*Theorem 4.1*]{} and give some corollaries from it. First of all we need an appropriate form of the so- called “continuity principle” for the extension of meromorphic mappings.
For the notion of meromorphic mapping from a domain $U$ in a complex manifold into a complex manifold (or space) $Y$ we refer to \[Rm\]. We only point out here that meromorphic mappings into $Y = \cc\pp^1$ are exactly meromorphic functions on $U$, see \[Rm\].
Definition 10.1.1. *A Hermitian complex space $Y$ is called disk-convex if, for any sequence $(C_n,u_n)$ of smooth curves over $Y$ parameterized by the same surface $\Sigma $, such that*
1\) $\area(u_n(C_n))$ are uniformly bounded and
2\) $u_n$ $C^1$-converges in the neighborhood of $\d C_n $,
*there is a compact $K\subset Y$ which contains all $u_n(C_n)$.*
This definition obviously carries over to the case where $Y$ is a symplectic manifold. In this case one should consider $(C_n,u_n)$ as $J_n$-holomorphic curves, with $J_n$ converging to some $J$ (everything in $C^1$-tology) and all structures being tamed by a given symplectic form.
Let $U$ now be a domain in the complex manifold $X$, and $Y$ is a complex space .
Definition 10.1.2. *An envelope of meromorphy of $U$ relative to $Y$ is the largest domain $(\hat U_Y,\pi )$ over $X$, which contains $U$ (i.e., there exists an imbedding $i:U\to \hat U_Y$ with $\pi \scirc i={\sl Id}$) such that every meromorphic map $f:U\to Y$ extends to a meromorphic map $\hat f:\hat U_Y\to Y$.*
Using the Cartan-Thullen construction for the germs of meromorphic mappings of open subsets of $X$ into $Y$ instead of germs of holomorphic functions, one can prove the existence and uniqueness of the envelope.
**Proposition 10.1.1. *For any domain $U$ in the complex space $X$ and for any complex space $Y$ there exists a maximal domain $(\hat U_Y,\pi )$ over $X$, containing $U$ such that every meromorphic mapping $f:U\longrightarrow Y$ extends to a meromorphic mapping $\hat f:\hat U_Y\longrightarrow Y$. Such a domain is unique up to a natural isomorphism.***
See \[Iv-1\] for details.
Theorem 10.1.2. (Continuity Principle-I). *Let $X$ be a disk-convex complex surface and $Y$ a disk-convex Kähler space. Then the envelope of meromorophy $(\hat U_Y,\pi )$ of $U$ relative to $Y$ is also disk-convex with respect to the pulled-back Kähler form.*
This result can be reformulated in more familar terms as follows.
Let $\{ (C_t,u_t)\}_{t\in [0, 1]}$ be a continuous (in a Gromov topology) family of complex curves over $X$ with boundaries, parameterized by a unit interval. More precisely, for each $t\in [0, 1[$ a smooth Riemann surface with boundary $(C_t, \d C_t)$ is given together with the holomorphic mapping $u_t: C_t\longrightarrow X$, which is $C^1$-smooth up to the boundary. Note that $C_1$ is not supposed to be smooth, i.e., it can be a nodal curve!
Suppose that in the neighborhood $V$ of $u_0(C_0)$ a meromorphic map $f$ into the complex space $Y$ is given.
**Definition 10.1.3. We shall say that the map $f$ meromorphically extends along the family $(C_t,u_t)$ if for every $t\in [0, 1]$ a neighborhood $V_t$ of $u_t(C_t)$ is given, and given a meromorphic map $f_t:V_t
\longrightarrow
Y$ such that**
a\) $V_0 = V$ and $f_0= f$;
b\) if $V_{t_1}\cap V_{t_2}\not=\emptyset $ then $f_{t_1}\ogran_{V_{t_1}\cap
V_{t_2}} = f_{t_2}\ogran _{V_{t_1}\cap V_{t_2}}$.
**Theorem 10.1.3. (*Continuity Principle-II) *Let $U$ be a domain in the complex surface $X$. Let $\{ (C_t,u_t)\} _{t\in [0, 1]} $ be a continuous family of complex curves over $X$ with boundaries in $U_1$, a relatively compact subdomain in $U$. Suppose moreover that $u_0(C_0)\subset U$ and that $C_t$ for $t\in [0,1[$ are smooth. Then every meromorphic mapping $f$ from $U$ to the disk-convex Kähler space $Y$ extends meromorphically along the family $(C_t,u_t)$.****
Taking an image manifold $Y$ complex line $\cc $ or $\cc \pp ^1$ we obtain the continuity principles for holomorphic or meromorphic functions. When it is necessary to emphasize that we are considering the mappings into a certain manifold $Y$, we shall refer to the statement above as the *continuity principle relative to $Y$ or *c.p. for the meromorphic mappings into $Y$.**
The discussion above leads to the following
**Corollary 10.1.4. *If we have the domain $U$ in a complex surface $X$, a Kähler space $Y$ and a family $\{ (C_t,u_t)\} $ satisfying the conditions of the “continuity principle”, then the family $\{ (C_t,u_t)\} $ can be lifted onto $\hat
U_Y$, [*i.e.,*]{} there exists a continuous family $\{ (C_t,\hat u_t) \} $ of complex curves over $\hat U$ such that $\pi \scirc \hat u_t = u_t$ for each $t$.***
Of course the point here is that the mapping can be extended to the neighborhood of $u_1(C_1)$, which is a reducible curve having in general compact components. This makes our situation considerably more general than the classical one, i.e., when $X$ is supposed to be Stein; compare with \[Ch-St\].
Remark. Let us explain the meaning of this theorem by an example. Let $X$ be disk-convex and $U\subset X$ some domain. Furthermore, let $f:U \to Y$ be a meromorphic map and $\{(C_n, u_n)\}$ a sequence of stable complex curves over $X$, converging in the Gromov topology to $(C_\infty, u_\infty)$. Suppose that the images of the boundaries $u_n (\d C_n)$ and $u_\infty (\d C_\infty)$ are contained in $U$. Suppose also that $f$ extends along every curve $u_n (C_n)$. This means that there exists a complex surface $V_n$, containing $C_n$, and a locally biholomorphic map $u'_n:V_n \to X$ such that $u'_n|_{C_n} = u_n$ and $f$ meromorphically extend from $u'_n{}\inv U$ onto the whole $V_n$.
The latter is equivalent to the lift of the curves $(C_n, u_n)$ into the envelope $\hat U$, i.e., to the existence of holomorphic mappings $\hat u_n: C_n \to \hat U$ such that $\hat \pi \scirc \hat u_n =u_n$. In other words, one can take as $V_n$ a neighborhood of the lift of $C_n$ into $\hat U$. One easily sees that the curves $(C_n,\hat u_n)$ are stable over $\hat U$, have uniformly bounded areas and converge in the neighborhood of the boundary $\d C_n$. By [*Theorem 10.1.2*]{} and by the Gromov compactness theorem , some subsequence $(C_n,\hat u_n)$ converges to the $\hat{U}$-stable curve $(C_\infty,\hat u_\infty)$, and $\hat\pi \scirc \hat u_\infty
=u_\infty$. This means that $f$ extends along the curve $u_\infty (
C_\infty)$. Thus, [*Theorem 4.2*]{} is a generalization of the E. Levi continuity principle.
[*10.2. Proof of the Continuity Principle.*]{}
Suppose that there is a subsequnce of our sequence, which we still denote by $(C_n,u_n)$ and which is not contained in any compact subset of the envelope $(\hat U_Y, \pi )$. Put $v_n=\pi \scirc u_n$ and consider our sequence as a sequence $(C_n,v_n)$ of stable curves over $X$. While the areas are bounded, we can suppose that this sequence converges in the Gromov topology by the disk-convexity of $X$. Denote by $(C_0,v_0)$ its limit. For $N$ sufficiently large take a holomorphic family $(\calc ,\pi ,\calv )$ joining $(C_N,v_N)$ with $(C_0,v_0)$ as in [*Lemma 9.6.1.*]{}
Let $f:\hat U_Y \to Y$ be some meromorphic mapping of our envelope into a disk-convex Kähler space $Y$. Composing $f$ with $\calv$, we obtain a meromorphic map $h:= f\scirc \calv :
\bigcup U_j\cup \pi^{-1}(V(\lambda_N))\to Y$. Here $V(\lambda_N)$ is a suffuciently small neighborhood of $\lambda_N \in \Delta $.
Denote by $W$ the maximal connected open subset of $\Delta $ containing $V(\lambda_N)$ such that $h$ meromorphically extends onto $\bigcup U_j\cup \pi^{-1}(W)$. We want to prove that $W=\Delta
$. Denote by $\Gamma_{f_{\lambda }}$ the graph of the restriction $f\mid_{C_{\lambda }}$. Fix some Hermitian mertic on $\calc $.
Lemma 10.2.1. *For any compact $K\subset \Delta $ there is a constant $M_K$ such that ${\sl area}(\Gamma_{f_{\lambda }})\le M_K$ for all $\lambda
\in W\cap K$.*
*Proof. Shrinking $\calc $ if nessessary, we can suppose that $f$ has only a finite number of points of indeterminancy in $\bigcup_jU_j$. Denote by $S_W$ the discrete subset in $W$, which consists of points $s\in W$ such that either the fiber $C_s$ is singular or $s$ is the projection of the indeterminancy points of $f\mid_{\calc \mid_W}$. Fix a point $\lambda_1\in W\setminus S_W$. Take a path $\gamma :[0,1]\to
W\setminus
S_W$ connecting $\lambda_1$ with some $\lambda_2\in W\setminus S_W$. Take some relatively compact in $W\setminus S_W$ neighborhood $V$ of $\gamma ([0,1])$.*
Recall that a Kähler metric on a complex space $Y$ consists of a locally finite covering $\{ V_{\alpha }\} $ of $Y$ and strictly plurisubharmonic functions $\phi_{\alpha }$ on $V_{\alpha }$ such that $\phi_{\alpha }-\phi_{\beta }$ are pluriharmonic on $V_{\alpha }\cap V_{\beta }$. The sets $\{ f^{-1}(V_{\alpha })\} $ form a covering of $\pi^{-1}(V)$ and $dd^cf^*\phi_{\alpha }=w$ is a correctly- defined semi-positive closed form on $\pi^{-1}(V)$. We have $${\sl area}(\Gamma_{f_{\lambda }})={\sl area}(C_{\lambda })+
\int_{C_{\lambda }}w.$$ By the Stokes formula $$\int_{C_{\gamma (0)}}w-
\int_{C_{\gamma (1)}}w =
\int_{\cup_tC_{\gamma (t)}}dw -
\int_{\cup_t{\d C_{\gamma (t)}}}w=$$ $$=
\int_{\cup_t{\d C_{\gamma (t)}}}w.$$ This is obviously uniformly bounded on $\lambda_2\in K\setminus S_W$, which implies the uniform bound in $K$ by the Gromov (or Bishop, in this case) compactness theorem.
Lemma 10.2.2. *Mapping $h$ meromorphically extends onto $\calc $.*
*Proof. Let us prove first that $h$ meromorhpically extends onto $\calc $ minus the singular fiber $C_0$, i.e., that $W\supset \Delta \setminus \{ 0\} $.*
The same arguments as in \[Iv-2\] (the only difference being that in \[Iv-2\] curves $C_
{\lambda }$ were disks) show that $(\Delta \cap \d W)\setminus (\{ 0\})
$ is an analytic variety. Now, using a Thullen-type extension theorem of Siu, see \[Si-1\], we can extend $h$ onto $\calc \setminus C_0$.
Denote by $\hat C_0$ the union of compact components of $C_0$. We shall show that $\hat C_0$ contracts to a finite number of normal points. To prove this we can suppose that $\hat C_0$ is connected. Otherwise, we can apply the same arguments as below to the connected components. We shall now prove that $\hat C_0$ contracts to a normal point. Denote by $E_1,...,E_n$ the irreducible components of $\hat C_0$. All we must prove is that the matrix $(E_i,E_j)$ is negatively defined, see \[Gra\]. Denote by $l_i$ the multiplicity of $E_i$. Thus $\hat C_0=\Sigma_{i=1}^nl_i\cdot E_i$. Denote by $M$ the $\zz $-module generated by $E_1,...,E_n$ with scalar product defined by intersection of divisors. Put $D=\hat C_0$. Then we have
1\) $E_i\cdot E_j\ge 0$ for $i\not= j$;
2\) $D\cdot E_i\le 0$ for all $i$, because this number is not more than the intersection of $E_i$ with the nonsingular fiber $C_{\lambda }$.
By Proposition 3 from \[Sh\], v.1, Appendix I, we have $A\cdot A\le 0$ for all $A\in M$ and $A\cdot A=0$ iff $A$ is proportional to $D$. But $D\cdot D <\hat C_0\cdot C_{\lambda }=0$, where $C_{\lambda }$ is a smooth fiber. This proves that $(E_i\cdot E_j)$ is negatively defined.
Therefore, all that is left to prove is that the normal point is a removable singularity for the meromorphic mappings into a disk-convex Kähler space.
Let $(\calc ,s)$ be a germ of two-dimensional variety with an isolated normal singularity $s$. Let a meromorphic map $h:\calc \setminus \{ s\} \to Y$ be given. Realize $\calc $ as a finite proper analytic cover over a bi-disk $\Delta^2$ with $s$ being the only point over zero. Denote by $p:\calc \to \Delta^2$ this covering.
The composition $h\scirc p^{-1}$ is a multivalued meromorphic map from $\Delta^2 \setminus \{ 0\} $ to $Y$. This can be extended to the origin, see \[Si-1\]. Consider the following analytic set in $\calc\times Y$: $$\Gamma = \{ (x,y)\in \calc \times Y: (p(x),y)\in \Gamma_{h\scirc p^{-1}} \} .$$ The irreducible component of $\Gamma$, which contains $\Gamma_h$, will be the graph of the extension of $h$ onto $\calc $.
Let us turn to the proof of the theorem. We have a holomorphic family $\calc \to \Delta $ of stable curves $(C_{\lambda },
v_{\lambda })$ over $X$ such that:
1\) $(C_{\lambda_0},v_{\lambda_0})=(C_0,v_0)$;
2\) $h=f\scirc \calv $ extends meromorphically onto $\calc $, where $\calv :\calc \to X$ is an evaluation map.
We want to lift this family to the envelope $(\hat U_Y, \pi )$. Denote by $E_1,...,E_l$ the set of all irreducible curves in $\calc $ which are contracted by $\calv :\calc \to X$ to the points. It follows that $\calc \setminus
\bigcup_{j=1}^lE_j$ lifts to the envelope by the Cartan-Thullen construction. Note further that $E_j$ do not intersect $\bigcup_{j=1}^mU_j$. That is why either $E_j\subset \hat C_0$ is a compact component of singular fiber or projects surjectively on $\Delta $ under the projection $\pi :\calc \to \Delta $. In the second case it intersects $C_0$ and the point $u(C_0 \cap E_j)$ lifts to the envelope. This proves that $\calc \setminus \hat C_0$ lifts to the envelope. More precisely, we have shown that $\calc \setminus
\hat C_0'$ lift to the envelope, where $\hat C_0'$ is a union of all compact components of the singular fiber mapped by $\calv $ into the points.
Denote by $\hat C$ some connected subset of $\hat C_0'$ and choose its neighborhood $V$ not to intersect other components of $\hat C_0'$. Then $\calv $ maps $V$ onto the neighborhood of the point $\calv (\hat C)$ in $X$. Now it is clear that this point lifts to the envelope.
Thus, we have proved that $(C_0,v_0)$ lifts to the envelope $(\hat U_Y,\pi )$ of $U$. This implies that $(C_n,u_n)$ should lie in a compct subset of $\hat U_Y$. This is the desired contradiction.
Let us give one corollary of the Continuity Principle just proved.
Corollary 10.2.3. *Let $X$ be a complex surface with one singular normal point $p$. Let $D$ be a domain in $X$, $\d D\ni p$. Suppose there is a sequence $(C_n,u_n)$ of stable curves over $X$ converging to $(C_0,u_0)$ in a Gromov topology and such that*
a\) there is a compact $K\subset D$ with $u_n(\d C_n)\subset K$ for all $n$;
b\) $p\in u_0(C_0)$.
Then every meromorphic function from $D$ extends to the neighborhood of $p$.
To our knowledge this statement is new also for holomorphic functions.
*10.3. Construction of Envelopes - I.*
We are now ready to give the proof of [*Theorem 4.1*]{}, first under some technical assumptions to clarify the main idea. Then in the next paragraph the general case will be considered.
Let a Kähler surface $(X,\omega )$ be fixed and some symplectic immersion $u:\ss^2\to X$, with only positive double intersections be given. Our technical assumptions for the moment will be the following:
1\. We suppose that $c_1(X)[M]$ is not only positive but moreover bigger than the sum $\delta $ of double points of $M:=u(\ss^2)$.
2\. $X$ is supposed to be “positive” in the sense that for any almost-complex structure $J\in \calj_{\omega }$ and $J$-complex sphere $C$ satisfies $[C]^2\ge 0$.
As we have already mentioned, those conditions will be eliminated in the next paragraph.
Now let $f$ be a meromorphic function in some neighborhood $U$ of $M$. Fix some relatively compact $U_1\Subset U$ containing $M$. Let $\{ J_t\}_{t\in [0,1]}$ be a family of $C^1$-smooth almost complex structures on the $4$-manifold $X$ constructed in [*Lemma 1.4.2*]{}. We lift the structures $J_t$ onto the envelope $(\hat U,\pi )$ in such a way that those liftings are standard on $\hat U\setminus i(U_1)$, where $i:U\to \hat U$ is a canonical imbedding.
For $t=1$ our $J_1$-holomorphic sphere $M=M_1$ is given. By assumption $c_1(X)[M]$ $\delta_1 =p \ge1$. Here $\delta_1=\delta $ is the geometric self-intersection of $M_1$, i.e., the number of double points. From [*Theorem 8.4.1*]{} we get a family $M_t=f(J_t)$ of $J_t$-holomorphic curves for $t$ sufficiently close to $1$, where $f:V\subset \calj \to
\cals \times \calj_S \times \calj$ is a local section of $\pr_\calj$ constructed in [*Theorem 8.4.1*]{}.
Definition 10.3.1. *We say that the family $\{M_t\}_{t\in(\hat t,1]}$ of (possibly reducible) $J_t$-holomorphic curves is semi-continuous if there exist a (maybe infinite) partition of the interval $(\hat t,1]$ of the form $1=t_0 >t_1>\ldots>t_n>\ldots$ with $t_n\searrow \hat t$, natural numbers $1=N_0 \le N_1 \le\ldots \le N_n \le\ldots$, $J_t$-holomorphic maps $u_t: \bigsqcup_{j=1}^{N_i} S^j_i\to X$ for $(t_{i+1}, t_i]$ with $M_t\deff u_t(\bigsqcup_{j=1}^{N_i} S^j_i)$ such that*
$u_t: (t_{i+1}, t_i] \times \bigsqcup_{j=1}^{N_i} S^j_i\to X$ is a continuous map;
$\area(M_t)$ are uniformly bounded from above;
$\calh$-$\lim_{t\searrow t_{i+1}} M_t \deff \overline
M_{t_{i+1}} \supset M_{t_{i+1}}$.
The inclusion $\overline M_{t_{i+1}} \supset M_{t_{i+1}}$ means that $M_{t_{i+1}}$ has no other components than those of $\overline M_{t_{i+1}}$.
We say that $\{M_t\}$ is a family of spheres if all $S^j_i$ are spheres.
Let $T$ be the infimum of such $\hat t$, for which there is a semi-continuous family $\{M_t\}_{t\in (\hat t, 1]}$ of spheres such that for all irreducible components $M^1_t,\ldots M^{N_i}_t$ of $M_t$, $t\in (t_{i+1}, t_i]$, one has
[*a)*]{} $c_1(X)[M^j_t] -\delta_t^j -\varkappa_t^j=p_t^j \ge1$;
[*b)*]{} $\sum_{j=1}^{N_i} p_t^j \ge p$.
Here $\varkappa_t^j$ is the sum of the Milnor numbers of cusp points of $M_t^j$, see Appendix 2. We allow the existence of multiple components, [*i.e.,*]{} that some of $M_t^j$ can coincide.
The set $T$ is open by [*Theorem 8.4.1*]{}. To prove the closeness of $T$, we note that since all $J_t$ are tamed by the same form $\omega$, the areas of a $J_t$-holomorphic curve are uniformly in $t$ estimated from above and below by $\int_{M_t} \omega$. Moreover, since $\overline M_{t_{i+1}} \supset
M_{t_{i+1}}$, we obtain $\int_{M_{t_i}} \omega \le \int_{M_{t_{i+1}}}
\omega$. In particular, this implies that the sequence $\{N_i\}$ is bounded from above, and hence stabilizes for $i$ big enough. From Gromov’s compactness theorem and disk-convexity of the envelope (Theorem 2.2.2)we obtain that for every $j={1,\ldots,N_i}$ the sequence $\{M^j_i\}_{i=1}^\infty$ has a subsequence, still denoted in the same way, which converges to a $J_{\hat t}$-holomorphic curve $\overline
M^j\subset \hat U_Y$.
To simplify the notations, we drop the upper index $j$ from now on and write $\overline M$ instead $\overline M^j$, having in mind that we can do the same constructions for all $j=1,\ldots,N_i$. Note that all irreducible components of $\overline M$ are $J_{\hat t}$-holomorphic spheres. We write $\overline M = \bigcup_{k=1}^d m_k\cdot M_k$, where $M_k$ are distinct irreducible components of $\overline M$ with the multiplicities $m_j\ge1$. The genus formula for $\overline M$ now has (because of multiplicities!) the following form:
$$0=\msmall{[\overline M]^2 - c_1(X)[\overline M] \over2}
+ \sum_{k=1}^d m_k - \sum_{k=1}^d m_k(\delta_k +\varkappa_k)
-\sum_{k < l}m_k\cdot m_l [M_k]\cdot [M_l]$$ $$-\sum_{k=1}^d \msmall{m_k(m_k -1) \over2 } [M_k]^2.
\eqno(10.3.1)$$
The formula can be obtained by taking the sum of the genus formulas for each $M_k$ and then completing the sum $\sum_{k=1}^d m_k[M_k]^2$ to $[\overline M]^2$. Here $\delta_k$ and $\varkappa_k$ denote self-intersection and Milnor numbers of $M_k$. We also have $$\varkappa^j_i + \delta^j_i = \msmall{ [M^j_i]^2 - c_1(X)[M^j_i] \over2} +1.
\eqno(10.3.2)$$ Note that $[\overline M]^2 = [M^j_i]^2$ and $c_1(X)[\overline M]=
c_1(X)[M^j_i]$ for $i$ sufficiently big, and that $\sum_{k<l} m_k m_l
\,[M_k][M_l] \ge \sum_{k=1}^d m_k -1$, because the system of curves $\bigcup_{k=1}^d M_k$ is connected. From (10.3.1) and (10.3.2) we obtain $$\sum_{k=1}^d m_k(\delta_k + \varkappa_k) +
\sum_{k=1}^d \msmall{m_k(m_k-1) \over2 } [M_k]^2
\le \delta^j_i + \varkappa^j_i$$ and $$\sum_{k=1}^d m_k c_1(X) [M_k] = c_1(X)[M^j_i].$$ Using positivity of $(X,\omega)$, [*i.e.,*]{} the property $[M_k]^2 \ge0$, we obtain $$\sum_{k=1}^d m_k \biggl( c_1(X) [M_k] -\delta_k -\varkappa_k\biggr)
\ge c_1(X)[M^j_i] -\delta^j_i -\varkappa^j_i =p_j.$$ Thus we can choose among $M_1,\ldots,M_d$ the subset with properties [*a)*]{} and [*b)*]{}.
Returning to the old notations, we can “correct” all $\overline M^j$, choosing an appropriate subset $M^j \subset \overline M^j$. Now applying [*Theorem 8.4.1*]{} to all $M^j$, we can extend our family $\{M_t\}$ in a neighborhood of $\hat t$, and hence define it for all $t\in [0,1]$.
For $t=0$ we get the next statement.
The next rigidity property of symplectic imbeddings is a straightforward corollary from this theorem.
Corollary 10.3.2. *Let $M_1$ and $M_2$ be two symplectically imbedded spheres in $\cc \pp^2$. Then any biholomorphism of a neighborhood of $M_1$ onto a neighborhood of $M_2$ is fractional linear.*
[10.4. Construction of Envelopes - II.]{}
Let us now give the proof in the general case. In fact, the proof works not only for the meromorphic functions but also for the meromorphic mappings into any disk-convex complex Kähler space $Y$.
Theorem 10.4.1.
*Let $u:\ss^2\to X$ be a symplectic immersion of the sphere $\ss^2$ into a disk-convex Kähler surface $X$ such that $M=u(\ss^2)$.*
Let $u:S^2\to X$ be a symplectic immersion of a sphere, having also positive self-intersections. Let $U$ be a relatively compact domain in $X$, which contains $M:=u(S^2)$. Denote by $(\hat U_Y,
\hat\pi_Y)$ its envelope of meromorphy relative to $Y$.
*Step 1. There exists an ($\omega$-tamed) $J_0\in \calj_U$ such that $M$ is $J_0$-holomorphic.*
This was proved in [*Lemma 1.1.2*]{} of \[I-S\]. Moreover, there exists a smooth homotopy $h :[0,1]\to \calj_U$ joining $J_0=h(0) $ and $J\st =h(1)$. Put $M_0:=M$ and denote as in [*Lemma 8.3.2*]{} by $\calm_h(M_0,J_0)$ a component of $\calm_h$ through the point $(M_0,J_0)$.
*Step 2. Suppose that the component $\calm_h(M_0,J_0)$ is not compact.*
Then by (iii) of [*Lemma 8.3.2*]{} there is a continuous path $\gamma :[0,1]\to \calm_h$ starting from $(M_0,J_0)$ with property (b), see (iii) of [*Lemma 2.5*]{}. Consider $J_n$-holomorphic spheres $M_n$ in $\calm_h(M_0,J_0)$, which are discrete there with $J_n\to J^*\in \calj_U$ as in (b) of [*Lemma 2.5*]{}.
If for some $n$ $M_n\cap U=\emptyset$, then, because $J_n=J\st $ on $\hat U\setminus U$, this $M_n$ will be the rational curve we are looking for.
If not, then some sequence, still denoted by $M_n$, will converge in the Gromov topology to a reducible curve $M^{(1)}$. If $M^{(1)}$ has an irreducible component $M^{(1)}_0$, lying outside of $U$ and such that $c_1(X)[M^{(1)}_0]>0$, then $M^{(1)}_0$ is our rational curve.
If not, there exists a component $M^{(1)}_0$ of the limit curve $M^{(1)}$ such that $c_1[M^{(1)}_0]>0$. Repeat Step 2 for $M^{(1)}_0$ instead of $M_0$. Since the area of a complex spheres is bounded from above (see \[G\]), after a finite number of steps we obtain a needed rational curve in the envelope $\hat U_Y$, or arrive at
*Step 3. $\calm_h(M_0,J_0)$ is compact.*
From of [*Theorem 8.3.3*]{} and [*Corollary 8.3.4*]{} we immediatly see that there are continuous (in fact, piecewise $C^k$) paths $(M^n_t,J^n_t)$ such that
1\) $M^n_0=M_0$ for all $n$;
2\) $J^n_0=J_0$ for all $n$;
3\) $J^n_1\to J\st $.
By the Gromov compactness theorem some subsequence from $M^n_1$ , still denoted by $M^n_1$, is converging to a $J\st $-holomorphic nodal curve $C^*$ with $\pi^*c_1(X)[C]>0$. Take some irreducible component $C$ of the curve $C^*$. Then $C$ is the rational curve in $\hat U_Y$ we are looking for.
*10.5. Examples.*
Here we discuss a few more examples concerning envelopes of meromorphy.
Example 1. Let $(X,\omega) = (\cc\pp^1 \times \cc\pp^1,\omega_{FS}
\oplus \omega_{FS} )$, where $\omega_{FS}$ denote the Fubini-Studi metric on $\cc\pp^1$. Note that $c_1(X)= 2[\omega]$. Let $J$ be an $\omega$-tame almost complex structure on $X$ and $C$ be a $J$ - complex curve on $X$. Denote by $e_1$ and $e_2$ the standard generators of $\sfh_2(X,\zz)= \zz^2$ and write $[C]=a\cdot e_1 + b\cdot e_2$. Then we get $a+b=\int_C\omega \ge 1$ and $c_1(X)[C]= 2(a+b)$. Furthermore, by the genus formula $0\le g(C) + \delta
+\varkappa = (2ab -2(a+b))/2 +1 =(a-1)(b-1)$. Thus, we conclude that both $a$ and $b$ are non-negative and $[C]^2=2ab \ge0$. So $\cc\pp^1 \times \cc\pp^1$ is non-negative in our sense.
Let $M$ be an imbedded symplectic sphere in $X$. Then $(a-1)(b-1)=0$ by the genus formula. Therefore, we can assume that $a=1$ and $b\ge0$. Now one concludes that the following holds:
Corollary 10.5.1. *Let $M$ be a symplectic sphere in $(\cc\pp^1
\times \cc\pp^1,\omega_{FS} \oplus \omega_{FS})$. Then the envelope of meromorphy of any neighborhood of $M$ contains a graph of a rational map of degree $0\le d \le b$ from $\cc\pp^1$ to $\cc\pp^1$.*
Example 2. $(X,\omega) = (\cc\pp^2,\omega_{FS})$. Let $M$ be a symplectic surface in $X$ of degree $m\deff \int_M \omega$ with [*positive*]{} self-intersections. Then obviously $c_1(X)[M]=3m$. Note that we proceed to the construction of a family $\{M_t\}$ if the condition $c_1(X)[M^j_t] > \varkappa(M^j_t)$ is satisfied for all irreducible components $M^j_t$ of $M_t$. By the genus formula one has $\varkappa(C) \le (d-1)(d-2)/2$ for every complex curve $C$ of degree $d$. So we can proceed if $3m > (m-1)(m-2)/2$, which is equivalent to $1\le m \le8$. Thus, we have the following
Corollary 10.5.2. *Let $M$ be a symplectic surface in $\cc\pp^2$ of degree $m\le 8$ with positive self-intersections. Then the envelope of meromorphy of any neighborhood of $M$ coincides with $\cc\pp^2$ itself.*
Remark. Note that the examples include all imbedded symplectic surfaces in $\cc\pp^2$ of genus $g\le 21$.
Example 3. Let $X$ be a ball in $\cc^2$ and $w$ is a standard Euclidean form. Blow up the origin in $\cc^2$ and denote by $E$ the exceptional curve. By $\hat X$ denote the blown-up ball $X$. The blow-up of $\cc^2$ is also Kähler, and we denote by $w_0$ some Kähler form there. Consider a sufficiently small $C^1$-perturbation of $E$. This will be a $w_0$-symplectic sphere in $\hat
X$, which is denoted by $M$. The Chern class of the normal bundle to $M$ is equal to that of $E$ and thus is $-1$. Therefore, $c_1(\hat X)[M]=1$ and the proof of Theorem 1 applies. One should only note that in the process of deformation $M_t$ cannot break into irreducible or multiply covered components in this special case. The only rational curve in $\hat X$ is $E$. Thus, we see that
*the envelope of meromorphy of any neighborhood of $M$ contains $E$.*
One can then blow down the picture to obtain downstairs a sphere $M_1$-image of $M$ under the blown-down map. This $M_1$ is homologous to zero, so cannot be symplectic, and for this $M_1$ our Theorem 1 cannot be applied.
**Example 4. Chirka in \[C\] proved the following “local version” of our Theorem 1, which he called “a generalized Hartogs’ lemma”. Denote by $\Gamma $ a graph of continuous function $f:\bar\Delta\to \cc $. Consider the following “generalized Hartogs’ figure” in $\cc^2$: $H_\Gamma := \d \Delta \times \bar\Delta\cup \Gamma $. Chirka showed that every holomorphic function in the neighborhood of $H_{\Gamma }$ extends holomorphically onto the unit bidisk $\Delta^2$. This is a corollary of our Theorem 4.1 (as is explained in \[C\]). Really, denote by $(z,w)$ coordinates in $\cc^2$. Any function $f$ holomorphic in the neighborhood of $\d \Delta \times \bar\Delta $ is a sum of a function holomorphic in the bidisk and a function $f_{-}$ holomorphic in $(\cc\pp^1
\setminus \bar\Delta )\times \Delta $. This $f$ is also holomorphic in the neighborhood of $\Gamma $. We need only to extend $f_{-}$. Extending $\Gamma $ “inside” $(\cc\pp^1\setminus \bar\Delta )\times \Delta $ we obtain $\tilde\Gamma $ - a sphere, homologous to the $\{ pt\} \times \cc\pp^1$ in $\cc\pp^1\times \cc\pp^1$. Rescailing the variable $w$ we can make $\tilde\Gamma $ symplectic and find ourselves in the conditions of [**Example 2**]{}.**
The proof of Chirka is roughly the same (i.e., uses the perturbation of the structure), but somewhat simpler. It uses, instead of Gromov’s techniques, the results of Vekua on so- called generalized analytic functions.
Answering the question, posed by Chirka, Rosay in \[R\] constructed an example showing that a “generalised Hartogs’ lemma” does not hold in $\cc^3$.
Example 5. This example was explained to us by E. Chirka, and it shows that our Continuity Principle is not valid when the complex dimension of the manifold $X$ is more than two.
Take as $X$ the total space of the bundle $\calo(-1)\oplus \calo (-1)$ over $\cc\pp^1$. Denote an affine coordinate on $\cc\pp^1$ by $z$, coordinates on the fibers by $\xi_1$, $\xi_2$ and $\eta_1$, $\eta_2$ such that $\eta_1 = z\xi_1$ and $\eta_2 = z\xi_2$. Identify $\cc\pp^1$ with the zero section of the bundle. Consider the meromorphic function $f=
e^{\xi_2 /\xi_1}$. The set of essential singularities of $f$ is $\{\xi_1=0 \}$, which contains the zero section $\cc\pp^1$.
Consider the following sequence of analytic disks $C_n$ in $U\deff X \bs
\{\xi_1=0 \}$, $C_n \deff \{ \xi_2 =0, |z|\le n, \xi_1 = {z\over n} \}$. Then the function $f$ is defined in a neighborhood of every $C_n$. On the other hand, the limit curve of the sequence is $C_0 = \cc\pp^1 \cup \Delta_\infty$, where $\Delta_\infty \deff \{ \eta_2=0, z=\infty, |\eta_1| \le 1\}$. In particular, $f$ does not extend in a neighborhood of $C_0$.
[Appendix IV.]{}
[Complex Points and Stein Neighborhoods.]{}
**The condition $c_1(X)[M]>0$ (in the case of imbedded surfaces, i.e., when $\delta =0$) in Theorem 4.1 cannot be dropped. In \[N-1\] Nemirovsky, using results of Eliashberg-Kharlamov and Forstneric, showed that any imbedded complex curve $C$ with $c_1(X)[M]\le 0$ can be perturbed to an imbedded surface $M$ which has a basis of Stein neighborhoods.**
*A4.1. Complex Points of Real Surfaces in Complex Surfaces.*
Let $S\subset X$ be a real surface embedded in a complex surface. A point $p\in S$ is said to be [*complex*]{} if the tangent plane $T_p S\subset T_p X$ is a complex line.
If locally $S=\Gamma_f$ is the graph of a smooth complex valued function $f$, then a point $(z,f(z))\in S$ is complex if and only if ${\d f\over \d\bar z}(z)=0$. It is therefore not difficult to show that for generic embeddings the complex points are isolated, and near each complex point there exists a complex system of coordinates $(z,w)$ such that $S$ is given by the equation $$w=|z|^2+{\gamma\over 2}(z^2+\bar z^2)+\bar o(|z|^2),\quad\gamma\in\R.
\eqno (A4.1.1)$$ If $|\gamma|<1$, then $p=(0,0)$ is called an [*elliptic*]{} complex point, and if $|\gamma|>1$, then it is called [*hyperbolic*]{}. (The ‘parabolic’ case $|\gamma|=1$ is not generic.) Note that the property of being elliptic or hyperbolic is invariant under biholomorphic transformations.
It is easy to check that if $S$ is a graph $\Gamma_f$, then elliptic and hyperbolic complex points correspond to positive and negative zeros of ${\d f\over \d\bar z}$, respectively.
Assume now that an embedded real surface $S\subset X$ in the general position is compact and oriented. Then there are two orientations of $T_pS$ at each complex point $p\in S$. The first one is the orientation of $S$ and the second one is the canonical complex orientation of $T_pX$. A complex point is called positive if these two orientations coincide; otherwise, it is called negative.
Denote by $e_\pm=e_\pm(S)$ and $h_\pm=h_\pm(S)$ the numbers of positive and negative elliptic and hyperbolic complex points of $S$. Introduce [*Lai’s indices*]{} $I_\pm:=e_\pm-h_\pm$. It turns out that $I_\pm$ are topological invariants of the embedding.
Let $S\subset X$ be a compact oriented real surface in the general position in a complex surface $X$. Then $$\matrix
I_+ +I_-&=\chi(S)+S^2,\\
I_+ -I_-&=c_1(X)\cdot[S],
\endmatrix
\eqno(A4.1.2)$$ where $\chi(S)=2-2g(S)$ is the Euler characteristic of $S$, $S^2=[S]\cdot[S]$ is the self-intersection index of $S$ in $X$, and $c_1(X)\in \sfh^2(X, \R)$ is the first Chern class of $X$.
Proof. We shall only sketch the proof. For the first formula, choose a tangent vector field $\xi\in TS$, twist it by the complex structure $J$ on $X$ and consider the projection $\eta=\pi(J\xi)\in NS$ to the normal bundle of $S$. The zeros of $\eta$ are exactly the zeros of $\xi$ and the complex points of $S$. Careful inspection of the signs (using local equations $(*)$) yields the first formula.
To prove the second formula, pick a non-vanishing 2-form $\omega$ on $S$ and observe that for a local frame $\xi_1,\xi_2\in TS$, the expression $\omega(\xi_1,\xi_2)^{-1}\xi_1\wedge_\C\xi_2$ gives a well-defined section of the linear bundle $\Lambda^2_\C(TX)|_S$. The zeros of this section are the complex points of $S$, because at these points $\xi_1$ and $\xi_2$ are linearly dependent over $\C$. It remains to calculate the signs of these zeros and to use the definition of $c_1(X)$.
Sometimes it is convenient to write Lai’s formulae in the following way: $$I_\pm(S)={1\over 2}(\chi(S)+S^2\pm c_1(X)\cdot[S]).\eqno(A4.1.3)$$ For example, if $C\subset X$ is a non-singular compact complex curve, then after a small real perturbation we obtain an embedded real surface $S$ that obviously has only positive complex points. Thus, $I_-(S)=0$ and therefore $-\chi(C)=C^2-c_1(C)\cdot[C]$, which is precisely the adjunction formula for $C$.
One can also give the following geometric intepretation of Lai’s indices. Consider the Grassmanian $Gr(X,\R,2)$ of oriented real 2-planes in $TX$. It has two natural oriented submanifolds ${\frak I}_\pm$ formed by complex lines with the complex and anti-complex orientation, respectively. Then $I_\pm=\tau(S)\cdot {\frak I}_\pm$, where $\tau:S\to Gr(X,\R,2)$ is the tangential Gauss map. This again can be proved by calculating local signs.
*A4.2. Cancellation of Complex Points. Lai’s indices $I_\pm$ are essentially the only topological invariants of the position of embedded real surfaces with respect to the complex structure. The following theorem of Kharlamov and Eliashberg shows that it is possible to cancel an elliptic and a hyperbolic complex point with the same sign. In particular, if $I_+=I_-=0$, then the surface is isotopic to a totally real one. This can be regarded as a partial case of Gromov’s $h$-principle for $CR$-embeddings.*
Let $S\subset X$ be an embedded real surface in a general position, and assume that $e_+(S)h_+(S)>0$. Then there exists an isotopic embedded surface $S'\subset X$ such that $e_+(S')=e_+(S)-1$ and $h_+(S')=h_+(S)-1$.
Join the pair of points that we want to remove by a smooth path in $S$ and choose a finite covering of this path by coordinate neighborhoods such that in each chart $S$ there is the graph of a smooth function with natural orientation. Now to move complex points along this path and to eventually ‘cancel’ them, it suffices to prove the following local result.
Let $f:\bar\Delta\to\C$ be a smooth function in a neighborhood of the unit disk. Assume that ${\d f\over \d\bar z}$ has two ordinary zeros $a,b\in\Delta$ of opposite sign. Then there exists a function $g:\bar\Delta\to\C$ that is $C^0$ close to $f$ in the disk, coincides with $f$ near the boundary and has a non-vanishing ${\d g\over \d\bar z}$.
Conversely, if ${\d f\over \d\bar z}$ does not vanish, then $g$ can be chosen so that ${\d g\over \d\bar z}$ has two zeros of opposite sign at the given points $a,b\in\Delta$.
Proof. By obvious topological reasons there exists a smooth function $\psi$ supported in a thin neighborhood of the segment $[a,b]$ such that ${\d f\over \d\bar z}+\psi$ does not vanish in $\Delta$. Note also that we may choose $\psi$ so that $\|\psi\|_{C^0}\le 2\|{\d f\over\d\bar z}\|_{C^0}$. Hence, if the area of $\supp(\psi)$ is small enough, then the Cauchy-Green transform $T_{CG}\psi $ is $L^{1,p}$-small by Calderon-Zygmund inequality and the graph of $f+T_{CG}\psi $ is totally real. It remains to multiply $T_{CG}\psi$ by a cut-off function $\phi $ so that $\phi T_{CG}\psi \equiv 0$ near the boundary of $\Delta$. This will not spoil $\d(f+\phi T_{CG}\psi)\over \d\bar z$, because $T_{CG}\psi$ is holomorphic near the boundary and $C^0$-small.
The second part of the lemma is proved in exactly the same way.
The following statement immediately follows from Lai’s formulae and Cancellation theorem.
Corollary A4.2.2. [*Let $\ss^2\subset \cc^2$ be a symplectically imbedded two-sphere in complex projective plane, which is homologous to the projective line. Then there is an imbedded two-sphere $\tilde \ss^2\subset
\cc^2$, which is $C^0$-close to $\ss^2$ and has precisely three points with complex tangents. Moreover, those points are positive and elliptic.* ]{}
*A4.3. Neighborhoods of Real Surfaces. Now we can give a construction of Stein neighborhoods of deformations of real surfaces satisfying certain topological conditions.*
Let $S\subset X$ be an embedded real compact oriented surface in a complex surface. If $S$ satisfies both inequalities $$2I_\pm=\chi(S)+S^2\pm c_1(X)\cdot[S]\leq 0, \eqno(A4.3.1)$$ then there exists an isotopic embedded real surface with a basis of strictly pseudoconvex Stein neighborhoods.
This result was proved in [@F] for $X=\C^2$. The case where $S$ is a smooth deformation of a complex curve was used in [@N1] and the general case in [@N2].
Proof. By the Cancellation Theorem there exists an isotopic real surface, denoted by $S$, that has only [*hyperbolic*]{} complex points.
Let $p\in S$ be a totally real or a hyperbolic complex point. It is not difficult to see from the local equations of $S$ that there exist coordinate neighborhoods $U\supset\!\supset V\ni p$ and a non-negative smooth function $\psi$ in $U$ with the following properties:
“1)” $S\cap U=\{x\in U\mid \psi(x)=0\}$;
“2)” $d\psi\equiv 0$ on $S\cap U$;
“3)” $\psi$ is strictly plurisubharmonic in $U\setminus(S\cap V)$.
Choose a finite covering of $S$ by $V_j\subset\!\subset U_j$. Let, furthermore, $\chi_j$ be smooth non-negative functions supported in $U_j$ such that
“1)” $\chi_j\equiv 1$ on $\bar{V_j}$;
“2)” $\displaystyle\sum_{|\alpha|\leq 2} |D^\alpha\chi_j|\leq A\chi_j$ everywhere in $U_j$ for some constant $A$. Such functions indeed exist: one may first pick functions satisfying the first condition and then raise them to the third power.
Now by the Leibnitz rule we have $$L(\chi_j\psi_j)=i\,\d\bar{\d}(\chi_j\psi_j)=
i\bigl(\psi_j\d\bar{\d}\chi_j+\d\psi_j\wedge\dbar\chi_j
+\d\chi_j\wedge\dbar\psi_j+\chi_j\d\bar{\p}\psi_j\bigr).$$ It readily follows that the Levi form $L(\chi_j\psi_j)$ is non-negative in some neighborhood $W_j\supset S\cap U_j$ and positive in $(W_j\setminus S)\cap V_j$.
Consider the function $\Phi:=\displaystyle\sum_j\chi_j\psi_j$. By construction
“1)” $\Phi\equiv 0$ on $S$ and $\Phi>0$ outside of $S$;
“2)” $\Phi$ is strictly plurisubharmonic in a punctured neighborhood of $S$. Thus, the sets $U_\eps=\{\Phi<\eps\}$, where $\eps>0$ is a regular value of $\Phi$, form a basis of strictly pseudoconvex Stein neighborhoods of $S$.
It is worth observing that an embedded real surface $S\subset X$ with elliptic complex points cannot have a Stein neighborhood basis. Indeed, according to Bishop for an elliptic point $p\in S$, there exists a non-trivial continuous family of holomorphic discs $f_t:(\Delta,\d\Delta)\to (X,S)$ with $f_0\equiv p$. By the continuity principle every holomorphic function in a neighborhood of $S$ extends holomorphically along this family, and therefore a sufficiently small neighborhood of $S$ cannot be Stein.
In fact, it can be proved that, at least when $S^2\geq 0$, a surface $S$ that ‘potentially’ has elliptic complex points (which means that one of the indices $I_\pm(S)$ is positive) does not admit a (not necessarily small) Stein neighborhood (see [@N2] and [@N3]).
Now, consider a complex surface $X$ and a smooth, compact complex curve $C\subset X$.
Corollary A4.3.2. [*If $c_1(X)\cdot[C]\le 0$ then a generic embedded real surface $S$ irotopic to $C$ has a basis of Stein neighborhoods. In particular such $S$ has no non trivial envelope of meromorphy.* ]{}
Proof. [If $c_1(X)\cdot[C]\le 0$, then a generic embedded real surface $S$ isotopic to $C$ satisfies $I_\pm(S)\le 0$, because of the adjunction formula and Lai’s formulae. The statement now follows from the Theorem A4.3.1. ]{}
Appendix V. Seiberg-Witten Invariants and Envelopes.
[A5.1. The Genus Estimate.]{} Throughout this Appendix $X$ will denote a compact, real, oriented manifold of dimension $4$. Fix some $c\in H^2(X, \rr )$. Let $M$ be a closed, oriented surface of genus $g(M)$ imbedded in $X$. For a riemannian metric $g$ on $X$ denote by $s_g$ the scalar curvature of $g$, i.e. $s_g=g_{ij}R^{ij}={\sl tr}R_g$-trace of the Ricci curvature. In the case when $X=M$ is a compact surface the Gauss-Bonnet theorem tells us that
$${1\over 4\pi }\int_Ms_gd\mu_g = 2-2g(M), \eqno(A5.1.1)$$ where $\mu_g$ is the area form associated to $g$.
The following statement is due to P. Kronheimer.
Theorem A5.1.1.
*Assume that $[M]^2=0$ and that for every riemannian metric $g$ on $X$ one can find a two-form $\omega $ representing $c$ such that*
$$\Vert \omega \Vert_{L^2_g}\le {1\over 4\pi }\Vert s_g\Vert_{L^2_g} + K(X),
\eqno(A5.1.2)$$ where constant $K(X)$ depends only on the topology of $X$. Then
$$\vert c[M]\Vert \le \max \bigl\{ 2g(M)-2,0\bigr\} .\eqno(A5.1.3)$$
Proof.
Take some neighborhood $U$ of $M$ which is diffeomorphic to $M\times D$, where $D$ is a disk in $\rr^2$ with coordinates $x,y$. Denote by $g_M$ the riemannian metric on $M$ of constant scalar curvature $s_g$ equal to $4\pi (2-2g(M))$.
(A5.1.1) tells us that if $M$ is not a torus, then the ${\sf area}_{g_M}M=1$. In the case of torus we just take the flat metric with total area one. For real positive $R$ consider the metrics $g_{D,R} = R(dx^2 + dy^2)$ on $D$.
Take now the following metric on $X$:
$$ g\_R = g\_Mg\_[D,R]{} & on $U$ g & away from $U$ $$
Let $\omega_R$ be the harmonic representative of $c$ with respect to the metric $g_R$. Harmonic form minimases the $L^2$-norm in its cohomology class. So by the assumption (A5.1.2) of our theorem
$$\Vert \omega_R\Vert_{L^2_{g_R}}\le (2g(M)-2)\sqrt{\pi }R + O(1). \eqno(A5.1.4)$$ Really, $\Vert \omega_R\Vert_{L^2_{g_R}}\le {1\over 4\pi }\vert s_{g_R}\Vert
+ O(1) $ $= \bigl(\int_{M\times D}[4\pi (2g-2)]^2d\mu_{g_R}\bigr)^{{1\over 2}}=
{1\over 4\pi }\bigl(\pi R^2\cdot 4\pi (2g(M)-2)\bigr)^{{1\over 2}} + O(1) $ $ = \sqrt{\pi } \cdot R\cdot (2g(M)-2) + O(1)$.
Further
$$\Vert \omega_R\Vert_{L^2_{g_R}}\ge \vert c[M]\vert \sqrt{\pi }R.
\eqno(A5.1.5)$$ Really, $\Vert \omega_R\Vert^2_{L^2_{g_R}}\ge \int_D\Vert \omega\mid_{M\times
(x,y)}\Vert^2_{L^2_{g_M}}R^2dxdy$ because $\omega\mid_{M\times (x,y)}$ is only one of the coefficients of $\omega $. $\Vert \omega\mid_{M\times
(x,y)}\Vert^2_{L^2_{g_M}}$ is bounded from below by the $L^2$-norm of the corresponding harmonic representative. Note that the harmonic representative of $\omega\mid_{M\times(x,y)}=c\mid_M$ is just $c[M]*_{g_M}1$ because Hodge $*_{g_M}$ commutes with the Laplacian: $\Delta_{g_M}: \Delta_{g_M}(*_{g_M}1)=*_{g_M}(\Delta_{g_M}1)=0$. So the $L^2$-norm of this harmonic representative and thus of $\omega\mid_{M\times (x,y)}$ is bounded from below by $c[M]$. Inequality (A5.1.5) follows.
Inequalities (A5.1.4) and (A5.1.5) clearly imply inequality (A5.1.2).
This theorem reduces the problem of minimazing the genus of an imbedded surfaces in a fixed homology class to the a priori estimate (A5.1.2). This estimate can be obtained if the Seiberg-Witten invariant $SW(X,c)$ for our cohomology class $c$ is not zero.
[A5.2. The use of Seiberg-Witten Invariant.]{}
The intersection form on the second cohomology space $H^2(X,{\Bbb R})$ of a compact $4$-manifold $X$ is symmetric, and hence it has Sylvester indices $b^\pm(X)$. By definition, the signature of $X$ is the signature $\sigma(X)=b^+ -b^-$ of its intersection form.
If $X$ is a Kähler complex surface, then the maximal positive subspace in $H^2(X,{\Bbb R})$ is spanned by the class of the Kähler form $\omega$ and by the classes of the real and imaginary parts of holomorphic $2$-forms on $X$. In particular, $b^+(X)=1+2\dim_\cc\Omega^2(X)$ is an odd positive integer.
Suppose now that for our $4$-manifold $X$ we have that $b^+(X)>1$. Let us pick a characteristic class $c\in H^2(X,{\Bbb Z})$, that is, an integral cohomology class which is congruent $\mathop{\roman{mod}} 2$ to the second Stiefel–Whitney class $w_2(X)\in H^2(X,{\Bbb Z}_2)$. (For example, the first Chern class of an almost-complex structure $J$ on $X$ is characteristic.)
In this situation one can define the [*Seiberg–Witten invariant*]{} $SW(X,c)\in{\Bbb Z}/2{\Bbb Z}$ of the smooth oriented manifold $X$ (if $\psi$ is an orientation-preserving diffeomorphism of $X$, then $SW(X,\psi^*c) =SW(X,c)$).
Roughly speaking, the invariant $SW(X,c)$ is non-zero if the moduli space of gauge equivalence classes of solutions of the Seiberg–Witten differential equations for a Riemannian metric on $X$ has non-zero homology class in the configuration space. For the precise definition of the invariant, the reader is referred to the original paper \[Wt\] and to the books \[Wg\] and \[Mr\].
Theorem of Witten - I.
*Let $X$ be a compact Kähler surface with $b^+>1$. Then the moduli space of solutions of the (suitably perturbed) Seiberg–Witten equations can be identified with the space of all effective divisors that are Poincaré dual to the cohomology class $\frac12\bigl(c-c_1(X)\bigr)$.*
In particular, $SW(X,c_1(X))=1$.
Really, if $c$ is chosen to be equal to $c_1(X)$ there is only one divisor Poincaré dual to the zero class, namely the zero divisor. Therefore the moduli space of the solutions of $SW$-equations in this case is a singleton.
Remark. [In the opposite direction, the Seiberg–Witten invariants of $X$ can be used to produce complex curves with given cohomology class, namely, if $SW(X,c)\ne 0$, then there is a complex curve dual to $\frac12\bigl(c-c_1(X)\bigr)$. This statement is due to Taubes, see \[Ta\]. ]{}
Suppose that, for a our pair $(X,c)$ the invariant $SW(X,c)$ is non-zero. Then, by definition, for every metric $g$ on $X$, there is a solution of the Seiberg–Witten equations. Using the properties of the equations, one can obtain [*a priori*]{} estimates of the solutions in terms of the metric $g$. We do not write out the equations themselves and only state the key estimate of this type, which is essentially due to Witten (see also \[Mr\] and \[Kr\]).
Theorem of Witten - II. [*If $SW(X,c)\ne 0$ then for every Riemannian metric $g$ on $X$ with there is a closed two-form $\eta$ representing this class in the de Rham cohomology and such that $$\Vert \eta \Vert_{L_g^2}\le {1\over 4\pi }\Vert s_g\Vert_{L^2_g} + K(X),
\eqno(A5.2.1)$$ where $K(X)$ is a constant depending only on the topology of $X$.* ]{} As a consequence we see that
Corollary A5.2.1. [*The number of classes $c\in H^2(X,\zz )$ with non-trivial Seiberg–Witten invariants $SW(X,c)$ is finite.* ]{} Proof.
Indeed, for a fixed metric $g$, the harmonic representatives of these integer classes are uniformly bounded.
Corollary A5.2.2 [*Let $X$ be a compact Kähler surface with $b^{+} >1$. Let $M\subset X$ be an embedded real surface with non-negative self-intersection $M^2\ge0$. Then $$\left| c_1(X)\cdot M \right|+M^2\le \max\{0,2g(M)-2\}. \eqno(A5.2.2)$$* ]{} Proof.
Let us choose an orientation on $M$ so that $c_1(X)\cdot M\le 0$. We blow up the surface $X$ at $d=M^2$ distinct points and denote the blown-up surface by $\widetilde X$, the exceptional curves by $E_j$, and the corresponding blow-down by $\pi:\tilde X\to\nomathbreak X$. Then $c_1(\tilde X)=\pi^*c_1(X)-\sum_j [E_j]$.
Let $\Sigma$ be the interior connected sum of $M$ with $\overline{E_j}$ in $\widetilde X$. Then $[\Sigma]=\pi^*[M]-\sum_j [E_j]$, and therefore $\Sigma^2=0$. Observe that the genus of $\Sigma$ is equal to that of $M$.
By Theorem A5.1.1 we have that $$\max\{0,2g(M)\} \ge\bigl|c_1(\tilde X)\cdot\Sigma\bigr|
=\left| c_1(X)\cdot M-M^2\right|=\left| c_1(X)\cdot M \right|+M^2.$$
Corollary A5.2.2, as it is stated, makes no difference between tori and spheres, i.e. $\max\{0,2g(M)\} =0$ in both cases. However for spheres it can be improved. We shall derive now the following result of Kronheimer and Mrowka \[Kr-M\] and Morgan, Szabó, and Taubes \[M-S-T\].
Corollary A5.2.3. [ *An embedded oriented real surface $M\subset X$ with non-negative self-intersection index in a compact Kähler surface with $b^+(X)>1$ satisfies the inequality $$M^2+\bigr|c_1(X)\cdot M\bigl|\le 2g(M) - 2 \eqno(A5.2.3)$$ provided that $M$ is not a $2$-sphere with trivial real homology class.* ]{} Proof.
The case $g(M)\ge 1$ is covered by (A5.2.2). So let $M$ be a sphere. (A5.2.2) tells that $M^2=0$ and $c_1(X)\cdot M=0$.
Let us blow up the surface $X$ at one point. Let $E\subset \tilde X$ be the exceptional curve. For every $n\ge 1$, we consider the interior connected sum $M_n=E\#nM$. This is an embedded two-sphere such that $M_n^2=-1$ and $c_1(\tilde X)\cdot M_n=1$.
Let us use the existence of an orientation-preserving diffeomorphism $\psi\:\tilde X\to\tilde X$ (with support in a neighbourhood of $M_n$) that maps $M_n$ into itself with opposite orientation. We have $\psi^*c_1(\tilde X)=c_1(\tilde X)+2[M_n]$. If the class $[M]\in H^2(X,\rr )$ were non-zero, this would give us an infinite set of cohomology classes on $\tilde X$ with non-zero Seiberg–Witten invariant, which contradicts Corollary A5.2.1.
Remark. [ By the usual adjunction formula, a complex curve $C$ with $C^2\ge 0$ violates our inequality if and only if $c_1(X)\cdot C>0$. This situation occurs in $\cc\pp^2$ and other complex surfaces with negative canonical class indeed. For this reason, it is impossible to drop the assumption $b^+>1$ in the above theorem. ]{}
In A5.3 we shall repeatedly use the fact that complex curves in complex surfaces with $c_1(X)>0$ violate the “general” form of the adjunction inequality.
[A5.3. The Genus Estimate on Stein Surfaces and Envelopes.]{}
We start from the following
Theorem of Stout. [ *Let $K\subset Y$ be a compact set in a Stein manifold. Then there is a neighbourhood $V$ of $K$ and biholomorphic imbedding $h$ of $V$ onto an open subset in an affine algebraic variety $X$. Moreover, if $Z\subset Y$ is a complex submanifold, then one can find such $h$ that $Z\cap V$ will be mapped onto the intersection of an algebraic subvariety in $X$ with $h(V)$.* ]{}
We shall see now that imbedded real surfaces in Stein surfaces remind real surfaces in compact Kähler surfaces with $b^+>1$.
Theorem A5.3.1.
*Let $M\subset Y$ be an imbedded oriented real surface in a Stein surface with non-negativ self-intersection. Suppose that $M$ is not a $2$-sphere with trivial homology class. Then $$S^2+\bigr|c_1(X)\cdot S\bigl|\le 2g-2. \eqno(A5.3.1)$$*
Proof.
We may always assume that the homology class of $M$ is non-zero because our inequality becomes trivial if $[M]=0$ and $M$ is not a sphere. Hence, there is a non-singular (non-compact) complex curve $L\subset Y$ such that $L\cdot M\ne 0$. Indeed, $\sfh_2(X,\zz)$ is torsion-free for any Stein surface, and hence by the Poincaré duality we can find a class $c\in H^2(Y, \zz )$ with $c\cdot M\ne 0$. Since $Y$ is a Stein surface, it follows that $c=c_1({\Cal L})$ for a holomorphic line bundle $\Cal L$. Then $L$ is the divisor of a generic holomorphic section of $\Cal L$.
Using the algebraic approximation theorem of Stout we obtain an open imbedding of a neighbourhood of $M$ into an affine part $X_a$ of a compact algebraic surface $X$ such that $L$ is mapped to an algebraic curve $C\subset X$ so that $C\cdot Y=L\cdot M\ne 0$. Therefore, the real homology class of $M$ remains non-trivial after the algebraization.
Denote by $H=X\setminus X_a$-“the curve at infinity”. We have non-zero classes $[M]$ and $[H]$ in $H^2(Y,\rr )$ such that $[M]^2\ge 0$, $[H]^2>0$, and $[H]\cdot [M]=0$. Since the intersection form is non-degenerate, it immediately follows that $b^+(Y)>1$. Hence, the desired inequality is a consequence of Corollary A5.2.3.
Remarks.
**1. It follows from Theorem A4.3.1 that the inequality A5.3.1 is sharp, namely, an imbedded real surface satisfying this inequality is isotopic to a surface with a basis of Stein neighbourhoods.**
[**2.**]{} This theorem is also true for the surfaces $M$ with negative self-intersection, see \[N-4\].
[**3.**]{} Concerning a real surface with non-negative self-intersection index the Theorem A5.3.1 still holds for non-Stein strictly pseudoconvex domains as well. This can proved by the same argument (see \[N-2\] and \[N-3\]) by using the Lempert algebraic approximation theorem \[Le\] for such domains.
[**4.**]{} [*Immersed Surfaces*]{} We assume that a surface $S$ is immersed with $\kappa_+$ positive and $\kappa_-$ negative ordinary double points. Then, after embedding a neighbourhood of $S$ in an algebraic surface $Y$, we replace each positive double point with an embedded handle and perform a blow-up at each negative double point. This operation increases the genus by $\kappa_+$ and does not change $S^2$ and $c_1(Y)\cdot S$. Hence, Theorem A5.3.1 holds with $g$ replaced by $g+\kappa_+$.
Now we shall apply the adjunction inequalities (or Genus estimates) for imbedded real surfaces in Stein surfaces to the study of envelopes of holomorphy and meromorphy of domains in complex surfaces.
The main idea is fairly simple. Suppose that we can solve the Levi problem for domains over a complex surface $X$, that is to prove that if $D$ is a pseudoconvex domain over $X$ then $D$ is a Stein domain unless some very special case occurs. I.e. for example $D=X$ (this is the case of $X=\cc\pp^2$) or $X$ is foliated by complex curves (case of $X=\cc\pp^1\times \cc\pp^1$. Then Theorem A5.3.1 yields then topological restrictions on real surfaces in $D$. The below are due to Nemirovski, see \[N-2\].
Let $M\subset\cc\pp^2$ be an embedded oriented real surface in the complex projective plane. By definition, the degree of $M$ is the intersection index of $M$ with complex lines, $d=M\cdot [\cc\pp^1]$. We can choose the orientation of $M$ so that $d\ge 0$ and note that $M$ is homologous to zero if and only if $d=0$.
For an embedded surface of degree $d$, we have $M^2=d^2$ and $c_1(\cc\pp^2)\cdot M=3d$. In particular, if $S$ is a smooth complex curve, then the classical genus formula reads $g(M)=g_\Bbb C(d)=\frac12(d^2-3d+2)$.
Corollary A5.3.2.
*Let $M\subset \cc\pp^2$ be an imbedded oriented real surface of positive degree $d>0$ and genus $g$. If there is a non-constant holomorphic function in a neighbourhood of $M$, then $$g\ge\frac12(d^2+3d+2)=g_\Bbb C(d)+3d.$$*
Conversely, if $M$ satisfies this inequality, then it is isotopic to an embedded surface with a basis of Stein neighbourhoods.
Corollary: Vitushkin’s conjecture.
*If an embedded two-sphere in $\cc\pp^2$ is not homologous to zero, then every holomorphic function in a neighbourhood of this sphere is constant.*
Proof.
Let us consider a neighbourhood $U\supset S$. If there is a non-constant holomorphic function in $U$, then, by Theorem 1, the envelope of holomorphy $\tilde U\supset U$ is a Stein domain because it cannot coincide with $\cc\pp^2$. Hence, the desired inequality immediately follows from Theorem A5.3.1.
Conversely, if the inequality of the theorem is satisfied, then, by Theorem 4, there is an isotopic surface with a basis of Stein neighbourhoods.
Corollary A5.3.3.
*If $M\subset \cc\pp^2$ be an imbedded oriented real surface of positive degree $d>0$ and genus $g$ with*
$$g < \frac12(d^2+3d+2)=g_\Bbb C(d)+3d.$$ then the envelope of meromorphy of any neighborhood of $M$ consides with $\cc\pp^2$.
For a real oriented surface $M\subset\cc\pp^1\times\cc\pp^1$, we consider its “bidegree” $d=(d_1,d_2)$, that is, the pair of intersection indices with horizontal and vertical complex lines. Then $c_1(\cc\pp^1\times\cc\pp^1)\cdot S=2(d_1+d_2)$ and $S^2=2d_1d_2$.
A smooth complex curve $C\subset\cc\pp^1\times\cc\pp^1$ has “positive” bidegree $d\in{\Bbb Z}^2_+\setminus\{0\}$. The genus of $C$ is given by the formula $g_\cc(d)=d_1d_2-d_1-d_2+1$.
Corollary A5.3.4.
*Let $S\subset\cc\pp^1\times\cc\pp^1$ be an embedded oriented real surface of non-zero bidegree $d=(d_1,d_2)\in{\Bbb Z}^2\setminus\{0\}$ and genus $g$. If there is a non-constant holomorphic function in a neighbourhood of $S$, then either $$ gd\_1d\_2+|d\_1+d\_2|+1,d\_1d\_2>0, $**$ $$ or $d_1d_2=0$ and the envelope of $S$ contains a vertical or a horizontal line.*
Conversely, if $S$ satisfies $(**)$, then it is isotopic to an embedded surface with a basis of Stein neighbourhoods.
Remark. [In particular, it is impossible to deform a complex curve in $\cc\pp^1\times\cc\pp^1$ other than a vertical or a horizontal line into a real surface with holomorphic functions in its neighbourhoods. ]{}
Proof.
By the Theorem of Fujita the envelope of holomorphy of a domain $U\subset\cc\pp^1\times\cc\pp^1$ either is a Stein domain or contains one of the fibers. We also note that if $\tilde U$ contains, say, a vertical line and intersects the other vertical lines, then $\tilde U$ coincides with the entire $\cc\pp^1\times\cc\pp^1$ by the usual Hartogs Theorem.
It remains to apply Theorems A5.3.1 and A4.3.1 in the same way as above.
Let us consider the direct product $X=\cc\pp^1\times Y$, where $Y$ is a non-compact Riemann surface. We recall that $H^2(Y,{\Bbb Z})=0$, and hence $H^2(X,{\Bbb Z})\cong H^2(\cc\pp^1,{\Bbb Z})\cong{\Bbb Z}$ by the Künneth formula. For a real oriented surface $S\subset X$, let $d=\deg\pi_{\cc\pp^1}$ be the topological degree of its projection to the $\cc\pp^1$-factor. Then $S^2=0$ and $c_1(X)\cdot S=2d$. We also note that we can orient $S$ so that $d\ge 0$.
By the classical Behnke–Stein theorem, $Y$ is a Stein manifold. Hence, Theorem A4.3.1 gives the following description of envelopes of holomorphy over $X$. The envelope of a domain $U\subset X$ either is a Stein domain or contains a fiber $\cc\pp^1\times\{y\}$. In the latter case, holomorphic functions in $U$ are constant in the $\cc\pp^1$-direction, and therefore the envelope is of the form $\cc\pp^1\times\pi_Y(U)$.
Using the same argument as above, we obtain the following result.
Corollary A5.3.5.
*Let $S\subset\cc\pp^1\times Y$ be an embedded oriented real surface of genus $g$ and positive degree $d>0$. If $g<d+1$, then every holomorphic function in a neighbourhood of $S$ extends to the set $\cc\pp^1\times\pi_Y(S)$.*
If $g\ge d+1$, then $S$ is isotopic to a surface with a Stein neighbourhood basis.
Remark. The results of this section are much more general then the [*Theorem 4.1*]{} in the case of $X=\cc\pp^2, \cc\pp^1\times \cc\pp^1$. For example, if a symplectic sphere $M\subset \cc\pp^2$ is in the homology class of the line, then the procedure of the proof of the Theorem 4.1 gives an isotopy of $M$ to a complex line. However there are knoted two-spheres in $\cc\pp^2$ in the homology class of the line (but they are of course non symplectic). From the other hand the Theorem 4.1 is valid for the spheres not nessessarily in compact Kähler surfaces, where the envelopes of holomorphy or meromorphy are not nessessarily Stein. And for a non-Stein envelope the existence of a compact curve is also a strong conclusion.
To our understanding of the subject a statement containing the both type of results is not avaiable by neither Gromov no Seiberg-Witten techniques.
[A5.4. Two-spheres in $\cc^2$.]{} In this section we apply Theorem A5.3.1 to embedded spheres in $\cc^2$ and its blow-ups. This enables us to show that an embedded 2-sphere in $\cc^2$ is homologous to zero in its envelope of holomorphy.
It follows from Theorem A5.3.1 that an embedded sphere $S$ with self-intersection index $S^2=0$ in a Stein complex surface $X$ is homologous to zero in $\sfh_2(X,\zz)$. For an embedded two-sphere $S\subset\cc^2$, the condition $S^2=0$ holds because $\sfh_2(\cc^2,\zz)=0$. Hence, $S$ is homologous to zero in every Stein domain over $\cc^2$ that contains a neighbourhood of $S$. Since the envelope of holomorphy of every domain in $\cc^2$ is a Stein domain, we obtain the following result \[N-3\].
Corollary A5.4.1. [ *An embedded two-sphere $S\subset\cc^2$ is homologous to zero in the envelope of holomorphy of every neighbourhood $U\supset S$.* ]{}
One can note that the theorem does not work for [*immersed*]{} spheres. Indeed, a simple modification of the results in Appendix IV for immersed surfaces (we must replace $S^2$ in Lai’s formulas by the Euler number of the normal bundle) shows that there is an immersed totally real 2-sphere in $\cc^2$ with a single positive double point. This sphere has a basis of Stein neighbourhoods in each of which this sphere is certainly not homologous to zero.
[A5.5. Attaching complex disks to strictly pseudoconvex domains.]{}
Let $U$ be a relatively compact domain with smooth strictly pseudoconvex boundary in a Stein surface. We say that an embedded analytic disk $D\subset X$, smooth up to the boundary, is attached to $U$ (from the outside) if $D\subset X\setminus U$ and $\partial D\subset\partial U$. We shall always assume that $D$ is transversal to $\partial U$.
For example, the disk $\{|z|\le1,\,w=0\}\subset{\Bbb C}^2$ is attached to the boundary of the $\eps$-neighbourhood of the circle $\{|z|=1+\eps,\,w=0\}$. We note that the boundary of this disk is not homologous to zero in $U$.
A more intricate example can be obtained as follows. Let us consider a totally real torus $T\subset{\Bbb C}^2$. Reversing the proof of the Cancellation Theorem, we create an elliptic and a hyperbolic point in $T$. Let us replace a neighbourhood of the elliptic point by a complex disk and observe that this disk is attached to a strictly pseudoconvex neighbourhood of the remaining part of the torus. The boundary of this disk bounds a real surface with one handle inside the domain. An important observation (due to Forstnerič \[F\]) is that, by Lai’s formulas, this construction does not work for a sphere. In fact, it is impossible to find an analytic disk and a disk with only hyperbolic complex points with common boundary and without common interior points.
We generalize these examples following \[N-3\] and prove that the boundary of a holomorphic disk attached to a strictly pseudoconvex domain in ${\Bbb C}^2$ is never sliced by a smooth disk inside the domain.
Theorem A5.5.1. [ *Suppose that an analytic disk $D$ is attached to a strictly pseudoconvex domain $U$ in a Stein surface $X$ with $\sfh_2(X,\rr)=0$. Then there is no smooth disk inside of $U$ that has the same boundary.* ]{}
Proof.
We assume that there is a smooth disk in $U$ with boundary $\partial D$. Smoothing the union of our two disks, we obtain an embedded sphere $S\subset X$. The naïve idea is to paste a tubular neighbourhood of $D$ to $U$ so that the result is a Stein domain. Then $S$ is not homologous to zero in this domain, which contradicts Theorem 9. The following basic example shows first that such a pasting is impossible and then provides a remedy for this problem.
Let us consider the standard disk $D=\{w=0,\,|z|\le 1\}$ in ${\Bbb C}^2$ with a strictly pseudoconvex “collar” $C=\{ |w|^2+(|z|-a)^2\le (a-1)^2,\,|z|\le 1+\eps\}$ for some $a>1$ and $\eps>0$. If we glue a small neighbourhood of $D$ to the collar, then the interior of this set contains a Hartogs figure with appropriate “walls,” and therefore this interior is not a Stein domain. However, we can construct a Stein neighbourhood “pinched” at the origin.
To see this, we note that, in logarithmic coordinates $\xi=\ln|z|$, $\eta=\ln|w|$, the domain on the $|z|>1$ side of $C$ corresponds to the subgraph of a smooth convex function $\eta=f(\xi)$, $\xi\in(0,\ln(1+\eps)]$. The $\eta$-axis is clearly a vertical asymptote of $f$. Let us choose a point $\xi_0\in(0,\ln(1+\eps)]$ such that the slope of $T_{\xi_0}\Gamma_f$ is a positive integer $n\ge 1$, and replace $f$ by this tangent for all $\xi\le\xi_0$. The subgraph of this new function defines a neighbourhood of $D\setminus\{0\}$ that is logarithmically convex, and hence a Stein domain. Near the origin, this neighbourhood is of the form $|w|\le K|z|^n$, $K>0$.
If we blow up the origin in ${\Bbb C}^2$, that is, make the change of variables $z=z'$, $w=w'z'$, then the proper preimage of our neighbourhood becomes $|w'|\le K|z'|^{n-1}$. Thus, after $n$ blow-ups, we obtain an ordinary neighbourhood of the proper preimage of $D\cup C$. By construction, this domain is locally pseudoconvex and intersects only the last exceptional curve by a disk. Hence, it is a Stein domain by the remark after Takeuchi’s.
The situation of the theorem is reduced to this model by a suitable choice of coordinates in a pinched neighbourhood of $D$ and by a small perturbation of $U$. Thus, we have the following construction of Stein neighbourhoods.
Lemma A5.5.2. [ *Suppose that an analytic disk $D$ is attached from the outside to a strictly pseudoconvex domain $U$ in a Stein surface $X$. Then the proper preimage of $D\cup U$, after several blow-ups at a point $p\in D$, is contained in a Stein domain.* ]{}
To complete the proof of the theorem, we recall that $\sfh_2(X,\rr)=0$, and thus $S^2=0$ and $c_1(X)\cdot S=0$. Consequently, for the proper preimage $\tilde S\subset\tilde X$ of the sphere $S$, after $n\ge 1$ blow-ups at $p\in S$ we have ${\tilde S}^2=-n$ and $|c_1({\tilde X})\cdot{\tilde S}|=n$, which contradicts the inequality of Theorem A5.3.1.
Corollary: Vitushkin’s conjecture. [ *It is impossible to attach an analytic disk from the outside to a strictly pseudoconvex domain in ${\Bbb C}^2$ diffeomorphic to the ball.* ]{}
Proof. The idea is to “reflect” the disk inside the domain by inversion.
Let $U\subset{\Bbb C}^2$ be a strictly pseudoconvex domain in ${\Bbb C}^2$ diffeomorphic to the ball. More precisely, we assume that there is a diffeomorphism $\psi\:\bar U\to\bar B$ of manifolds with boundary. Then $\psi$ extends to a diffeomorphism $\Psi$ of ${\Bbb C}^2$ that maps $U$ to the standard ball $B$.
If a holomorphic disk $D$ is attached to $U$ from the outside, then its image $\Psi(D)$ is a smooth disk attached to $B$. Applying the inversion with respect to $\partial B$, we obtain a smooth disk $D'$ in $B$ with the same boundary. The preimage $\Psi^{-1}(D')$ is a smooth disk in $U$ with boundary $\partial D$, which contradicts the previous theorem.
plus3.5pt minus 1.5pt plus4pt minus 2pt =cmcsc10 =msbm10 =3.5=-\#1\#2\#3\#4
[=2 0pt to [**\[\#1\]**]{}[*[\#3]{}*]{}\#4]{} =cmbx12
References.
. Springer-Verlag (1980).
Geometric Complex Analysis, Proc. Conf., ed. J. Noguchi et all (1966), World Scientific.
Invent. math. [**16**]{}, 335-341 (1973).
J. Math. Pures Appl., [**36**]{}, 235–339, (1957).
J. Amer. Math. Soc., [**4**]{}, 623-646, (1991).
Amer. J. Math. [**105**]{}, 975-1009 (1983).
Springer Verlag, (1984).
Astérisque [**107–108**]{}, 87–161 (1982).
To appear in the volume dedicated to B.V.Shabat, PHASIS, Moscow.
In “Topics in Complex Analysis”, Banach Center Publications, [31]{}, 143–150, (1995).
Int. Symp. “Complex geometry and analysis” in Pisa/Italy, 1988; Lect. Notes in Math., [**1422**]{}(1990), 37–44.
IHES Math. Publ., [**36**]{}(1969), 75–109.
London Math. Soc. Lecture Notes, [**151**]{}, Geometry of low dimensional manifolds, (1991).
Turk. J. Math. v.19, N 2, 145–157 (1995).
Int. Math. Res. Notes [**1**]{}, 17-22 (1995).
Duke Math. J., [**67**]{}, 353–376, (1992).
J. Math. Soc. Japan, [**15**]{}, 443–473, (1963).
3-rd edition, Nauka, Moscow (1977).
Math.Ann. [**146**]{}, 331–368 (1962).
John Wiley & Sons, N.-Y., (1978).
Invent. math. [**82**]{}, 307–347 (1985).
Acta Math. [**125**]{}, 39-56 (1970).
Amer. J. Math., [**75**]{}, 449–476, (1953).
J. of Geom.Anal., [**7**]{}, 149–159, (1998).
Birkhäuser (1997).
In proceedings of Conference on Complex Analysis and Math. Phys., 53–61, Krasnojarsk, (1988).
Invent. Math. [**109**]{}, 47–54 (1992).
Invent. Math. [**136**]{}, 571-602 (1999).
Russ. Math. Sbornik. [**189**]{}, 1295-1333 (1998).
math.DG/9903047 .
Preprint, Bochum (1995), available as e-print math.CV/9804014.
Vol.II, Interscience Publishers, (1969).
proc. Conf. “The moduli spaces of curves” on Texel Island, Netherland. Birkhäuser prog. Math., [**129**]{}(1995), 335–368.
Comm. Math. Phys., [**164**]{}(1994), 525–562.
Preprint, Harvard (1996).
Math. Res. Lett. [**1**]{}, 797–808 (1994).
Trans. AMS, v.172, 1-33 (1972).
Invent. Math., v.121, N 2, 335-354 (1995).
Invent. Math. v.129, N 3, 509-525 (1997).
Ann. Math. [**139**]{}, 35-85 (1994).
J.Diff.Geom. [**34**]{}, 143-164 (1991).
Invent. math., [**89**]{}, 13–36 (1987).
Mathematical Notes 44, princeton University press, princeton (1996).
J. Differential Geom. l44, N4, 706-788 (1996).
Springer Verlag, (1966).
Lecture Notes in Math.1629, Springer, Berlin-New York (1996).
proc.Amer. Math. Soc., 28, 289-294 (1971).
Russian Math. Notes, [**60**]{}, 295-298 (1996).
Math. Notes [**63**]{}, 599-606 (1998).
Doklady RAN , v. 362, (1998).
Russ. Math. Surveys, [**54**]{}, N 4, 47-74 (1999).
Springer, Berlin (1970).
Math. Ann. [**133**]{}, 328–370 (1957).
Publish or perish, [**N7**]{}, (1976).
Michigan Math. Journal [**45**]{}, 529-535 (1998).
J. Diff.Geom., [**44**]{}(1996), 595–633.
J. Geom. Anal., [**3**]{}(1993), 63–98.
Annal. Math., [**113**]{}(1981), 1–24.
In “Holomorphic curves in symplectic geometry”. Edited by M. Audin, J. Lafontaine. Birkhäuser, progress in Mathematics v. 117, Ch.V, 165-189 .
Math. Z., [**226**]{}, 359–373, (1997).
Annals of Math. [**102**]{}, 421–462 (1975).
Invent.math. [**38**]{}, 89–100 (1976).
Amer. J. Math. [**87**]{}, 861-866 (1965).
Springer-Verlag.
Soviet Math. Doklady, [**49**]{}, 629–632, (1991).
Banach Algebras and several complex variables. Proc. Conf. New Haven/Conn. (1983), Contemp. Math. (1984) [**32**]{}, 259–266.
J. Math. Soc. Japan, [**16**]{}, 159–181 (1964).
J. Amer. Math. Soc. [**9**]{}, n 3, 845-918 (1996).
Springer Verlag, (1978).
Math. Res. Lett. [**1**]{}, 769–796 (1994).
2 Institute of Applied Problems &Institute of Applied problemsof Mechanics and Mathematics &of Mechanics and MathematicsUkrainian Acad. Sci., &Ukrainian Acad. Sci.,vul. Naukova 3b, 290053 L’viv &vul. Naukova 3b, 290053 L’viv Ukraine &Ukraine U.F.R. de Mahtématiques & Ruhr-Universität Bochum Université de Lille-I & Mathematisches Institut
Villeneuve d’Ascq Cedex & Universitätsstrasse 150 59655 France & NA 4/67 44780 Bochum Germany ivachkovgat.univ-lille1.fr & sewacplx.ruhr-uni-bochum.de
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'ECML PKDD is the main European conference on machine learning and data mining. Since its foundation it implemented the publication model common in computer science: there was one conference deadline; conference submissions were reviewed by a program committee; papers were accepted with a low acceptance rate. Proceedings were published in several Springer Lecture Notes in Artificial (LNAI) volumes, while selected papers were invited to special issues of the Machine Learning and Data Mining and Knowledge Discovery journals. In recent years, this model has however come under stress. Problems include: reviews are of highly variable quality; the purpose of bringing the community together is lost; reviewing workloads are high; the information content of conferences and journals decreases; there is confusion among scientists in interdisciplinary contexts. In this paper, we present a new publication model, which will be adopted for the ECML PKDD 2013 conference, and aims to solve some of the problems of the traditional model. The key feature of this model is the creation of a journal track, which is open to submissions all year long and allows for revision cycles.'
author:
- Hendrik Blockeel
- Kristian Kersting
- Siegfried Nijssen
- Filip Zelezny
bibliography:
- 'refs.bib'
title: A Revised Publication Model for ECML PKDD
---
Introduction
============
We propose a new conference/journal publication model, the main feature of which is the creation of a journal track that allows all-year submissions to the conference. We would like to adopt this model at ECML PKDD 2013, but hope it will be useful beyond this conference. We start this text with arguments why a new model is needed. Next, we summarize the goals we wish to achieve, and the basic ideas underlying our model. We end with a detailed description of how we intend to implement it.
Motivation
==========
Computer Science is atypical as a scientific field, in that it focuses on publishing at conferences, rather than journals. There is a growing discontent, internationally and in various subfields of computer science, with this tradition. Many argue that the conference-oriented publication system has reached its limits, and is breaking down. See the Appendix \[arguments\] and the bibliography for detailed arguments. In particular the ideas of Halpern and Parkes (2011) are very similar to ours. Briefly, in its current form, the system leads to reviews of highly variable quality; proliferation of conferences (such that the main purpose of conferences, bringing the community together, is lost); high reviewing workloads; slower reviewing for journals; decrease of information content of conferences and journals alike; lengthy journal articles; confusion among scientists in interdisciplinary contexts; unfair evaluations of academics.
The model adopted in most other domains of research, where conferences serve as community meetings and journals as the main publication channel, suffers less from the above shortcomings. Full papers are reviewed more thoroughly, with multiple reviewing rounds when necessary; there is less confusion about the expected standards; there is no artificial limit on the number of submissions that can be accepted; the total reviewing workload is lower (fewer different versions of papers are reviewed); journal articles are more concise (they need not “significantly extend” previous full-length versions). The main problem with journals, in computer science, seems to be the processing time: journal reviewing tends to take much longer than conference reviewing.
There have been several initiatives to address these issues. ECMLPKDD has experimented with selecting papers for exclusive publication in journals; this resolves some of the mentioned issues, but not all. SIGGRAPH and ICLP have taken similar initiatives to directly publish conference papers in journals. IJCAI’13 will have a track in which Journal of Artificial Intelligence Research articles will be presented that were not published in any form at a conference. Several conferences (AAAI, IJCAI, ECMLPKDD’12) are introducing NECTAR tracks in which important papers published in related communities can be presented. VLDB has moved to a journal-reviewing-like model. Changes towards different publication models have been considered for ICML and NIPS, but have not been implemented up till the 2012 editions. ACML’12 has introduced a system with two yearly submission deadlines for the conference track, allowing for resubmissions. The ACL conference on computational linguistics has moved to the VLDB system from 2012.
Goals
=====
The model that we propose for ECML PKDD 2013, aims at achieving four main goals.
1. [**Further increase the quality of full paper presentations at the conference**]{}, by improving the quality of reviewing and providing more visibility to top contributions at the conference.
2. [**Further increase the quality of journals**]{}, by evaluating a new model for journal reviewing that combines quality with speed.
3. [**Make the conference more inclusive**]{}, by allowing for a broader set of contributions to be presented at the conference, albeit not as journal publications.
4. [**Reduce the “conference versus journal” dilemma**]{} that authors face.
Even though our goals are ambitious, we would like the implement the new system by the smallest possible changes to the current ECMLPKDD system. These changes are
1. [**safe**]{}: the current procedures for conference and journal reviewing remain in place, we only install an alternative track;
2. [**small**]{}: the alternative track is quite similar in spirit to previous ideas of selecting papers for inclusion into a journal directly;
3. [**advantageous**]{}, both to the ECML/PKDD conference as to the journals taking part in the initiative (MLj, DMKD).
The model is similar to the VLDB publication model, which is increasingly known within the computer science community.
Concepts of the Proposed System
===============================
The starting point of our proposal is that journal reviewing systems are the standard for high quality reviewing. The journal system has the following key distinguishing features:
- articles can go through [**iterations of revisions**]{} before being accepted;
- there is [**no submission deadline**]{}: articles can be submitted, evaluated, and accepted all year long;
- reviewing for journals is typically done by [**senior researchers**]{}, who are rewarded for this by the prestige of being member of an editorial board.
We would like to see these principles applied to good conference publications as well. The easiest way to do this is to simply use the journal reviewing system for the conference. That is: authors who submit to the journal, can indicate that they would like to have their article considered for presentation at the conference. Submissions that are accepted, are then automatically accepted for the conference as well. The authors can thus draw additional attention to their work. Submissions may be considered not mature enough for publication in the journal, but interesting enough for presentation at the conference; these are again automatically accepted for the conference, without additional reviewing.
However, to make this successful, we need to ensure that the revision and reviewing cycles are short. Indeed, the duration of journal reviewing is a major reason, for many researchers, for preferring conference publication. We intend to achieve this by introducing some efficiency-increasing ideas from conference reviewing into the journal reviewing process. These arguments could entail a solution in which ECMLPKDD publishes all contributions in one or more scientific journals. That is the VLDB model. We believe this may be a too big step for ECMLPKDD at this point; it may require increasing the capacity of existing journals, or the creation of new ones. Our initiative is intended to be a first, easily achievable step in this direction.
Technical Details
=================
The proposed ECML PKDD submission system will have two tracks:
- a proceedings track, which will have one deadline as usual, and whose proceedings will be published in lecture notes; a [*program committee*]{} will review these papers;
- a journal track, which will allow for submissions all year long, and whose articles will be published either in the machine learning or data mining journals, or in the lecture notes; a [*guest editorial board*]{} will review these papers.
At this moment, ECML PKDD’s proceedings track will not be modified significantly. The main change is the introduction of a journal track. Only papers submitted to the journal track can enter the machine learning and data mining journals. Papers that are not of journal quality, but of good conference quality, will be considered for inclusion in the proceedings even if submitted to the journal. Accepted papers are immediately published online, either using the journal’s [*Online First*]{} facility, or on the conference website (until the proceedings are there).[^1] We are considering to mark papers that end up in the proceedings after journal quality reviewing. The benefits for submitting to the journal track are the following:
- papers can be submitted all year long; the possibility to resubmit gives authors a higher chance of presenting their work at ECML PKDD, while the repeated reviewing cycles should ensure the quality of the papers;
- multiple iterations of the same paper will be reviewed by the same reviewers, which should reduce the reviewing efforts;
- papers will receive higher quality reviews, as the reviewing process is carried out by senior researchers;
- papers are available online almost immediately after acceptance (whether for journal or conference).
In detail, we propose the following setup.
#### Guest Editorial Board
A Guest Editorial Board (GEB) will be appointed for the duration of one year, consisting of members of the editorial board of the machine learning and data mining journals, as well as additional senior researchers. GEB members will agree to the following:
- timely reviewing of a limited number of articles that will be sent to them over the course of one year;
- monitoring of their performance.
Monitoring reviewer’s performance is expected to be important to ensure a timely reviewing process. The process will however be lenient to reviewers as long as the chairs are notified in advance. If reviewers do not meet their deadlines and do not notify the chairs well in advance, the assumption is that they will not be invited for next year’s GEB. In the following, asterisks indicate points that will count towards positive or negative evaluation of GEB members.
Note that we explicitly do not limit the size of the GEB in advance. Even though we will start with a relatively small GEB, we intend to grow the editorial board as needed, depending on the popularity of the journal track. Our aim is to limit the number of articles that each GEB member has to review to at most 5.
#### Submission Procedure
We envision the following submission procedure:
1. There is a deadline on Sunday night, every two weeks.
2. The list of submissions is sent immediately to the GEB on Monday morning. Each member will receive a personalized mail that contains the following:
- an overview of the number of articles he/she reviewed;
- an average of the number of articles other editorial board members reviewed (once this number is high enough);
- a list of submitted articles (titles & abstracts), each with a direct link on which the GEB member can click to show interest in this article.\*
Furthermore, each GEB member will have the opportunity to post a comment on a submitted paper; these comments will not be shown to the authors, but can be seen as suggestions to the final reviewers; for instance, such a comment could be: “this reminds me of the work by author A, the reviewers should check this”.\*
3. Based on these bids, the paper is manually assigned to reviewers on Friday, taking into account possible conflicts of interests. Papers that did not receive a sufficient number of bids, will be allocated a reviewer.
4. The GEB member should accept or decline the task within the next week.\*
5. The GEB member has 4 weeks for reviewing after being invited to review the paper, the deadline being on Friday.
6. On Wednesday morning before the deadline, the GEB member is reminded by email of the approaching deadline. The GEB member can request an extension of one week by reacting to this mail before Friday.
7. Any paper for which no response is received\* is immediately allocated to another reviewer.
8. The chairs aim to decide within one week whether the article [*can*]{} be published in the journal. If positive, this decision is communicated to the editor-in-chief or a responsible action editor, who has to agree before the authors are notified. Accepted papers will be published online as soon as possible.
Overall, this would result in a turn around time of only 6 to 8 weeks. The VLDB experience suggests that this is realistic.
The last deadline that will allow for publication in the ECML PKDD 2013 journal track will be approximately 7 weeks before the deadline of the conference track. However, the chairs will not promise that reviewing will be finished within 7 weeks – authors are recommended to submit sufficiently early.
Note that the ECML PKDD program chairs will decide on accepted papers both for the conference and the journal track. They will ensure that papers in both tracks have the desired quality; the editors-in-chief of the journals cannot autonomously decide to accept a paper in the conference track.
#### Ratings
Reviewers will be given the opportunity to mark papers as follows:
1. Accept immediately in the journal.
2. Accept for the conference proceedings, but request revisions for the journal track.
3. Request revisions and allow resubmission to the journal track.
4. Reject.
In case a paper is rated (2), the authors will have to indicate whether they intend to resubmit to the journal track. If not, the paper will enter the procedure for a final version in the conference proceedings.
In case a paper is rated (3), the authors are also allowed to resubmit their paper to the conference track instead of the journal track. After the last ECML PKDD 2013 journal track deadline has passed, this will be the only possible way to publish their results at ECML PKDD 2013. Such submissions will be reviewed by the GEB members that reviewed the original submission.
In case a paper is rated (4), the paper will also not be accepted to the conference track.
#### Format of Submissions
Articles submitted to the journal track will have to be of journal quality. Hence, if experimental in nature, they will have to contain a sufficient number of experiments; if theoretical in nature, they will have to provide full proofs of their claims.
It will be possible to submit articles of any length to the journal track. However, only articles that adhere to a page limit of 20 pages in journal (DAMI/MLJ) style will be evaluated using the above described process, and will hence receive short review times. Articles beyond 20 pages will be treated as normal submissions to the corresponding journals; they cannot be forwarded to the proceedings and will typically not be reviewed in a short time.
By not imposing an overall page limit, we ensure that the quality of journal articles is not limited. At the same time, we do not believe it is reasonable to expect from reviewers that they provide high quality reviews for long papers within a short time.
We will encourage all authors to publish their full versions also in arXiv/CoRR, and aim to have permissions for this from Springer.
#### Relationships to the Proceedings Track
The proceedings track and the journal track clearly are not completely independent of each other. In particular, articles that are not revised in time for the journal track, may be resubmitted to the proceedings track. However, to reduce the reviewing load, such resubmissions from the journal track will be reviewed by the GEB members that reviewed the original submission. The number of reviews carried out by an GEB member will be recorded and taken into account when allocating proceedings submissions to reviewers; reviewers that reviewed a significant number of journal articles will not be asked to review proceedings submissions. Nevertheless, they will also be listed as members of the program committee; the program committee will automatically include all GEB members.
#### Further Changes to the Proceedings Track
There will be a separate deadline for the proceedings track. The setup of the proceedings track, such as its acceptance rate and its reviewing procedures (whether or not to have author response, ...) are in principle independent from the setup of the journal track. It would however be in the spirit of our proposal to accept a larger number of papers to the proceedings track and to reduce the selective pressure for this track. Other tracks, such as a demo track, an industrial track, or Nectar track, are under consideration.
Future Directions
=================
We envision that the system can evolve in several directions after ECMLPKDD’13.
#### A specialized journal
ECMLPKDD can adopt the VLDB model and found a journal specific for ECMLPKDD, preferably open access in nature. This journal could be used for all full paper submissions; this journal may be indexed by conference indexing services to track the impact of ECMLPKDD. The present conference track reduces to a session for posters and short presentations on ongoing work.
#### Additional conferences join the system
The ML and DMKD journals continue to receive submissions for ECMLPKDD, but other conferences start using ECMLPKDD’s procedure, effectively establishing an ECMLPKDD reviewing system in collaboration with these journals. At each resubmission the authors can indicate to which conference they wish to resubmit, allowing papers that not were not finished in time for ECMLPKDD to be considered in other conferences.
#### Additional journals join the system
To increase the number of papers that ECMLPKDD can publish in established journals, it may be an option that more journals are allowed to join the system (a special purpose ECMLPKDD journal possibly being one of them). Based on the reviews, the chairs of the reviewing system allocate papers to journals; this is similar to conferences in bioinformatics, which nowadays often have special issues in many journals.
#### The Old System
When our proposal is a failure, the journal track can be abolished and the conference track’s acceptance rate can be decreased again.
#### {#section .unnumbered}
Key to our proposal is that it can evolve in each of these directions, based on the evolution of the field as a whole in the coming years.
Arguments against the Conference System {#arguments}
=======================================
The tables below summarize the arguments made against the conference system by several well-known computer scientists.
: @fortnow\
[**Credentials author**]{}:
Founding editor-in-chief of the ACM Transaction on Computation Theory
Chair of ACM SIGACT
Chair of the IEEE Conference on Computational Complexity
[**Observations**]{}:
Rating researchers based on conferences is too random
The number of conferences is so large each individual one no longer brings together communities
Conferences can no longer accept all high-quality work
For papers on the margin, there is now a bias for “safe” papers (incremental and technical) and certain subareas (researchers from top CS departments dominate PCs and set the agendas)
Collaboration with researchers in other fields is hard due to having different publication procedures
Papers get rejected due to simple misunderstandings
------------------------------------------------------------------------
: @vardi1 [@vardi2]\
[**Credentials author**]{}:
Editor-in-chief of the Communications of the ACM
[**Observations**]{}:
The computer science conference system compromises one of the cornerstones of scientific publication: peer review
The reviewing process performed by program committees is done under extreme time and workload pressure, and does not rise to the level of careful refereeing
The long turnaround times for journal articles is one important reason why computer science does not switch to a journal-oriented model; in other areas, the turnaround time is much lower
The conference-focused publication culture can not be separated from the sluggish journal editorial process. Roles as editors and referees are not taken as seriously as PC memberships
------------------------------------------------------------------------
: @grudin\
[**Credentials author**]{}:
Former editor-in-chief of the ACM Transactions on Computer-Human Interaction
Associate Editor of ACM Computing Surveys
[**Observations**]{}:
A focus on conference publication has led to deadline-driven short-term research at the expense of journal publication, a reviewing burden that can drive off prominent researchers, and high rejection rates that favor cautious incremental results over innovative work
In other fields, journals focus on identifying and improving research quality; large conferences focus on community building: people don’t retain quite the same warm feeling when their work is rejected
With pressure to reject $\approx75\%$ and differing views of what constitutes significant work, the minor flaws or literature omissions that inevitably accompany novel work become grounds for exclusion
It is increasingly difficult to evolve conference papers into journal articles. By Grudin’s estimation, no more than 15% of the work published in highly selective HCI conferences later appears in journals
\[table:args1\]
[**Reference**]{}: @birman\
[**Credentials authors**]{}:
Former editor-in-chief of the ACM Transactions on Computer Systems
Former editor-in-chief of Distributed Computing
Associate editor-in-chief IEEE Security and Privacy
[**Observations**]{}:
Major conferences are overwhelmed by submissions
The conference system leads to more publications per researcher and per project, even though the aggregate scientific content of all these papers is likely the same as one journal article
Authors submit almost any paper to almost any conference, because acceptance will advance their research and career goals; rejection does them virtually no harm
The more innovative papers are the most likely to be either completely misunderstood or underappreciated by an increasingly error-prone process
There is a risk of a “death spiral” as senior people cease to review. Young researchers often feel more comfortable identifying minor flaws and are less comfortable in declaring that work is more or less important
------------------------------------------------------------------------
: @naughton\
[**Credentials author**]{}:
Keynote speaker at ICDE’10
Associate Editor of VLDB Journal
Associate Editor of ACM Transactions on Database Systems
[**Observations**]{}:
The combination of pressure to publish lots of papers, low acceptance rates, and bad reviewing, is sucking the air out of our community
The conference system is often seen as an evaluation system for authors; in this sense, it can be compared to studying for and taking college entrance exams: ‘students’ (authors) are not that interested in the questions, the ‘graders’ (reviewers) are even less interested in the answers, no-one else is interested either, caring only about the scores
Emphasis on paper count can be somewhat ameliorated by increasing acceptance rate: if it is easier to publish papers, publishing lots of them will be perceived as less impressive; shifts the focus from paper count to paper quality
SIGMOD 2010 received 350 submissions, 1 with uniform “accept” recommendations, 4 with average “accept”
Receiving dysfunctional reviews begets writing dysfunctional reviews
Due to the absence of face-to-face PC meetings, there is less accountability pressure; there is fewer coaching and mentoring
If you get a ‘killer’ reviewer, you are dead; so you resubmit until you have three non-killers
: @jagadish\
[**Credentials author**]{}:
Founding editor-in-chief of the Proceedings of the Very Large Database Endowment (PVLDB)
Program Chair of the International Conference on Intelligent Systems for Molecular Biology (ISMB), 2005
Editor for the database section of the Computing Research Repository (CoRR)
[**Observations**]{}:
The number of submissions to many conferences has sky-rocketed, leading to a downward spiral in reviewing quality and author satisfaction
The enormous size of program committees leads to huge variances in reviewing. An individual PC member sees only a very small piece of the set of submissions; with large PCs we have lost the normalization across accept decisions that a PC-based decision allows
Many rejected papers are resubmitted; the high [*re*]{}submission rate may be a leading cause of the high submission rate
Two recent developments, “rollover” (resubmission “with memory” between conferences) and author feed-back, have received mixed appreciation
Jagadish proposes founding a “Journal of Data Management Research” that publishes papers for participating conferences
------------------------------------------------------------------------
: @lecun\
[**Credentials author**]{}:
General chair and organizer of the yearly “Learning Workshop” in Snowbird, Utah
Associate editor of PLoS ONE
Associate editor of the International Journal on Computer Vision
Program chair for NIPS ’95, ’94, ’90, and many other conferences
[**Observations**]{}:
Conference reviews tend to favor papers containing incremental improvements on well-established methods, and tend to reject papers with truly innovative ideas
Reviewers get very little credit for doing a good job with reviews and can do a bad job with few adverse consequences
The current system is breaking down due to the enormous number of papers submitted and the impossibility of getting papers reviewed properly
The current evaluation system favors citations to well-known authors
Lengthy and faulty evaluations is what currently holds back the dissemination of good papers
: @jl\
[**Credentials author**]{}:
Program chair of ICML’12
This text summarizes arguments of a discussion at NIPS’09, where [**observations**]{} were:
Reviewers are overloaded, boosting the noise in decision making
A new system should run with as little built-in delay and friction to the process of research as possible
It is bad to take double blind so seriously as to disallow publishing on arxiv or personal webpages
Any new system should appear to outsiders as if it is the old system, or a journal, because it is already hard enough to justify CS tenure cases to other disciplines
There should not be a big change with a complex bureaucracy, but rather smaller changes which are obviously useful or at least worth experimenting with
------------------------------------------------------------------------
: @halpern\
[**Credentials authors**]{}:
Editor-in-chief of the Journal of the ACM
Chairman of ACM Preprint
Computing Research Repository (CoRR) coordinator
[**Observations**]{}:
Outside CS, there are two reasons for publishing in journals: certification and publicity
It is rare that conference reviewers review proofs as thoroughly as journal reviewers; for theoretical work certification in journals remains important
The approaches taken by SIGGRAPH and ICLP to directly publish conference papers in journals suffer from the same problems that conference pbulications suffer from: paper are subject to page restrictions and paper submission deadlines
The CS publication model complicates interdisciplinary research
Experimentation with publication models is needed
A system is proposed in which papers are submitted to public archives; journals are the main “certification” authorities that take papers from these archives; journals are much faster in reviewing papers. To achieve fast reviewing, page limits, better coordination between conferences and journals and imposing “costs” on certification are possible ideas
[**1) Review quality is low and highly variable.**]{} The inclusion of inexperienced reviewers in program committees, as well as the need to review many papers within a short amount of time, leads to submissions being rejected for the wrong reasons. Papers with good ideas get rejected because of minor flaws. This very competitive context tends to favor incremental contributions.
[**2) The number of papers submitted for reviewing is too large.**]{} An important reason for this is clearly the need for researchers to publish to get funding (“publish or perish”). However, the situation is significantly worsened by the conference system: papers that are rejected in one conference will usually be resubmitted to another conference, where another set of reviewers is expected to evaluate the paper again; contributions are split over multiple (incremental) publications to meet page limits.
[**3) Acceptance for conferences is highly random.**]{} Low acceptance percentages, combined with variable review quality, causes acceptance to be highly random. This leads to frustration among authors, and decreases the quality of the conference, not only because less good work may be presented, but even more because it misses out on good work.
[**4) Journal papers are often unnecessarily long.**]{} Their extension over a (series of) conference paper(s) often consists of additions made only to make the paper “sufficiently different” from the previous version(s). Often, all crucial information was already in the conference version(s) (otherwise it would not have been accepted there).
[**5) Conference papers in CS are undervalued outside and inside the field.**]{} Our system is not only misunderstood by many people outside the field (including, but not limited to, people who evaluate our research), but also increasingly within the field. Some areas in CS, such as bio-informatics, statistical learning, …are moving towards the standard model, due to the influence of biologists, physicists, …This biases selection of papers at conferences, as well as selection of candidates for postdoc and tenure track positions.
[**6) The status of CS journals is too low.**]{} Journals in computer science do not publish the work with the highest impact: by some estimates only 15% of results in computer science ends up in journals. The slow reviewing cycles lead to low impact factors, as few articles that build on earlier journal work will be published within the 2 years that are needed to increase the most well-known impact factors. Consequently, journals do not have the same visibility as conferences.
[**7) Authors are faced by a “conference versus journal” dilemma.**]{} Authors have the choice between: submitting directly to a journal (less visibility), submitting only to a conference, not to a journal (disadvantageous in countries where evaluation is based on journal publications), submitting first to a conference, then (an extended version) to a journal (problematic because journal articles must “significantly extend” conference papers)
[**8) Deadline-driven research leads to worse papers**]{}, which are not as polished as one might hope. One might argue that these will be rejected anyway, but due to the variability in reviewing this is not necessarily so. In any case, submission of unpolished papers causes a waste of time and effort on the authors’, reviewers’, and possibly readers’ side; furthermore, such papers will typically lead to incremental follow-up papers, which increase the reviewing load later on as well.
[**9) Proliferation of conferences.**]{} When there is a higher demand for papers to be published than there are conferences, the number of separate conferences grows. People lack the time and/or money to visit all the conferences. As a result, the role of conferences as places where the whole community gets together is diluted.
[**10) Researchers are increasingly overloaded with reviewing tasks.**]{} This discourages the best researchers in the field to stay actively involved. As a result, reviewing is often delegated to younger researchers, even PhD students. This lowers the quality of the reviews to an unacceptable level.
[^1]: This idea is subject to approval from Springer.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present results concerning the nature of the cosmological big bounce (BB) transition within the loop geometry underlying loop quantum cosmology (LQC). Our canonical quantization method is an alternative to the standard LQC. An evolution parameter we use has clear interpretation both at classical and quantum levels. The physical volume operator has discrete spectrum which is bounded from below. The minimum gap in the spectrum defines a quantum of the volume. The spectra of operators are parametrized by a free parameter to be determined.'
address: |
Department of Theoretical Physics, Institute for Nuclear Studies,\
Hoza 69, 00-681 Warsaw, Poland
author:
- 'W[ł]{}odzimierz Piechocki[^1]'
title: 'Non-standard loop quantum cosmology'
---
Introduction {#sect1}
============
There are two alternative methods of quantization of a Hamiltonian system with constraints: (i) Dirac’s method - ‘first quantize, then impose constraints’, and (ii) non-Dirac’s method - ‘first solve constraints, then quantize’. We have two corresponding methods in quantization of cosmological models of general relativity (GR) which make use of the so-called loop geometry: (i) standard LQC - Dirac’s method [@abl; @boj], and (ii) non-standard LQC - method proposed recently [@ppw1; @ppw2]. The latter method corresponds to the reduced phase space quantization of loop quantum gravity [@gies].
In what follows we present quantization of flat FRW model with massless scalar field by making use of the non-standard LQC. Available results are the following: (1) within standard LQC - classical Big-Bang is replaced by quantum Big-Bounce due to strong quantum effects at the Planck scale [@ash1; @Dzierzak:2008dy; @Bojowald:2008ik]; (2) within non-standard LQC - modification of GR by loop geometry is responsible for the resolution of the singularity, quantization may lead to discrete spectra of physical observables [@Malkiewicz:2009zd; @Malkiewicz:2009xz].
Classical Level {#sect2}
===============
Modified Hamiltonian
--------------------
The gravitational part of the Hamiltonian of the flat FRW universe with massless scalar field (in special gauges) is found to be [@ppw1] $$\label{hamG}
H_g = - \gamma^{-2} \int_{\mathcal V} d^3 x ~N
e^{-1}\varepsilon_{ijk}
E^{aj}E^{bk} F^i_{ab}\, ,$$ where $\gamma$, Barbero-Immirzi parameter; $\mathcal V\subset
\Sigma$, elementary cell; $N$, lapse function; $\varepsilon_{ijk}$, alternating tensor; $E^a_i $, density weighted triad; $ F^k_{ab} =
\partial_a A^k_b - \partial_b A^k_a + \epsilon^k_{ij} A^i_a
A^j_b$, curvature of $SU(2)$ connection $A^i_a$; $e:=\sqrt{|\det
E|}$;
Modification by loop geometry means approximation of $ F^k_{ab}$ as follows [@ppw1] $$\label{finite}
F^k_{ab}(\lambda) \approx
-2\;Tr\;\Big(\frac{h^{(\lambda)}_{B_{ij}}-1}{\lambda^2
V_o^{2/3}}\Big)\;{\tau^k}\; ^o\omega^i_a \;
^o\omega^j_a,\;\;\;\;\;\;\;F^k_{ab}= \lim_{\lambda\,\rightarrow
\,0}\, F^k_{ab}(\lambda),$$ where the holonomy of the connection around the square loop $
B_{ij}$, with sides length $\mu V_0^{1/3}$, reads $$\label{box}
h^{(\mu)}_{B_{ij}} = h^{(\mu)}_i
h^{(\mu)}_j (h^{(\mu)}_i)^{-1}
(h^{(\mu)}_j)^{-1},\;\;\;\;\;h^{(\mu)}_k (c) = \cos (\mu c/2) +
2\,\sin (\mu c/2)\;\tau_k,$$ where $\tau_k = -i \sigma_k/2\;$ ($\sigma_k$ are the Pauli spin matrices).
Making use of Thiemann’s identity leads finally to $$\label{hamR}
H_g = \lim_{\lambda\rightarrow \,0}\; H^{(\lambda)}_g ,$$ where $$\label{hamL}
H^{(\lambda)}_g = - \frac{sgn(p)}{2\pi G \gamma^3 \lambda^3}
\sum_{ijk}\,N\, \varepsilon^{ijk}\, Tr \Big(h^{(\lambda)}_i
h^{(\lambda)}_j (h^{(\lambda)}_i)^{-1} (h^{(\lambda)}_j)^{-1}
h_k^{(\lambda)}\{(h_k^{(\lambda)})^{-1},V\}\Big),$$ and where $V= |p|^{\frac{3}{2}}= a^3 V_0$ is the volume of the elementary cell $\mathcal{V}$. Variables $ c$ and $p$ determine connections $A^k_a$ and triads $E^a_k$: $A^k_a =
\,^o\omega^k_a\,c\,V_0^{-1/3} \,$ and $\,E^a_k =
\,^oe^a_k\,\sqrt{q_o}\,p\,V_0^{-2/3} $, where $\,c = \gamma
\,\dot{a}\,V_0^{1/3}$ and $\,|p| = a^2\,V_0^{2/3}$, $\{c,p\} = 8
\pi G \gamma /3$.
The total Hamiltonian for FRW universe with a massless scalar field $\phi$ reads $$\label{ham}
H = H_g + H_\phi,$$ where $H_g$ is defined by (\[hamR\]) and $H_\phi = p^2_\phi
|p|^{-\frac{3}{2}}/2$, and where $\phi$ and $p_\phi$ are elementary variables satisfying $\{\phi,p_\phi\} = 1$. The relation $ H \approx 0$ defines the physical phase space.
Making use of (\[box\]) we calculate (\[hamL\]) and get the modified total Hamiltonian corresponding to (\[ham\]) $$\label{regH}
H^{(\lambda)}/N = -\frac{3}{8\pi G \gamma^2}\;\frac{\sin^2(\lambda
\beta)}{\lambda^2}\;v + \frac{p_{\phi}^2}{2\, v},\;\;\;\;\beta :=
\frac{c}{|p|^{1/2}},\;\;\;v := |p|^{3/2} .$$
Equation (\[regH\]) presents a modified classical Hamiltonian.
Observables
-----------
A function, $\mathcal{O}: \mathcal{F}_{kin}^{(\lambda)}\rightarrow
R$, is a Dirac observable if $$\label{dirac}
\{\mathcal{O},H^{(\lambda)}\}= 0,\;\;\;\;\;\;\; \{\cdot,\cdot\}:=
4\pi G\gamma\;\bigg[ \frac{\partial \cdot}
{\partial \beta} \frac{\partial \cdot}{\partial v} -
\frac{\partial \cdot}{\partial v} \frac{\partial \cdot}{\partial \beta}\bigg] +
\frac{\partial \cdot}{\partial \phi} \frac{\partial \cdot}{\partial p_\phi} -
\frac{\partial \cdot}{\partial p_\phi} \frac{\partial \cdot}{\partial
\phi}.$$ Thus, $\mathcal{O}$ is solution to the equation $$\label{dir}
\frac{\sin(\lambda\beta)}{\lambda}\,\frac{\partial
\mathcal{O}}{\partial\beta} - v \cos(\lambda\beta)\,\frac{\partial
\mathcal{O}}{\partial v} -
\frac{\kappa\gamma\,\textrm{sgn}(p_{\phi})}{4 \pi
G}\,\frac{\partial \mathcal{O}}{\partial\phi} = 0.$$ Solutions to (\[dir\]) are found to be [@ppw1] $$\label{obser1}
\mathcal{O}_1:= p_{\phi},\;\;\;\;\;\mathcal{O}_2:= \phi -
\frac{\textrm{sgn}(p_{\phi})}{3\kappa}\;\textrm{arth}\big(\cos(\lambda
\beta)\big),\;\;\;\;\; \mathcal{O}_3:= \textrm{sgn}(p_{\phi})\,v\,
\frac{\sin(\lambda \beta)}{\lambda}.$$ Observables satisfy the Lie algebra $$\label{ala1}
\{\mathcal{O}_2,\mathcal{O}_1\}=
1,\;\;\;\;\;\{\mathcal{O}_1,\mathcal{O}_3\}= 0,\;\;\;\;\;
\{\mathcal{O}_2,\mathcal{O}_3\}= \gamma\kappa .$$ Due to the constraint $H^{(\lambda)}=0$, we have $\mathcal{O}_3=
\gamma \kappa \,\mathcal{O}_1.$ Thus, in the physical phase space, $\mathcal{F}_{phys}^{(\lambda)}$, we have only two observables which satisfy the algebra $$\label{alg1}
\{\mathcal{O}_2,\mathcal{O}_1\}= 1.$$
In what follows we consider functions which can be expressed in terms of observables and an evolution parameter $ \phi$ so they are not observables. They do become observables for each fixed value of $ \phi$, since in such case they are only functions of observables:
The energy density of matter field is found to be [@ppw1] $$\label{rho2}
\rho(\lambda,\phi) =
\frac{1}{2}\,\frac{1}{(\kappa\gamma\lambda)^2\,
\cosh^2 3\kappa (\phi- \mathcal{O}_2)}.$$ The volume operator may be expressed as [@ppw1] $$\label{vol}
v(\phi,\lambda) = \kappa\gamma\lambda\,
|\mathcal{O}_1|\,\cosh3\kappa (\phi-
\mathcal{O}_2).$$
Quantum Level
=============
The energy density operator has been considered recently in [@Malkiewicz:2009zd]. The spectrum of the quantum operator corresponding to $\rho$ turns out to coincide with (\[rho2\]). The energy density operator is bounded and has continuous spectrum.
In what follows we present quantization of the volume observable [@Malkiewicz:2009xz]. The classical volume operator, $v$, reads $$\label{vvol}
v = |w|,~~~w :=
\kappa\gamma\lambda\;\mathcal{O}_1\;\cosh3\kappa(\phi-
\mathcal{O}_2).$$ Thus, quantization of $v$ reduces to the quantization problem of $w$: $$\label{c1}
\hat{w}\,f(x) :=
\kappa\gamma\lambda\,\frac{1}{2}\,\big(
\widehat{\mathcal{O}}_1\,\cosh3\kappa (\phi-
\widehat{\mathcal{O}}_2)
+ \cosh3\kappa (\phi-
\widehat{\mathcal{O}}_2)\;\widehat{\mathcal{O}}_1\big) f(x),$$ where $f \in L^2 (R)$. For $\mathcal{O}_1$ and $\mathcal{O}_2$ we use the Schrödinger representation $$\label{rep1}
\mathcal{O}_1 \longrightarrow \widehat{\mathcal{O}}_1 f(x):=
-i\,\hbar\,\partial_x f(x),\;\;\;\;\; \mathcal{O}_2
\longrightarrow \widehat{\mathcal{O}}_2 f(x):= \widehat{x} f(x) :=
x f(x).$$ Thus, an explicit form of $\hat{w}$ is $$\label{repp1}
\hat{w}= i\,\frac{\kappa\gamma\lambda\hbar}{2}\big(
2 \cosh3\kappa(\phi-x)\;\frac{d}{dx}
-3\kappa\sinh3\kappa
(\phi-x)\big).$$
An explicit form of $\hat{w}$ is $$\label{repp1}
\hat{w}= i\,\frac{\kappa\gamma\lambda\hbar}{2}\big(
2 \cosh3\kappa(\phi-x)\;\frac{d}{dx}
-3\kappa\sinh3\kappa
(\phi-x)\big).$$ Solution to the eigenvalue problem $$\label{eq4}
\hat{w}\, f_a (x) = a\,f_a (x),~~~a \in R ,$$ is found to be [@Malkiewicz:2009xz] $$\label{eq5}
f_a (x):= \frac{\sqrt{\frac{3\kappa}{\pi}}\exp\big(i \frac{2
a}{3\kappa^2 \gamma\lambda\hbar}\arctan
e^{3\kappa(\phi-x)}\big)}{\cosh^{\frac{1}{2}}3\kappa(\phi-x)},\;\;\;\;\;\;\;\;
a = b + 8\pi G\gamma\lambda\hbar\, m,$$ where $b \in R$ and $m\in Z$. Completion of the span of $$\label{set1}
\mathcal{F}_b:=\{~f_a\;|\; a = b + 8\pi G\gamma\lambda\hbar\,
m\}\subset L^2(R),$$ in the norm of $L^2(R)$ leads to $\infty$-dim separable Hilbert space $\mathcal{H}_b$. One may show [@ppw2] that the operator $\hat{w}$ is essentially self-adjoint on each orthonormal space $\mathcal{F}_b$.
Due to the the relation (\[vol\]) and the spectral theorem on self-adjoint operators we get the solution of the eigenvalue of the volume operator $$\label{sp1}
v = |w|~~~\longrightarrow~~~\hat{v} f_a := |a| f_a .$$ The spectrum is bounded from below and discrete. There exists the minimum gap $ \bigtriangleup := 8\pi G\gamma\hbar\,\lambda\;$ in the spectrum, which defines a quantum of the volume. In the limit $ \lambda \rightarrow 0$, corresponding to the classical FRW model, there is no quantum of the volume.
The discreteness may translate at the semi-classical level into a foamy structure of spacetime. Our results suggest that the foamy structure of space is a real property of the Universe so its identification via astro-cosmo observations has sound motivation and is important for the fundamental physics.
Conclusions
============
Modification of gravitational part of classical Hamiltonian, realized by making use of the loop geometry (parameterized by $
\lambda$), turns big bang into big bounce (BB). Since there is no specific $0< \lambda \in R$, there is no specific energy density at BB, and no specific quantum of the volume. The spectra of operators are parametrized by a free parameter $\lambda$ that has not been determined theoretically, but is expected to be fixed by the data of observational cosmology.
The author would like to thank the organizers for inspiring atmosphere at the Meeting.
[10]{}
A. Ashtekar, M. Bojowald and J. Lewandowski, *Adv. Theor. Math. Phys.* [**7**]{}, 233 (2003).
M. Bojowald, *Living Rev. Rel.* [**8**]{}, 11 (2005).
P. Dzierzak, P. Malkiewicz and W. Piechocki, “Turning big bang into big bounce: Classical dynamics,” arXiv:0907.3436 \[gr-qc\].
P. Malkiewicz and W. Piechocki, “Turning big bang into big bounce: Quantum dynamics,” arXiv:0908.4029 \[gr-qc\].
K. Giesel and T. Thiemann, “Algebraic Quantum Gravity (AQG) IV. Reduced Phase Space Quantisation of Loop Quantum Gravity,” arXiv:0711.0119 \[gr-qc\].
A. Ashtekar, T. Pawlowski and P. Singh, “Quantum nature of the big bang”, Phys. Rev. Lett. [**96**]{}, 141301 (2006) \[arXiv:gr-qc/0602086\].
P. Dzierzak, J. Jezierski, P. Malkiewicz and W. Piechocki, “Conceptual issues concerning the Big Bounce” arXiv:0810.3172.
M. Bojowald, “Consistent Loop Quantum Cosmology”, Class. Quant. Grav. [**26**]{}, 075020 (2009) \[arXiv:0811.4129 \[gr-qc\]\].
P. Malkiewicz and W. Piechocki, “Energy Scale of the Big Bounce,” Phys. Rev. D [**80**]{}, 063506 (2009) \[arXiv:0903.4352\].
P. Malkiewicz and W. Piechocki, “Foamy structure of spacetime,” arXiv:0907.4647 \[gr-qc\].
[^1]: E-mail: [[email protected]]{}, Phone: (+48 22) 55 32 275, Fax: (+48 22) 62 16 085
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '$\alpha$ decay is a common and important process for natural radioactivity of heavy and superheavy nuclei. The $\alpha$ decay half-lives for even-even nuclei from Z=62 to Z=118 are systematically researched based on the two-potential approach with a quasi-stationary state approximation. To describe the deviations between experimental half-lives and calculated results due to the nuclear shell structure, a hindrance factor related with $\alpha$ particle preformation probability is introduced. Our results can well reproduce the experimental data equally to the density-dependent cluster model and the generalized liquid drop model. We also study the isospin effect of nuclear potential in this work. Considering the isospin effect the calculated results improved about 7.3$\%$.'
author:
- 'Xiao-Dong Sun'
- Ping Guo
- 'Xiao-Hua Li[^1]'
title: 'Systematic study of $\alpha$ decay half-lives for even-even nuclei within a two-potential approach'
---
Introduction
============
Transuranic elements follow the trend that their half-lives decrease as atomic numbers increase until the next nuclear shell appears. Synthesized atoms of the most recently discovered 117 element have lasted some tens of microseconds [@Oga12; @Khu14], gradually approaching the island of stable superheavy element [@Hof00]. One of useful ways to confirm the superheavy elements is to discriminate specific $\alpha$ particles emanated from itself as well as its $\alpha$ decay chain nuclei. Landing the island of stability and other interesting discoveries, such as the triplet shape coexistence [@And00] and extremely long $\alpha$ decay half-life nuclide $^{209}Bi$ [@De03] and so on, make experimental and theoretical researches on $\alpha$ decay becoming one of hot topics again.
In 1928, Gamow and Condon [@Gam28] and Guerney [@Gur28] had independently put forward the quantum tunnel theory, which successfully estimates the probability of an $\alpha$ particle tunnelling through the Coulomb barrier. The process of barrier tunnelling (penetration) is one important assumption for $\alpha$ decay. The other is that $\alpha$ particle cluster is prone to forming on the surface of the parent nucleus. The existing problem is that before the $\alpha$ particle in bound state collided with the barrier, we know little of the $\alpha$ particle how to form and motion inside the parent nucleus. The difficulties come from the complicated structure of the quantum many body system, e.g. the collective deformation, the fundamental excitations and the nuclear shell closure, and the uncertain of nuclear potential between the $\alpha$ particle and remaining nucleus.
The traditional methods [@Poe79; @Gur87; @Buc90; @Gon93; @Roy00; @Xu06; @Zha06; @Moh06; @Den15; @San15; @Poe12; @Zde13; @Tav05; @Guo15; @Del10]such as the WKB approximation, and empirical formulas [@Gei11; @Ren12; @Del09; @Wan15] are constantly evolving. Among these methods the microscopic double-folding model adopting density dependent M3Y force and the liquid drop model adopting the proximity potential have been researched frequently. Other methods are also developed for $\alpha$ decay, e.g. the coupled-channel method is used to interpret the fine structure of $\alpha$ decay [@Ni10]. The first empirical formula for $\alpha$ decay, Geiger-Nuttall(GN) law in 1911 [@Gei11], relates the $\alpha$ decay half-lives with the decay energy $Q$, and its microscopic interpretation will improves the accuracy of GN law [@Qi14]. These calculations are very successful for $\alpha$ decay. In general, the absolute $\alpha$ decay constant is determined by the preformation probability, the assault frequency and the penetration probability. It is arduous to obtain the actual wave function of parent nucleus and decay state, thus the preformation probability is ambiguous. The shell effect controls the trend that the preformation probability abruptly decreases in the vicinity of the nucleon magic number. Fortunately on the one hand the effective preformation factor can be extracted from the ratios of the experimental $\alpha$ decay half-lives to the calculated penetration probability [@Zha09; @Qi09; @Gan09; @Qia13; @Guo15; @Sei15]. On the other hand a microscopic shell model plus cluster component can provides the preformation probability successfully [@Var92; @Lov98; @Del13; @Bet12; @War15].
In this article we focus on predicting $\alpha$ decay half-life more accurately and studying isospin effect of the nuclear potential. We adopt the two potential approach with a quasi-stationary state approximation [@Gur87], and draw on the analytic expression for $\alpha$ particle preformation probability in Ref. [@Guo15] to estimate variation of the preformation probabilities with the number of valence nucleon. The model parameters are obtained by fitting 164 $\alpha$ decay half-lives of even-even nuclei taken from the newest nuclear property table NUBASE2012 [@Aud12NUBASE]. Based on the above model we systematically calculate the half-lives of even-even nuclei, thereafter we study the relations of depth and diffuseness of the nuclear potential to isospin by fitting the experimental data.
This article is organized as follows. In Sec.II the theoretical framework of the calculation of $\alpha$ decay half-lives and analytic expression for $\alpha$ particle preformation probability are briefly described. In Sec.III we present numerical results, discussion for the hindrance factor and the isospin effect of nuclear potential. A brief summary is given in Sec.IV.
Theoretical framework
=====================
The half-life $T_{1/2}$ for $\alpha$ decay could be determined by $\alpha$ decay width $\Gamma$ or decay constant $\lambda$. It can be written as $$\begin{aligned}
\label{1}
T_{1/2}=\frac{\hbar ln2}{\Gamma}=\frac{ln2}{\lambda}.\end{aligned}$$ The decay constant $\lambda$ depending on the $\alpha$ particle preformation probability $P_{\alpha}$, the penetration probability $P$ and the normalized factor $F$, which represents the collision probability or assault frequency, can be expressed as $$\begin{aligned}
\label{2}
\lambda=\frac{P_{\alpha}FP}{h},\end{aligned}$$ where $h=\frac{T^{exp}_{1/2}}{T^{cal}_{1/2}}$ is defined as hindrance factor. The superscript $exp$ and $cal$ represent experimental data and calculated values, respectively. The normalized factor $F$, which is given by the integration over the internal region [@Gur87], can be written as $$\begin{aligned}
\label{3}
F\int_{r_1}^{r_2}\frac{\mathit{d}r}{2k(r)}=1,\end{aligned}$$ where $r$ is the mass center distance between the preformed $\alpha$ particle and the daughter nucleus. The $r_1$, $r_2$ and following $r_3$ are the classical turning points. $k(r)=\sqrt{\frac{2\mu}{\hbar^2}\mid Q_\alpha-V(r)\mid}$ is the wave number. $\mu$ is the reduced mass of the $\alpha$ particle and daughter nucleus in the center of mass coordinate. $V(r)$ and $Q$ are the height of $\alpha$-core potential and $\alpha$ decay energy, respectively. The penetration probability $P$, which is calculated by WKB approximation, can be expressed as $$\begin{aligned}
\label{4}
P=exp[-\frac{2}{\hbar}\int_{r_2}^{r_3}k(r)\mathit{d}r].\end{aligned}$$ The classical turning points satisfy the condition $V(r_1)=V(r_2)=V(r_3)=Q$. In the inner region $(r_1<r<r_2)$ the strong interaction commands the state of the preformed $\alpha$ particle, while in the outer region $(r_2<r<r_3)$ the electromagnetic interaction plays a major role.
The potential between the preformed $\alpha$ particle and the daughter nucleus, including nuclear, Coulomb and centrifugal potential barrier, can be written as $$\begin{aligned}
\label{5}
V(r)=V_N(r)+V_C(r)+V_l(r).\end{aligned}$$ Where $V_N(r)$ represents nuclear potential, which is critical and uncertain for $\alpha$ decay. In this work, we choose a type of $cosh$ parameterized form for nuclear potential [@Buc92]. It can be expressed as $$\begin{aligned}
\label{6}
V_N(r)=-V_0\frac{1+cosh(R/a)}{cosh(r/a)+cosh(R/a)},\end{aligned}$$ where $V_0$ and $a$ are parameters of the depth and diffuseness for the nuclear potential, respectively. $V_C(r)$ is the Coulomb potential and is taken as the potential of a uniformly charged sphere with sharp radius $R$, which can be expressed as $$\begin{aligned}
\label{7}
V_C(r)= \left \{
\begin{aligned}
\frac{Z_dZ_{\alpha}\mathit{e}^2}{2R}[3-(\frac{r}{R})^2]~~~r<R\\
\frac{Z_dZ_{\alpha}\mathit{e}^2}{2r}~~~~~~~~~~~~~~~~r>R,
\end{aligned}
\right.\end{aligned}$$ where $Z_d$ and $Z_{\alpha}$ are proton number of the daughter nucleus and the $\alpha$ particle, respectively. The sharp radius of interaction $R$ is given by $$\begin{aligned}
\label{8}
R=1.28A^{1/3}-0.76+0.8A^{-1/3}.\end{aligned}$$ This empirical formula is commonly used to calculate $\alpha$ decay half-lives [@Roy00], which derived from the nuclear droplet model and the proximity energy. $V_l(r)=\frac{l(l+1) \hbar^2}{2\mu r^2}$ is centrifugal potential, where $l$ is the orbital angular momentum taken away by $\alpha$ particle. In general, only the favored transitions ($l=0$) take place for $\alpha$ decay of even-even nuclei [@Guo15], and then $V_l(r)=0$.
The hindrance factor $h$ reflects the deviations between the calculated half-lives $T_{1/2}^{cal}$ with constant preformation probability $P_\alpha$ and experimental half-lives $T_{1/2}^{exp}$. It will systematically varies due to the nuclear shell effect. The trend of hindrance factor $h$ can be estimated by the simple formula with five parameters proposed by Zhang *et al* to research the preformation probaiblity of $\alpha$ particle varies in the different nuclear shells [@Zha09; @Guo15]. The hindrance factor can be given by $$\begin{gathered}
\label{9}
log_{10}h=a+b(Z-Z_1)(Z_2-Z)+c(N-N_1)(N_2-N)\nonumber\\
+dA+e(Z-Z_1)(N-N_1),\end{gathered}$$ where $Z$, $N$ and $A$ are the proton, neutron and mass numbers of parent nucleus. $Z_1$ and $Z_2$ ($N_1$ and $N_2$) are the proton (neutron) magic numbers around $Z$ ($N$). $a$, $b$, $c$, $d$ and $e$ are the adjustable parameters. In eq.(9), the first and fourth terms describe the magnitude and the trend of the preformation probability with the increasing mass number, the second and third terms show a parabolic dependence of $log_{10}h$ as a function of the valence proton (neutron) number, the last term relates to the integrated valence neutron-proton interaction strength [@Guo15; @Gan09].
Results and discussions
=======================
Systematic calculation of half-lives
------------------------------------
We calculate the half-lives for even-even nuclei $\alpha$ transition between ground states of parent nuclei and daughter nuclei within the two-potential approach. The experimental data of $\alpha$ decay energy $Q_\alpha$ and half-lives $T^{exp}_{1/2}$ are taken from AME2012 [@Wan12; @Aud12AME] and NUBASE2012 [@Aud12NUBASE], respectively. The adjustable parameters, depth $V_0$ and diffuseness $a$ of nuclear potential and average value of preformation probability $P_\alpha$, are fitted by minimizing the total square deviation $\Delta$, which is defined as $$\begin{aligned}
\label{10}
\Delta=\sum_{i=1}^{N}(log_{10}T_{1/2}^{cal}-log_{10}T_{1/2}^{exp})^2.\end{aligned}$$ All the experimental data of 164 even-even nuclei from Z=62 to Z=118, listed in Table \[Tab2\], are chosen as the database for parameter fitting. Using the method of genetic algorithms [@Car01], a set of parameters is obtained, i.e. $a=0.5654 fm, V_0=189.53 MeV, P_{\alpha}=0.7$. And the RMS deviation is $\sqrt{\frac{\Delta}{164}}=0.350$.
Systematic variations of the half-lives as a function of the neutron numbers $N$ of the parent nuclei are drawn on Fig. \[Fig1\]. The black squares and red circles represent experimental half-lives and calculated ones, respectively. As we can see that the theoretical results can well reproduce the experimental data of $\alpha$ decay half-lives, although the magnitude of half-lives vary in a very wide range from $10^{-7}$ s to $10^{22}$ s. This shows that our parameters are effective, and besides, the assumption that $\alpha$ particle preformation probability keeps constant is satisfactory. It noticed that $\alpha$ decay half-life $T_{1/2}$ is extremely sensitive to decay energy $Q_{\alpha}$, for example the decay energy increasing 2.6 times results in 18 order of magnitude shorter half-life for N=84 isotones. A decrease in symmetry energy may be responsible for the stability weakened [@Wan14].
![(Color online) Logarithmic of half-lives as a function of neutron numbers of parent nuclei. The black squares and red circles represent the experimental data and calculated results, respectively.[]{data-label="Fig1"}](half-lives1.eps){width="50.00000%"}
Furthermore it is obvious that the size of deviation between experimental half-lives and calculated results roughly increases with the increasing valence nucleon especially for nuclide with N=82-126. In order to more clearly show the results, the deviations for 164 nuclei with neutron numbers N larger than 82 are plotted as a colour-map on Fig. \[Fig2\]. The complete area is divided into four regions by spherical magic number Z=82, N=126 and deformed magic number N=152 [@Zha09]. The deviations in the area close to the magic number are greater than 1, indicating the calculated half-life is small, suggesting predicted $\alpha$ particle preformation probability in this region is too large. This rule is significant in particular for spherical magic numbers Z=82 and N=126, where clustering induced by the pairing mode is inhibited [@Qi09]. In general, the predicted $\alpha$ particle preformation probabilities of nuclei in Reg. I and III are small, while the predictions in Reg. II and IV are large. In other words, the preformation probabilities in Reg. I and III are greater than those in Reg. II and IV. The reason why the preformation probabilities in Reg. II are small is that the nuclei are close to magic number Z=82 and N=126. And the reason for small preformation probabilities in Reg. IV may be the nucleus are approaching the next proton and neutron shell closure, such as the doubly magic spherical nuclei at (Z=114, N=184), (Z=120, N=172), or at (Z=126, N=184) depending on different parameters [@Rut97], and the shell effect appears again. In conclusion, the preformation probabilities vary with the distance from the magic number.
![(Color online)Logarithmic of deviation on a colour-map as a function of neutron numbers N and proton numbers Z of parent nuclei[]{data-label="Fig2"}](hfcolmap1.eps){width="50.00000%"}
Hindrance factor
----------------
Based on above methods, the calculated results are greater than the experimental data for some nuclei especially around the magic nuclei, and smaller for other nuclei. To describe the deviations in size the hindrance factor is introduced, which varies due to the nuclear shell structure.
\[Tab1\] Reg. number $Z_1$ $Z_2$ $N_1$ $N_2$ $\overline{\sigma^2}$
--------------- -------- ------- ------- ------- ------- -----------------------
I 61 50 82 82 126 0.0358
II 26 82 126 82 126 0.0159
II 26 82 120 82 126 0.0159
II 26 82 114 82 126 0.0160
III 59 82 126 126 152 0.0495
III 59 82 120 126 152 0.0496
III 59 82 114 126 152 0.0498
IV 13 82 126 152 184 0.0131
IV 13 82 120 152 172 0.0129
IV 13 82 114 152 184 0.0129
: The fitness of magic number for superheavy nuclei to the $\alpha$ particle preformation probabilities
We fit to extracted hindrance factors using the analytic expression of eq.(9), which has taken into account the nuclear shell effect and proton-neutron interaction. Z=50, 82, N=82, 126, 184 are well known magic number for neutron and proton. However, we do not know exactly the magic number for superheavy nuclei, and the predicted proton and neutron magic numbers for superheavy nuclei depend on the models and force parameters. According to the investigation within various parametrizations of relativistic and nonrelativistic nuclear mean-field models and prediction in ref.[@Rut97], we fit the preformation probabilitite to the extracted ones with different protons magic number and the results are listed in table \[Tab1\]. The fitness is defined as $\overline{\sigma^2}=\frac{\sum_{i=1}^{n}(y_{fit}-y_{data})^2}{n}$. The fitness of (Z=114, N=184) is the worst in Reg. II and III, indicating that Z=114 is unlikely to be the next protons magic number in this region of mass, which is consistent with the result in ref.[@Don11].
\[Tab2\] Reg. $a$ $b$ $c$ $d$ $e$
--------------- ---------- --------- --------- --------- ---------
I 1.7828 -0.0017 -0.0015 -0.0087 0.0011
II 9.9252 -0.0054 -0.0029 -0.0417 0.0033
III 15.803 0.0012 -0.0004 -0.0744 0.0052
IV -19.5004 0.0042 -0.0010 0.0686 -0.0019
: The parameters of hindrance factor for even-even nuclei from four different regions. Reg. I is $50 < Z \leqslant 82$ and $82 < N \leqslant 126$, Reg. II is $82 < Z \leqslant 126$ and $82 < N \leqslant 126$, Reg. III is $82\leqslant Z \leqslant 126$ and $126 < N \leqslant 152$, Reg. IV is $82 < Z \leqslant 126$ and $152 < N \leqslant 184$.
In this work, Z=126 is our choice and the obtained parameters are listed in table \[Tab2\]. When the hindrance factor increases, and instead the preformation probability decreases. The extracted hindrance factors and fitted ones as a function of neutron numbers $N$ are drawn on Fig. \[Fig3\]. As we can see the trend for fitted hindrance factors is similar to the extracted ones. Based on the improvement of hindrance factor, our results can reproduce the experimental half-lives within a factor of 2 for most nuclei, and the RMS deviation drop to $\sqrt{\frac{\Delta}{164}}=0.205$.
![(Color online) Logarithmic of hindrance factor as a function of neutron numbers of parent nuclei. The black squares and red circles represent the extracted data and fitted results, respectively.[]{data-label="Fig3"}](hindrance_factor.eps){width="50.00000%"}
As can be seen from table \[Tab2\], the absolute value of parameters $b$ and $c$ in Reg. II are both maximum, showing the preformation probabilities of nuclei in Reg. II are strong parabolic due to these nuclei sandwiched between magic numbers Z=82 and N=126. And it is interested that parameter $d$ is negatively correlated with parameter $a$ and $e$. When the term of nuclei mass $A$ makes greater contributions to the hindrance factors in different regions the contributions from constant term and integrated valence neutron-proton term are smaller, which shows that these three terms are related. A physical explanation of this relationship closely associated with the nucleons clustering in heavy nuclei is worth exploring in the future.
In Table \[Tab3\], we list the experimental half-lives and calculated results of even-even nuclei with proton number Z=62-118. The first and second columns denote the parent nucleus and their neutron numbers $N$, respectively. The third column is the experimental half-lives of $\alpha$ decay in unit of second. The next three columns are calculated half-lives within our work, the generalized liquid droplet model and the density-dependent cluster model, respectively. The lastest column is the calculated results of our work taking into account isospin-dependent nuclear potential of eq.(11). In general, our results are better than others, especially for the nuclei around the shell closure, with the help of analytic expression for hindrance factors. For example, the Radon isotopes stride across the neutron shell in N=126, and the preformation probability will abruptly decreases. It is shown that our results overcome the shortcoming of the shorter predicted half-life of $^{212}$Rn and $^{210}$Rn.
![(Color online)Logarithmic of deviation on a colour-map as a function of neutron numbers N and proton number Z of parent nuclei considering hindrance factor[]{data-label="Fig4"}](hfcolmap2.eps){width=".5\textwidth"}
To evaluate the role of hindrance factors, we plot the deviations between the calculated and experimental half-lives again in Fig. \[Fig4\]. Compared with Fig. \[Fig2\], the abrupt increasing of the deviations close to the magic number has been got over.
Isospin effect of phenomenological nuclear potential
----------------------------------------------------
From the term of symmetry energy in Bethe-Weizsäcker mass formula [@Wan10] to asymmetry-dependent components in nucleon-nucleon optical model physics [@Kon03], they both show that isospin effect play a role in nuclear potential. The excited analogue state in nuclei and exotic phenomenon of neutron-proton pairing could be also isospin related [@War06]. Neutron and proton been treated as being different charge states of the same particles, but the fact that the strong interaction is independent on charge, lead to confusedness of origin and uncertainty of the isospin effect [@Li08; @Ste05].
The accurate calculations of $\alpha$ decay half-lives has been given within two potential approach and correction of hindrance factor. Now we introduce two extra parameters to indicate the isospin effect of nuclear potential. The new depth $V_0$ and diffuseness $a$ of nuclear potential could be given as $$\begin{aligned}
\label{12}
V_0&=193.57-75.61\frac{N-Z}{A}~MeV\nonumber\\
a&=0.5598+0.0014A^{1/3}~fm.\end{aligned}$$ The database includes 164 even-even nuclei with Z=62-118. The RMS deviation drop to $\sqrt{\frac{\Delta}{164}}=0.190$, which improved $\frac{0.205-0.190}{0.205}=7.3\%$. The calculated results are listed in Table \[Tab3\].
Summary
=======
In summary, we systematically calculate $\alpha$ decay half-lives for even-even nuclei with proton number from Z=62 to Z=118 within two potential approach based on phenomenological nuclear potential and correction of hindrance factor. A set of new parameters of nuclear potential and analytic expression for hindrance factors is obtained by fitting to the experimental half-lives. Numerical results can well reproduce the experimental half-lives compared with the DDCM and GLDM, eliminating the shortcomings that calculated results deteriorated in the vicinity of the magic number. Finally quantitative results of isospin effect on $\alpha$-core mean nuclear potential have also been given.
This work is supported in part by the National Natural Science Foundation of China (Grant No.11205083), the construct program of the key discipline in hunan province, the Research Foundation of Education Bureau of Hunan Province,China (Grant No.15A159 ),the Natural Science Foundation of Hunan Province,China (Grant No.2015JJ3103),the Innovation Group of Nuclear and Particle Physics in USC, Hunan Provincial Innovation Foundation For Postgraduate (Grant No.CX2015B398).
[54]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.109.162501) [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.112.172501) [****, ()](http://journals.aps.org/rmp/abstract/10.1103/RevModPhys.72.733) [****, ()](http://www.nature.com/nature/journal/v405/n6785/abs/405430a0.html) [****, ()](http://www.nature.com/nature/journal/v422/n6934/abs/nature01541.html) [****, ()](http://link.springer.com/article/10.1007/BF01343196) [****, ()](http://www.nature.com/physics/looking-back/gurney/index.html) [****, ()](http://iopscience.iop.org/article/10.1088/0305-4616/5/10/005/meta) [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.59.262) [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.65.2975) [****, ()](http://www.iaea.org/inis/collection/NCLCollectionStore/_Public/25/065/25065235.pdf) [****, ()](http://iopscience.iop.org/article/10.1088/0954-3899/26/8/305/meta) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.74.014304) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.74.017304) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.73.031301) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.92.014602) [****, ()](http://www.sciencedirect.com/science/article/pii/S0375947414006393) [****, ()](http://iopscience.iop.org/article/10.1088/0954-3899/39/1/015105/meta) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.87.024308) [****, ()](http://iopscience.iop.org/article/10.1088/0954-3899/31/2/005/meta) [****, ()](http://www.sciencedirect.com/science/article/pii/S0375947414006320) in @noop [**]{} (, ) pp. [****, ()](\doibase 10.1080/14786441008637156) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.85.044608) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.80.024310) [****, ()](http://iopscience.iop.org/article/10.1088/0954-3899/42/5/055112/meta) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.81.064318) [****, ()](http://www.sciencedirect.com/science/article/pii/S0370269314003761) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.80.057301) [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.103.072501) [****, ()](http://iopscience.iop.org/article/10.1088/0954-3899/36/9/095105/meta) [****, ()](http://link.springer.com/article/10.1007/s11433-013-5159-5) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.91.014322) [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.69.37) [****, ()](http://www.sciencedirect.com/science/article/pii/S0370157397000495) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.87.041302) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.86.034338) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.92.014314) [****, ()](http://iopscience.iop.org/article/10.1088/1674-1137/36/12/001/meta;jsessionid=D1F49E999345CE44B5D0E76618764772.c2.iopscience.cld.iop.org) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.45.2247) [****, ()](http://stacks.iop.org/1674-1137/36/i=12/a=003) [****, ()](http://iopscience.iop.org/article/10.1088/1674-1137/36/12/002/meta) @noop [ ()]{} [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.89.047301) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.56.238) [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.107.012501) [****, ()](http://www.sciencedirect.com/science/article/pii/S0375947413007835) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.78.044329) [****, ()](http://www.sciencedirect.com/science/article/pii/S0375947405009553) [****, ()](http://journals.aps.org/prc/abstract/10.1103/PhysRevC.81.044322) [****, ()](http://www.sciencedirect.com/science/article/pii/S0375947402013210) [****, ()](http://www.nature.com/nphys/journal/v2/n5/abs/nphys291.html) [****, ()](http://www.sciencedirect.com/science/article/pii/S0370157308001269) [****, ()](http://www.sciencedirect.com/science/article/pii/S0370157305001043)
[^1]: Corresponding author:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'By using a single formalism to handle charm, strange and light valence quarks in full lattice QCD for the first time, we are able to determine ratios of quark masses to 1%. For $m_c/m_s$ we obtain 11.85(16), an order of magnitude more precise than the current PDG average. Combined with 1% determinations of the charm quark mass now possible this gives $\overline{m}_s(2{\rm GeV}) =$ 92.4(1.5) MeV. The MILC result for $m_s/m_l = 27.2(3) $ yields $\overline{m}_l(2{\rm GeV})$ = 3.40(7) MeV for the average of $u$ and $d$ quark masses.'
author:
- 'C. T. H. Davies'
- 'C. McNeile'
- 'K. Y. Wong'
- 'E. Follana'
- 'R. Horgan'
- 'K. Hornbostel'
- 'G. P. Lepage'
- 'J. Shigemitsu'
- 'H. Trottier'
title: Precise charm to strange mass ratio and light quark masses from full lattice QCD
---
[*Introduction*]{}. – The masses of $u$, $d$ and $s$ quarks are some of the least well-known parameters of the Standard Model. Even the most inaccurate lepton mass (that of the $\tau$) is known to better than 0.01% and yet errors on light quark masses of 30% are quoted in the Particle Data Tables [@pdg09]. The reason for the mismatch is the confinement property of the strong force that obscures the connection between the properties of the quark constituents and the hadron physics that is accessible to experiment. To make this connection requires accurate calculations in QCD and accurate experimental results for appropriate hadronic quantities. A method particularly well-suited to this is lattice QCD. Here we will demonstrate its use by determining $m_c/m_s$ to 1% and obtaining as a result 1.5% errors for light quark masses, which brings them almost into line with those of heavy quarks.
Heavy quark masses, $m_Q$, can be determined accurately because $\alpha_s(m_Q)$ is relatively small. 1% errors for charm and bottom quark masses have recently become possible using $\cal{O}$$(\alpha_s^3)$ calculations in QCD perturbation theory for the heavy quark vacuum ‘bubble’ [@pert] and therefore for the energy-derivative (or time) moments of correlation functions for a heavy quark-antiquark pair at zero momentum. Since the scale of $\alpha_s$ is naturally related to the relevant heavy quark mass, the expressions can be evaluated accurately. To extract the quark mass the perturbative result is compared to a nonperturbative determination containing information from experiment. For a $1^{--}$ $Q\overline{Q}$ configuration moments of the experimentally measured cross-section for ($e^+e^- \rightarrow \gamma^* \rightarrow {\mathrm{hadrons}}$) can be used after isolating the heavy quark contribution and using dispersion relations [@kuhn09]. Alternatively the time-moments for heavy quark current-current correlation functions of various $J^{PC}$ can be directly determined in lattice QCD calculations that have been tuned so that a charmonium or bottomonium mass agrees with experiment [@mcjj; @mylat08]. The time moments must be extrapolated to the zero lattice spacing (continuum) limit before the comparison to QCD perturbation theory. These two methods give results that agree, with 1% errors for $m_c(3 {\mathrm{GeV}})$ in the $\overline{MS}$ scheme. The more traditional ‘direct’ lattice QCD method, although somewhat less accurate, also gives results in good agreement [@mchighbeta]. We can conclude from this that $m_c$ is now accurately known.
The strange quark mass, $m_s$, being much smaller, cannot be determined this way and is poorly known at present. Instead of a direct determination of $m_s$, however, we can use the leverage of an accurate result for the ratio $m_c/m_s$ combined with the accurate $m_c$ above [@gasserleutwyler]. But simple ratios of hadron mass differences give unreliable estimates of $m_c/m_s$. Two such estimates: $$\frac{m(B_c)-m(B_u)}{m(B_s)-m(B_u)} = 11;
\frac{m(\Sigma_c)-m(N)}{m(\Sigma)-m(N)} = 6$$ differ by almost a factor of 2. The ratio of $m_s/m_l$ (where $m_l = (m_u+m_d)/2$) is known to about 10% from ratios of squared masses of $K$ and $\pi$ mesons using SU(3) chiral perturbation theory [@pdg09]. Clearly neither ratio is determined well enough this way to provide the accuracy we need, because the relationship between hadron mass and well-defined running quark mass is more complicated than these simple ratios must assume.
Lattice QCD, on the other hand, can give very accurate results for the ratio of two quark masses but only if the same formalism is used for both quarks. This has already been used to give accurate results for $m_s/m_l$, although neither $m_s$ nor $m_l$ is very well determined. Here, for the first time, we give an accurate result for $m_c/m_s$ by using the same formalism for charm, strange and light quarks and this enables us to cascade the accuracy of the heavy quark mass down to the light quarks.
[*The Lattice QCD calculation.*]{} – Lattice QCD gives direct access to quark masses through the lattice QCD Lagrangian. Tuning of the masses is done by calculating an appropriate hadron mass and adjusting the quark mass until the hadron mass agrees with experiment. Experimental measurements of appropriate hadron masses are extremely accurate in most cases, with errors at the level of tenths or hundredths of a percent. To make maximum use of this precision we need to calculate the hadron mass in lattice QCD with small statistical and systematic errors. In particular it requires the full effect of sea quarks in the hadron to be included. This is now possible in lattice QCD [@ratio]. Fixing the four quark masses ($m_l$, $m_s$, $m_c$, $m_b$) from four ‘gold-plated’ hadrons ($\pi$, $K$, $\eta_c$, $\Upsilon$) enables other quantities to be calculated with errors of a few percent and agreement with experiment is obtained [@ratio; @lp07]. This is an important test that QCD, with only one scale parameter and one mass parameter per quark flavor, describes the full range of hadron physics consistently.
The lattice quark mass is a perfectly well-defined running quark mass. However, it is scheme-dependent and so varies with the discretisation of the Dirac equation used in the lattice calculation. For wider applicability it is more useful to convert the lattice quark mass to a standard continuum scheme such as $\overline{MS}$. This renormalization has been a major source of systematic error in previous determinations of light and strange quark masses. The best existing result for $m_s(2 {\mathrm{GeV}})$, with a 7% error, uses the direct method of converting the tuned quark mass in the lattice QCD Lagrangian to the $\overline{MS}$ scheme using $\alpha_s^2$ lattice QCD perturbation theory [@ourstrange]. The error is dominated by the error in the renormalization and it is the error that we will remove here, by instead determining $m_c/m_s$ accurately. The Highly Improved Staggered Quark action [@hisq; @fds] allows us to use the same discretization of QCD for both charm and strange quarks because it is a fully relativistic ‘light quark’ action that can also be used for charm quarks. Then the mass renormalisation factor cancels in the quark mass ratio.
We work with eight different ensembles of gluon field configurations provided by the MILC collaboration. These include the effect of $u$, $d$ and $s$ sea quarks using the improved staggered quark (asqtad) formalism using the fourth root ‘trick’. This procedure, although ‘ugly’, appears to be a valid discretization of QCD [@sharpe; @milcreview; @kronfeld; @golterman]. Tests include studies of the Dirac operator and comparisons to effective field theories. Configurations are available with large spatial volumes ($> 2.4({\rm fm})^3$) at multiple values of the light sea masses (using $m_u = m_d = m_l$) and for a wide range of values of the lattice spacing, $a$. We use configurations at five values of $a$ between 0.15 fm and 0.05 fm with parameters as listed in Table \[tab:params\].
Set $\beta$ $r_1/a$ $au_0m_{0l}^{asq}$ $au_0m_{0s}^{asq}$ $L/a$ $T/a$ $N_{conf}\times N_{t}$
----- --------- ----------- -------------------- -------------------- ------- ------- ------------------------ --
1 6.572 2.152(5) 0.0097 0.0484 16 48 $631 \times 2$
2 6.586 2.138(4) 0.0194 0.0484 16 48 $631 \times 2$
3 6.76 2.647(3) 0.005 0.05 24 64 $678 \times 2$
4 6.76 2.618(3) 0.01 0.05 20 64 $595 \times 2 $
5 7.09 3.699(3) 0.0062 0.031 28 96 $566 \times 4$
6 7.11 3.712(4) 0.0124 0.031 28 96 $265 \times 4$
7 7.46 5.296(7) 0.0036 0.018 48 144 $201 \times 2$
8 7.81 7.115(20) 0.0028 0.014 64 192 $208 \times 2 $
: Ensembles (sets) of MILC configurations used, with gauge coupling $\beta$, size $L^3 \times T$ and sea masses ($\times$ tadpole parameter $u_0$) $m_{0l}^{asq}$ and $m_{0s}^{asq}$. The lattice spacing values in units of $r_1$ after ‘smoothing’ are given in column 3 [@milcreview]. Column 8 gives the number of configurations and time sources per configuration used for calculating correlators. []{data-label="tab:params"}
Set $am_{0c}$ $1+\epsilon$ $am_{\eta_c}$ $am_{0s}$ $am_{\eta_s}$
----- ----------- -------------- --------------- ----------- ---------------
1 0.81 0.665 2.19381(16) 0.061 0.50490(36)
0.825 0.656 2.22013(15) 0.066 0.52524(36)
0.85 0.641 2.26352(15) 0.080 0.57828(34)
2 0.825 0.656 2.21954(13) 0.066 0.52458(35)
3 0.65 0.762 1.84578(8) 0.0537 0.43118(18)
4 0.63 0.774 1.80849(11) 0.0492 0.41436(23)
0.66 0.756 1.86674(19) 0.0546 0.43654(24)
0.72 0.72 1.98114(15) 0.05465 0.43675(24)
0.753 0.70 2.04293(10) 0.06 0.45787(23)
0.063 0.46937(24)
5 0.413 0.893 1.28057(7) 0.0337 0.29413(12)
0.43 0.885 1.31691(7) 0.0358 0.30332(12)
0.44 0.88 1.33816(7) 0.0366 0.30675(12)
0.45 0.875 1.35934(7) 0.0382 0.31362(14)
6 0.427 0.885 1.30731(10) 0.03635 0.30513(20)
7 0.273 0.951 0.89932(12) 0.0228 0.20621(19)
0.28 0.949 0.91551(9) 0.024 0.21196(13)
8 0.195 0.975 0.67119(6) 0.0165 0.15484(14)
0.018 0.16209(17)
: Results for the masses in lattice units of the goldstone pseudoscalars made from valence HISQ charm or strange quarks on the different MILC ensembles enumerated in Table \[tab:params\]. Columns 2 and 3 give the corresponding bare charm quark mass, and Naik coefficient respectively. Column 6 gives the bare strange quark mass ($\epsilon=0$ in that case). []{data-label="tab:charmmass"}
On these configurations we have calculated quark propagators for charm quarks, strange quarks and light quarks (again $m_u=m_d=m_l$) using the HISQ action. The numerical speed of HISQ means that we have been able to use several nearby quark masses for charm and strange to allow accurate interpolation to the correct values. Table \[tab:charmmass\] gives masses for the goldstone pseudoscalar mesons made from either a charm quark-antiquark pair or a strange one (the $\eta_c$ and the $\eta_s$), which are used for tuning. In the charm case, as well as the quark mass, we list the coefficient of the ‘Naik’ term in the HISQ action that corrects for discretisation errors through $(am_{0c})^4$. The quark propagators are generated from random wall sources and the goldstone mesons have good signal/noise properties so the meson masses can be determined to high precision using a standard multi-exponential fit [@bayesfits].
![ Grey points show the raw data for every ratio of $m_c/m_s$ on each ensemble (Table \[tab:charmmass\]); these ratios are fit to eq. \[eq:fit\]. The dashed line and associated grey error band (and red point at $a=0$) show our extrapolation of the resulting tuned $m_c/m_s$ to the continuum limit. Blue points with error bars are from a simple interpolation, separately for each ensemble, to the correct $m_c/m_s$, and are shown for illustration.[]{data-label="fig:mcmsplot"}](mcms.pdf){width="8.5cm"}
The meson masses can be converted to physical units with a determination of the lattice spacing. On an ensemble by ensemble basis this is taken from a parameter in the heavy quark potential called $r_1$. Values for $r_1/a$ determined by the MILC collaboration [@milcreview] are given in Table \[tab:params\]. They have errors of 0.3-0.5%. The physical value for $r_1$ must then be obtained by comparing to experimentally known quantities and we use the value 0.3133(23) fm obtained from a set of four such quantities, tested for consistency in the continuum limit [@newr1; @massnote].
Using the information about meson masses that we have on each ensemble we can interpolate to the correct ratio for $am_{0c}$ and $am_{0s}$ using appropriate continuum values for the masses of the $\eta_c$ and $\eta_s$. We correct the experimental value of $m_{\eta_c}$ of 2.9803 GeV to $m_{\eta_c, \mathrm{phys}}$ = 2.9852(34) GeV. This allows for electromagnetic effects (2.4 MeV) [@newr1] and $\eta_c$ annihilation to gluons (2.5MeV) [@hisq], both of which are missing from our calculation, so increasing the $\eta_c$ mass. We take a 50% error on each of these corrections and also increase the experimental error to 3 MeV to allow for the spread of results from different $\eta_c$ production mechanisms [@pdg09]. Since the total shift is only around 0.2% of the $\eta_c$ mass it has a negligible effect as can be seen from our error budget below.
The $\eta_s$ is not a physical particle in the real world because of mixing with other flavor neutral combinations to make the $\eta$ and $\eta^{\prime}$. However, in lattice QCD, the particle calculated (as here) from only ‘connected’ quark propagtors does not mix and is a well-defined meson. Its mass must be determined by relating its properties to those of mesons such as the $\pi$ and $K$ that do appear in experiment. From an analysis of the lattice spacing and $m_l$-dependence of the $\pi$, $K$, and $\eta_s$ masses we conclude that the value of the $\eta_s$ mass in the continuum and physical $m_l$ limits is 0.6858(40) GeV [@newr1].
The connection between the $\overline{MS}$ mass at a scale $\mu$ and the lattice bare quark mass is given by [@mctree; @ourstrange]: $$\begin{aligned}
&&\overline{m}(\mu) = \frac{am_0}{a} Z_m(\mu a, m_0a), \\ \nonumber
Z_m &=& 1 + \alpha_s(-\frac{2}{\pi}\log(\mu a) + C + b(am_0)^2 + \ldots) + \ldots.
\label{eq:zm}\end{aligned}$$ From these two equations it is clear that $$\frac{\overline{m}_c(\mu)}{\overline{m}_s(\mu)} = \left. \frac{am_{0c}}{am_{0s}} \right|_{{\mathrm {phys}}},$$ where ${\mathrm{phys}}$ denotes extrapolation to the continuum limit and physical sea quark mass limit.
On each ensemble the ratios we have for $am_{0c}/am_{0s}$ then differ from the physical value because of three effects: mistuning from the correct physical meson mass; finite $a$ effects that need to be extrapolated away and effects because the sea light quark masses are not correct. We incorporate these into our fitting function: $$\begin{aligned}
\label{eq:fit}
\left.\frac{m_{0c}}{m_{0s}}\right|_\mathrm{lat} &=
\left.\frac{m_{0c}}{m_{0s}}\right|_\mathrm{phys} \times
\left(
1 + d_\mathrm{sea}\frac{\delta m^\mathrm{sea}_\mathrm{tot}}{m_s}
\right) \\ \nonumber
&\times
\left(
1 + \sum_{i,j,k,l} c_{ijkl} \,\delta_c^i\,\delta_s^j
\left(\frac{am_{\eta_c}}{2}\right)^{2k}
(am_{\eta_s})^{2l}
\right).\end{aligned}$$ $$\delta_c = \frac{m_{\eta_c, MC} - m_{\eta_c, \mathrm{phys}}}{m_{\eta_c,\mathrm{phys}}};\,
\delta_s = \frac{m^2_{\eta_s, MC} - m^2_{\eta_s, \mathrm{phys}}}{m^2_{\eta_s,\mathrm{phys}}}$$ are the measures of mistuning, where $MC$ denotes lattice values converted to physical units. The last bracket fits the finite lattice spacing effects as a power series in even powers of $a$. These can either have a scale set by $m_c$ (for which we use $am_{\eta_c}/2$) or by $\Lambda_{QCD}$ (for which we use $am_{\eta_s}$). $i,j,k,l$ all start from zero and are varied in the ranges: $i, j \le 3$, $k \le 6$, $l \le 2$ with $i+j+k+l \le 6$. Doubling any of the upper limits has negligible effect on the final result. The prior on $c_{ijkl}$ is set to 0(1). $\delta m^{\mathrm{sea}}_{\mathrm{tot}}$ is the total difference between the sea-quark masses used in the simulation and the correct value for $2m_l+m_s$ [@newr1]. This has a tiny effect and we simply use a linear term (adding higher orders has negligible effect). The prior for $d_{\mathrm{sea}}$ is 0.0(1). Figure \[fig:mcmsplot\] shows the results of the fit, giving $m_c/m_s$ in the continuum limit as 11.85(16) ($\chi^2/{\mathrm{dof}}$ = 0.42). The error budget is given in Table \[tab:errormcms\].
$m_c/m_s$ $m_s/m_l$
--------------------------------------- ----------- -----------
overall $r_1$ uncertainty 0.4% 0.1%
$r_1/a$ uncertainties 0.2 -
continuum $M_{\eta_c}$ 0.2 -
continuum $M_{\eta_s}$ 1.1 -
Finite volume - 0.3
$a^2$ extrapolation, $m_q$ interpolns 0.4 0.8
sea-quark mass extrapolation 0.0 0.2
statistical errors 0.3 0.4
Total 1.3% 1.0%
: Error budgets for $m_c/m_s$ and $m_s/m_l$. []{data-label="tab:errormcms"}
$m_s/m_l$ is known to 1% from lattice QCD as a byproduct of standard chiral extrapolations of $m_{\pi}^2$ and $m_K^2$ to the physical point [@milcm]. MILC quote 27.2(3) using asqtad quarks [@milcreview]. Our HISQ analysis in [@fds] gave a result in agreement at 27.8(3), using a Bayesian fit to a function including terms from chiral perturbation theory up to third order in $m_l$ and allowing for discretisation errors up to and including $a^4$ and for mixed terms (i.e $m_l$-dependent discretisation errors). A full error budget is given in Table \[tab:errormcms\]; the data are given in [@newr1].
[*Conclusions*]{}. – Our $m_c/m_s$ can be used with any value for $m_c$ to give $m_s$. The best existing result [@mcjj] (converted from $n_f$=4 to 3) is $\overline{m}^{(3)}_c(2.0{\rm GeV})$ = 1.095(11) GeV or $\overline{m}^{(3)}_c(3.0{\rm GeV})$ = 0.990(10) GeV. Dividing by 11.85(16) gives $\overline{m}^{(3)}_s(2.0{\rm GeV})$ = 92.4(1.5) MeV and $\overline{m}^{(3)}_s(3.0{\rm GeV})$ = 83.5(1.4) MeV.
![ Our results for the 4 lightest quark masses compared to the current PDG evaluations (shaded bands) [@pdg09]. Each mass is quoted in the $\overline{MS}$ scheme at its conventional scale: 2 GeV for $u$, $d$, $s$ ($n_f=3$); $m_c$ for $c$ ($n_f=4$).[]{data-label="fig:qmass"}](qmass.pdf){width="6.5cm"}
Using the MILC values for $m_s/m_l$ and $m_u/m_d$ (0.42(4) [@milcreview]) we can then obtain: $\overline{m}^{(3)}_l(2.0{\rm GeV})$ = 3.40(7) MeV and $\overline{m}^{(3)}_l(3.0{\rm GeV})$ = 3.07(6) MeV; $\overline{m}^{(3)}_u(2.0{\rm GeV})$ = 2.01(14) MeV and $\overline{m}^{(3)}_d(2.0{\rm GeV})$ = 4.79(16) MeV. The values for all four quark masses are plotted in Figure \[fig:qmass\] in comparison to the current evaluations from the Particle Data Tables [@pdg09].
Thus our high accuracy on $m_c/m_s$ allows us to leverage 2% accurate values for $m_s$ and $m_l$ that are completely nonperturbative in lattice QCD, for the first time. Our $m_s$ mass is higher, by around $1 \sigma$, than our previous value of $\overline{m}_s(2 {\rm GeV})$ = 87(6) MeV which used 2-loop lattice QCD perturbation theory [@ourstrange]. Then the error was dominated by unknown $\alpha_s^3$ terms. Our new result, which does not have this limitation, has an error almost five times smaller. Our new error is almost an order of magnitude smaller than other lattice QCD results from full QCD [@rbc; @jlqcd]. These use direct methods of converting the lattice mass to the $\overline{MS}$ mass, and have 10% errors.
[**[Acknowledgements]{}**]{} We are grateful to MILC for configurations and to H. Leutwyler for useful discussions. Computing was done at the Ohio Supercomputer Centre, USQCD’s Fermilab cluster and at the Argonne Leadership Computing Facility supported by DOE-AC02-06CH11357. We used chroma for some analysis. We acknowledge support by the Leverhulme Trust, the Royal Society, STFC, SUPA, MICINN, NSF and DoE.
[99]{} Particle Data Group, http://pdg.lbl.gov/. Y. Kiyo [*et al*]{}, Nucl. Phys. B[**823**]{}:269 (2009) \[arXiv:0907.2120\]; A.Hoang [*et al*]{}, Nucl. Phys. B[**813**]{}:349 (2009) \[arXiv:0807.4173\]. J. H. Kühn, M. Steinhauser and C. Sturm, Nucl. Phys. B[**778**]{}, 192 (2007) \[arXiv:hep-ph/0702103\]; K. G. Chetyrkin [*et al*]{} \[arXiv:0907.2110\]. I. Allison [*et al*]{}, HPQCD + K. G. Chetyrkin, J. H. Kühn, M. Steinhauser and C. Sturm, Phys. Rev. D[**78**]{}:054513 (2008) \[arXiv:0805.2999\]. C. T. H. Davies [*et al*]{}, HPQCD, PoS(LATTICE 2008)118 \[arXiv:0810.3548\]. I. Allison [*et al*]{}, HPQCD, PoS(LATTICE 2008)225 \[arXiv:0810.0285\]. J, Gasser and H. Leutwyler, Phys. Rept. [**87**]{} (1982) 77. C. T. H. Davies [*et al*]{}, HPQCD/Fermilab/MILC, Phys. Rev. Lett. [**92**]{}:022001 (2004) \[arXiv:hep-lat/0304004\]. E. Gregory [*et al*]{}, HPQCD, Phys. Rev. Lett. [**104**]{}:022001 (2010) \[arXiv:0909.4462\]. Q. Mason [*et al*]{}, HPQCD, Phys. Rev. D[**73**]{}:114501 (2006) \[arXiv:hep-ph/0511160\]. E. Follana [*et al*]{}, HPQCD, Phys. Rev. D[**75**]{}:054502 (2007) \[arXiv:hep-lat/0610092\]. E. Follana [*et al*]{}, HPQCD, Phys. Rev. Lett.[**100**]{}:062002 (2008) \[arXiv:0706.1726\]. S. Sharpe, PoSLAT2006:022 \[arXiv:hep-lat/0610094\]. A. Bazavov [*et al*]{}, arXiv:0903.3598. A. S. Kronfeld, PoSLAT2007:016 \[arXiv:0711.0699\]. M. Golterman, PoSCONF8:014 \[arXiv:0812.3110\]. G. P. Lepage [*et al*]{}, Nucl. Phys. B (Proc. Suppl. [**106**]{}), 12 (2002) \[arXiv:hep-lat/0110175\]. C. T. H. Davies [*et al*]{}, HPQCD, arXiv:0910.1229, Phys. Rev.D (in press). Note that [@mcjj] determined $m_c$ using an earlier and less accurate value of $r_1$ (0.321(5) fm). We have checked that using the new value does not change $m_c$ significantly. The mass parameter $am_{0c}$ in equation \[eq:zm\] should really be the perturbative pole mass at tree level [@hisq], which differs from $am_{0c}$ by $\approx 0.04(am_{0c})^4$, but gives the same extrapolated $m_c/m_s$ ratio. C. Aubin [*et al*]{}, MILC, Phys. Rev. D[70]{}:114501 (2004) \[arXiv:hep-lat/0407028\]. C. Allton [*et al*]{}, RBC-UKQCD, Phys. Rev. D[**78**]{}:114501 (2008) \[arXiv:0804.0473\]. T. Ishikawa [*et al*]{}, CP-PACS/JLQCD, Phys. Rev. D[**78**]{}:011502R (2008) \[arXiv: 0704.1937\].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Muhammad Farhan\
Department of Computer Science\
School of Science and Engineering\
Lahore University of Management Sciences\
Lahore, Pakistan\
`[email protected]`\
Juvaria Tariq\
Department of Mathematics\
School of Science and Engineering\
Lahore University of Management Sciences\
Lahore, Pakistan\
`[email protected]`\
Arif Zaman\
Department of Computer Science\
School of Science and Engineering\
Lahore University of Management Sciences\
Lahore, Pakistan\
`[email protected]`\
Mudassir Shabbir\
Department of Computer Science\
Information Technology University\
Lahore, Pakistan\
`[email protected]`\
Imdad Ullah Khan\
Department of Computer Science\
School of Science and Engineering\
Lahore University of Management Sciences\
Lahore, Pakistan\
`[email protected]`\
bibliography:
- 'bibliography\_CameraReady.bib'
title: Efficient Approximation Algorithms for String Kernel Based Sequence Classification
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we give a characterization for a class of edge-transitive Cayley graphs, and provide methods for constructing Cayley graphs with certain symmetry properties. Also this study leads to construct and characterise a new family of half-transitive graphs.'
address: |
Beijing International Center for Mathematical Research\
Peking University\
Beijing, 100871\
P. R. China
author:
- Lei Wang
title: 'Tetravalent edge-transitive Cayley graphs of Frobenius groups'
---
Introduction
============
Graphs considered in this paper are assumed to be finite, simple, and unless stated otherwise, connected and undirected. For a graph $\Ga$, let $V\Ga$, $E\Ga$ and $\Aut\Ga$ denote its vertex set, edge set and the full automorphism group, respectively. If there exists a subgroup $X\leqslant\Aut\Ga$ is transitive on $V\Ga$ or $E\Ga$, then the graph $\Ga$ is said to be $X$-vertex transitive or $X$-edge transitive, respectively. A sequence $v_0,v_1,\ldots, v_s$ of vertices of $\Ga$ is called an $s$-arc if $v_{i-1}\not=v_{i+1}$ for $1\leqslant i\leqslant s-1$, and $\{v_i,v_{i+1}\}$ is an edge for $0\leqslant i\leqslant s-1$. The graph $\Ga$ is called $(X, s)$-arc-transitive, if $X$ is transitive on the $s$-arcs of $\Ga$; if in addition $X$ is not transitive on the $(s+1)$-arcs, then $\Ga$ is said to be $(X,s)$-transitive. In particular, a $1$-arc is simply called an arc, and an $(X,1)$-arc-transitive graph is called $X$-arc transitive.
A graph $\Ga$ is called a Cayley graph if there exist a group $G$ and a subset $S\subset G\setminus \{1\}$ with $S=S^{-1}{:}=\{g^{-1}\mid g\in S\}$ such that the vertices of $\Ga$ may be identified with the elements of $G$ in such a way that $x$ is adjacent to $y$ if and only if $yx^{-1}\in S$. The Cayley graph $\Ga$ is denoted by $\Cay(G, S)$. Throughout this paper, denote by [**1**]{} the vertex of $\Cay(G,S)$ corresponding to the identity of $G$.
It is well-known that a graph $\Ga$ is a Cayley graph of a group $G$ if and only if the full automorphism group $\Aut\Ga$ contains a subgroup which is regular on vertices and isomorphic to $G$. In particular, a Cayley graph $\Cay(G,S)$ is vertex-transitive of order $|G|$. However, a Cayley graph is of course not necessarily edge-transitive. Thus, characterizing the Cayley graphs which are edge-transitive is a current hot topic in algebra graph theory. For instance, see [@Xin; @gui; @Hua; @Zhang; @Xiuyun1; @Ming; @Yao] for those with valency $4$, see [@Zai; @Ping; @Lu] for a classification of connected edge-transitive tetravalent Cayley graphs of square-free order, and [@Corr] for a classification of normal edge-transitive Cayley graphs of Frobebius groups of order a product of two primes. In this paper, a characterization is given of tetravalent edge-transitive Cayley graphs of a class of primitive Frobenius groups. This study provides a method for constructing edge-transitive graphs of valency 4, and is then applied to construct a new family of half-transitive graphs. To state this result, we need more definitions.
For an $X$-vertex-transitive graph $\Ga$ and a normal subgroup $N\lhd X$, the normal quotient graph $\Ga_N$ induced by $N$ is the graph which has vertex set $V\Ga_N=\{u^N\div u\in V\Ga\}$ such that $u^N$ and $v^N$ are adjacent if and only if $u$ is adjacent in $\Ga$ to some vertex in $v^N$. Furthermore, if the valency of $\Ga_N$ equals the valency of $\Ga$, then $\Ga$ is called a normal cover of $\Ga_N$.
For an integer $m\geqslant3$, we denote by $\C_{m[2]}$ the lexicographic product of the empty graph $2\bf K_1$ of order $2$ by a cycle $\C_m$ of size $m$, which has vertex set $\{(i,j)\div 1\leqslant i\leqslant m, 1\leqslant j\leqslant 2\}$ such that $(i,j)$ and $(i',j')$ are adjacent if and only if $i-i'\equiv\pm1\,(\mod m)$.
A group $G$ is said to be a [*Frobenius group*]{} if and only if $G$ has the form $G=W{:}H$ such that each non-identity element of $H$ centralises no non-identity element of $W$, that is, $xy\not=yx$ for any $x\in W\setminus\{1\}$ and $y\in H\setminus\{1\}$. In particular, $G$ is called a [*primitive Frobenius group*]{} if $H$ acts irreducibly on $W$, refer to [@DM-book].
Let $\FF$ be a field, $R$ be a group and $V$ be an $\FF R$-module. Suppose that $V=V_1\oplus\cdots\oplus V_r$ $(r>1)$, where $V_i$ are subspaces of $V$ which are transitively permuted by the action of $R$. We call $R$ imprimitive on $V$ if there exists such decomposition. Otherwise, $R$ is called primitive on $V$.
\[soluble\] Let $G=W{:}H\cong\ZZ_p^d{:}\ZZ_n$ be a primitive Frobenius group, where $d,n$ are integers, and $p$ is a prime. Assume that $\Ga$ is a connected tetravalent $X$-edge-transitive Cayley graph of $G$, where $G\leqslant X\leqslant \Aut\Ga$. If $X$ is soluble, then one of the following statements holds:
- $G$ is normal in $X$, and $X_1\leqslant\D_8$;
- $G\cong\D_{2p}$, $\Ga\cong\C_{p[2]}$, and $\Aut\Ga\cong\ZZ_2^p{:}\D_{2p}$;
- $X=W{:}((N{:}H).O)$ with $\soc(X)=W\times L$, and $X_1=N.O$, where $N\cong\ZZ_2^l$ with $2\leqslant l\leqslant d$, $L\cong1$ or $\ZZ_2$, and $O\cong1$ or $\ZZ_2$, satisfying the following statements:
- there exist $x_1,\ldots,x_d\in W$ and $\tau_1,\tau_2,\ldots,\tau_d\in N$ such that $W=\l x_1,x_2,\ldots,x_d\r$, $\l x_i,\tau_i\r\cong\D_{2p}$ and $N=\l\tau_i\r\times\C_N(x_i)$ for $1\leqslant i\leqslant d$;
- $H$ does not centralise $N$, and $H$ is imprimitive on $W$;
- $X/(WN)\cong\ZZ_n$ or $\D_{2n}$, and $\Ga$ is $X$-arc-transitive if and only if $X/(WN)\cong\D_{2n}$;
- $\Ga_W\cong\C_{\frac{n}{2}[2]}$, $\Ga$ is a cover of $\Ga_W$ and $X=W{:}((NH).O)$ such that
- $X_1\leqslant N.O$, $N\cap H\cong\ZZ_2$, and $H$ normalizes $N$, but $H$ does not centralise $N$, where $N\cong\ZZ_2^l$ with $2\leqslant l\leqslant \frac{n}{2}$, $4$ divides $n$, and $O\cong1$ or $\ZZ_2$;
- $W$ is the unique minimal normal subgroup of $X$, and $H$ is imprimitive on $W$;
- $X/(WN)\cong\ZZ_{\frac{n}{2}}$ or $\D_n$, and $\Ga$ is $X$-arc-transitive if and only if $X/(WN)\cong\D_n$;
- $X=((WN){:}H).O$ and $\Ga$ is $X$-arc-transitive if and only if $X/(WN)\cong\D_{2n}$, where $W\cong\ZZ_2^d$, $N$ is a $2$-group, and $O\cong1$ or $\ZZ_2$.
0.1in [**Remarks on Theorem \[soluble\].**]{}
- The Cayley graph $\Ga$ in part (1), called a normal edge-transitive graph, is studied in [@Praeger]. Furthermore, if $X=\Aut\Ga$, then $\Ga$ is called a normal Cayley graph, introduced in [@Xu].
- $H$ acts irreducibly on $W$ if and only if $n$ does not divide $p^m-1$ for any proper divisor $m$ of $d$ (such $n$ is called a [*primitive divisor*]{} of $p^d-1$), refer to [@DF Proposition 2.3].
- Lemma \[Lee\] and Lemma \[Le\] show that every group $X$ satisfies part (3) or part (4) if and only if $H$ is imprimitive on $W$, see Constructions \[im\] and \[yin\]. In addition, $H$ is imprimitive on $W$ if and only if there exists some prime $k$ dividing $d$ such that $n$ divides $k(p^{\frac{d}{k}}-1)$, see [@DF Proposition 2.8].
0.1in
\[isoluble\] Using the notation defined in Theorem $\ref{soluble}$, if $X$ is insoluble, then one of the following holds:
- $G\cong\ZZ_p^4{:}\ZZ_5$, $X=W.\overline X$ and $\Ga_W\cong\K_5$, where $\soc(\ov X)\cong\A_5$, and $\Ga$ is constructed as in Construction $\ref{AA}$;
- $G\cong\ZZ_p^4{:}\ZZ_{10}$, $X=W.(\overline X\times\ZZ_2)$, and $\Ga_W\cong\K_{5,5}-5\K_2$, where $\soc(\ov X)\cong\A_5$, and $\Ga$ is constructed as in Construction $\ref{AAAAA}$;
- $\Ga$ is isomorphic to one of the graphs listed in Table $1$.
0.07in
**TABLE 1: [Graphs which are not normal edge-transitive.]{}**
$$\begin{array}{|c|c|c|c|c|}\hline
\Aut\Ga& G & (\Aut\Ga)_1 & \Ga \\ \hline
\PSL(3,3){:}\ZZ_2 &\D_{26} & \ZZ_3^2{:}\GL(2,3) & \mbox{Example}~\ref{TA}\\ \hline
\PGL(2,7) & \D_{14} & \S_4 & \mbox{Example}~\ref{TA}\\\hline
\PGL(2,7) & \ZZ_7{:}\ZZ_3 & \D_{16}& \mbox{Example}~\ref{A} \\\hline
\PGL(2,7) & \ZZ_7{:}\ZZ_6 & \D_8 & \mbox{Example}~\ref{B}\\\hline
\PSL(2,23) &\ZZ_{23}{:}\ZZ_{11} & \S_4 & \mbox{Example}~\ref{C}\\ \hline
\PGL(2,11) &\ZZ_{11}{:}\ZZ_{5} & \S_4 & \mbox{Example}~\ref{C}\\ \hline
\PGL(2,11)\times\ZZ_2 &\ZZ_{11}{:}\ZZ_{10} & \S_4 & \mbox{Example}~\ref{D}\\\hline
\end{array}$$
0.2in A graph is said to be half-transitive if its automorphism group acts transitively on the vertex set and edge set but intransitively on the arc set. Constructing and characterising half-transitive graphs was initiated by Tutte (1965), and is a currently active topic, see [@Maru; @D.Maru; @D.Mar; @DD] for references. Theorem \[soluble\] provides a method for characterising some classes of half-transitive graphs of valency $4$. The following theorem is such an example.
\[solubles\] Let $G=W{:}\langle h\rangle\cong\ZZ_p^d{:}\ZZ_n$ be a primitive Frobenius group, where $d>1$ is odd, $p$ is an odd prime, and $n$ is an integer. Let $\Ga$ be a connected tetravalent edge-transitive Cayley graph of $G$. Assume that $\l h\r$ is primitive on $W$. Then $\Aut\Ga=G{:}\ZZ_2$, $\Ga$ is half-transitive, and $\Ga\cong\Ga_i=\Cay(G,S_i)$, where $1\leqslant i\leqslant\lfloor\frac{n-1}{2}\rfloor$, $(n,i)=1$, and $$S_i=\{ah^i, a^{-1}h^i, (ah^i)^{-1}, (a^{-1}h^i)^{-1}\}, where\,\, a\in W\setminus \{1\}.$$ Moreover, if $p^ri\equiv\pm j\ (\mod n)$ for some $r\geqslant 0$, then $\Ga_i\cong\Ga_j$.
Preliminary results
===================
In this section, we quote some preliminary results, which will be used in the subsequent sections.
0.07in For a core-free subgroup $H$ of $X$ and an element $g\in X\setminus H$, let $[X{:}H]:=\{Hx\mid x\in X\}$, and define the coset graph $$\Ga=\Cos(X,H, H\{g,g^{-1}\}H)$$ with vertex set $[X{:}H]$ such that $Hx$ and $Hy$ are adjacent whenever $yx^{-1}\in H\{g,g^{-1}\}H$. Then $\Ga$ is well-defined, and $X$ induces a subgroup of $\Aut\Ga$ acting on $[X{:}H]$ by right multiplication, namely, $\a : Hx\to Hxa$ for $x,a\in X$. Label $v,w$ the two vertices of $\Ga$ corresponding to $H$ and $Hg$, respectively. Then we have the following lemma.
\[Cos\]
For a coset graph $\Ga=\Cos(X,H, H\{g,g^{-1}\}H)$, we have
- $\Ga(v)=\{Hgh|h\in H\}\cup\{Hg^{-1}h|h\in H\}$;
- $\Ga$ is $X$-edge-transitive and $X$ is transitive on the vertices of $\Ga$;
- $\Ga$ is connected if and only if $X=\l H,g\r$;
- $H\cap H^g=X_{vw},$ the stabilizer of the arc $(v,w)$, where $H^g$ is the conjugate of $H$ by $g$;
- the valency of $\Ga$ equals
$${val(\Ga)=}
\begin{cases}
|H{:}H\cap H^g| & \text{if $HgH=Hg^{-1}H$,}\\
2|H{:}H\cap H^g| &\text{otherwise};
\end{cases}$$
- $\Ga$ is $X$-arc-transitive if and only if $HgH=Hg^{-1}H,$ which yields that $HgH=HoH$ for some $(2$-element $)$ $o\in\N_{X}(X_{vw})\setminus H$ with $o^2\in X_{vw}$ (refer to [@Zai; @Ping; @Lu]). (An element o in the group [**$X$**]{} is a $2$-element if its order is a power of [**$2$**]{}).
Moreover, for any $X$-edge-transitive graph $\Sigma$, if $X$ is transitive on $V\Sigma$, then the map $u^x\to Hx$ with $x\in X$ gives an isomorphism from $\Sigma$ to $\Cos(X, H, H\{g, g^{-1}\}H)$, where $u\in V\Sigma$, $H = X_u$ and $g\in X \setminus H$ with $u^g\in\Ga(u)$.
The vertex stabilizer for $s$-arc-transitive graphs of valency $4$ is known (refer to [@Weiss]).
\[Lu\] Let $\Ga=(V\Ga,E\Ga)$ be a connected $(X,s)$-transitive graph of valency $4$. Then $s$ and the stabilizer $X_1$ are listed in the following table, $$\begin{array}{|c|c|c|c|c|c|c|}\hline
s & 2 & 3 & 4 & 7 \\ \hline
X_1 & \A_4,\,\S_4 & \ZZ_3\times\A_4, (\ZZ_3\times\A_4).\ZZ_2,\,\S_3\times\S_4 &\ZZ_3^2{:}\GL(2,3)& [3^5]{:}\GL(2,3) \\
\hline
\end{array}$$ where $[3^5]$ is a $3$-group of order $3^5.$
Let $\Ga=(V\Ga,E\Ga)$ be a connected graph. Assume that $X\leqslant\Aut\Ga$ is transitive on both $V\Ga$ and $E\Ga$. Then we have an important conclusion in the next lemma.
\[insoluble\] Let $N\lhd X$. If $\Ga$ is of valency $4$ and $X/N$ is insoluble, then $\Ga$ is a normal $N$-cover of $\Ga_N$.
[*Proof. *]{}
Pick any vertex $u\in V\Ga$. Let $B$ be an orbit of $N$ acting on $V\Ga$, which contains $u$. By Lemma \[Lu\], the stabilizer $X_u$ is a $\{2,3\}$-group. In particular, $X_u$ is soluble. Let $K$ be the kernel of $X$ acting on $\Ga_N$. Then $K_u\unlhd X_u$, so $K_u$ is soluble. Since $N$ is transitive on $B$, we have $K=NK_u$. Note that $K/N\cong NK_u/N\cong K_u/(N\cap K_u)$, $K/N$ is soluble. Then $X/K\cong (X/N)/(K/N)$ is insoluble because $X/N$ is insoluble. So $\Aut\Ga_N$ is also insoluble, hence $\Ga_N$ is not a cycle. Since $\Ga$ is connected and the valency of $\Ga_N$ is a divisor of the valency of $\Ga$, we conclude that $\Ga_N$ is of valency $4$, and the lemma holds.
For a normal edge-transitive Cayley graph $\Ga=\Cay(G,S)$, let $\Aut(G,S)=\{\a\in\Aut(G)\div S^\a=S\}$, we have a simple lemma.
\[G\] Let $G=W{:}\l h\r\cong\ZZ_p^d{:}\ZZ_n$ be a primitive Frobenius group, where $d, n$ are integers, and $p$ is a prime. Let $\Ga=\Cay(G,S)$ be connected of valency $4$. Assume that $\Aut\Ga$ has a subgroup $X$ such that $\Ga$ is $X$-edge-transitive and $G\unlhd X$. Then $X_1\leqslant\D_8$.
[*Proof. *]{}
Since $\Ga$ is connected, we have $\l S\r=G$, and so $\Aut(G,S)$ acts faithfully on $S$. Hence $\Aut(G,S)\leqslant\S_4$. By [@Godsil Lemma 2.1], we obtain $X\leqslant\N_{\Aut\Ga}(G)=G{:}\Aut(G,S)$. So $X_1\leqslant\Aut(G,S)\leqslant\S_4$. Suppose that $3$ divides $|X_1|$. Then $X_1$ is $2$-transitive on $S$. Hence $\Ga$ is $(X,2)$-arc-transitive, and all elements in $S$ are involutions, see for example [@flag].
Pick any $s\in S$. Write $s=\s h^i$ where $\s\in W$ and $i$ is an integer. Recall that $s$ is an involution, we obtain that $h^{2i}=1$. By [@Doerk Proposition 12.10], $\Aut(G)\cong\ZZ_p^d{:}\GammaL(1,p^d)$. For a finite group $T$, it is known that the action of $\Aut(T)$ on $T/\Z(T)$ is permutationally isomorphic to the conjugation action of $\Aut(T)$ on $\Inn(T)$. Since $G\cong\Inn(G)$, it follows from the above fact that we may identify $G$ with $\Inn(G)$ a normal subgroup of $\Aut(G)$. Then write $\Aut(G)=W{:}M.L$, where $M\cong\ZZ_{p^d-1}$, and $L\cong\ZZ_d$. Without loss of generality, we may assume that $h$ belongs to $ M$, refer to [@DF Proposition 2.5]. For that case, $(h^i)^\eta=h^i$ for any $\eta\in M.L$. Take any $\theta\in\Aut(G)$, $\theta$ has the form $xyz$, where $x\in W$, $y\in M$, and $z\in L$. By easy calculations, we have $s^\theta=\ov s h^i$, where $\ov s\in W$. It follows that for each $a\in S$, $a$ has the form $\ov a h^i$ with $\ov a\in W$ because $X_1$ is transitive on $S$. Recall that $\l S\r=G$, we have $h=(\s_1h^i)(\s_2h^i)\cdots(\s_mh^i)$ where $\s_j\in W$ for each $j$. Since $W\unlhd G$ and $h\not=1$, we obtain $h=h^i$. Consequently, $\l h\r\cong\ZZ_2$. By the definition of $G$, we have $G\cong\D_{2p}$, and thus $\Aut(G)\cong\ZZ_p{:}\ZZ_{p-1}$. However, since $X_1$ is $2$-transitive on $S$, we conclude that $X_1\cong\A_4$ or $\S_4$, which is impossible. Therefore, $X_1\leqslant\D_8$.
Finally, we quote a result about simple groups, which will be used later.
[([Kazarin[@L.Kazarin])]{}]{}\[Kazarin\] Let $T$ be a non-abelian simple group which has a $2'$-Hall subgroup. Then $T=\PSL(2,p)$, where $p=2^e-1$ is a prime. Furthermore, $T=GH,$ where $G=\ZZ_p{:}\ZZ_{\frac{p-1}{2}}$ and $H=\D_{p+1}=\D_{2^e}$.
existence of graphs satisfying Theorem $\ref{soluble}$ and Theorem $\ref{isoluble}$
===================================================================================
In this section, we first construct some examples of graphs satisfying Theorem \[soluble\].
0.07in The following construction produces normal edge-transitive Cayley graphs admitting a group $X$ satisfying part (1) of Theorem \[soluble\].
\[BBB\]
Let $p\geqslant5$ be a prime such that $p$ is a primitive divisor of $2^{p-1}-1$. Let $G=W{:}\l h\r\cong\ZZ_2^{p-1}{:}\ZZ_p$ be a Frobenius group. By [@Doerk Proposition 12.10], we have $\Aut(G)\cong\ZZ_2^{p-1}{:}\GL(1,2^{p-1}).\ZZ_{p-1}$, where $\ZZ_{p-1}$ is the group of Frobenius automorphisms. Arguing similarly as Lemma \[G\], we may write $\Aut(G)=W{:}M{:}L$, and $h$ belongs to $M$, where $M\cong\ZZ_{2^{p-1}-1}$, and $L\cong\ZZ_{p-1}$. For this case, we may identify $W$ with a field $\FF{:}=\FF_{2^{p-1}}$ of order $2^{p-1}$ and there exists $\a\in\FF$ of order $p$ such that $\l h\r$ acts on each $x\in W$ by $h : x=\a x$. By the definition, $G$ is a primitive Frobenius group.
Let $\FF^{\#}=\l\o\r$, and let $\s$ be a Frobenius automorphism of order $2$. Then $\o^\s=\o^{2^{\frac{p-1}{2}}}$. Let $X=G{:}\l\s\r$, and let $g=\o^{2^{\frac{p-1}{2}}+1} h$. Set $$\Ga(2,p-1,p)=\Cos(X,\l\s\r,\l\s\r\{g,g^{-1}\}\l\s\r).$$
\[BB\] Let $\Ga=\Ga(2,p-1,p)$ be a graph constructed in Construction $\ref{BBB}$. Then $\Ga$ is a connected normal $X$-edge-transitive Cayley graph of $G$ of valency $4$.
[*Proof. *]{}
By the definition, $\l\s\r$ is core-free in $X$, and hence $X\leqslant\Aut\Ga$. Now $X=G\l\s\r$ and $G\cap\l\s\r=1$, and thus $G$ is regular on the vertex set $[X{:}\l\s\r]$. So $\Ga$ is a Cayley graph of $G$, which has order $2^{p-1}p$.
Let $Y=\l g,\s\r$. Noting that $p\geqslant 5$, we conclude that $2^{\frac{p-1}{2}}\not\equiv-1(\mod 2^{p-1}-1)$. It implies that $\o^\s\not=\o^{-1}$. However, since $\a^\s=\a^{-1}$, we have that $h^\s=h^{-1}$. Furthermore, $\s$ induces an automorphism of $G$. Then we have $$g^\s=(\o^{2^{\frac{p-1}{2}}+1} h)^\s=(\o^\s)^{2^{\frac{p-1}{2}}+1} h^\s=\o^{2^\frac{p-1}{2}+1}h^{-1}.$$ Let $\ov g=g^\s g$. Then $\ov g=\o^{2(2^{\frac{p-1}{2}}+1)}\a$. Denote by $\ell$ the integer $2^{\frac{p-1}{2}}-1$. Recall that $p$ is a primitive divisor of $2^{p-1}-1$, we conclude that $(p,\ell)=1$. Thus $\ov g^\ell=\a^\ell$ belongs to $Y$. So does $\a$. Consequently, $\o^{2(2^{\frac{p-1}{2}}+1)}$ belongs to $Y$, and so $h$ belongs to $Y$. Since $\l h\r$ acts irreducibly on $W$, we obtain that $X=Y$. Thus $\Ga$ is connected. It is straightforward to show that $\l\s\r\cap\l\s\r^g=1$, and hence $\l\s\r\cap\l\s\r^g$ has index $2$ in $\l\s\r$. Since $X\leqslant\Aut\Ga$, it follows that $\Ga$ is not a cycle. By Lemma $\ref{Cos}$, $\Ga$ is connected, $X$-edge-transitive and of valency $4$.
[**Remark.**]{} In fact, the graphs in Construction $\ref{BBB}$ really exist. For example, $p=5,11,13,19$, and so on.
0.07in
The following construction produces edge-transitive graphs admitting a group $X$ satisfying part (3) of Theorem \[soluble\] with $L\cong\ZZ_2$, and $O=1$.
\[im\]
Let $X=W{:}(N{:}\l h\r)\cong\ZZ_p^d{:}(\ZZ_2^d{:}\ZZ_n)$, where $p=2^\ell m+1$ be an odd prime, $m\geqslant3$ is an odd number, and $\ell\geqslant1$, such that $W\cong\ZZ_p^d$, $N\cong\ZZ_2^d$, and $\l h\r\cong\ZZ_n$ satisfy
- $d>1$, $d$ divides $m$, and $2md$ is a primitive divisor of $p^d-1$;
- $W=\prod_{i=1}^d\l x_i\r$, where $x_i=(1,\ldots,1,x,1,\ldots,1)$ with $o(x)=p$ for each $i$;
- $N=\prod_{i=1}^d\l \t_i^{\frac{p-1}{2}}\r$, where $\t_i=(1,\ldots,1,\t,1,\ldots,1)$ with $x^{\t}=x^r$ and $r^{p-1}\equiv1\,(\mod p)$ for each $i$;
- $h=c_1\t_1^{\frac{p-1}{2m}}(12\ldots d)$, where $c_1=(c,1,\ldots,1)$ with $x^c=x$, $\t^x=\t$, and $o(c)=2$.
Let $y=(x_1h)^{-1}$. Set $$\Ga(p,2,n)=\Cos(X,N,N\{ y,y^{-1}\}N).$$
\[Liu\] Let $\Ga=\Ga(p,2,n)$ be a graph constructed in Construction $\ref{im}$, and let $G=W{:}\l h\r$. Then $\Ga$ is a connected tetravalent $X$-edge-transitive Cayley graph of Frobenius group $G$, and $G$ is not normal in $X$.
[*Proof. *]{}
By the definition, $N$ is core-free in $X$, and hence $X\leqslant\Aut\Ga$. Now $X=GN$ and $G\cap N=1$, and thus $G$ acts regularly on the vertex set $[X{:}N]$. So $\Ga$ is a Cayley graph of $G$. Obviously, $G$ is not normal in $X$.
Let $H=\l h\r$. Suppose that $C{:}=\C_H(W)\not=1$. Then $C=\l h^\ell\r$, where $\ell$ divides $2md$. Write $\ell=l_1d+l$, where $0\leqslant l_1<2m$, and $0\leqslant l<d$. Let $\ov\t=\t_1\t_2\cdots\t_d$, and $\ov c=c_1\cdots c_d$, where $c_i=(1,\ldots,1,c,1,\ldots,1)$ for each $i$. If $l=0$, then $h^\ell=\ov c~\ov \t^{\frac{p-1}{2m}l_1}$ and so $x_1^{h^\ell}\not=x_1$, a contradiction. Thus $l\not=0$. Then $h^\ell=\ov c~\ov\t^{\frac{p-1}{2m}l_1}h^{l}$. Let $(l,d)=k$. Let $k'=l/k$, and $d'=d/k$. Relabeling if necessary, we may rewrite $\{1,\ldots,d\}=\{1_1,\ldots,i_j,\ldots,k_{d'}\}$. Without loss of generality, we may assume that $h^l=h_1\cdots h_k$, where $h_i=(c_{i_1}\t_{i_1}^{\frac{p-1}{2m}})(c_{i_2}\t_{i_2}^{\frac{p-1}{2m}})\cdots(c_{i_{k'}}\t_{i_{k'}}^{\frac{p-1}{2m}})(i_1 i_2\cdots i_{d'})$. Then $x_{1_1}^{h^\ell}=x_{1_2}^{r^{\frac{p-1}{2m}(l_1+1)}}\not=x_{1_1}$ because $1_1\not=1_2$, a contradiction. Thus $H$ acts faithfully on $W$.
We claim that $H$ is fixed-point-free on $W$. Let $U=\l w\div w^h=w, w\in W\r$. If otherwise, then $U$ is a proper subgroup of $W$. By Maschke’s Theorem, $V$ can be decomposed as $W=U\times V$ such that $H$ normalises both $U$ and $V$. By the definition of $U$, $H$ is fixed-point-free on $V$. Let $k=\dim(V)$. Then $k<d$. By the above paragraph, we conclude that $2md$ divides $p^k-1$, contrary to our assumption. This establishes the claim. So $G$ is a primitive Frobenius group.
For $y$ defined in Construction \[im\], let $z=y^{-1}y^{\ov\t^{\frac{p-1}{2}}}=x_1^2$. Then $x_1$ belongs to $\l N,y\r$. So does $x_i$ for $1\leqslant i\leqslant d$. It follows that $\l N,y\r=X$. Thus $\Ga$ is connected. It furthermore implies that $\l\ov c\r$ belongs to $X$, and so $\soc(X)=W\times\l\ov c\r$.
Let $\s_i=\t_i^{\frac{p-1}{2}}$ where $1\leqslant i\leqslant d$. Finally, as $\s_i^y=\s_{i-1}$ for $3\leqslant i\leqslant d$, $\s_1^y=\s_d$ and $\s_2^y=x_1^2\s_1$, we obtain that $N\cap N^y=\l\s_2,\s_3,\ldots,\s_d\r\cong\ZZ_2^{d-1}$. That is to say, $N\cap N^y$ has index $2$ in $N$. Since $X\leqslant\Aut\Ga$, $\Ga$ is not a cycle. By Lemma \[Cos\], $\Ga$ is connected, $X$-edge-transitive and of valency 4.
0.07in [**Remark.**]{} The normal quotient $\Ga_W$ induced by $W$ is a cycle (see Lemmas \[Lee\] and \[Le\]).
0.1in As a matter of fact, there are several groups which are primitive Frobenius groups and satisfy Construction \[im\]. For example, $G=\ZZ_7^3{:}\ZZ_{18}$, $\ZZ_{13}^3{:}\ZZ_{18}$, $\ZZ_{41}^5{:}\ZZ_{50}$, and so on. 0.1in
The following construction produces edge-transitive graphs admitting a group $X$ satisfying part (4) of Theorem \[soluble\] with $O=1$.
\[yin\] Using the notation in Construction $\ref{im}$. Assume $\ell\geqslant 2$. Let $N=\prod_{i\not=3}\l\t_i^{\frac{p-1}{2}}\r\cong\ZZ_2^{d-1}$, and $h=\t_1^{\frac{p-1}{4m}}(12\cdots d)$. Let $X=W{:}\l N,h\r$, and let $G=W{:}\l h\r$. Set $y=(x_2h)^{-1}$, and $$\Ga(p,n)=\Cos(X,N,N\{y,y^{-1}\}N).$$
Let $\Ga=\Ga(p,n)$ be a graph constructed in Construction $\ref{yin}$. Then $\Ga$ is a connected tetravalent $X$-edge-transitive Cayley graph of Frobenius group $G$, and $G$ is not normal in $X$.
[*Proof. *]{}
Obviously, $G$ is not normal in $X$. Let $\s_i=\t_i^{\frac{p-1}{2}}$ where $1\leqslant i\leqslant d$. By easy calculations, $\s_1^y=\s_d$, $\s_2^y=\s_1$, and $\s_i^y=\s_{i-1}$ for $4\leqslant i\leqslant d$. It follows that $\s_3$ belongs to $\l N,y\r$, and $N\cap N^y=\l\s_1,\s_4,\ldots,\s_d\r\cong\ZZ_2^{d-2}$. At the same time, we obtain $\s_3^y=x_2^2\s_2$, and hence $x_2$ belongs to $\l N,y\r$. So does $x_i$ for each $i$. It implies that $\l N,y\r=X$. Consequently, $\Ga$ is connected. Arguing similarly as Lemma \[Liu\], we can obtain that $G$ is a Frobenius group, and $\Ga$ is $X$-edge-transitive Cayley graph of $G$ and of valency 4, the statement follows.
0.07in [**Remark.**]{} Clearly, $\l h\r$ does not normalise $N$. In other words, $X$ can’t satisfy the properties in part (a) of Lemma \[Lee\]. However, $h$ normalises $\l N,h^{\frac{n}{2}}\r$, namely, $X$ satisfies the properties in part (ii) of Lemma \[Le\]. Thus $\Ga_W\cong\C_{\frac{n}{2}[2]}$, where $N$, $W$, and $\Ga$ appear in Construction \[yin\], (see Lemma \[Lee\] and Lemma \[Le\]).
0.07in The following construction produces edge-transitive graphs admitting a group $X$ satisfying part (5) of Theorem \[soluble\] with $O=1$.
0.07in
Let $n=3p_1^{l_1}\cdots p_s^{l_s}$ be an odd number, where $3$, $p_1$, $\ldots$, $p_s$ are pairwise distinct primes, and $l_i\geqslant1$ for each $i$. Let $G_1=W_1{:}H_1\cong\ZZ_2^d{:}\ZZ_n$ be a primitive Frobenius group. Let $G_2$ be a subgroup of $ G_1$ such that $G_2=W_2{:}H_2\cong\A_4$. Write $H_1=\l h_1\r$, and $H_2=\l h_2\r$ where $h_2=h_1^{\frac{n}{3}}$.
Let $H=\l h\r$ where $h=(h_1,h_2)$. Let $V=W_1\times W_2$, and $W=\{(w,1)\div w\in W_1\}$. Set $$\mbox{$G=W{:}H$\, and\,
$X=V{:}H$}.$$ By the definition, it is easy to show that $G$ is a primitive Frobenius group.
\[D[2p]{}\] [Let $R=\l (w,w), (w,w)^h\r$, where $1\not=w\in W_2$. Set $$\Ga(2,d,n)=\Cos(X, R, R\{h,h^{-1}\}R).$$ ]{}
Let $\Ga=\Ga(2,d,n)$ be a graph constructed in Construction $\ref{D{2p}}$. Then $\Ga$ is a connected tetravalent $X$-edge-transitive Cayley graph of $G$, and $G$ is not normal in $X$. In particular, $\Ga_W$ is a cycle.
[*Proof. *]{}
We first prove that $R$ is core-free in $X$. By the definition of $R$, we have $R\cong\ZZ_2^2$. Assume $K\leqslant R$, and $1\not=K\unlhd X$. Then $\Aut(K)$ is isomorphic to a subgroup of $\S_3$. So we conclude that $\C_H(K)\not=1$, which contradicts the fact that $G_1$ is a Frobenius group. Thus $R$ is core-free in $X$. We observe that $R\cap G=1$, it follows that $|X|=|R||G|$, and so $X=RG$. It implies that $G$ is regular on the vertex set $[X{:}R]$, and hence $\Ga$ is a Cayley graph of $G$.
By the definition of $G_1$, we conclude that $w^{h_1^3}\not=w$ for any $1\not=w\in W_1$. It implies that $(w,w)^{h^3}(w,w)\not=1$, namely, $(w^{h_1^3}w,1)\not=1$. Since $H$ is irreducible on $W_1$, implying that $W_1$ belongs to $\l R,h\r$. So does $V$. Thus $\Ga$ is connected. Arguing similarly as above, $G$ is not normal in $X$. Clearly, $R\cap R^h=\l (w,w)^h\r\cong\ZZ_2$. It follows that $R\cap R^h$ has index $2$ in $R$. As $X\leqslant\Aut\Ga$, $\Ga$ is not a cycle, so that by Lemma \[Cos\], we obtain $\Ga$ is $X$-edge-transitive and of valency 4. Note that $\Ga_W$ is a Cayley graph. By [@Baik Theorem 1.2], we conclude that $\Ga_W$ is a cycle.
By Constructions \[BBB\]-\[yin\], each case of Theorem \[soluble\] occurs.
0.07in We now construct some examples of graphs appearing in Theorem $\ref{isoluble}$.
0.07in Based on several previous known results, arc-transitive elementary abelian covers of the complete graph $\K_5$ were classified by Bo$\check{s}$tjan Kuzman [@Kuzman]. However, for the completeness, we present here a distinct and independent construction.
0.07in Let $p$ be a prime such that $5$ is a primitive divisor of $p^4-1$. Set $$V=\l e_1\r\times\cdots\times\l e_5\r\cong\ZZ_p^5.$$ We define an action of $\A_5$ on $V$ as follows: $$\mbox{$(\prod_{i=1}^5e_i^{\lambda_i})^{g}=\prod_{i=1}^5 e_{i^{g^{-1}}}^{\lambda_i}$,
where $g\in\A_5$, and $0\leqslant\lambda_i\leqslant 4$ for each $i$.}$$ By this definition, $\A_5$ acts naturally on $V$. Let $\ov e_i=e_5e_i^{-1}$ for $1\leqslant i\leqslant 4$. Set $$W=\l\ov e_1\r\times\l\ov e_2\r\times\l\ov e_3\r\times\l\ov e_4\r.$$ It is straightforward to show that $\A_5$ acts faithfully on $W$.
\[AA\] [Let $G=W{:}\l h\r$ with $h=(12345)$, and let $X=W{:}N=\ZZ_p^4{:}\A_5$. Let $R=\Alt\{2,3,4,5\}\cong\A_4$, and let $g=\ov e_1(15)(24)$. Set $$\Ga(p,4,5)=\Cos(X, R, RgR).$$ ]{}
\[A\_5\] Let $\Ga=\Ga(p,4,5)$ be a graph constructed in Construction $\ref{AA}$. Then $\Ga$ is a connected tetravalent $(X,2)$-arc-transitive Cayley graph of Frobenius group $G$, and $G$ is not normal in $X$. In particular, $\Ga$ is a cover of $\Ga_W$, and $\Ga_W\cong\K_5$.
[*Proof. *]{}
Let $H=\l h\r$. By definition of $W$, we conclude $H$ is fixed-point-free on $W$. Since $5$ is a primitive divisor of $p^4-1$, $H$ acts irreducibly on $W$. That is to say, $G$ is a primitive Frobenius group. Clearly, $N$ has a decomposition $HR$. It implies that $R$ is core-free in $X$, and hence $X\leqslant\Aut\Ga$. Now $X=GR$ and $G\cap R=1$, and so $G$ is regular on the vertex set $[X{:}R]$. Thus $\Ga$ is a Cayley graph of $G$. Obviously, $G$ is not normal in $X$.
Denote by $u$ and $v$ the vertices $R$ and $Rg$, respectively. Then $X_u=R$ and $X_v=R^g$. Let $r=(234)$. Since $X_{uv}=X_u\cap X_v$, a small calculations show $X_{uv}=\l r\r$. By Lemma \[Cos\], $\Ga$ is of valency $4$. It is clear that $g$ has order $2$, and $r^g=r^{-1}$. So $g\in\N_X(X_{uv})$. Let $\ov R=\l R,g\r$. Since $(15)(24)=(12345)(25)(34)$, we conclude that $\ov e_1h$ belongs to $\ov R$. Let $a=(25)(34)$ and $b=(23)(45)$. By easy calculations, we obtain $$\mbox{$\ov e_1h(\ov e_1h)^a=\ov e_1\ov e_2\ov e_3^{-1}$,
$(\ov e_1\ov e_2\ov e_3^{-1})^b=\ov e_1\ov e_2^{-1}\ov e_3\ov e_4^{-1}$, and
$(\ov e_1\ov e_2^{-1}\ov e_3\ov e_4^{-1})^{ab}=\ov e_1\ov e_2^{-1}\ov e_4^{-1}$}.$$ Combining the above three equations, we conclude that $\ov e_3$ belongs to $\ov R$. So does $\ov e_i$ for $i=1,2$, $4$. Consequently, $W\leqslant\ov R$. Recall that $\ov e_1h$ is inside in $\ov R$, it follows that $h$ belongs to $\ov R$, forcing $\l R,g\r=X$. Thus $\Ga$ is connected. Since $X/W$ is insoluble, by Lemma \[insoluble\], $\Ga$ is a cover of $\Ga_W$. Clearly, $\Ga_W$ is a Cayley graph of $G/W$. By [@Baik Theorem 1.2], we obtain $\Ga_W\cong\K_5$.
Let $p$ be a prime such that $10$ is a primitive divisor of $p^4-1$. Let $$V_1=\l e_1\r\times\cdots\times\l e_5\r,\, \mbox{and}\,\,\,
V_2=\l e_{1'}\r\times\cdots\times\l e_{5'}\r$$ such that $V_1\cong V_2\cong\ZZ_p^5$. Set $T=\l(12345)(1'2'3'4'5'),(12)(1'2')\r$. It is straightforward to show that $T\cong\S_5$. Then, for any $g\in T$, $g$ acts on $V_i(i=1,2)$ as follows: $$\begin{array}{l}
(\prod_{i=1}^5e_i^{\lambda_i})^{g}=\prod_{i=1}^5{e_{{i}^{g^{-1}}}^{\lambda_i}}, \mbox{where}\,\ 0\leqslant \lambda_i\leqslant 4\,\, \mbox{for each}\,\, i,\\
(\prod_{i=1}^5e_{i'}^{\lambda_{i'}})^{g}=\prod_{i=1}^5e_{{i'}^{g^{-1}}}^{\lambda_{i'}}, \mbox{where}\,\ 0\leqslant \lambda_{i'}\leqslant 4\,\, \mbox{for each}\,\, i'.
\end{array}$$
Let $\ov e=\prod_{i=1}^5e_i$, and $\ov e'=\prod_{i=1}^5e_{i'}$. Let $\ov e_i= e_i\l\ov e\r$ and $\ov e_{i'}=e_{i'}\l \ov e'\r$ for $1\leqslant i\leqslant 5$. Set $$\mbox{$W=\l w_1\r\times\l w_2\r\times\l w_3\r\times\l w_4\r$ where $w_i=\ov e_i\ov e_{i'}^{-1}$.}$$ Then $W\cong\ZZ_p^4$. Note that $T$ fixes each element of $\l\ov e\r$ and $\l\ov e'\r$. So $T$ induces a faithful action on $W$. Without loss of generality, we may assume that $T$ is a subgroup of $\GL(W)$. Let $g=(11')\cdots(55')$. Obviously, $g$ inverts each non-identity element of $W$.
\[AAAAA\] [Let $G=W{:}H$ where $H=\l h, g\r$ with $h=(12345)(1'2'3'4'5')$. Set $X=W{:}N=W{:}(T\times \l g\r)\cong\ZZ_p^4{:}(\S_5\times\ZZ_2)$. Let $R=\l (1234)(1'2'3'4'),(12)(1'2')\r\cong\S_4$, and let $y=w_1w_5(15)(1'5')g$. Set $$\Ga(p,4,10)=\Cos(X, R,RyR).$$ ]{}
Arguing similarly as Lemma \[A\_5\], we have the following conclusion in next lemma. 0.07in
Let $\Ga=\Ga(p,4,10)$ be a graph constructed in Construction $\ref{AAAAA}$. Then $\Ga$ is a connected tetravalent $(X,2)$-arc-transitive Cayley graph of Frobenius group $G$, and $G$ is not normal in $X$. In particular, $\Ga$ is a cover of $\Ga_{W}$, and $\Ga_W\cong\K_{5,5}-5\K_2$.
0.07in
Here are a few of graphs whose automorphism groups are almost simple.
\[TA\]
Let $\FF=\GF(p)$ be a finite field of order $p$. Let $U$ and $V$ consist of 1-subspaces and 2-subspaces of $\FF^3$, respectively.
0.07in [**Case 1[:]{}**]{}Let $p=2$. Define a bipartite graph $\Ga$ with bipartite $U$ and $V$ such that $u\in U$ and $v\in V$ are adjacent if and only if $u+v=\FF^3$. This is the point-line non-incidence graph of the Fano plane $\PG(2,2)$. Furthermore, $\Aut\Ga=\PGL(3, 2).\ZZ_2$, and $\Ga$ is a Cayley graph of $G=\D_{14}$. For example, refer to [@p].
0.07in [**Case 2[:]{}**]{}Let $p=3$. Define a bipartite graph $\Ga$ with bipartite $U$ and $V$ such that $u\in U$ and $v\in V$ are adjacent if and only if $u$ is a subspace of $v$. Then $\Ga$ is the point-line incidence graph of the projective plane $\PG(2,3)$. Furthermore, $\Aut\Ga=\PGL(3,3).\ZZ_2$, and $\Ga$ is a Cayley graph of $G=\D_{26}$. Refer to [@Li; @flag], for example.
\[A\] [Let $X=\PGL(2,7)$. By the Atlas [@atlas], $X$ has a maximal subgroup $H\cong\D_{16}$. Pick a subgroup $K\leqslant H$ with $K\cong\ZZ_2^2$. Then $\D_8\cong\N_{H}(K)\leqslant\N_{X}(K)\cong\S_4$. Choose an involution $o\in\N_{H}(K)\setminus K$ and an element $z\in\N_X(K)$ of order $3$ such that $z^o=z^{-1}$. Then $\langle o, z\rangle\cong\S_3$, and $o(oz)=2$. Since $H$ is a maximal subgroup of $X$, it follows that $\langle H, oz\rangle=X$. Let $\Ga=\Cos(X, H, HozH)$. By the choices of $o$ and $z$, we conclude that $|H{:}H\cap H^{oz}|=4$, namely, $\Ga$ is a connected $X$-arc-transitive graph of valency $4$. By MAGMA [@Magma], we have that $X=GH$ where $G=\ZZ_7{:}\ZZ_3$, and thus $\Ga$ is a connected $X$-arc-transitive Cayley graph of $G$ of valency $4$. By Li et al.[@Zai; @Ping; @Lu], $\Aut\Ga=\PGL(2,7)$. ]{}
\[B\] [Let $X=\PGL(2,7)$. Then $T=\soc(X)\cong\PSL(2,7)$. Take $H\leqslant T$ such that $H\cong\D_8$. Choose an involution $o$ such that $o$ is not in the center of $H$. It is simple to check that $\N_{H}(\langle o\rangle)\cong\ZZ_2^2$, $\N_{T}(\langle o\rangle)\cong\D_8$ and $\N_X(\langle o\rangle)\cong\D_{16}$. Let $\N_{X}(\langle o\rangle)=\N_{T}(\langle o\rangle){:}\langle z\rangle$ for some involution $z\in X\setminus T$. Take $y\in\N_{T}(\langle o\rangle){:}\langle z\rangle$ of order $4$. Set $\Ga=\Cos(X, H, HxH)$, where $x=z$ or $yz$. By Li et al.[@Zai; @Ping; @Lu], $\Ga$ is a connected tetravalent arc-transitive Cayley graph of $\ZZ_7{:}\ZZ_6$, and $\Aut\Ga=\PGL(2,7)$. ]{}
\[C\] [Let $X=\PGL(2,11)$ or $\PSL(2,23)$. By the Atlas [@atlas], $X$ has a maximal subgroup $H\cong\S_4$. Let $L\cong\S_3$ be a subgroup of $H$. Checking the subgroups of $X$ in the Atlas [@atlas], we conclude that $\N_X(L)=\l o\r\times L\cong\D_{12}$, where $o\in \N_X(L)\setminus H$ is an involution. Set $\Ga=\Cos(X,H,HoH)$. Since $H$ is a maximal subgroup of $X$, $\langle o, H\rangle=X$. It is straightforward to check that $|H{:}H\cap H^o|=4$. Then $\Ga$ is a connected tetravalent $X$-arc-transitive graph. Moreover, $X$ has a subgroup $G$ which is regular on the vertices, where $G\cong\ZZ_{11}{:}\ZZ_5$ or $\ZZ_{23}{:}\ZZ_{11}$, respectively. We denote by $P_{11,5}$ and $P_{23,11}$ the graphs associated with $\PGL(2,11)$ and $\PSL(2,23)$, respectively. By Li et al.[@Zai; @Ping; @Lu], $\Aut P_{11,5} =\PGL(2,11)$ and $\Aut P_{23,11} =\PSL(2,23)$. ]{}
\[D\] [Let $\Ga=(V\Ga, E\Ga)$ be a connected arc-transitive Cayley graph. The standard double cover $\Ga^{(2)}$ is the graph with vertex set $V\Ga\cup\{u'|u\in V\Ga\}$ such that $\{u,v'\}\in E\Ga^{(2)}$ whenever $\{u,v\}\in E\Ga$. For each $x\in\Aut\Ga$, define $\tilde x : u\to u^x$, $u'\to (u^x)'$. Then $\Aut\Ga$ can be viewed as a subgroup of $\Aut\Ga^{(2)}$ in this way. Define $\epsilon : u\to u'$, $u'\to u$. Then $\epsilon\in\Aut\Ga^{(2)}$. Set $X=\l\Aut\Ga,\epsilon\r$. Then $X=\Aut\Ga\times\l\epsilon\r$. So $\Ga^{(2)}$ is an $X$-arc-transitive Cayley graph. By Li et al.[@Zai; @Ping; @Lu], $P_{11,5}^{(2)}$ is a Cayley graph of $\ZZ_{11}{:}\ZZ_{10}$, and $\Aut P^{(2)}_{11,5}\cong\PGL(2,11)\times\ZZ_2$. ]{}
Soluble automorphism groups
===========================
In this section, let $G=W{:}H\cong\ZZ_p^d{:}\ZZ_n$ be a primitive Frobenius group. Let $\Ga=\Cay(G,S)$ be a connected $X$-edge-transitive tetravalent Cayley graph, where $G\leqslant X\leqslant \Aut\Ga$. Denote by $F$ the Fitting subgroup of $X$. If $X$ is solvable, then an important property of its Fitting subgroup is that it is self-centralized, that is, $\C_X(F)\leqslant F$. In what follows, we will determine the graph $\Ga$ for the case where $X$ is solvable.
\[un\] Assume that $F$ is a $r$-group, where $r$ is a prime. If $\Ga$ is a cover of $\Ga_F$, then $F=W$.
[*Proof. *]{}
Note that $W$ is minimal and normal in $G$. Then either $W\leqslant F$ or $F\cap G=1$. If $W\leqslant F$, then $F=W$ as $\Ga$ is a cover of $\Ga_F$. Thus we assume that $F\cap G=1$. Let $\ov G=G\Phi(F)/\Phi(F)$, and let $\ov F=F/\Phi(F)$. Since $\Phi(F)$ $F$, we obtain $\ov G$ can act on $\ov F$ by conjugation. Clearly, $\ov G\cong G$. In what follows, write $\ov G=\ov W{:}\ov H\cong W{:}H$.
Assume first that $r\not=p$. If $\ov W$ acts trivially on $\ov F$, then $W$ induces the identity on $\ov F$. From [@Gorenstein p.174, Theorem 1.4], it follows that $W$ acts trivially on $F$. So $W\leqslant\C_X(F)\leqslant F$, against our assumption. Thus $\ov W$ acts nontrivially on $\ov F$.
Let $M=\ov F{:}\ov G$. Let $$1\unlhd M_1\unlhd M_2\unlhd\cdots\unlhd M_{m-1}\unlhd M_m=\ov F$$ be the normal series of $\overline{F}$ such that each $M_i/M_{i-1}$ is a minimal normal subgroup of $M/M_{i-1}$ for $1\leqslant i\leqslant m$, where $M_0=1$.
It is straightforward to show that $\ov G$ normalizes $\C_{M_1}(\ov W)$, and hence $\C_{M_1}(\ov W)\unlhd M$. By the minimality of $M_1$, we conclude that either $\C_{M_1}(\ov W)=1$ or $\C_{M_1}(\ov W)=M_1$. If $\C_{M_1}(\ov W)=1$, so that by [@Evgenii; @I Theorem 2.7] we have $|M_1|=|\C_{M_1}(\ov H)|^{|\ov H|}$. However, $\Ga$ is a normal cover of $\Ga_F$, we conclude that $|M_1|$ divides $|\ov H|$, a contradiction occurs. Thus $\C_{M_1}(\ov W)=M_1$, that is, $\ov W\leqslant\C_M(M_1)$. Repeating the above argument for $M_{i}/M_{i-1}$ and $\ov G M_{i-1}/M_{i-1}$, we obtain that $[M_i,\ov W]\subseteq M_{i-1}$ for $1\leqslant i\leqslant m$. It follows that $\ov W$ stabilizes the normal series of $\ov F$, and hence $\ov W$ centralizes $\ov F$, refer to [@Gorenstein p.178, Theorem 3.2]. It implies that $W$ induces the identity on $\ov F$, and so $W$ is trivial on $F$ (see [@Gorenstein p.174, Theorem 1.4]), namely, $W\leqslant\C_X(F)$, again against our assumption.
Assume now that $r=p$. Then $|F|\leqslant|W|$. Denote by $\Sigma$ the normal quotient graph $\Ga_F$. If $|F|=|W|$, then $W$ fixes each vertex of $\Sigma$, and hence $W\leqslant F$, which is impossible. Thus $|F|<|W|$. Set $\ov X=X/F$. Let $F_{\ov X}$ be the Fitting subgroup of $\ov X$. It is known that $F_{\ov X}$ is a $p'$-group. Let $\tilde G=GF/F\cong G$. Write $\tilde G=\tilde W{:}\tilde H$. For that case, we conclude $F_{\ov X}\cap \tilde G=1$. It follows that $\Ga$ is $(X,2)$-arc-transitive, and so $\Sigma$ is $(\ov X,2)$-arc-transitive. By [@CE Theorem 4.1], $\Sigma$ is a cover of $\Sigma_{F_{\ov X}}$ or $\Sigma_{F_{\ov X}}=K_2$. For the former, arguing as above, we also obtain $\tilde W\leqslant F_{\ov X}$, which contradicts $F_{\ov X}\cap \tilde G=1$. For the latter, we obtain $p=2$, and $|G|$ divides $2^53^6$. Since $G$ is a primitive Frobenius group, we have $G\cong\ZZ_2^2{:}\ZZ_3$. So $F\cong\ZZ_2$, and thus $F\leqslant\Z(X)$, again a contradiction. Therefore, $F=W$.
For a group $T$ and a prime $q$, by $T_q$ we mean a Sylow $q$-subgroup of $T$.
\[Lei\] Use the notation defined above, rewrite $\Ga=(V\Ga,E\Ga)$. Then we have:
- If $p$ is an odd prime, then either $G\cong\D_{2p}$ or $W\unlhd X$;
- If $p=2$, then $F=O_2(X)$, and
- $W<F$, $\Ga_F$ is a cycle, and $X=(F{:}H).\calO$, where $\calO=1$ or $\ZZ_2$;
- $W=F$ and $W\unlhd X$.
[*Proof. *]{}
Suppose that $G\ncong\D_{2p}$. We first claim that $W\leqslant F$ and $F\cap H=1$. Suppose, by way of contradiction, that $G\cap F=1$. Since $X=GX_1$, we obtain $|F|$ divides $|X_1|$. From Lemma $\ref{Lu}$, it follows that each prime divisor of $|F|$ is either $2$ or $3$.
Let $K$ be the kernel of $X$ acting on $\Ga_F$. Then $X/K\leqslant\Aut\Ga_F$. Recall that $p$ is a prime divisor of $|W|$. Suppose that $p>3$. Let $B$ be an orbit of $F$ acting on $V\Ga$. So $|B|$ divides $|F|$. If $G\cap K\not=1$, then $W\leqslant K$. Let $\bigtriangleup$ be an orbit of $W$ acting on $V\Ga$, which is contained in the block $B$. Then $|\bigtriangleup|$ divides $|B|$, which is impossible. So $G\cap K=1$. Let $\ov G=GK/K$. Then $\ov G\cong G$ is a Frobenius group. Write $\ov G=\ov W{:}\ov H\cong G$. If $\Ga_K$ is a cycle, then $d=1$, and so $\ov G\cong\D_{2p}$, against our assumption. Thus $\Ga$ is a cover of $\Ga_K$, and then $K=F$. By Lemma \[Lu\], $\ov G_B\leqslant \ov H$. In view of Lemma $\ref{un}$, $F$ is a $\{2,3\}$-group. So is $\ov G_B$ because $|\ov G_B|=|F|$. Note that $\Ga_F$ is $\ov G$-vertex-transitive. By [@Cai; @heng Lemma 2.1], we conclude that $\ov G_B^{\Ga(B)}$ is a cyclic group of order $2^i3^j$ for $i,j\geqslant1$. However, $\ov G_B^{\Ga(B)}$ is isomorphic to a subgroup of $\S_4$, which is a contradiction. Thus $p=2$ or $3$.
If $F=O_2(X)$, by Lemma $\ref{un}$, $\Ga_F$ is a cycle. So is $\Ga_K$. By the assumption, we conclude that $p=2$, and $W\leqslant K$. It follows that $K=W.K_1$, where $K_1$ is a $2$-group. Since $K\unlhd X$, we have $K\leqslant F$, which contradicts the fact that $F\cap G=1$.
If $F=O_3(X)$, then $\Ga$ is $(X,2)$-arc-transitive. From Lemma \[un\], it follows that $\Ga_F=K_2$. It implies that $G\cong\D_6$, against our assumption.
If $F=O_2(X)\times O_3(X)$, then $\Ga$ is $(X,2)$-arc-transitive. By [@CE Theorem 4.1], $\Ga$ is a cover of $\Ga_F$, $\Ga_F=K_2$ or $F$ is transitive on $V\Ga$. Assume first that $\Ga$ is a cover of $\Ga_F$. Let $Y=\Aut\Ga_F$ and $\ov X=X/F$. Then $\ov X$ is a subgroup of $Y$. Recall that $F\cap G=1$, we have $|\ov G_{B}|=|F|$, and hence $\ov G_{B}$ is a Frobenius group of $Y_{B}$. Since $\Ga_F$ is $\ov G$-vertex-transitive, we conclude that $\ov G_{B}^{\Ga_F(B)}=\A_4$, $\S_3$ or $\S_4$, refer to [@Cai; @heng Lemma 2.1]. Assume $\ov G_{B}^{\Ga_F(B)}=\A_4$ or $\S_4$. Then $\Ga_F$ is $(\ov G,2)$-arc-transitive. From Lemma \[Lu\], it follows that $\ov G_{B}=\A_4$, which implies that $|F|=12$. For this case, it is easy to show that $W\leqslant\C_X(F)\leqslant F$, a contradiction occurs. Assume $\ov G_B^{\Ga_F(B)}=\S_3$. Let $\ov G_B^{[1]}$ be the kernel of $\ov G_B$ acting on $\Ga_F(B)$. Recall that $\ov G_{B}$ is a Frobenius group, we conclude $\ov G_B^{[1]}$ is a $3$-group, and hence $O_2(X)\cong\ZZ_2$. It further implies that $W$ is a $3$-group. Let $\ov F$ be the Fitting subgroup of $\ov X$. Then $\ov F\cap \ov G=1$. It follows that $\ov F$ is a $2$-group. Since $|F.\ov F|$ divides $2^43^6$, we have $|\ov F|\leqslant 8$, and thus $\ov W\leqslant\C_{\ov X}(\ov F)$, again a contradiction.
Assume now that $\Ga_F=K_2$. Then $|V\Ga|$ divides $2^53^6$. Since $\Ga$ is a cover of $\Ga_{F_2}$ and $\Ga_{F_3}$, we conclude that $|F_2|=\frac{|G_2|}{2}$ or $|G_2|$, and $|F_3|=|G_3|$. When $p=3$, we have $|F_3|=|W|$. Note that $W$ is minimal and normal in $G$. Then $W$ fixes each vertex of $\Ga_{F_3}$, and thus $W\leqslant F$, which contradicts $G\cap F=1$. For $p=2$, and $|F_2|=|G_2|$, we also obtain the same contradiction. When $p=2$ and $|F_2|=\frac{|G_2|}{2}$. Since $G$ is a $\{2,3\}$-group, we conclude that $G\cong\ZZ_2^2{:}\ZZ_3$, and hence $|F|=6$. For that case, we easily obtain that $W\leqslant F$, again a contradiction. Similarly, we also exclude the case where $F$ is transitive on $V\Ga$.
Summarizing the above discussion, we obtain $W\leqslant F$. Since $G$ is a Frobenius group, we have $F\cap H=1$, as claimed. Next we process our analysis by several cases.
0.07in [**Case 1:**]{} If $p>3$, then $W\unlhd X$.
By the previous discussion, we have $W\leqslant F_p$. By Lemma \[Lu\], we conclude that $W=F_p$, and hence $W\unlhd X$.
0.07in [**Case 2:**]{} If $p=3$, then $W\unlhd X$.
If $W<F_3$, then $\Ga$ is $(X,2)$-arc transitive. For this case, $\Ga_{F_3}=K_2$, and so $G\cong\D_6$, contrary to our assumption. Thus $W=F_3$, and then $W\unlhd X$.
0.07in [**Case 3:**]{} If $p=2$, then either $\Ga_F$ is a cycle, or $W\unlhd X$.
Assume that $W<F_2$. Since $F\cap H=1$, we know that $|FH|=|F||H|$. Note that $|FH|$ divides $|X|$, it follows that $|F|$ divides $|W||X_1|$, and hence $|F_{2'}|$ divides $|X_1|$. So $F_{2'}$ is a $3$-group. If $F_{2'}\not=1$, then $\Ga$ is $(X,2)$-transitive, and thus it follows from [@CE Theorem 4.1] that $\Ga$ is a cover of $\Ga_{F_2}$, a contradiction. So $F_{2'}=1$. By Lemma \[un\], $\Ga_F$ is a cycle because $|H|$ is an odd number. Recall that $B$ is a vertex of $\Ga_F$. Since $\Ga$ is a Cayley graph of $G$, we obtain $W$ is regular on $B$. So $K=F=WK_1$. Consequently, $X=(F{:}H).\calO$ where $\calO\cong1$ or $\ZZ_2$. This completes the proof of Lemma \[Lei\].
For the group $G\cong\D_{2p}$ where $p$ is an odd prime. Applying [@Zai; @Ping; @Lu Theorem 1.1], we have the following conclusions.
\[D\_2p\] Let $G\cong\D_{2p}$, and let $\Ga$ be a connected edge-transitive tetravalent Cayley graph of $G$. Then we have
- $\Ga$ is arc-regular, and $\Aut\Ga\cong\D_{2p}{:}\ZZ_4$;
- $\Ga\cong\C_{p[2]}$, and $\Aut\Ga\cong\ZZ_2^p{:}\D_{2p}$.
In the remainder of this section assume that $G\not\cong \D_{2p}$ with $p$ an odd prime, unless specified otherwise.
0.01in
Recall that the [*socle*]{} of a finite group $R$ (denoted by $\soc(R)$) is the product of all minimal normal subgroups of $R$. Evidently, $\soc(R)$ is a characteristic subgroup of $R$.
0.01in
We next treat the case where $W\unlhd X$, and the normal quotient $\Ga_W$ is a cycle.
\[Lee\] Let $K$ be the kernel of $X$ acting on $\Ga_W$. Then the following statements hold:
- $X=((WK_1){:}H).\calO$, and $W\cong\ZZ_2^d$, where $\calO\cong1$ or $\ZZ_2$;
- Assume $p$ is an odd prime. Then we have
- $G$ is normal in $ X$, or
- $G$ is not normal in $X$, and
- $X=W{:}((K_1{:}H).\calO)$, and $H$ does not centralise $K_1$ where $K_1\cong\ZZ_2^l$ with $2\leqslant l\leqslant d$, and $\calO\cong1$ or $\ZZ_2$;
- there exist $x_1,x_2,\ldots,x_d\in W$ and $\t_1,\t_2,\ldots,\t_d\in K_1$ such that $W=\l x_1,\ldots,x_d\r$, $\l x_i,\t_i\r\cong\D_{2p}$, and $K_1=\l\t_i\r\times\C_{K_1}(x_i)$ for $1\leqslant i\leqslant d$;
- $\soc(X)=W\times L$, where $L\cong 1$ or $\ZZ_2$;
- $H$ is imprimitive on $W$.
[*Proof. *]{}
Let $B$ be a vertex of $\Ga_W$. Since $\Ga$ is a Cayley graph of $G$, we obtain $W$ is regular on $B$. Thus $K=WK_1$, and $K\cap H=1$, where $K_1$ is a $2$-group. For that case, $\Ga_W$ is a connected Cayley graph. Recall that $H$ is of order $n$, $\Ga_W$ is a cycle of size $n$, say. It follows that $X/K\cong\ZZ_n$ or $\D_{2n}$. Further, $\Ga$ is $X$-arc-transitive if and only if $X/K\cong\D_{2n}$.
Assume first that $p=2$. Since $G\leqslant X$ and $(|K|,|H|)=1$, we conclude that $K{:}H\leqslant X$. Noting that $X/K$ is isomorphic to a subgroup of $\D_{2n}$, it follows that $X=(K{:}H).\calO$ with $\calO\cong1$ or $\ZZ_2$, so we have part (i).
Assume now that $p$ is an odd prime. Furthermore, we assume that $G$ is not normal in $X$. If $K_1=1$, then $K=W$, and hence $G\lhd X$, which contradicts the assumption. Thus $K_1\not=1$.
Let $U=\N_X(K_1)$. Since $K_1\ntrianglelefteq X$, it follows that $U\not=X$. Noticing that $(|W|, |K_1|) = 1$, we obtain that $\N_{X/W}(K/W) =\N_{X/W}(WK_1/W) =\N_X(K_1)W/W = UW/W$. As $K/W\unlhd X/W$, implying that $X = WU$. Since $W\lhd X$, we have that $W\cap U\lhd U$. Furthermore, $W\cap U\lhd W$ since $W$ is abelian. Then $W\cap U\lhd\l U,W\r= UW=X$. If $W\leqslant U$, then $K=WK_1=W\times K_1$, and hence $K_1\lhd X$, which is impossible. Thus $W\cap U<W$. Furthermore, note that $W$ is a minimal normal subgroup of $X$, we obtain that $W\cap U=1$, and so $K\cap U=WK_1\cap U=(W\cap U)K_1=K_1$. Now $X/K=UW/K=UK/K\cong U/(K\cap U)=U/K_1$, and hence $U=(K_1.\hat H).\calO$, where $\hat H\cong\ZZ_n$ and $\calO\cong1$ or $\ZZ_2$. Noting that $G$ belongs to $X$ and $G=W{:}H$, there exists some $H^z\leqslant U$ such that $G=W{:}H^z$ is regular on $V\Ga$, where $z\in W$. Without loss of generality, we may assume that $U=(K_1{:}H).\calO$. Then $X_1=K_1.\calO$. Furthermore, since $G$ is not normal in $X$, we conclude that $H$ does not centralise $K_1$.
Set $Y=W{:}(K_1{:}H)$. Then $Y$ has index at most $2$ in $X$, and $\Ga$ is $Y$-edge-transitive. It is obvious that $\Ga$ is not $Y$-arc-transitive. Hence $\Ga=\Cos(Y,K_1,K_1\{y, y^{-1}\}K_1)$, where $y\in Y$ is such that $\l K_1,y\r=Y$ and $K_1\cap K_1^y$ has index $2$ in $K_1$. We may choose $y\in W{:}H=G$ such that $H=\l h\r$ and $y=hx$ where $x\in W$. Then $K_1\cap K_1^y=K_1\cap K_1^x$ has index $2$ in $K_1$.
We claim that $K_1\cap K_1^x=\C_{K_1}(x)$. For any $\sigma\in K_1\cap K_1^x$, we have that $\sigma^{x^{-1}}\in K_1$, and hence $\sigma^{-1}\sigma^{x^{-1}}\in K_1$. Since $x\in W$ and $W\lhd WK_1$, we obtain that $\sigma^{-1}\sigma^{x^{-1}}=(\sigma^{-1}x\sigma)x^{-1}\in W$. So $\sigma^{-1}\sigma^{x^{-1}}\in W\cap K_1=1$, and then $\sigma^{x^{-1}}=\sigma$. Thus $\sigma$ centralises $x$. It follows that $K_1\cap K_1^x\leqslant \C_{K_1}(x)$. Clearly, $\C_{K_1}(x)\leqslant K_1\cap K_1^x$. So $\C_{K_1}(x)=K_1\cap K_1^x$ as required.
Recall that $W$ is a minimal normal subgroup of $X$ and $X=WU$, we obtain that $W=\l x\r\times\l x^{\sigma_2}\r\times\cdots\times\l x^{\sigma_d}\r$ where $\sigma_i\in U$. Then $\C_{K_1}(x^{\sigma_i})=(\C_{K_1}(x))^{\sigma_i}<K_1^{\sigma_i}=K_1$. The intersection $\cap_{i=1}^d\C_{K_1}(x^{\sigma_i})\leqslant\C_K(W)=W$, and hence $\cap_{i=1}^d\C_{K_1}(x^{\sigma_i})=1$. Since each $\C_{K_1}(x^{\sigma_i})$ is a maximal subgroup of $K_1$, the Frattini subgroup $\Phi(K_1)\leqslant\cap_{i=1}^d\C_{K_1}(x^{\sigma_i})=1$. Hence $K_1$ is an elementary abelian $2$-group, that is, $K_1\cong\ZZ_2^l$ for some $l\geqslant1$. Recall that $\cap_{i=1}^d\C_{K_1}(x^{\sigma_i})=1$, it follows that $l\leqslant d$. Assume that $l=1$. Then $K_1\cong\ZZ_2$ and so $K_1\leqslant\C_X(H)$. Thus $G\lhd X$, which contradicts the fact that $G$ is not normal in $X$. Hence $l>1$, as in part (a).
Since $\C_{K_1}(x)$ has index $2$ in $K_1$, there exists some $\tau_1$ belonging to $ K_1$ such that $K_1=\l \tau_1\r\times\C_{K_1}(x)$. Set $x_1=x^{-1}x^{\tau_1}$. Then $x_1\not=1$, $x_1^{\tau_1}=x_1^{-1}$ and $\C_{K_1}(x)=\C_{K_1}(x_1)$, and hence $K_1=\l \tau_1\r\times\C_{K_1}(x_1)$. Noticing that $W$ is a minimal normal subgroup of $X=WU$, there exist $\mu_1=1,\mu_2,\ldots,\mu_d\in U$ such that $W=\l x^{\mu_1}\r\times \l x^{\mu_2}\r\times\cdots\times\l x^{\mu_d}\r$. Let $x_i=x_1^{\mu_i}$ and $\tau_i=\tau_1^{\mu_i}$, where $i=1,\ldots,d$. Then $\ZZ_2^{l-1}\cong(\C_{K_1}(x_1))^{\mu_i}=\C_{K_1^{\mu_i}}(x_1^{\mu_i})=\C_{K_1}(x_i)$, and $K_1=K_1^{\mu_i}=\l\tau_i\r\times\C_{K_1}(x_i)$. Furthermore, $x_i^{\tau_i}=x_1^{\tau_1\mu_i}=(x_1^{-1})^{\mu_i}=x_i^{-1}$, and so $\l x_i,\tau_i\r\cong\D_{2p}$, as in part (b).
Recall that $W\cong\ZZ_p^d$ for an odd prime $p$. Since $G$ is not normal in $X$, we conclude that $d>1$. Assume that $X$ has a minimal normal subgroup $L\not=W$. Then $W\cap L=1$, and $LK/K\lhd X/K\leqslant\D_{2n}$. It follows that either $L\leqslant K$, or $L\cap K=1$. If $L\leqslant K$, then $L$ is a $2$-group. Since $K_1$ is a Sylow $2$-subgroup of $K$, there exists some $w\in W$ such that $L^w\leqslant K_1$. It follows that $L\unlhd K_1$, and then $L=1$, which is impossible. Hence $L\cap K=1$, and so $L\leqslant K_1H$, and $L\cong\ZZ_2$. Thus $\soc(X)=W\times L$, as in part (c).
By the above paragraph, we obtain that $\C_{X}(W)=W\times L$. Let $\ov X=X/L$, and $\ov G=GL/L\cong G$. Let $\ov K_1=K_1L/L\cong K_1$. Write $\ov G=\ov W{:}\ov H$. Since $H$ normalizes $K_1$, we conclude that $\ov H$ normalizes $\ov K_1$. Note that $\ov K_1 \ov H$ acts irreducibly and faithfully on $\ov W$. By Clifford’s Theorem, $\ov W$ can be decomposed as $\ov W=e(U_1\oplus U_2\oplus\cdots\oplus U_t)$ such that $\ov K_1$ normalises each $U_i$, and all $U_i$ are pairwise non-equivalent and irreducible with respect to the action of $\ov K_1$. Recall that $K_1$ is of order at least $4$. It implies that $t\geqslant 2$. Let $V_i=eU_i$ for each $i$. Rewrite $\ov W=V_1\oplus V_2\oplus\cdots\oplus V_t$. Now $\ov H$ normalises $\ov K_1$, we conclude that $\ov H$ preserves such decomposition. Since the maximal subgroup preserving such decomposition in $\GL(\ov W)$ is $\GL(V_1)\wr\S_t$, implying that $\ov H$ belongs to $\GL(V_1)\wr\S_t$, forcing $\ov H$ is imprimitive on $\ov W$. By [@DF Proposition 2.8], we are done, as in part (d).
We now determine the graph $\Ga$ for the case where $W\unlhd X$, and $\Ga$ is a normal cover of $\Ga_W$.
\[Le\] Assume that $\Ga$ is a normal cover of $\Ga_W$. Then we have
- $G$ is normal in $X$, or
- $G$ is not normal in $X$, and
- $\Ga_W\cong\C_{\frac{n}{2}[2]}$, and $n\equiv 0\,(\mod 4)$;
- $X=W{:}((NH).\calO)$, $X_1\leqslant N.\calO$, $N\cap H\cong\ZZ_2$, and $H$ normalizes $N$, but $H$ does not centralise $N$, where $N\cong\ZZ_2^l$ with $2\leqslant l\leqslant \frac{n}{2}$, and $\calO\cong1$ or $\ZZ_2$;
- $W$ is unique and minimal in $X$, and $H$ is imprimitive on $W$;
- $X/(WN)\cong\ZZ_{\frac{n}{2}}$ or $\D_n$, and $\Ga$ is $X$-arc-transitive if and only if $X/(WN)\cong\D_n$.
[*Proof. *]{}
Let $\ov H=G/W$ with $\ov H=\l\ov h\r$. Since $\Ga$ is a Cayley graph of $G$, $\Ga_W$ is a Cayley graph of $\ov H$. By [@Baik Theorem 1.2], either $\Ga_W$ is a normal Cayley graph or $\Ga_W=\C_{{\frac{n}{2}}[2]}$, and $4$ divides $n$. It follows that either $G$ is normal in $X$ or $\Ga_W\cong\C_{{\frac{n}{2}}[2]}$. Suppose that $G$ is not normal in $X$. Then $\Ga_W\cong\C_{{\frac{n}{2}}[2]}$, as in part (a).
Clearly, $\Aut\Ga_W\cong\ZZ_2^{\frac{n}{2}}{:}\D_n$. Let $\ov K\unlhd \Aut\Ga_W$ such that $\ov K\cong\ZZ_2^{\frac{n}{2}}$. Then we may write $\Aut\Ga_W=\ov K~\ov HO$, where $O\cong\ZZ_2$. Let $B$ be a vertex of $\Ga_W$, and $1\in B$. Choose $\ov M\leqslant \ov K$ such that $M\cong\ZZ_2^{\frac{n}{2}-1}$ and $(\Aut\Ga_W)_B=\ov MO$.
Let $\ov X=X/W$. Since $\ov X~\ov K/\ov K\cong \ov H\calO/(\ov H\calO\cap \ov K)$ where $\calO=1$ or $O$, we conclude that $\ov X=(\ov X\cap \ov K)\ov H\calO$, and $\Ga$ is $X$-arc-transitive if and only if $\calO=O$. Let $\hat K=\ov X\cap \ov K$. Then $\hat K\unlhd\ov X$, and $\hat K\cap \ov H\cong\ZZ_2$. Thus $X=W.((\hat K\ov H).\calO)$. Let $K$ be the preimage of $\hat K$, under $X\to X/W$. Note that $G$ is a Frobenius group, we conclude that the order of $W$ is odd. By Hall’s Theorem, $K=W{:}N$, where $N\cong \hat K$. It further implies that $N\cong\ZZ_2^l$, where $l\leqslant \frac{n}{2}$.
Now $(|N|,|W|)=1$, we get $X/W=\N_{X/W}(NW/W)=\N_X(N)W/W$, and so $X=W\N_X(N)$. Since $H\leqslant X$, it follows that $H^w$ belongs to $\N_X(N)$ for some $w\in W$. Without loss of generality, we may assume that $H$ belongs to $\N_X(N)$. Thus $X=W{:}((NH).\calO)$. By comparing the order, we conclude that $N\cap H\cong\ZZ_2$. If $l=1$, then $NH=H$, and so $G\unlhd X$, which contradicts the assumption that $G$ is not normal in $X$. Then $l\geqslant 2$. Thus $2\leqslant l\leqslant \frac{n}{2}$.
Set $Y=W{:}(NH)$. Clearly, $G$ is a subgroup of $Y$. Then $Y=GY_1$. Note that $|Y|=\frac{|W||H||N|}{|H\cap N|}$, and $|Y|=|G||Y_1|$, we obtain $|Y_1|=\frac{|N|}{|N\cap H|}=\frac{|N|}{2}$. Let $\ov Y=Y/W$. Since $Y_1W/W=\ov Y_B$, we have $Y_1W/W\leqslant \ov Y\cap \ov M\leqslant\hat K$, and hence $Y_1W/W\leqslant NW/W$. Consequently, $Y_1\leqslant N^{\ov w}$ for some $\ov w\in W$. For simplicity, we may assume that $Y_1\leqslant N$. For that case, $Y_1$ has index $2$ in $N$, and hence $X_1\leqslant N.\calO$, as in part (b).
Let $C{:}=\C_{NH}(W)$. Assume that $C\not=1$. Clearly, $C$ is normal in $Y$. Without loss of generality, $C$ is minimal in $Y$. Since $H$ acts fixed-point-freely on $W$, we have $C\cap H=1$. Let $\ov C$ be the image of $C$ under $X\to X/W$. Then $\ov C$ is normal and minimal in $\ov Y$, and hence $\ov C$ is a subgroup of $\hat K$. It implies that $C\cong\ZZ_2^{\ell}$ for some $\ell$.
Let $\ov K=\prod_{i=1}^{\frac{n}{2}}\l\s_i\r$. Note that $\ov H$ acts on $\ov K$ by permuting transitively on all $\s_i$. Relabeling if necessary, we may assume $\ov h=\s\pi$, where $\s\in\ov K$, and $\pi=(12\cdots\frac{n}{2})^{-1}$. Let $\ov K_B=\prod_{i\not=1}\l\s_i\r$. Choose $\ov B,\tilde B\in\Ga(B)$ such that $\ov K_{\ov B}=\prod_{i\not=2}\l\s_i\r$ and $\ov K_{\tilde B}=\prod_{i\not=\frac{n}{2}}\l\s_i\r$. Pick some $x\in C$ such that $x=\s_{i_1}\cdots\s_{i_k}$, where $i_1=2$, and $2\leqslant i_j\leqslant i_{j+1}\leqslant\frac{n}{2}$. Then $\l x,x^{\ov h^{\frac{n}{2}-i_k}}\r\leqslant \ov K_B\cap C$. It follows that $\Ga_W$ is $\ov C{:}\ov H$-edge-transitive Cayley graph of valency $4$.
Let $Z=(W\times C){:}H$. By the above paragraph, $\Ga$ is $Z$-edge-transitive. However, $\Ga$ is not $Z$-arc-transitive. By Lemma \[Cos\], $\Ga=\Cos(Z,Z_1,Z_1\{g,g^{-1}\}Z_1)$. It is obvious that $Z_1\cong\ZZ_2^\ell$. Now we may assume $Z_1=\l\t_1,\t_2,\ldots,\t_{\ell}\r$. Write $h=\l(h_1,h_2)\r$. If $\t_i=(1,c)$ for some $i$, since $C$ is minimal in $Z$, we conclude that $Z_1=C$, which is impossible. Thus each $\t_i$ has the form $(1,c)h^{\frac{n}{2}}$ or $(u,c)h^{\frac{n}{2}}$, where $u\in W\setminus\{1\}$, and $c\in C$. Note that $n$ is divisible by $4$. Then we may write $g$ as $(v,c')h$, where $v\in W$, and $c'\in C$. Assume $\t_{i_0}=(u_0,c)h^{\frac{n}{2}}$ for some $i_0$, where $u_0\not=1$. Then all $\t_i$ have the form $(u_0,c')h^{\frac{n}{2}}$, where $c'\in C$. If $\t_i^g=\t_j$ for some two $i,j$, it is easy to show that $\l Z_1,g\r$ is a subgroup of $C{:}\l g\r$, which is a contradiction. Thus all $\t_i^{g}$ do not belong to $Z_1$, which leads to the valency of $\Ga$ is greater than $4$, again a contradiction. Similarly, we also exclude the other case. Thus $C=1$, namely, $NH$ acts faithfully and irreducibly on $W$. That is to say, $W$ is the unique minimal normal subgroup of $X$. Arguing similarly as Lemma \[Lee\], we obtain that $H$ is imprimitive on $W$, as in part (c).
Clearly, $\Ga_{WN}$ is a cycle. Since $X/(WN)$ is transitive on $V\Ga_{WN}$, we conclude that $X/(WN)\cong\ZZ_{\frac{n}{2}}$ or $\D_n$. For that case, $\Ga$ is arc-transitive if and only if $X/(WN)\cong\D_n$, as in part (d).
With the above preparation, we are ready to embark on the proof of Theorem \[soluble\].
0.1in [**Proof of Theorem \[soluble\]:**]{} If $G\lhd X$, then by Lemma \[G\], we have $X_1\leqslant \D_8$, as in Theorem \[soluble\] (1). In what follows, we assume that $G$ is not normal in $X$.
Assume first that $p$ is an odd prime. By Lemmas \[D\_2p\]-\[Le\], if $W$ is not normal in $X$, we obtain that $\Ga\cong\C_{p[2]}$, and $\Aut\Ga\cong\ZZ_2^p{:}\D_{2p}$, as in Theorem \[soluble\] (2). If $W$ is normal in $X$, and $\Ga_W$ is a cycle, by Lemma \[Lee\], part (3) of Theorem \[soluble\] occurs. If $W$ is normal in $X$, and $\Ga$ is a cover of $\Ga_W$, from Lemma \[Le\], it follows that part (4) of Theorem \[soluble\] holds.
Assume now that $p=2$. By Lemmas \[Lei\] and \[Lee\], Theorem \[soluble\](5) occurs.
Insoluble automorphism groups
=============================
Let $G=W{:}H\cong\ZZ_p^d{:}\ZZ_n$ be a primitive Frobenius group. Assume that $\Ga=(V\Ga,E\Ga)$ is a connected $X$-edge-transitive tetravalent Cayley graph of $G$, where $G\leqslant X\leqslant\Aut\Ga$. In this section, we study the case where the automorphism group $X$ is insoluble.
For a finite group $R$, the socle of $R$, denoted by $\soc(R)$, is the subgroup generated by all minimal normal subgroups of $R$. The group $R$ is said to be almost simple if its socle $\soc(R)$ is a non-abelian simple group.
0.1in
**TABLE 2: [Almost simple automorphism groups.]{}**
$$\begin{array}{|c|c|c|c|c|}\hline
X& G & X_1 \\ \hline
\PSL(3,3){:}\ZZ_2 &\D_{26} & \ZZ_3^2{:}\GL(2,3)\\ \hline
\PGL(2,7) & \D_{14} & \S_4 \\\hline
\PGL(2,7) & \ZZ_7{:}\ZZ_3 & \D_{16} \\\hline
\PGL(2,7) & \ZZ_7{:}\ZZ_6 & \D_8 \\\hline
\PSL(2,23) &\ZZ_{23}{:}\ZZ_{11} & \S_4\\ \hline
\PSL(2,11) &\ZZ_{11}{:}\ZZ_{5} & \A_4 \\ \hline
\PGL(2,11) &\ZZ_{11}{:}\ZZ_{5} & \S_4 \\ \hline
\PGL(2,11) &\ZZ_{11}{:}\ZZ_{10} & \A_4\\ \hline
\end{array}$$
0.1in We now determine the structure of insoluble group $X$. Denote by $R(X)$ the maximal solvable normal subgroup of $X$. We first treat the case where $R(X)=1$.
\[Z(X)\] Let $N$ be minimal and normal in $X$. If $R(X)=1$, then $\C_X(N)=1$.
[*Proof. *]{}
Note that $N$ is minimal in $X$. Since $R(X)=1$, we have $N\cong T^k$, where $T$ is a nonabelian simple group, and $k$ is an integer. Clearly, $\Z(N)=1$. Let $C{:}=\C_X(N)$. Since $N\unlhd X$, we have $C\unlhd X$. Suppose that $C\not=1$. By our assumption, $C$ is insoluble. Notice that $N\cap G\unlhd G$, we conclude that $N\cap G=1$ or $W\leqslant N\cap G$. For the former, $|N|$ divides $|X_1|$, and so $N$ is soluble, contrary to our assumption. Thus $W\leqslant N\cap G$. Similarly, $W\leqslant C\cap G$. It follows that $W\leqslant N\cap C$, a contradiction. Thus $C=1$.
\[X\] If $R(X)=1$, then $X$ is almost simple.
[*Proof. *]{}
By Frattini argument, we have that $X=GX_u$, where $u\in V\Ga$. By Lemma $\ref{Lu}$, either $X_u$ is a $2$-group or $|X_u|$ divides $2^43^6$. Let $N$ be a minimal normal subgroup of $X$. By our assumption, $N$ is unsolvable. So $N=T_1\times T_2\times\cdots\times T_k$, where $T_i\cong T$ is a nonabelian simple group for any $i$. By [@L.Kazarin], we obtain that $T$ is one of the following: $$\PSL(2,q)(q>3),\, \PSL(3,q)(q<9),\,\PSL(4,2),\, \PSp(4,3),\ \PSU(3,8),\ \mbox{or}\ \M_{11}.$$
In what follows, suppose that $k\geqslant2$. Since $W$ is minimal and normal in $G$ and $N\cap G\not=1$, we conclude that $W\leqslant N$. Let $r>3$ be a prime divisor of $|T|$. Since $r$ divides $|X|$ and $(|W|,|H|)=1$, we conclude that $r$ divides either $|W|$ or $|H|$. Suppose first that $r$ divides $|W|$. Then $T_i\cap W\not=1$ for each $i$. Let $W_i=T_i\cap W$ for $1\leqslant i\leqslant k$. Assume that $N\cap H=1$. Then $\frac{|N|}{|W|}$ divides $|X_u|$. So does $\prod_{i=1}^k\frac{|T_i|}{|W_i|}$. An inspection of the above simple groups shows $T=\A_5$ and $k=2$. By Lemma \[Z(X)\], $X\lesssim\S_5\wr\ZZ_2$. For this case, $\frac{|N|}{|W|}=144$. By Lemma \[Lu\], the only possibility is that $X\cong\A_5\times\A_5$. Clearly, this is a contradiction.
Thus $N\cap H\not=1$. Let $\ov H=N\cap H$. Since $G$ is a Frobenius group, it follows that $\ov H$ is a diagonal subgroup of $N$. Write $\ov H=\l\s_1\s_2\cdots\s_k\r$ where $\l \s_i\r\cong\ov H$, and $\s_i\in T_i$ for each $i$. Let $H_i=\l\s_i\r$ where $1\leqslant i\leqslant k$. Then $G_i=W_i{:}H_i$ is a Frobenius group. For this case, we obtain that $\frac{|N|}{|W||\ov H|}$ divides $|X_u|$.
Let $T=\PSL(2,q)$ where $q>3$. By [@Suzuki Theorem 6.25], we conclude that $ G_i\leqslant [q]{:}[\frac{q-1}{d}]$, $G_i\leqslant \D_{\frac{2(q\pm1)}{d}}$, $G_i\leqslant\A_5$, or $G_i\leqslant\PGL(2,r)$, where $d=(2,q-1)$, and $r\div q$. Suppose $ G_i\leqslant [q]{:}[\frac{q-1}{d}]$. Then $W{:}\ov H\leqslant [q]^k{:}[\frac{q-1}{d}]$, namely, $N\cap G\leqslant [q]^k{:}[\frac{q-1}{d}]$. Since $|N{:}N\cap G|$ divides $|X_u|$, we conclude that $\frac{d|N|}{q^k(q-1)}$ divides $|X_u|$. If $q$ is even, then $(q+1)^k(q-1)^{k-1}\div2^43^6$, which is a contradiction because $q+1$ and $q-1$ are two distinct odd numbers. If $q$ is odd, then $(q+1)^k(\frac{q-1}{2})^{k-1}\div 2^43^6$. By easy calculations, $q=5$ and $k=2$. For this case, the only possibility is that $G\cong\ZZ_5^2{:}\ZZ_8$ and $X\cong\S_5\wr\ZZ_2$. By Lemma \[Lu\], we have that $X_u\cong\S_3\times\S_4$. Without loss of generality, we may assume that $X=(\S_5\times\S_5){:}\l \s\r$, where $\s$ permutes the first and second coordinates. Note that $G\cap X_u=1$. By MAGMA [@Magma], there is an element $G$ (up to conjugate) in $X$, and there are two elements $X_u$ (up to conjugate) in $X$ such that their intersections equal to 1. Choose $w,h\in \S_5$ such that $o(w)=5$, $o(h)=4$ and $w^h=w^2$. For this case, write $G=W{:}H$ with $W=\l(w,1),(1,w)\r$ and $H=(h,1)\s$. Meanwhile, we choose $X_u=\l((123),1),((12),1),(1,(1234),(1,(12)))$ or $\l((123),1),((12)(45),1),(1,(1234)),(1,(12))\r$. It is simple to show that, for the above two choices, $X_u$ belongs to different conjugate classes of $X$, and $X_u\cap G=1$. Choose $v\in\Ga(u)$. By Lemma \[Cos\], write $\Ga=\Cos(X,X_u,X_uoX_u)$, where $o\in\N_X(X_{uv})\setminus X_{u}$ and $o^2\in X_{uv}$. Since $\Ga$ is $X$-arc-transitive graph, we conclude $|X_u{:}X_{uv}|=4$, and hence $|X_{uv}|=36$. In such two cases, again by MAGMA [@Magma], there is no $o\in\N_X(X_{uv})$ such that $\l X_u,o\r=X$, namely, $\Ga$ is not connected. Suppose that $G_i\leqslant\D_{\frac{2(q\pm1)}{d}}$ or $G_i\leqslant\A_5$ where $1\leqslant i\leqslant k$. Arguing similarly as above, we conclude that $q=4$, $k=2$, and $G_i\cong\D_{10}$ for each $i$. For this case, the only possibility is that $G\cong\ZZ_5^2{:}\ZZ_8$ and $X\cong\S_5\wr\ZZ_2$, which is impossible by the above discussion. Thus $G_i\leqslant \PGL(2,r)$. Write $q=p^n$ where $p$ is a prime, and $n$ is a natural number. Then $r=p^m$, where $m\div n$. Let $n=ms$ with $s\geqslant2$. Note that $|N{:}N\cap G|$ divides $|X_u|$. So does $|T{:}\PGL(2,r)|^k$. It follows that $\frac{1}{2}p^{m(s-1)}(\Sigma_{i=1}^sp^{m(s-i)})(\Sigma_{j=1}^sp^{2m(s-j)})$ divides $2^43^6$. By easy calculations, there are no $p,n$ and $m$ satisfying the above equation. Thus this case is excluded.
Let $T=\PSL(3,q)$ with $q<9$. Assume that $q=2$. By Atlas [@atlas], we have $G_i\cong\ZZ_7{:}\ZZ_3$ where $1\leqslant i\leqslant k$. Clearly, $\frac{|N|}{|W||\ov H|}$ does not divide $|X_u|$. Similarly, we can exclude the other cases. Let $T=\PSL(4,2)$. Again by Atlas [@atlas], we conclude that $35\notdiv|G_i|$ where $1\leqslant i\leqslant k$. It implies that $5$ or $7$ divides $|X_u|$, which is impossible. Arguing similarly as above, we can conclude $T$ can not equal to other simple groups.
Suppose now that $r$ divides $|H|$. If $r\notdiv |H_i|$, then $r$ divides $|X_u|$, which is impossible. Thus $r$ divides $ |H_i|$ for each $i$. Recall that $\frac{|N|}{|W||\ov H|}$ divides $|X_u|$, we conclude that $r$ divides $|X_u|$, again a contradiction. Therefore, $X$ is almost simple.
\[ll\] Let $X$ be an almost simple group with $\soc(X)=\PSL(2,7)$. Assume that $\Ga$ is not $(X,2)$-arc-transitive. Then either $X=\PGL(2,7)$, $X_1=\D_8$ and $G=\ZZ_7{:}\ZZ_6$, or $X=\PGL(2,7)$, $X_1=\D_{16}$ and $G=\ZZ_7{:}\ZZ_3$.
[*Proof. *]{}
By Frattini argument, we have that $X=GX_u$, where $u\in V\Ga$. Since $\Ga$ is not $(X,2)$-arc-transitive, it follows that $X_u$ is a $2$-group. Note that $G$ is a Frobenius group. Checking the subgroups of $\PGL(2,7)$ in the Atlas [@atlas], we obtain $G=\ZZ_7{:}\ZZ_6$, or $\ZZ_7{:}\ZZ_3$.
Denote by $T$ the socle $\soc(X)$. Assume first that $G=\ZZ_7{:}\ZZ_6$. Since $\ZZ_7{:}\ZZ_3$ is maximal in $T$, we have $X=\PGL(2,7)$. It follows that $X_u\cong\D_8$. Assume now that $G=\ZZ_7{:}\ZZ_3$. Furthermore, assume that $X=\PSL(2,7)$. Then $\Ga$ is a connected tetravalent $X$-edge-transitive Cayley graph, and $X_u\cong\D_8$ is a Sylow $2$-subgroup of $X$. Choose $v\in\Ga(u)$. Then $|X_u{:}X_{uv}|=2$ or $4$. Since $\Ga$ is vertex transitive graph, we write $\Ga$ as coset graph $\Cos(X,H,H\{x,x^{-1}\}H)$, where $H=X_u\cong\D_8$, and $x\in X$ is such that $\langle H, x\rangle=X$; in particular, $x\notin H$.
Suppose that $|X_u{:}X_{uv}|=4$. Then $\Ga$ is $X$-arc-transitive. By Lemma \[Cos\], we choose $x$ such that $(u,v)^x=(v,u)$, resulting $x\in \N_X(X_{uv})\cong\D_8$. In particular, $\N_X(X_{uv})\not=X_u$. Then $|\N_{X_u}(X_{uv})|=4$. Hence $\N_{X_u}(X_{uv})$ is normal in both $\N_{X}(X_{uv})$ and $X_u$, and so $\N_{X_u}(X_{uv})\unlhd\langle X_u,\N_{X}(X_{uv})\rangle$. Checking the subgroups of $\PSL(2,7)$, we obtain that $\langle X_u,\N_{X}(X_{uv})\rangle\cong\S_4$, which contradicts the fact $\langle X_u, x\rangle=X$.
Suppose that $|X_u{:}X_{uv}|=2$. Then $|X_{uv}|=4$, and hence $X_{uv}\unlhd M=
\langle X_u, X_v\rangle$, so $M\cong\S_4$. By Lemma \[Cos\], we may choose $x$ such that $u^x=v$. It is clear that $X_u$ and $X_v$ are two Sylow $2$-subgroup of $M$, there exists some $y\in M$ such that $X_u^y=X_v=X_u^x$. Hence $xy^{-1}\in\N_X(X_u)=X_u$, so $\langle X_u, x\rangle\leqslant\langle X_u, xy^{-1}, y\rangle\leqslant M$, again a contradiction. Thus $X=\PGL(2,7)$.
Lemma \[X\] tells us that if $X$ is insoluble and $R(X)=1$, then $X$ is almost simple. The next two lemmas determine $\Ga$ for the case where $X$ is almost simple.
\[al\] If $X$ is almost simple, then, for $u\in V\Ga$, the triple $(X, G, X_u)$ is one of the triples listed in Table $2$.
[*Proof. *]{}
By Frattini argument, we have that $X=GX_u$. Since $\Ga$ is a connected Cayley graph of valency $4$, we obtain that $X_u$ is a $\{2,3\}$-group. It follows that $X$ is decomposed as a product of two solvable subgroups. Denote by $T$ the socle $\soc(X)$. By [@L.Kazarin], we conclude that $T$ appears in Lemma $\ref{X}$. In what follows, we process our analysis by two cases.
0.07in [**Case 1:**]{} Assume that $\Ga$ is $(X,2)$-arc-transitive. By Lemma $\ref{Lu}$, $|X_u|$ divides $2^43^6$.
Assume first that $T=\PSL(2,q)$ with $q>3$. If $q=5$, then $G=\ZZ_5$ and $X_u=\A_4$ or $\S_4$, a contradiction. If $q=7$, by the Atlas [@atlas], we conclude that $X=\PGL(2,7)$, $G=\D_{14}$, and $X_u=\S_4$. Suppose $q=11$. Checking the subgroups of $\PGL(2,11)$ in the Atlas [@atlas], we know that $X=\PSL(2,11).\calO$, $G=\ZZ_{11}{:}(\ZZ_5\times\calO_1)$, and $X_u=\A_4.\calO_2$, where $\calO=\ZZ_2$, and $\calO_1\calO_2=\calO$, (refer to [@Maru Theorem 1.5]). Suppose that $q=23$. By [@Zai; @Ping; @Lu Theorem 1.1], $X=\PSL(2,23)$, $G=\ZZ_{23}{:}\ZZ_{11}$ and $X_u=\S_4$.
In what follows, we assume that $q\not=4,5,7,11,23$. Then, by [@Xia Proposition 4.1], interchanging $G$ and $X_u$ if necessary, $G\cap T\leqslant\D_{2(q+1)/d}$ and $[q]\unlhd T\cap X_u\leqslant [q]{:}[\frac{q-1}{d}]$ where $d=(2,q-1)$. Let $T_u=T\cap X_u$. Assume that $G\cap T\leqslant\D_{2(q+1)/d}$, and $[q]\unlhd T_u\leqslant [q]{:}[\frac{q-1}{d}]$. Then $\frac{q(q-1)}{2}$ divides $|T{:}T\cap G|$. Since $|T{:}T\cap G|=|TG{:}G|$ divides $|X_u|$, we conclude that $\frac{q(q-1)}{2}\div 2^43^6$. By easy calculations, $q=9$. It follows that $T_u^{\Ga(u)}\cong\ZZ_3$ or $\S_3$. However, $T_u^{\Ga(u)}\unlhd X_u^{\Ga(u)}\leqslant \S_4$, and so $X_u^{\Ga(u)}\cong \S_3$, which is a contradiction because $X_u^{\Ga(u)}$ is $2$-transitive on $\Ga(u)$. Thus $[q]\leqslant T\cap G\leqslant [q]{:}[\frac{q-1}{d}]$ and $T_u\leqslant \D_{2(q+1)/d}$. Then $T_u^{\Ga(u)}\cong\S_3$ or $\ZZ_3$, and so $ X_u^{\Ga(u)}\cong \S_3$, again a contradiction.
Assume now that $T=\PSL(3,q)$ ($q<9$), $\PSp(4,3)$, $\PSL(4,2)$, $\PSU(3,8)$ or $\M_{11}$. It is clear that $T\cap G\not=1$. Then $W\leqslant T$. Suppose that $T=\PSp(4,3)$. Using [@Xia Proposition 4.1], $T\cap G=\ZZ_2^4{:}\ZZ_5$ and $T\cap X_u=3_{+}^{1+2}{:}2\A_4$. For that case, it is clear that $G\cap X_u\not=1$, a contradiction. Suppose that $T=\M_{11}$. Then $X=\M_{11}$. Again by [@Xia Proposition 4.1], $G=\ZZ_{11}{:}\ZZ_5$, and $X_u=\M_9.2$, again a contradiction, refer to Lemma \[Lu\]. Suppose that $T=\PSL(3,4)$. By Atlas [@atlas], we conclude that $35\notdiv |T\cap G|$. It implies that $5$ or $7$ divides $|X_u|$, which is impossible. Similarly, $T$ does not equal $\PSL(3,q)$ with $5\leqslant q\leqslant 8$, $\PSL(4,2)$, or $\PSU(3,8)$. If $T=\PSL(3,3)$, we obtain that $X=\PSL(3,3){:}\ZZ_2$, $G=\D_{26}$ and $X_u=\ZZ_3^2{:}\GL(2,3)$, refer to [@Xia Proposition 4.1].
0.07in [**Case 2:**]{} Assume that $\Ga$ is not $(X,2)$-arc-transitive. By Case 1, we only need deal with the case where $T=\PSL(2,q)$ with $q>3$. Since $\Ga$ is not $(X,2)$-arc-transitive, $X_u$ is a 2-group. Since $X=GX_u$ and $G\cap X_u=1$, $G$ contains a $2'$-Hall subgroup of $X$. Then $G\cap T$ contains a $2'$-Hall subgroup of $T$. By Lemma \[Kazarin\], we have that $T=\PSL(2,q)$, $T\cap G=\ZZ_q{:}\ZZ_{\frac{q-1}{2}}$, and $T_u=\D_{q+1}$, where $q=2^e-1$ is a prime (see [@Xia Proposition 4.1]). Suppose that $e=3$. Then $q=7$. By Lemma \[ll\], the statement follows. In what follows, we assume that $e\geqslant5$.
Note that $\ZZ_q{:}\ZZ_{\frac{q-1}{2}}$ is maximal in $T$. By [@Hua; @Zhang Theorem 1.1], we conclude that $G=\ZZ_q{:}\ZZ_{q-1}$, and hence $X=\PGL(2,q)$, and $X_u=T_u=\D_{q+1}$. Let $v\in\Ga(u)$. By Lemma \[Cos\], $T_{uv}$ has index $2$ or $4$ in both $T_u$ and $T_v$. Since $e\geqslant5$, $T_{uv}$ contains a subgroup $C\cong\ZZ_4$. It is easily shown that $C$ is normal in both $T_u$ and $T_v$, and so $C\lhd L{:}=\l T_u,T_v\r$. In view of [@Suzuki p.417], each Sylow $2$-subgroup of $T$ is maximal in $T$. Note that $T_u$ is a Sylow $2$-subgroup of $T$, it follows that $L=T_u=T_v$. By the connectedness of $\Ga$, we obtain $L$ fixes each vertex of $\Ga$, which is impossible. This completes the proof of Lemma \[al\].
0.07in
By [@Zai; @Ping; @Lu Theorem 1.1], we have the following lemma.
\[T\] Let $X$ be in Table $2$, and let $T=\soc(X)$. Then we have:
- if $T=\PSL(3,3)$, then $\Ga$ is isomorphic to the graph given in Example $\ref{TA}$;
- if $T=\PSL(2,7)$, then $\Ga$ is isomorphic to a graph given in Examples $\ref{TA}$, $\ref{A}$ and $\ref{B}$;
- if $T=\PSL(2,23)$, then $\Ga$ is isomorphic to the graph given in Example $\ref{C}$;
- if $T=\PSL(2,11)$, then $\Ga$ is isomorphic to a graph given in Examples $\ref{C}$-$\ref{D}$.
We now begin with treating the case where $R(X)\not=1$.
\[R\] If $R(X)\cap G=1$, then $G=\ZZ_{11}{:}\ZZ_{10}$ and $X=\PGL(2, 11)\times\ZZ_2$.
[*Proof. *]{}
Let $\ov X_1=X_1R(X)/R(X)$, $\ov G=GR(X)/R(X)$, and $\ov X=X/R(X)$. Since $X=GX_1$, we conclude that $\ov X=\ov G~\ov X_1$. Since $R(X)\cap G=1$, implying that $\ov G\cong G$ is a Frobenius group. Since $\ov X$ is insoluble, $\Ga$ is a cover of $\Ga_{R(X)}$, refer to Lemma \[insoluble\]. Let $B$ be a vertex of $\Ga_{R(X)}$, where $1\in B$. By Frattini argument, $\ov X=\ov G~\ov X_B$. Clearly, $\ov X_1\leqslant \ov X_B$, and $\ov G\cap\ov X_B\not=1$.
Assume that $\ov X$ is not almost simple. Let $\ov N$ be a minimal normal subgroup of $\ov X$. Arguing similarly as Lemma \[Z(X)\], we obtain that $\C_{\ov X}(\ov N)=1$. Suppose that $\soc(\ov X)=\A_5\times\A_5$. By the definition of $G$, we conclude that $\ov G\cong\ZZ_5^2{:}\ZZ_8$. For this case, $\ov X=((\A_5\times\A_5){:}\ZZ_2){:}\ZZ_2$ is a subgroup of $\S_5\wr\S_2$, and $\ov X_B\cong\S_3\times\S_4$. It implies that $\ov G\cap\ov X_B\cong\ZZ_2$. By MAGMA [@Magma], there are two elements $\ov G$ (up to conjugate) in $\ov X$, and there is an element $\ov X_B$ (up to conjugate) in $\ov X$ such that their intersections are isomorphic to $\ZZ_2$. Choose $\ov B\in\Ga(B)$. By Lemma \[Cos\], write $\Ga_{R(X)}=\Cos(\ov X,\ov X_B,\ov X_Bo\ov X_B)$, where $o\in\N_{\ov X}(\ov X_{B\ov B})\setminus \ov X_{B}$ and $o^2\in \ov X_{B\ov B}$. Since $\Ga$ is $\ov X$-arc-transitive graph, we conclude $|\ov X_B{:}\ov X_{B\ov B}|=4$, and hence $|X_{B\ov B}|=36$. Again by MAGMA [@Magma], for each choice of $\ov G$ and $\ov X_B$, there is no $o\in\N_{\ov X}(\ov X_{B\ov B})$ such that $\l \ov X_B,o\r=\ov X$, namely, $\Ga_{R(X)}$ is not connected. For other cases, arguing similarly as Lemma \[X\], we exclude these possibilities. Thus $\ov X$ is almost simple.
Let $\ov T=\soc(\ov X)$. By [@L.Kazarin], we obtain that $\ov T$ is one of the following: $$\PSL(2,q)(q>3),\, \PSL(3,q)(q<9),\,\PSL(4,2),\, \PSp(4,3),\ \PSU(3,8),\ \mbox{or}\ \M_{11}.$$ Suppose first that $\ov T=\PSL(2,q)$ where $q=5,7,11,23$. If $q=5$, the only possibility is that $\ov G\cong\ZZ_5{:}\ZZ_4$, refer to [@Zai; @Ping; @Lu Theorem 1.1]. For this case, $\Ga$ is a Cayley graph of order $20$, by [@Pan Theorem 5.3], $G$ is normal in $X$, which is a contradiction. If $q=7,11$ or $23$, then $|G|$ is square-free, refer to Atlas [@atlas]. Again by [@Zai; @Ping; @Lu Theorem 1.1], $X=\PGL(2,11)\times\ZZ_2$, $G=\ZZ_{11}{:}\ZZ_{10}$, $X_1=\S_4$ and $\Ga$ is isomorphic to the graph given in Example \[D\]. Arguing similarly as Lemma \[al\] with $\ov X=\ov G~\ov X_B$ in the place $X=GX_1$, we can exclude other cases. This completes the proof.
\[Rr\] If $R(X)\cap G\not=1$, then we have:
- $G\cong\ZZ_p^4{:}\ZZ_5$, $X=W.\overline X$, and $\Ga_W=\K_5$, where $\soc(\overline X)\cong \A_5$;
- $G\cong\ZZ_p^4{:}\ZZ_{10}$, $X=W.(\overline X\times\ZZ_2)$, and $\Ga_W=\K_{5,5}-5\K_2$, where $\soc(\overline X)\cong\A_5$.
[*Proof. *]{}
Let $R=R(X)\cap G$. Assume that $R(X)\cap G\not=1$. From the minimality of $W$ in $G$, it follows that $R\geqslant W$. Since $X/R(X)$ is insoluble, by Lemma $\ref{insoluble}$, $\Ga$ is a normal cover of $\Ga_{R(X)}$. Hence $GR(X)/R(X)\leqslant\Aut\Ga_{R(X)}$.
We claim that $R=R(X)$. Let $\ov H=HR(X)/R(X)$. Since $\Ga$ is a Cayley graph of $G$, implying that $\Ga_{R(X)}$ is a connected Cayley graph of $\ov H$. It follows that $|R(X)||\overline H|=|G|$, and hence $|R(X)|=|W||R(X)\cap H|$. So $R(X)\leqslant G$, and then $R=R(X)$, as claimed. Furthermore, it implies that $W$ is a normal subgroup of $X$. By [@Baik Theorem 1.2], we obtain that either $\Ga_W\cong\K_5$ and $H\cong\ZZ_5$ or $\Ga_W\cong\K_{5,5}-5\K_2$ and $H\cong\ZZ_{10}$. In the former case, we easily obtain $\Aut\Ga_W\cong\S_5$, and in the latter case, $\Aut\Ga_W\cong\S_5\times\ZZ_2$.
Let $\ov X=X/W$. Since $\Ga$ is a cover of $\Ga_W$, it follows that $\ov X$ is a subgroup of $\Aut\Ga_W$. Assume first that $\Ga_W\cong\K_5$. Notice that $\ov X$ is insoluble, we conclude that $\soc(\ov X)\cong\A_5$. Assume now that $\Ga_W\cong \K_{5,5}-5\K_2$. Since $H\cong\ZZ_{10}$, and $X$ is insoluble, we obtain $\ov X\cong L\times\ZZ_2$ where $\soc(L)\cong\A_5$.
Recall that $H$ is irreducible on $W$, we conclude that $d=4$, refer to [@Wenqin Lemma 3.1]. Hence $G\cong\ZZ_p^4{:}\ZZ_5$ or $\ZZ_p^4{:}\ZZ_{10}$.
0.1in The assertion of Theorem \[isoluble\] follows from Lemmas \[X\]-\[Rr\].
Half-transitive graphs
======================
In the last section, we aim to prove Theorem \[solubles\].
Let $p$ be an odd prime, and $d>1$ be an odd integer. Let $n$ be a primitive divisor of $p^d-1$, and let $$G=W{:}\langle h\rangle=\ZZ_p^d{:}\ZZ_n<\AGL(1,p^d).$$
\[construction\] [Let $i$ be coprime to $n$ such that $1\leqslant i\leqslant n-1$, and let $a\in W\backslash\{1\}$. Let $$\left\{
\begin{array}{ll}
S_i = \{ah^i, a^{-1}h^i, (ah^i)^{-1}, (a^{-1}h^i)^{-1}\},\\
\Ga_i = \Cay(G,S_i).
\end{array}
\right.$$ ]{}
0.1in [**Proof of Theorem \[solubles\]:**]{} Let $X=\Aut\Ga$. Let $\Ga=\Cay(G,S)$ be connected, edge-transitive and of valency $4$. Note that $\l h\r$ is primitive on $W$, $d>1$ is odd, and $p$ is an odd prime. By Theorem \[soluble\] and Theorem \[isoluble\], we obtain that $G$ is normal in $X$. In view of [@Godsil Lemma 2.1], we have $X=G{:}\Aut(G,S)$.
By Lemma $\ref{G}$, we have that $X_1=\Aut(G,S)\leqslant\D_8$. By [@Doerk Proposition 12.10], $\Aut(G)\cong\ZZ_p^d{:}(\ZZ_{p^d-1}{:}\ZZ_d)$. Since $d$ and $p$ are odd, $\Aut(G)$ has a cyclic Sylow $2$-subgroup. It follows that $X_1=\l\s\r\cong\ZZ_4$ or $\ZZ_2$. Thus $\s$ fixes an element of $G$ of order $n$, say $f\in G$ such that $o(f)=n$ and $f^\s=f$. Then $G=W{:}\l f\r$, and $X=G{:}\l\sigma\r=(W{:}\l f\r){:}\l\s\r$. Moreover, it implies that all involutions of $\Aut(G)$ are conjugate. Recall that $G$ is a Frobenius group, every involution of $\Aut(G)$ inverts all elements of $W$.
Since $\Ga$ is connected, $\l S\r=G$ and $\Aut(G,S)$ is faithful on $S$. Assume that $S$ contains an involution. Recall that $\Ga$ is $X$-edge-transitive, we conclude that $S$ consists of involutions. By the proof of Lemma \[G\], $G\cong\D_{2p}$, against our assumption. Hence $S$ does not contain an involution. For that case, we may write $S=\{x,x^{-1},y,y^{-1}\}$ such that either $o(\s)=2$ and $(x,y)^\s=(y,x)$, or $o(\s)=4$ and $(x,y)^\s=(y,x^{-1})$, refer to [@Praeger Proposition 1]. Now $x=af^i$ , where $a\in W$ and $i$ is an integer. Suppose that $o(\s)=4$. Then $y=x^{\s}=(af^i)^{\s}=a^\s f^i$, and $a'f^{-i}=f^{-i}a^{-1}=(af^i)^{-1}=x^{-1}=x^{\s^2}=a^{\s^2}f^i=a^{-1}f^i$. It follows that $f^{2i}=1$, and hence $f^i$ has order $1$ or $2$. If $f^i=1$, then $x=a$, and $y=x^\s=a^\s$, belonging to $W$, and so $\l S\r\leqslant W< G$, which is a contradiction. Thus $f^i$ has order $2$. Note that $f^i$ inverts each element of $W$, we conclude that $x$ has order $2$, again a contradiction. Thus $\s$ is an involution, and so $(x,y)^\s=(y,x)$, $x=af^i$, and $y=x^\s=a^\s f^i=a^{-1}f^i$. In particular, $\Ga$ is not arc-transitive, and $S=\{af^i,a^{-1}f^i,(af^i)^{-1},(a^{-1}f^i)^{-1}\}$.
Since $f\in G$ has order $n$, it follows from Hall’s theorem that there exists $b\in W$ such that $h^b\in\l f\r$. So $f^{b^{-1}}=h^r$ for some $r$. Let $\t=\s^{b^{-1}}$. Then $\t$ centralizes $\l h\r$, and $X=G{:}\l\t\r$. Moreover, $S^{b^{-1}}=\{ah^{ir},a^{-1}h^{ir},(ah^{ir})^{-1},(a^{-1}h^{ir})^{-1}\}$. Let $ir\equiv j\,(\mod n)$, and $1\leqslant j\leqslant n-1$. Then $S_j{:}=\{ah^j,a^{-1}h^j,(ah^j)^{-1},(a^{-1}h^j)^{-1}\}$. Note that $\Ga\cong\Cay(G,S_j)$ is connected. By [@Hua; @Zhang Lemma 6.2], $(j,n)=1$, $\Ga_i\cong\Ga_{n-i}$, and if $p^ki\equiv \pm j\,(\mod n)$ for some $k$, then $\Ga_i\cong\Ga_j$. This completes the proof of Theorem \[solubles\].
[99]{} Y. G. Baik, Y. Q. Feng, H. S. Sim, M. Y. Xu, On the normality of Cayley graphs of abelian groups, [*Algebra Colloq.*]{} [**5**]{} (1998) 297-304.
W. Bosma, C. Cannon, C. Playoust, The MAGMA algebra system I: The user language, [*J. Symbolic Comput.*]{} [**24**]{} (1997) 235-265.
K. Bo$\check{s}$tjan, Arc-transitive elementary abelian covers of the complete graph $K_5$, [*Linear Algebra Appl.*]{} [**433**]{} (2010) 1909-1921.
J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, P. Wilson, Atlas of finite Groups, Oxford University Press, Oxford, 1985.
B. P. Corr, C. E. Praeger, Normal edge-transitive Cayley graphs of Frobenius groups, [*J. Algebraic Combin.*]{} [**42**]{} (2015) 803-827.
A. S. Detinko, D. L. Flannery, Nilpotent primitive linear groups over finite fields, [*Comm. Algebra*]{}. [**33**]{} (2005) 497-505.
J. D. Dixon, B. Mortimer, Permutation Groups, Springer-Verlag, New York, 1996.
K. Doerk, T. Hawkes, Finite Soluble Groups, Walter de Gruyter Co., Berlin, 1992.
X. G. Fang, C. H. Li, M. Y. Xu, On edge-transitive Cayley graphs of valency four, [*European J. Combin.*]{} [**25**]{} (2004) 1107-1116.
C. D. Godsil, On the full automorphism group of a graph, [*Combinatorica*]{}. [**1**]{} (1981) 243-256.
D. Gorenstein, Finite Groups, Harper and Row, 1968.
L. S. Kazarin, Groups that can be represented as a product of two solvable subgroups, [*Comm. Algebra*]{}. [**14**]{} (1986) 1001-1066.
E. I. Khukhro, N. Yu. Makarenko, P. Shumyatsky, Frobenius groups of automorphisms and their fixed points, [*Forum Math.*]{} [**26**]{} (2014) 73-112.
C. H. Li, The finite vertex-primitive and vertex-biprimitive s-transitive graphs for $s\geqslant4$, [*Trans. Amer. Math. Soc.*]{} [**353**]{} (2001) 3511-3529.
C. H. Li, Finite $s$-arc transitive Cayley graphs and flag-transitive projective planes, [*Proc. Amer. Math. Soc.*]{} [**133**]{} (2005) 31-41.
C. H. Li, Semiregular automorphisms of cubic vertex transitive graphs, [*Proc. Amer. Math. Soc.*]{} [**136**]{} (2008) 1905-1910.
C. H. Li, Z. Liu, Z. P. Lu, The edge-transitive tetravalent Cayley graphs of square-free order, [*Discrete Math.*]{} [**312**]{} (2012) 1952-1967.
C. H. Li, Z. P. Lu, D. Maru$\check{s}$i$\check{c}$, On primitive permutation groups with small suborbits and their orbital graphs, [*J. Algebra*]{}. [**279**]{} (2004) 749-770.
C. H. Li, Z. P. Lu, H. Zhang, Tetravalent edge-transitive Cayley graphs with odd number of vertices, [*J. Combin. Theory Ser. B.*]{} [**96**]{} (2006) 164-181.
C. H. Li, B. Z. Xia, Factorizations of almost simple groups with a solvable factor, and Cayley graphs, arXiv:1408.0350v3 [**\[math.GR\]**]{}.
D. Maru$\check{s}$i$\check{c}$, Recent developments in half-transitive graphs, [*Discrete Math.*]{} [**182**]{} (1998) 219-231.
D. Maru$\check{s}$i$\check{c}$, R. Nedela, On the point stabilizers of transitive groups with non-self-paired suborbits of length $2$, [*J. Group Theory.*]{} [**4**]{} (2001) 19-43.
D. Maru$\check{s}$i$\check{c}$, R. Nedela, Maps and half-transitive graphs of valency $4$, [*European J.Combin.*]{} [**19**]{} (1998) 345-354.
J. M. Pan, Y. Liu, Z. H. Huang, C. L. Liu, Tetravalent edge-transitive graphs of order $p^2q$, [*Sci. China Math.*]{} [**57(2)**]{} (2014) 293-302.
P. Poto$\check{c}$nik, A list of 4-valent 2-arc-transitive graphs and finite faithful amalgams of index (4, 2), [*European J. Combin.*]{} [**30**]{} (2009) 1323-1336.
C. E. Praeger, An O’Nan-Scott theorem for finite quasiprimitive permutation groups and an application to $2$-arc transitive graphs, [*J. London. Math. Soc.*]{} [**47**]{} (1992) 227-239.
C. E. Praeger, Finite normal edge-transitive Cayley graphs, [*Bull. Aust. Math. Soc.*]{} [**60(2)**]{} (1999) 207-220.
M. Suzuki, Group Theory I, Springer-Verlag, Berlin, New York, 1982.
X. Y. Wang, Y. Q. Feng, Tetravalent half-edge-transitive graphs and non-normal Cayley graphs, [*J. Graph Theory*]{}. [**70(2)**]{} (2012) 197-213.
R. M. Weiss, $s$-transitive graphs, [*Algebraic Methods in Graph Theory*]{}. [**2**]{} (1981) 827-847.
M. Y. Xu, Half-transitive graphs of prime-cube order, [*J. Algebraic Combin.*]{} [**1(3)**]{} (1992) 275-282.
M. Y. Xu, Automorphism groups and isomorphisms of Cayley digraphs, [*Discrete Math.*]{} [**182**]{} (1998) 309-319.
W. Q. Xu, S. F. Du, J. H. Kwak, M. Y. Xu, $2$-Arc-transitive metacyclic covers of complte graphs, [*J. Combin. Theory Ser. B.*]{} [**111**]{} (2015) 54-74.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Exploiting symmetry in Gröbner basis computations is difficult when the symmetry takes the form of a group acting by automorphisms on monomials in finitely many variables. This is largely due to the fact that the group elements, being invertible, cannot preserve a term order. By contrast, inspired by work of Aschenbrenner and Hillar, we introduce the concept of [*equivariant Gröbner basis*]{} in a setting where a [*monoid*]{} acts by [*homomorphisms*]{} on monomials in potentially [*infinitely*]{} many variables. We require that the action be compatible with a term order, and under some further assumptions derive a Buchberger-type algorithm for computing equivariant Gröbner bases.
Using this algorithm and the monoid of strictly increasing functions ${{\mathbb N}}\to {{\mathbb N}}$ we prove that the kernel of the ring homomorphism $${{\mathbb R}}[y_{ij} \mid i,j \in {{\mathbb N}}, i > j] \to {{\mathbb R}}[s_i,t_i \mid i \in {{\mathbb N}}],\
y_{ij} \mapsto s_is_j + t_it_j$$ is generated by two types of polynomials: [*off-diagonal $3 \times 3$-minors*]{} and [*pentads*]{}. This confirms a conjecture by Drton, Sturmfels, and Sullivant on the Gaussian two-factor model from algebraic statistics.
address:
- |
Department of Mathematics and Computer Science\
Technische Universiteit Eindhoven\
P.O. Box 513, 5600 MB Eindhoven, The Netherlands\
- |
Department of Mathematics and Computer Science\
Technische Universiteit Eindhoven\
P.O. Box 513, 5600 MB Eindhoven, The Netherlands\
and Centrum voor Wiskunde en Informatica, Amsterdam, The Netherlands
author:
- 'Andries E. Brouwer'
- Jan Draisma
title: 'Equivariant Gröbner bases and the Gaussian two-factor model'
---
[^1]
Introduction and results
========================
Equivariant Gröbner bases {#equivariant-gröbner-bases .unnumbered}
-------------------------
Algebraic varieties arising from applications often have many symmetries. When analysing such varieties with tools from computational algebra, it is desirable to do so in an [*equivariant*]{} manner, that is, while keeping track of those symmetries, and if possible exploiting them. The notion of [*Gröbner basis*]{}, which lies at the heart of computational algebra, depends heavily on choices of [*coordinates*]{} and of a [*term order*]{}, which is a well-order on monomials in the coordinates. It is therefore natural, at least from a computational point of view, to study symmetries of ideals that preserve both the coordinates and the term order. Now the term [*symmetry*]{} is usually reserved for certain invertible maps, but it is easy to see that an invertible map cannot preserve a well-order; see Remark \[re:WhyMonoid\]. Hence we are led to relax the condition that symmetries be invertible. On the other hand, if a non-invertible map is to preserve the restriction of the term order to the set of coordinates, then that set better be infinite, in contrast with the usual set-up in computational commutative algebra.
In fact, there is another, more compelling reason for allowing infinitely many variables: many varieties from applications come in infinite families, and it is convenient to pass to a suitable limit. For example, the variety of symmetric $n \times n$-matrices of rank $2$ has a well-defined projective limit for $n$ tending to infinity, and so does the closely related two-factor model that we study in this paper. In both cases, the limit is not only stable under the union of all symmetric groups $S_n$ simultaneously permuting rows and columns, but also under the [*monoid*]{} ${\mathrm{Inc}({{\mathbb N}})}$ of all strictly increasing maps from ${{\mathbb N}}$ to itself. And while the union of the symmetric groups does not preserve any term order, the monoid ${\mathrm{Inc}({{\mathbb N}})}$ does preserve such an order; this fundamental observation allows us to do computations in Section \[sec:2f\].
This discussion leads to the following set-up, which we believe will have applications to numerous other problems. Let $X$ be a potentially infinite set, whose elements we call [*variables*]{}. The free commutative monoid generated by $X$ is denoted ${\mathrm{Mon}}$; its elements are called [*monomials*]{}. Suppose that we have
1. a term order, i.e., a well-order $\leq$ on ${\mathrm{Mon}}$ such that $m \leq m' \Rightarrow mm'' \leq m'm''$ for all $m,m',m'' \in {\mathrm{Mon}}$; and
2. a monoid $G$, i.e., a (typically non-commutative) semigroup with identity, acting on ${\mathrm{Mon}}$ by means of monoid homomorphisms ${\mathrm{Mon}}\to {\mathrm{Mon}}$ preserving the strict order: $\pi 1=1,\
\pi(mm')=(\pi m)(\pi m'),$ and $m < m' \Rightarrow \pi m< \pi m'$ for all $\pi \in G,\ m,m' \in {\mathrm{Mon}}$.
\[ex:AH\] The setting that Aschenbrenner and Hillar study in [@Aschenbrenner07] fits into this framework, and indeed inspired our set-up. There $X=\{x_1,x_2,\ldots\}$ and $G$ is the monoid ${\mathrm{Inc}({{\mathbb N}})}$ of all increasing maps $\pi: {{\mathbb N}}\to {{\mathbb N}}$ acting on $X$ by $\pi x_i=x_{\pi(i)}$ and on ${\mathrm{Mon}}$ by multiplicativity. As a term order one can choose the lexicographic order with $x_i>x_j$ if $i>j$. Aschenbrenner and Hillar have turned their proof of finite generation of ${\mathrm{Inc}({{\mathbb N}})}$-stable ideals in $K[x_1,x_2,\ldots]$ into an algorithm; see [@Aschenbrenner08].
Let $K$ be a field and let $K[X]=K{\mathrm{Mon}}$ be the polynomial $K$-algebra in the variables $X$, or, equivalently, the monoid $K$-algebra of ${\mathrm{Mon}}$. Then $G$ acts naturally on $K[X]$ by means of homomorphisms. A [*$G$-orbit*]{} is a set of the form $Gz=\{\pi z \mid \pi \in G\}$, where $z$ is in a set on which $G$ acts. The fact that $G$ acts by monoid homomorphisms on ${\mathrm{Mon}}$ implies that the ideal generated by a union of $G$-orbits in $K[X]$ is automatically [*$G$-stable*]{}, that is, closed under the action of $G$.
We use the notation ${\mathrm{lm}}(f)$ for the [*leading monomial*]{} of $f$, i.e., the $\leq$-largest monomial having non-zero coefficient in $f$. By the requirement that $G$ preserve the order, we have ${\mathrm{lm}}(\pi
f)=\pi\,{\mathrm{lm}}(f)$. Given an ideal $I$ of $K[X]$, ${\mathrm{lm}}(I)$ is an ideal in the monoid ${\mathrm{Mon}}$, that is, ${\mathrm{lm}}(I)$ is closed under multiplication with any element from ${\mathrm{Mon}}$. If $I$ is $G$-stable, then so is ${\mathrm{lm}}(I)$.
Let $I$ be a $G$-stable ideal in $K[X]$. A [*$G$-Gröbner basis*]{} of $I$ is a subset $B$ of $I$ for which ${\mathrm{lm}}(GB)(=\{{\mathrm{lm}}(\pi
b) \mid b \in B,\ \pi \in G\})$ generates the ideal ${\mathrm{lm}}(I)$ in ${\mathrm{Mon}}$. If $G$ is fixed in the context, then we also call $B$ an [*equivariant Gröbner basis*]{}. If $G=\{1\}$, then we call $B$ an [*ordinary Gröbner basis*]{}.
At MEGA 2009, Viktor Levandovskyy pointed out to the second author that our equivariant Gröbner bases are a special case of Gröbner $S$-bases in the sense of [@Drensky06], which were invented for analysing certain two-sided ideals in free associative algebras. The focus of the present article is on getting exactly the right set-up for doing machine computations of equivariant Gröbner bases in the commutative setting.
It is easy to see that if $B$ is a $G$-Gröbner basis of $I$, then $GB$ generates $I$ as an ideal; see Lemma \[lm:GBgenerates\].
\[ex:Rankk\] Let $X=\{y_{ij} \mid i,j \in {{\mathbb N}}\}$, let $k$ be a natural number, and let $I$ be the ideal of all polynomials in the $y_{ij}$ that vanish on all ${{\mathbb N}}\times {{\mathbb N}}$-matrices $y$ of rank at most $k$. Order the variables $y_{ij}$ lexicographically by the pair $(i,j)$, where $i$ is the most significant index; so for instance $y_{3,5}>y_{2,6}>y_{2,4}>y_{1,10}$. The corresponding lexicographic order on monomials in the $y_{ij}$ is a well-order. Let $G:={\mathrm{Inc}({{\mathbb N}})}\times {\mathrm{Inc}({{\mathbb N}})}$ act on $X$ by $(\pi,\sigma)y_{ij}=y_{\pi(i),\sigma(j)};$ this action preserves the strict order. The $G$-orbit of the determinant $D$ of the matrix $(y_{ij})_{i,j=1,\ldots,k+1}$ consists of all $(k+1)\times(k+1)$-minors of $y$, which by the results of [@Sturmfels90] form a Gröbner basis of the ideal $I$. As a consequence, $\{D\}$ is a $G$-Gröbner basis of $I$.
A $G$-stable ideal need not have a finite $G$-Gröbner basis. Indeed, if one requires that [*every*]{} $G$-stable ideal $I$ in $K[X]$ has a finite $G$-Gröbner basis, then this must in particular be true for [*monomial*]{} ideals. This implies that ${\mathrm{Mon}}$ does not have infinite antichains relative to the [*partial order*]{} on ${\mathrm{Mon}}$ defined by $m \preceq m':\Leftrightarrow \exists \pi \in G: \pi m | m'$. Observe that this is, indeed, a partial order: transitivity is straightforward, and antisymmetry follows from the fact that $\pi m|m \Rightarrow \pi m
\leq m$, while on the other hand $\pi m \geq m$ for all $\pi,m$; see Remark \[re:WhyMonoid\]. Conversely, if $({\mathrm{Mon}}, \preceq)$ does not have infinite anti-chains, then every ideal has a finite $G$-Gröbner basis. This is the case in the set-up of Example \[ex:AH\], which is generalised in [@Hillar09]; there equivariant Gröbner bases are called [*monoidal Gröbner bases*]{}.
We have not yet really used that ${\mathrm{Mon}}$ is the free commutative monoid generated by $X$. So far, we could have taken ${\mathrm{Mon}}$ any commutative monoid equipped with EGB1 and EGB2. This viewpoint, and a generalisation thereof, is adopted in [@Hillar09]. However, for doing computations we need that ${\mathrm{Mon}}$ has more structure; see conditions EGB3 and EGB4 below. This is why we have restricted ourselves to free monoids ${\mathrm{Mon}}$.
In the polynomial ring of Example \[ex:Rankk\] the set $\{y_{12}y_{21},y_{12}y_{23}y_{31},y_{12}y_{23}y_{34}y_{41},\ldots\}$ is an infinite $\preceq$-antichain of monomials, hence the ${\mathrm{Inc}({{\mathbb N}})}$-stable ideal generated by it does not have a finite ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner basis. But even in such a setting where not [*all*]{} $G$-stable ideals have finite $G$-Gröbner bases, ideals of interesting $G$-stable varieties may still have such bases. We will derive an algorithm for computing equivariant Gröbner bases under the following two additional assumptions:
1. for all $\pi \in G$ and $m,m' \in {\mathrm{Mon}}$ we have ${\operatorname{lcm}}(\pi
m,\pi m')=\pi {\operatorname{lcm}}(m,m')$; and
2. for all $f,h \in K[X]$ the set $Gf \times Gh$ is the union of a finite number of $G$-orbits (where $G$ acts diagonally on $K[X]
\times K[X]$), and generators of these orbits can be computed effectively.
Note that EGB3 is automatically satisfied if $G$ stabilises the set $X$ of variables. Although this is the only setting that we will need for the application to the two-factor model, future applications may need the greater generality where $X$ is not $G$-stable. Condition EGB4 is of computational importance, as will become clear in Section \[sec:GGB\]. There we also show in Examples \[ex:EGB3\] and \[ex:EGB4\] that these requirements are not redundant.
\[thm:GGB\] Under conditions EGB1, EGB2, EGB3, and EGB4 there exists an algorithm that takes a finite subset $B$ of $K[X]$ as input and that returns a finite $G$-Gröbner basis of the ideal generated by $B$, [*provided that it terminates*]{}.
In Section \[sec:GGB\] we derive this algorithm, and in Section \[sec:2f\] we apply it to a conjecture concerning a statistical model to be discussed now.
The Gaussian two-factor model {#the-gaussian-two-factor-model .unnumbered}
-----------------------------
The [*Gaussian $k$-factor model with $n$ observed variables*]{} consists of all covariance matrices of $n$ jointly Gaussian random variables $X_1,\ldots,X_n$, the [*observed variables*]{}, consistent with the hypothesis that there exist $k$ further variables $Z_1,\ldots,Z_k$, the [*hidden variables*]{}, such that the joint distribution of the $X_i$ and the $Z_j$ is Gaussian and such that the $X_i$ are pairwise independent given all $Z_j$. This set of covariance matrices turns out to be $$F_{k,n}:=\{D+SS^T \mid D \in M_{n}({{\mathbb R}}) \text{ diagonal and
positive definite, and } S \in M_{n,k}({{\mathbb R}})\},$$ where $M_{n,k}({{\mathbb R}})$ is the space of real $n \times
k$-matrices, and $M_n({{\mathbb R}})$ is the space of real $n \times
n$-matrices. In [@Drton07] this model is studied from an algebraic point of view. In particular, the ideal of polynomials vanishing on $F_{k,n}$ is determined for $k=2,3$ and $n \leq 9$. The case where $k=1$ had already been done in [@deLoera95]. The authors of [@Drton07] pose some very intriguing finiteness questions. In particular, one might hope that for fixed $k$ the ideal of $F_{k,n}$ stabilises, as $n$ grows, modulo its natural symmetries coming from simultaneously permuting rows and columns. For $k=1$ this is indeed the case, and for arbitrary $k$ it is true in a weaker, set-theoretic sense [@Draisma08b]. In this paper we prove that the ideals of $F_{2,n}$ stabilise at $n=6$. To state our theorem we denote by $y_{ij}$ the coordinates on the space of symmetric $n \times n$-matrices; we will identify $y_{ji}$ with $y_{ij}$. Recall from [@Drton07] that the ideal of $F_{2,5}$ is generated by a single polynomial $$P:=\frac{1}{10} \sum_{\pi \in {\operatorname{Sym}}(5)} \sgn(\pi)
y_{\pi(1),\pi(2)}y_{\pi(2),\pi(3)}y_{\pi(3),\pi(4)}y_{\pi(4),\pi(5)}y_{\pi(5),\pi(1)},$$ called the [*pentad*]{}. The normalisation factor is important only because it ensures that all coefficients are $\pm 1$—indeed, the stabiliser in ${\operatorname{Sym}}(5)$ of each monomial in the pentad is the dihedral group of order $10$. We consider $P$ an element of ${{\mathbb Z}}[y_{ij} \mid i \geq j]$. The ideal of $F_{2,6}$ contains another type of equation: the [*off-diagonal minor*]{} $$M:=\det(y[\{4,5,6\},\{1,2,3\}]) \in {{\mathbb Z}}[y_{ij} \mid i \geq j]$$ the determinant of the square submatrix of $y$ sitting in the lower left corner of $y$. If $f$ is any polynomial in ${{\mathbb R}}[y_{ij} \mid i \geq j]$ that vanishes on $F_{2,n}$ and if we regard $f$ as an element of ${{\mathbb R}}[y_{ij} \mid i>j][y_{11},\ldots,y_{nn}]$, then each of the coefficients of the monomials in the diagonal variables $y_{ii}$ is a polynomial in the off-diagonal variables that vanishes on $F_{2,n}$, as well. Therefore the following theorem settles the conjecture of Drton, Sturmfels, and Sullivant, that pentads and off-diagonal minors generate the ideal of $F_{2,n}$ for all $n$; see [@Drton07 Conjecture 26].
\[thm:Main\] For any field $K$ and any natural number $n \geq 6$ the kernel $I_n(K)$ of the homomorphism $K[y_{ij}\mid 1 \leq j < i \leq n] \to
K[s_1,\ldots,s_n,t_1,\ldots,t_n]$ determined by $y_{ij} \mapsto s_is_j +
t_it_j$ is generated, as an ideal, by the orbits of $P$ and $M$ under the symmetric group ${\operatorname{Sym}}(n)$.
In [@Drton08] it is proved that $F_{2,n}$ equals the set of all positive definite matrices with the property that every principal $6
\times 6$-minor lies in $F_{2,6}$. Theorem \[thm:Main\] implies an analogous statement for the Zariski closures of $F_{2,n}$ and $F_{2,6}$.
We sketch the proof of Theorem \[thm:Main\], which appears in Section \[sec:2f\]. We put a suitable elimination order on the monomials in $y_{ij},\ i,j \in {{\mathbb N}},\ i \geq j$, and report on a computation that yields a finite ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner basis for the determinantal ideal generated by all $3 \times 3$-minors of $y$. Intersecting this ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner basis with the ring in the off-diagonal matrix entries gives Theorem \[thm:Main\].
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Jan Willem Knopper and Rudi Pendavingh for motivating discussions on alternative computations that would prove Theorem \[thm:2factorGB\]. We also thank the referees for suggestions on improving the exposition.
An algorithm for equivariant Gröbner bases {#sec:GGB}
==========================================
We retain the setting of the introduction: $X$ is a potentially infinite set and ${\mathrm{Mon}}$ is the free commutative monoid generated by $X$, equipped with a term order (EGB1) preserved by the action of a monoid $G$ (EGB2) which also preserves least common multiples (EGB3). Condition EGB4 will be needed only later.
\[re:WhyMonoid\] Note that $G$ acts by injective maps on ${\mathrm{Mon}}$ by EGB2. It is essential that we allow $G$ to be a monoid rather than a group. Indeed, the image of $G$ in the monoid of injective maps ${\mathrm{Mon}}\to {\mathrm{Mon}}$ contains no other invertible elements than the identity: If $\pi \in G$ then $\pi m \geq m$ since otherwise $m>\pi m>\pi^2 m>\ldots$ would be an infinite strictly decreasing chain. But then if (the image of) $\pi$ is invertible, we have $\pi m>m>\pi^{-1}m>\pi^{-2}m>\ldots$, another infinite decreasing chain.
We set out to translate familiar notions from the setting of ordinary Gröbner bases to our equivariant setting. In what follows the coeffient in $f$ of ${\mathrm{lm}}(f)$, the [*leading coefficient*]{}, is denoted ${\mathrm{lc}}(f)$, and ${\mathrm{lt}}(f)={\mathrm{lc}}(f) {\mathrm{lm}}(f)$ is the [*leading term*]{} of $f$.
\[lm:GBgenerates\] If $I$ is $G$-stable and $B$ is a $G$-Gröbner basis of $I$, then $GB=\{\pi b \mid \pi \in G,\ b \in B\}$ generates the ideal $I$.
If not, then take an $f \in I \setminus {\langle}GB {\rangle}$ with ${\mathrm{lm}}(f)$ minimal. Take $b \in B$ and $\pi \in G$ with ${\mathrm{lm}}(\pi b)|{\mathrm{lm}}(f)$. Subtracting $({\mathrm{lt}}(f)/{\mathrm{lt}}(\pi b)) \pi b$ from $f$ yields an element in $I\setminus {\langle}GB {\rangle}$ with leading term strictly smaller than that of $f$, a contradiction.
Given $f \in K[X]$ and $B \subseteq K[X]$, proceed as follows: if $\pi
{\mathrm{lm}}(b) | {\mathrm{lm}}(f)$ for some $\pi \in G$ and $b \in B$, then subtract the multiple $({\mathrm{lt}}(f)/{\mathrm{lt}}(\pi b))\pi b$ of $\pi b$ from $f$, so as to lower the latter’s leading monomial. Do this until no such pair $(\pi,b)$ exists anymore. The resulting polynomial is called a [*$G$-remainder*]{} (or an [*equivariant remainder*]{}, if $G$ is fixed) of $f$ modulo $B$.
This procedure is non-deterministic, but necessarily finishes after a finite number of steps, since $\leq$ is a well-order. Any potential outcome is called an equivariant remainder of $f$ modulo $B$.
\[de:EqS\] Consider two polynomials $b_0,b_1$ with leading monomials $m_0,m_1$, respectively. Let $H$ be a set of pairs $(\sigma_0,\sigma_1) \in G \times G$ for which $Gb_0 \times Gb_1=\bigcup_{(\sigma_0,\sigma_1) \in H} \{(\pi
\sigma_0b_0,\pi \sigma_1b_1)
\mid \pi \in G\}$. For every element $(\sigma_0,\sigma_1) \in H$ we consider the ordinary S-polynomial $$S(\sigma_0b_0,\sigma_1b_1):=
{\mathrm{lc}}(b_1)\frac{{\operatorname{lcm}}(\sigma_0m_0,\sigma_1m_1)}{\sigma_0m_0}\sigma_0b_0 -
{\mathrm{lc}}(b_0)\frac{{\operatorname{lcm}}(\sigma_0m_0,\sigma_1m_1)}{\sigma_1m_1}\sigma_1b_1.$$ The set $\{S(\sigma_0b_0,\sigma_1b_1) \mid (\sigma_0,\sigma_1) \in H\}$ is called a [*complete set of equivariant S-polynomials*]{} for $b_0,b_1$. It depends on the choice of $H$. Under condition EGB4, $H$ can be chosen finite.
\[thm:Buchberger\] Under the assumptions EGB1, EGB2, and EGB3, let $B$ be a subset of $K[X]$ such that for all $b_0,b_1 \in B$ there exists a complete set of S-polynomials each of which has $0$ as a $G$-remainder modulo $B$. Then $B$ is a $G$-Gröbner basis of the ideal generated by $GB$.
We will first prove this for [*ordinary*]{} Gröbner bases, and from that deduce the theorem for equivariant Gröbner bases. The proof for the ordinary case is identical to the proof in the case of finitely many variables. We include it for completeness, and also because we have no reference where the result is stated for infinitely many variables.
We may and will assume that all elements of $B$ are monic. Let $I$ denote the ideal generated by $B$. If ${\mathrm{lm}}(B)$ does not generate the ideal ${\mathrm{lm}}(I)$ in ${\mathrm{Mon}}$ then there exists a polynomial of the form $$f=\sum_{b \in B} f_{b} b$$ with only finitely many of the $f_{b}$ non-zero, for which ${\mathrm{lm}}(f)$ is not in the ideal generated by ${\mathrm{lm}}(B)$. We may choose the expression above such that first, the [*maximum*]{} $m$ of ${\mathrm{lm}}(f_{b}
b)$ over all $b$ for which $f_{b}$ is non-zero is [*minimal*]{} and second, the number of $b$ with ${\mathrm{lm}}(f_{b} b)=m$ is also minimal. The maximum is then attained for at least two values $b_0,b_1$ of $b$, because otherwise $m$ would be the leading monomial of $f$. Write $m_i:={\mathrm{lm}}(b_i)$ for $i=0,1$, and let $t_0,t_1$ be such that ${\operatorname{lcm}}(m_0,m_1)=t_0m_0=t_1m_1$. Now $m={\mathrm{lm}}(f_{b_0})m_0={\mathrm{lm}}(f_{b_1})m_1$ is a multiple of both $m_0$ and $m_1$, and therefore ${\mathrm{lm}}(f_{b_0})$ is divisible by $t_0$; set $$A:=\frac{{\mathrm{lt}}(f_{b_0})}{t_0}.$$ Next consider $$S :=S(b_0,b_1) = t_0b_0-t_1b_1,$$ where we have used that $b_0$ and $b_1$ are monic. As $0$ is a remainder of $S$ modulo $B$ by assumption, we can write $S$ as a sum $\sum_{b \in B} s_{b}b$ with only finitely many non-zero terms that moreover satisfy ${\mathrm{lm}}(s_{b} b) \leq {\mathrm{lm}}(S) <
{\operatorname{lcm}}(m_0,m_1)$ for all $b$. Then we may rewrite $f$ as $$f=f-A(S - \sum_{b} s_{b} b) =
\sum_{b} (f_{b}+f'_{b}+f''_{b}) b$$ where $f'_{b} = A s_{b}$ and $$f''_{b} = \left\{
\begin{array}{ll}
-{\mathrm{lt}}(f_{b_0}) & \mbox{if $b= b_0$,} \\[2pt]
{\mathrm{lc}}(f_{b_0}){\mathrm{lm}}(f_{b_1}) &
\mbox{if $b = b_1$,} \\[2pt]
0 & \mbox{otherwise.}\\
\end{array}\right.$$ For all $b \in B$ we have $$\begin{aligned}
{\mathrm{lm}}((A s_{b})b)&={\mathrm{lm}}(A s_{b} b)
<\frac{{\mathrm{lm}}(f_{b_0})}{t_0} {\operatorname{lcm}}(m_0,m_1)\\
&= {\mathrm{lm}}(f_{b_0}) m_0=m,\end{aligned}$$ so for all $b$ we have ${\mathrm{lm}}(f'_{b} b) < m$. Moreover, ${\mathrm{lm}}((f_{b_0}+f''_{b_0})b_0)$ is strictly smaller than $m$. Finally, ${\mathrm{lm}}(f''_{b_1}b_1) = m$. We conclude that either $\max_{b}{\mathrm{lm}}((f_{b}+f'_{b}+f''_{b}) b)$ is strictly smaller than $m$, or else the number of $b$ for which it equals $m$ is smaller than the number of $b$ for which ${\mathrm{lm}}(f_{b}b)$ equals $m$. This contradicts the minimality of the expression chosen above.
We prove that $GB$ is an ordinary Gröbner basis of the ideal that it generates. By the ordinary Buchberger criterion it suffices to verify that for all $b_0,b_1 \in B$ and $\pi_0,\pi_1 \in G$ the S-polynomial $S(\pi_0 b_0,\pi_1 b_1)$ has $0$ as a remainder modulo $GB$. By assumption there exists a triple $(\sigma_0,\sigma_1,\pi_2)$ for which $(\pi_0 b_0,\pi_1 b_1)=(\pi_2 \sigma_0 b_0,\pi_2 \sigma_1
b_1)$ and for which $S(\sigma_0
b_0,\sigma_1 b_1)$ has $0$ as a $G$-remainder modulo $B$, which means that it has $0$ as an ordinary remainder modulo $GB$. Since $G$ preserves least common multiples (EGB3), we have $$S(\pi_2 \sigma_0 b_0, \pi_2 \sigma_1 b_1)=
\pi_2 S(\sigma_0 b_0,\sigma_1 b_1),$$ and applying $\pi_2$ to the entire reduction of $S(\sigma_0 b_0,\sigma_1 b_1)$ to $0$ modulo $GB$ yields a reduction of $S(\pi_0 b_0,\pi_1 b_1)$ to $0$, as claimed.
The following example shows that EGB3 is not a redundant assumption in Theorem \[thm:Buchberger\].
\[ex:EGB3\] Suppose that $X=\{x,y,z_1,z_2,\ldots\}$ and that the monoid $G$ is generated by ${\mathrm{Inc}({{\mathbb N}})}$ acting by $\pi z_i=z_{\pi i}$ and trivially on $x,y$, together with a single homomorphism $\sigma:{\mathrm{Mon}}\to {\mathrm{Mon}}$ determined by $\sigma x=x$, $\sigma y=x z_1$, and $\sigma z_i=z_{i+1}$ for all $i$. Then $G$ preserves the lexicographic order on ${\mathrm{Mon}}$ for which $z_{i+1}>z_i>y>x$ for all $i$. Now consider the set $B=\{y+1\}$. We have $$G(y+1) \times G(y+1)=G(y+1,y+1) \cup G(y+1,xz_1+1) \cup
G(xz_1+1,y+1),$$ so we may take $H$ from Definition \[de:EqS\] equal to $\{(1,1),(1,\sigma),(\sigma,1)\}.$ The S-polynomial $S(y+1,y+1)$ is zero, and the S-polynomials $S(y+1,xz_1+1)$ and $S(xz_1+1,y+1)$ reduce to zero modulo $y+1$ and $\sigma(y+1)=xz_1+1$. Hence we have a complete set of S-polynomials of $y+1$ with itself that all $G$-reduce to zero modulo $B$. Nevertheless, $B$ is not a $G$-Gröbner basis of the $G$-stable ideal that it generates, since that ideal also contains $S(xz_1+1,xz_2+1)=z_2-z_1$, which does not $G$-reduce to zero modulo $B$.
Here is an example where EGB4 is not fulfilled.
\[ex:EGB4\] Let $X=\{x,y\}$, let $G$ be the multiplicative monoid of the positive integers, where $m$ acts by $x \mapsto x^m$ and $y \mapsto y^m$. Now $$Gx \times Gy = \{(x^i,y^j) \mid i,j \in {{\mathbb Z}}_{>0}\},$$ while the diagonal $G$-orbit of the pair $(x^i,y^j)$ equals $\{(x^{ai},y^{aj}) \mid a \in {{\mathbb Z}}_{>0}\}$. Hence $Gx \times Gy$ is not the union of finitely many $G$-orbits.
However, if we do assume EGB4, then every pair $(b_0,b_1)$ has a finite and computable complete set of $S$-polynomials and have the following theoretical algorithm, alluded to in Theorem \[thm:GGB\]. We do not claim that it terminates, but if it does, then it returns a finite equivariant Gröbner basis by Theorem \[thm:Buchberger\].
Input
: a finite subset $B$ of $K[X]$.
Output (assuming termination)
: a finite equivariant Gröbner basis of the ideal generated by $GB$.
Procedure
:
1. $P:=B \times B$;
2. while $P \neq \emptyset$ do
1. choose $(b_0,b_1) \in P$ and set $P:=P \setminus
\{(b_0,b_1)\}$;
2. compute a finite complete set ${\mathcal{S}}$ of equivariant $S$-polynomials for $(b_0,b_1)$;
3. while ${\mathcal{S}}\neq \emptyset$ do
1. choose $f \in {\mathcal{S}}$ and set ${\mathcal{S}}:={\mathcal{S}}\setminus \{f\}$;
2. compute a $G$-remainder $r$ of $f$ modulo $B$;
3. if $r \neq 0$ then set $B:=B \cup \{r\}$ and $P:=P \cup (B \times r)$;
3. return $B$.
Note the order in which $B$ and $P$ are updated: one needs to add $(r,r)$ to $P$, as well. The proof of correctness of this algorithm is straightforward and omitted.
If the partial order $\preceq$ on ${\mathrm{Mon}}$ defined in the introduction does not admit infinite antichains, then the equivariant Buchberger algorithm always terminates. Indeed, suppose that the algorithm would not terminate, and let $r_1,r_2,r_3,\ldots$ be the sequence of remainders added consecutively to $B$. Then for all $i<j$ we have ${\mathrm{lm}}(r_i) \not
\preceq {\mathrm{lm}}(r_j)$ since $r_j$ is $G$-reduced modulo $r_i$. Using the fact that decreasing $\preceq$-chains are finite, one finds an infinite subsequence $r_{i_1}, r_{i_2}, \ldots $ with $i_1<i_2<\ldots$ such that ${\mathrm{lm}}(r_{i_a}) \not \preceq {\mathrm{lm}}(r_{i_b})$ holds not only for $a<b$ but also for $a>b$. This sequence contradicts the assumption that $\preceq$ does not have infinite antichains.
An equivariant Gröbner basis for the two-factor model {#sec:2f}
=====================================================
Theorem \[thm:Main\] will follow from the following result. Let $X=\{y_{ij}
\mid \ i,j \in {{\mathbb N}}, i \geq j\}$ be a set of variables representing the entries of a symmetric matrix. We consider the lexicographic monomial order on ${\mathrm{Mon}}$ in which the diagonal variables $y_{ii}$ are larger than all variables $y_{ij}$ with $i>j$, and apart from that $y_{ij} \geq y_{i'j'}$ if and only if $i>i'$ or $i=i'$ and $j \geq j'$. So for instance we have $$y_{2,2} > y_{1,1} > y_{5,2} > y_{4,3}.$$ Note that this monomial order is compatible with the action of the monoid ${\mathrm{Inc}({{\mathbb N}})}$ of all increasing maps ${{\mathbb N}}\to {{\mathbb N}}$. For any polynomial $p \in
K[X]$ let ${l(p)}$ denote the [*largest index of $p$*]{}, i.e., the largest index appearing in any of the variables in any of the monomials of $p$.
\[thm:2factorGB\] For any field $K$, let $J_{{\mathbb N}}(K)$ be the ideal in $K[X]$ generated by all $3 \times 3$-minors of the matrix $y$ (recall that we identify $y_{ji}$ for $j<i$ with $y_{ij}$). Relative to the monomial order $\leq$ the ideal $J_{{\mathbb N}}(K)$ has an ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner basis $B$ consisting of $42$ polynomials. The intersection $B \cap K[y_{ij} \mid i >j]$ is an ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner basis of $J_{{\mathbb N}}(K) \cap K[y_{ij} \mid i>j]$ consisting of $20$ polynomials. The largest indices and the degrees of the elements in these bases are summarised in Table \[tab:Bases\].
\[tab:Bases\]
[|l|lllllll|]{} ${l(p)}$ & 3 & 4 & 5 & 6 & 7 & 8 & 9\
$\# p \in B$ & 1 & 6 & 11 & 10 & 8 & 5 & 1\
degrees & $3^1$ & $3^6$ & $3^{10} 5^1$ & $3^5 5^5$ & $5^8$ & $5^5$ & $5^1$\
$\# p \in B \cap K[y_{ij} \mid i>j]$ & & & 1 & 5 & 8 & 5 & 1\
degrees & & & $5^1$ & $3^5 $ & $5^8$ & $5^5$ & $5^1$\
\[re:PentadMinor\] The polynomial with largest index $5$ in the ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner basis $B \cap K[y_{ij} \mid i >j]$ is the pentad $P$. The five degree-$3$ polynomials with largest index $6$ in that Gröbner basis form the ${\operatorname{Sym}}({{\mathbb N}})$-orbit of the off-diagonal minor $M$. All $19$ remaining polynomials are already in the ${\mathrm{Inc}({{\mathbb N}})}$-stable ideal generated by these polynomials; this latter statement also follows from the result in [@Drton07] that at least up to $n=9$ the ideal of the two-factor model is generated by pentads and off-diagonal minors. The complete ${\mathrm{Inc}({{\mathbb N}})}$-Gröbner bases of $J_{{\mathbb N}}(K)$ can be downloaded from the second author’s website.
A Gröbner basis of the ideal of the two-factor model $F_{2,n}$ relative to [*circular term orders*]{} was already found in [@Sullivant08]. The proof involves general techniques for determining the ideal of secant varieties, especially of toric varieties; see also [@Sturmfels05]. The Gröbner basis found there, however, does not stabilise as $n$ grows—and indeed, circular term orders are not compatible with the action of ${\mathrm{Inc}({{\mathbb N}})}$. It would be interesting to find a direct translation between Sullivant’s Gröbner basis and ours.
Theorem \[thm:2factorGB\] implies Theorem \[thm:Main\].
It is well known that the $(k+1) \times (k+1)$-minors of the symmetric matrix $(y_{ij})_{i,j=1,\ldots,n}$ generate the ideal of all polynomials vanishing on all rank-$k$ matrices (for a recent combinatorial proof of this fact, see [@Sturmfels05 Example 4.12]; in characteristic $0$ this fact is known as the Second Fundamental Theorem for the orthogonal group). Hence the ideal $I_n(K)$ of Theorem \[thm:Main\] is the intersection of the ideal $J_n$ generated by the $3 \times
3$-minors of $(y_{ij})_{i,j=1,\ldots,n}$ with the ring $K[y_{ij}
\mid i>j]$. Theorem \[thm:2factorGB\] implies that one obtains a Gröbner basis of $J_n$, relative to the restriction of the monomial order on $K[y_{ij} \mid i,j \in {{\mathbb N}}, i\geq j]$ to $K[y_{ij} \mid 1 \leq
j < i \leq n]$ by applying all increasing maps $\{1,\ldots,{l(p)}\}
\to \{1,\ldots,n\}$ to all $p \in B \cap K[y_{ij} \mid i >j] $ with ${l(p)} \leq n$. Such an increasing map can be extended to an element of ${\operatorname{Sym}}(n)$, and Remark \[re:PentadMinor\] concludes the proof.
We conclude with some remarks on the computation that proved Theorem \[thm:2factorGB\]. First we need to verify EGB4.
\[lm:FiniteS\] For all $b_0,b_1 \in K[y_{ij} \mid i,j \in {{\mathbb N}}, i \geq j]$ the set $({\mathrm{Inc}({{\mathbb N}})}b_0) \times ({\mathrm{Inc}({{\mathbb N}})}b_1)$ is the union of a finite number of ${\mathrm{Inc}({{\mathbb N}})}$-orbits.
Consider all pairs $(S_0,S_1)$ of sets $S_0,S_1 \subseteq {{\mathbb N}}$ with $|S_i|={l(b_i)}$ for which $S_0 \cup S_1$ is an interval of the form $\{1,\ldots,k\}$ for some $k$, which is then at most ${l(b_0)}+{l(b_1)}$. Note that there are only finitely many such pairs $(S_0,S_1)$. For each such pair let $(\pi_0,\pi_1)$ be a pair of elements of ${\mathrm{Inc}({{\mathbb N}})}$ such that $\pi_i$ maps $\{1,\ldots,{l(b_i)}\}$ onto $S_i$; it is irrelevant how $\pi$ acts on the rest of ${{\mathbb N}}$. Then we have $${\mathrm{Inc}({{\mathbb N}})}b_0 \times {\mathrm{Inc}({{\mathbb N}})}b_1 = \bigcup_{(S_0,S_1)} {\mathrm{Inc}({{\mathbb N}})}(\pi_0 b_0,\pi_1
b_1),$$ where the union is over all pairs $(S_0,S_1)$ as above.
The $42$ polynomials of $B$ were constructed by computing a Gröbner basis for $J_9({{\mathbb Q}})$ with [Singular]{} [@GPS05] and retaining only those polynomials $p$ for which the set of indices occurring in their variables form an interval of the form $\{1,\ldots,k\}$ with $k \leq 9$. All elements of $B$ are monic and have integral coefficients (in fact, equal to $\pm 1$ except for the $3 \times 3$-minor with largest index $3$, which has a coefficient $2$). By the equivariant Buchberger criterion and the proof of Lemma \[lm:FiniteS\], we need only ${\mathrm{Inc}({{\mathbb N}})}$-reduce modulo $B$ all S-polynomials of pairs $(\pi_0b_0,\pi_1b_1)$ with $b_0,b_1 \in B$ and $\pi_i:\{1,\ldots,{l(b_i)}\} \to {{\mathbb N}}$ increasing and such that ${\operatorname{im}}\pi_0 \cup {\operatorname{im}}\pi_1=\{1,\ldots,k\}$ for some $k$. For instance, for $b_0=b_1=b$ equal to the polynomial in $B$ with largest index $9$, we having to ${\mathrm{Inc}({{\mathbb N}})}$-reduce $S(\pi_0b,\pi_1b)$ modulo $B$ for all increasing maps $\pi_0,\pi_1:\{1,\ldots,9\} \to \{1,\ldots,18\}$ whose image union is an interval $\{1,\ldots,k\}$. However, if $k=17$ or $k=18$, then $\pi_0b$ and $\pi_1b$ turn out to have leading monomials with gcd $1$, so these cases can be skipped. This reduces the theorem to a finite computation involving polynomials with largest indices up to $16$, which we have implemented directly in [C]{}. Finally, to deduce the result for all base fields—and to speed up the computation—we used the following argument. Since ${\mathrm{Inc}({{\mathbb N}})}B \cap K[y_{ij} \mid 1 \leq j \leq i \leq
n]$ is a subset of the ideal of $3 \times 3$-minors, it is a Gröbner basis if and only if the ideal generated by ${\mathrm{lm}}(B)$ has the same Hilbert series as the ideal generated by $3 \times 3$-minors. Since this Hilbert series is known and does not depend on the field [@Conca94], we may do all our computations over one field and conclude that it holds over all fields. We have verified the equivariant Buchberger criterion over ${{\mathbb F}}_2$, which made the computation slightly faster than working over ${{\mathbb Q}}$.
[10]{}
Matthias Aschenbrenner and Christopher J. Hillar. Finite generation of symmetric ideals. , 359(11):5171–5192, 2007.
Matthias Aschenbrenner and Christopher J. Hillar. An algorithm for finding symmetric [G]{}r[ö]{}bner bases in infinite dimensional rings. 2008. Preprint available from `http://arxiv.org/abs/0801.4439`.
Aldo Conca. r[ö]{}bner bases of ideals of minors of a symmetric matrix. , 166(2):406–421, 1994.
Jesús A. de Loera, Bernd Sturmfels, and Rekha R. Thomas. Gr[ö]{}bner bases and triangulations of the second hypersimplex. , 15:409–424, 1995.
Jan Draisma. Finiteness for the k-factor model and chirality varieties. , 223:243–256, 2010. Preprint available from `http://arxiv.org/abs/0811.3503`.
Vesselin Drensky and Roberto La Scala. Gr[ö]{}bner bases of ideals invariant under endomorphisms. , 41(7):835–846, 2006.
Mathias Drton, Bernd Sturmfels, and Seth Sullivant. Algebraic factor analysis: tetrads, pentads and beyond. , 138(3–4):463–493, 2007.
Mathias Drton and Han Xiao. Finiteness of small factor analysis models. personal communication, 2008.
Gert-Martin Greuel, Gerhard Pfister, and Hans Sch[ö]{}nemann. — [A]{} computer algebra system for polynomial computations. 2005. `http://www.singular.uni-kl.de`.
Chris J. Hillar and Seth Sullivant. Finite [G]{}röbner bases in infinite dimensional polynomial rings and applications. preprint, available from `http://arxiv.org/abs/0908.1777`, 2009.
Bernd Sturmfels. Gröbner bases and [S]{}tanley decompositions of determinantal ideals. , 205(1):137–144, 1990.
Bernd Sturmfels and Seth Sullivant. Combinatorial secant varieties. , 2(3):867–891, 2006.
Seth Sullivant. A [G]{}roebner basis for the secant ideal of the second hypersimplex. 2008. Preprint, available from `http://arxiv.org/abs/0804.2897`.
[^1]: The second author is supported by DIAMANT, an NWO mathematics cluster.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $\,G/K\,$ be an irreducible Hermitian symmetric space and let $\,D\,$ be a $\,K$-invariant domain in $\,G/K\,$. In this paper we characterize several classes of $\,K$-invariant plurisubharmonic functions on $\,D\,$ in terms of their restrictions to a slice intersecting all $\,K$-orbits. As applications we show that $\,K$-invariant plurisubharmonic functions on $\,D\,$ are necessarily continuous and we reproduce the classification of Stein $\,K$-invariant domains in $\,G/K\,$ obtained by E. Bedford and J. Dadok in [@BeDa91].'
address: 'Dipartimento di Matematica, Università di Roma “Tor Vergata", Via della Ricerca Scientifica 1, I-00133 Roma, Italy'
author:
- Laura Geatti
- Andrea Iannuzzi
title: 'Invariant plurisubharmonic functions on non-compact Hermitian symmetric spaces.'
---
[^1]
[^2]
[^3]
Introduction
============
Let $\,G/K\,$ be an irreducible Hermitian symmetric space of rank $\,r$. By the polydisk theorem (cf. \[Wo72\], p.280) the space $\,G/K\,$ contains a closed subspace $\,\Delta^r$, biholomorphic to an $\,r$-dimensional polydisk, with the property that $\,G/K=K\cdot \Delta^r$. If $\,D\,$ is a $\,K$-invariant domain in $\,G/K$, one has $\,D=K\cdot R$, where $\,R:=D\cap \Delta^r\,$ is a Reinhardt domain in $\,\Delta^r$. The polydisk $\,\Delta^r\,$ and $\,R\,$ are invariant under the group $\,T\ltimes {\mathcal S}_r$, generated by rotations and coordinate permutations.
As the Reinhardt domain $\,R\,$ intersects all $\,K$-orbits in $\,D$, it encodes all information on $\,K$-invariant objects in $\,D$. In this paper we focus on $\,K$-invariant plurisubharmonic functions on $\,D$. When $\,D\,$ is Stein, we obtain the following characterization of the class $\,P^{{\infty},+}(D)^K\,$ of smooth, $\,K$-invariant, strictly plurisubharmonic functions on $\,D\,$:
*$\,f\,\in\,P^{{\infty},+}(D)^K\ \ \ $ if and only if $\ \ \ f|_R\, \in\,P^{{\infty},+}(R)^{T \ltimes \mathcal S_r}$,*
where $\,f|_R\,$ is the restriction of $\,f\,$ to $\,R$. Such result is extended to wider classes of plurisubharmonic functions as follows. Let $\,P^{\infty}(D)^K\,$ denote the class of smooth, $\,K$-invariant, plurisubharmonic functions and $\,P^+(D)^K\,$ (resp. $\,P(D)^K\,$) the class of $\,K$-invariant, strictly plurisubharmonic (resp. plurisubharmonic) functions on $\,D$. One has (Thm. \[PLURIREIN\]):
As a by-product (Cor. \[CHARACTER\]) we reproduce the classification of Stein $\,K$-invariant domains in $\,G/K\,$ obtained by Bedford and Dadok in [@BeDa91] by a direct computation on some classical cases (see also [@FeHu93]).
.
*Let $\,D\,$ be a $\,K$-invariant domain in $\,G/K$.*
- If $\,G/K\,$ is of tube type, then $\,D\,$ is Stein if and only if $\,R\,$ is Stein and connected.
- If $\,G/K\,$ is not of tube type, then $\,D\,$ is Stein if and only if $\,R\,$ is Stein and complete. In particular $\,R\,$ contains the origin and is connected.
Let $\,\g= \k\oplus \p\,$ be a Cartan decomposition of the Lie algebra $\,\g\,$ of $\,G$, let $\,\a \,$ be a maximal abelian subspace of $\,\p$, with Weyl group $\,W$, and let $\,G=K\exp \a \,K\,$ be the corresponding decomposition of $\,G$. Every $\,K$-invariant domain $\,D\,$ in $\,G/K\,$ is uniquely determined by a $\,W$-invariant domain $\, \mathcal D_\a\,$ in $\,\a\,$ by $$\mathcal D_\a \to K \exp \mathcal D_\a\,K/K=D\,.$$ Moreover, to every smooth $\,K$-invariant function $\,f\,$ on $\,D\,$ there corresponds a unique smooth $\,W$-invariant function $\,\tilde f\,$ on $\,\mathcal D_\a$, defined by $$\, \tilde f(H):= f (\exp (H) K)\,,\quad H\in\ \mathcal D_\a,$$ (cf. [@Fle78], [@Dad82]). The proof of the above results is carried out as follows. As a first step we explicitly compute the Levi form of $\,f\,$ in terms of the first and second derivatives of $\,\tilde f$. This is achieved in Proposition \[LEVI\] by means of a fine decomposition of the tangent bundle of $\,D$, induced by the restricted root decomposition of $\,\g\,$ (Rem. \[BASIS2\]), and a simple pluripotential argument which enable us to maximally exploit the involved symmetries.
The Levi form computation leads to the key ingredient of our results. Namely, the following characterization of smooth $\,K$-invariant strictly plurisubharmonic functions on a Stein $\,K$-invariant domain $\,D$ (Thm.\[BIJECTIVEK1\]):
*$\,f\,\in \,P^{{\infty},+}(D)^K\ \ \ $ if and only if $\ \ \ \tilde f\, \in\,LogConv^{\infty,+}(\mathcal D_\a)^{W}$,*
where the latter class consists of smooth $\,W$-invariant functions on $\,\mathcal D_\a\,$ satisfying an appropriate differential positivity condition. We also show that $\,\tilde f\,$ belongs to $\,LogConv^{\infty,+}(\mathcal D_\a)^{W}\,$ if and only if the corresponding $\,T \ltimes \mathcal S_r$-invariant function on $\,R\,$ is strictly plurisubharmonic (Prop.\[PSHPOLIDISC\]). This fact, which may be of independent interest in the context of Reinhardt domains, implies (i) in the above theorem. When extending the above characterization to the non-smooth setting (Thm.\[BIJECTIVEK\]), it turns out that $\,K$-invariant plurisubharmonic functions on $\,D\,$ are necessarily continuous.
Finally, in the appendix we explicitly determine a $\,K$-invariant potential of the Killing form on $\,G/K\,$ (Prop. \[POTENTIALK\]) and we observe that, up to an additive constant, it coincides with the logarithm of the Bergman kernel function.
We remark that our methods require no classification results, nor any distinction between classical and exceptional cases.
We wish to thank our colleague Stefano Trapani for several useful discussions and suggestions.
Preliminaries {#PRELIMINARIES}
=============
Let $\,\g\,$ be a non-compact semisimple Lie algebra and let $\,\k\,$ be a maximal compact subalgebra of $\,\g\,$. Let $\,\g=\k\oplus\p$ be the Cartan decomposition of $\g$ with respect to $\,\k$, with Cartan involution $\theta$. Let $\,\a\,$ be a maximal abelian subspace in $\,\p$. The dimension $r$ of $\,\a\,$ is by definition the [*rank*]{} of $\,G/K$. Let $\g= \m\oplus \a\oplus\bigoplus_{\alpha\in\Sigma}\g^\alpha $ be the restricted root decomposition of $\g$, where $\,\m\,$ is the centralizer of $\,\a\,$ in $\,\k$, the joint eigenspace $\,\g^\alpha=\{X\in\g~|~ [H,X]=\alpha(H)X, {\rm \ for\ all \ } H\in\a\}\,$ is the $\,\alpha$-restricted root space and the restricted root system $\,\Sigma\,$ consists of those $\,\alpha\in\a^*\,$ for which $\,\g^\alpha\not=\{0\} $. Denote by $B(\,\cdot\,,\,\cdot\,)$ the Killing form of $\g$, as well as its holomorphic extension to $\g^\C$ (which coincides with the Killing form of $\g^\C$). Denote by $W$ the Weyl group of $\a$, i.e. the quotient of the normalizer over the centralizer of $\a$ in $K$. In the Hermitian case $W$ acts on $\a$ by signed permutations.
For $\,\alpha\in\Sigma$, consider the $\,\theta$-stable space $\,\g[\alpha]:=\g^\alpha\oplus \g^{-\alpha}$, and denote by $\,\k[\alpha]\,$ and $\,\p[\alpha]\,$ the projections of $\,\g[\alpha]\,$ along $\,\p\,$ and $\,\k$, respectively. Let $\Sigma^+$ be a choice of positive roots in $\Sigma$. Then $$\label{DECO}\k=\m \oplus \bigoplus_{\alpha \in \Sigma^+} \k[\alpha]\, \qquad
{\rm and} \qquad \p= \,\a \oplus \bigoplus_{\alpha \in \Sigma^+} \p[\alpha] \,$$ are $\,B$-orthogonal decompositions of $\,\k\,$ and $\,\p$, respectively. The next lemma is an easy exercise.
\[BASIS\] Every element $\,X\,$ in $\,\p\,$ decomposes in a unique way as $$X_\a + \textstyle \sum_{\alpha \in \Sigma^+} P^\alpha,$$ where $\,X_\a\in \a\,$ and $\,P^\alpha\in \p[\alpha]$. The vector $\,P^\alpha\,$ can be written uniquely as $\,P^\alpha=X^\alpha-\theta X^\alpha$, where $\,X^\alpha\,$ is the component of $\,X\,$ in the root space $\,\g^\alpha$. Moreover, $[H,P^\alpha]=\alpha(H)K^\alpha$, where $\,K^\alpha\,$ is the element in $\,\k[\alpha]\,$ uniquely defined by $\,K^\alpha=X^\alpha+\theta X^\alpha$.
The restricted root system of a Lie algebra $\,\g\,$ of Hermitian type is either of type $\,C_r\,$ (if $\,G/K\,$ is of tube type) or of type $\,BC_r\,$ (if $\,G/K\,$ is not of tube type), i.e. there exists a basis $\,\{e_1,\ldots,e_r\}\,$ of $\,\a^*\,$ for which $$\Sigma^+=\{2e_j, ~1\le j\le r,~~e_k\pm e_l,~ 1\le k< l\le r\},\quad \hbox{ for type
$\,C_r$},$$ $$\Sigma^+=\{e_j,~2e_j,~1\le j\le r,~~e_k\pm e_l,~~1\le k<l\le r\},\quad \hbox{ for type
$\,BC_r$}\,.$$ With the above choice of a positive system $\,\Sigma^+$, the roots $$2e_1,\,\dots\,,\,\ 2e_r \,$$ form a maximal set of long strongly orthogonal positive restricted roots (i.e. such that $\,2e_k\pm 2e_l\not\in
\Sigma $, for $\,k\not=l$).
Denote by $\,I_0\,$ the $\,G$-invariant complex structure of $\,G/K$. For every $\,j=1,\ldots,r$, the root space $\, \g^{2e_j}\,$ is one-dimensional. Choose generators $\,\ E^j \in \g^{2e_j}\,$ such that the $\,\s \l (2)$-triples $\,\{E^j,~\theta E^j,~ A_j:=[\theta E^j,\, E^j]\}\,$ are normalized as follows $$\label{NORMALIZ1}
[ A_j,\, E^j]=2 E^j, \quad \hbox{for}\quad j=1,\ldots,r.$$ We also assume that $I_0(E^j-\theta E^j) = A_j$ (see [@GeIa13], Def. 2.1). By the strong orthogonality of $ 2e_1,\ldots, 2e_r$, the vectors $\, A_1,\ldots, A_r \,$ form a $\,B$-orthogonal basis of $\,\a\,$, dual to the basis $e_1,\ldots,e_r$ of $\a^*$, and the associated $\,\s \l (2)$-triples pairwise commute. For $j=1,\ldots,r$, define $$\label{KJPJ} K^j:=E^j+\theta E^j\quad\hbox{ and}\quad P^j:=E^j-\theta E^j.$$
On $\,\p \cong T_{eK} G/K$ the complex structure $I_0$ coincides with the adjoint action of the element $\,Z_0 \in Z(\k)\,$ given by $$\label{CENTER}
\textstyle Z_0=S_0+\frac{1}{2} \sum_{j=1}^r K^j \,,$$ for some element $\,S_0\,$ in a Cartan subalgebra $\,\s\,$ of $\m$. One has $\,S_0=0\,$ in the tube case (see [@GeIa13], Lem. 2.2). The complex structure $\,I_0\,$ permutes the blocks of the decomposition (\[DECO\]) of $\,\p$ (cf. [@RoVe73]), namely $$\label{CPLXBIS}
\,I_0\a=\bigoplus_{j=1}^r \p[2e_j], \quad I_0\p[e_j+e_l]=\p[e_j-e_l], \quad I_0 \p[e_j]=\p[e_j]\,.$$
In order to state the next result, we need to recall a few more facts. Let $ \g^\C=\h^\C\oplus \bigoplus_{\mu\in\Delta} \g^\mu$ be the root decomposition of $\g^\C$ with respect to the maximally split Cartan subalgebra $\h=\s\oplus \a$ of $\g$. Let $\sigma$ be the conjugation of $\g^\C$ with respect to $\g$. Let $\theta$ denote also the $\C$-linear extension of $\theta$ to $\g^\C$. One has $\theta\sigma=\sigma\theta$. Write $\overline Z:=\sigma Z$, for $Z\in\g^\C$. As $\sigma$ and $\theta$ stabilize $\h$, they induce actions on $\Delta$, defined by $\bar\lambda (H):=\overline{\lambda(H)}$ and $\theta\lambda(H):=\lambda(\theta(H))$, for $H\in\h$, respectively. Fix a positive root system $\Delta^+$ compatible with $\Sigma^+$, meaning that $\lambda|_\a =Re(\lambda) \in \Sigma^+$ implies $\lambda\in\Delta^+$. Then $\sigma\Delta^+= \Delta^+$.
The next lemma gives a more detailed description of the complex structure $I_0$ on $\p$.
\[COMPLEXSTRUCTURE\]
[$(a)$]{} For $j=1,\ldots,r$, let $A_j$ and $P^j$ be as in $($\[NORMALIZ1\]$)$ and $($\[KJPJ\]$)$. One has $I_0P^j =A_j$ and $I_0A_j=-P^j$.\
[$(b)$]{} Let $P=X-\theta X\in\p[e_j+e_l]$, where $X=Z^\mu+\overline Z^\mu$, with $Z^\mu\in\g^\mu$ and $\mu\in\Delta^+$ is a root satisfying $Re(\mu)=e_j+e_l$. If $\bar\mu=\mu$, we may assume $Z^\mu=\overline Z^\mu$ and set $X=Z^\mu$. Then $I_0P=Y-\theta Y$, where $Y=[Z_0,X]={1\over 2} [K^j+K^l,X]\in \g^{e_j-e_l}
\oplus \g^{\theta(e_j-e_l)}$.\
[$(c)$]{} Let $P=X-\theta X\in\p[e_j]$, where $X=Z^\mu+\overline Z^\mu$, $Z^\mu\in\g^\mu$ and $\mu$ is a root in $\Delta^+$ satisfying $Re(\mu)=e_j$ $($as $\dim \p[e_j]$ is even, necessarily $\bar\mu\not= \mu)$. Then $I_0P=Y-\theta Y$, where $Y=i(Z^\mu- \overline{Z^\mu})\in\g^{e_j}$.
\(a) follows by definition from (\[NORMALIZ1\]) and (\[KJPJ\]).
Observe that $Z_0 \in Z(\k)$ implies $$\label{AA}[Z_0,X]=-[Z_0, \theta X], \quad \hbox{ for every $X\in \g$}.$$ (b) By (\[CENTER\]), (\[CPLXBIS\]) and the fact that $[S_0,\g^\mu]\subset \g^\mu$, for every $\mu\in \Delta$, the action of $S_0$ is necessarily trivial on $\p[e_j+e_l]$. Moreover, if $X\in \g^{e_j+e_l}$, then $[K^i,X\pm \theta X]=0$, for all $i\not=j,l$, implying that $[Z_0,X\pm \theta X]={1\over 2}[ K^j+ K^l ,X\pm \theta X].$ Note that $[K^l,Z^\mu]\in\g^\lambda$ and $[K^j,Z^\mu]\in\g^{\theta \lambda}$, where $\lambda\in\Delta$ is the root with real part $e_j-e_l$ and the same imaginary part as $\mu$. Then, by equating terms in the same root spaces in (\[AA\]), one obtains the relations which guarantee that $Y=[Z_0,X]\in\g^{e_j-e_l} \oplus \g^{\theta(e_j-e_l)}$. Namely $$[K^l,Z^\mu]=-[K^j,\theta Z^\mu] \in \g^\lambda\, \qquad [K^l,\overline Z^\mu]=-[K^j,\theta \overline Z^\mu] \in\g^{\bar\lambda}\,,$$ $$[K^j,Z^\mu]=-[K^l,\theta Z^\mu] \in \g^{\theta\lambda}\, \qquad [K^j,\overline Z^\mu]=-[K^l,\theta \overline Z^\mu] \in \g^{\theta\bar\lambda}\,.$$
\(c) If $X\in \g^{e_j}$, then $[K^l,X\pm \theta X]=0$, for all $l\not=j$, implying that $[Z_0,X\pm \theta X]=[{1\over 2}K^j +S_0,X\pm \theta X].$ From (\[AA\]) one obtains $$\textstyle {1\over 2} [K^j,X]+[S_0,X]=-{1\over 2} [K^j,\theta X]-[S_0,\theta X],$$ and, by equating terms in the same root spaces, one obtains the relations $$\textstyle [S_0,Z^\mu]=-{1\over 2} [K^j,\theta Z^\mu]\in\g^\mu \qquad [S_0,\theta Z^\mu]=-{1\over 2} [K^j, Z^\mu]\in\g^{\theta\mu}$$ $$\textstyle [S_0,\overline Z^\mu]=-{1\over 2} [K^j, \theta \overline Z^\mu ]\in\g^{\bar\mu} \qquad [S_0,\theta\overline Z^\mu]=-{1\over 2} [K^j, \overline Z^\mu ]\in\g^{\theta\bar\mu}\,,$$ which imply $$[Z_0,X-\theta X]= 2(-[S_0,\theta Z^\mu]-[S_0,\theta\overline Z^\mu]+[S_0,Z^\mu]+[S_0,\overline Z^\mu]).$$ As $\mu(S_0)=:i\mu_0\in i\R$, the above expression becomes $$2\mu_0 i (Z^\mu-\overline Z^\mu-\theta (Z^\mu-\overline Z^\mu)).$$ From $I_0^2=-Id$, one obtains $\mu_0=\pm {1\over 2}$. Depending on the value $\mu_0$, the pairs of roots $\mu,~\bar\mu$ can be relabelled so that $I_0P$, has the desired expression.
\[BASIS2\] In view of Lemma \[COMPLEXSTRUCTURE\], one can choose a $I_0$-stable basis of $\p$, compatible with the decomposition $($\[DECO\]$)$.
$ (a)$ A basis of $\a\oplus\bigoplus_j \p[2e_j]$: take the elements $A_j, P_j=-I_0A_j $, for $j=1,\ldots,r$, normalized as in $($\[NORMALIZ1\]$)$ and $($\[KJPJ\]$)$;
$ (b)$ A basis of $\p[e_j+e_l]\oplus\,\p[e_j-e_l]$: take 4-tuples of elements $P,\, P',\, I_0P,\, I_0P'$, parametrized by the pairs of roots $\mu\not=\bar\mu \in\Delta^+$ satisfying $Re(\mu)=e_j+e_l $ $($with no repetition$)$. More precisely, $P=X-\theta X $ and $P'=X'-\theta X'$, where $X=Z^\mu+\overline Z^\mu$, $X'=i(Z^\mu-\overline Z^\mu)$, and $Z^\mu $ is a root vector in $\g^\mu$. For $\mu= \bar \mu$, we may assume $Z^\mu=\overline Z^\mu$. Then take the pair $P,\, I_0 P$.
$ (c)$ A basis of $\p[e_j]$ $($non-tube case$)$: take pairs of elements $P,\, I_0P$, parametrized by the pairs of roots $\mu\not=\bar\mu \in\Delta^+$ satisfying $Re(\mu)=e_j $ $($with no repetition$)$. More precisely, $P=X -\theta X,$ where $X=Z^\mu+\overline Z^\mu$, and $Z^\mu$ is a root vector in $\g^\mu$.
\[BRACKETS\] Let $\mu\in \Delta^+$ be a root satisfying $Re(\mu)=e_j+e_l$ and let $Z^\mu$ a root vector in $\g^\mu$. Let $X=Z^\mu+\overline Z^\mu \in \g^{e_j+e_l}$ and $Y=[Z_0,X]$. Then
$(a)$ $[Y,X]+\theta[Y,X]= r_1K^j +s_1K^l$, for $r_1,\, s_1 \in \R $;
$(b)$ $[Y,\theta X]+\theta [Y,\theta X]=r_2K^j +s_2K^l$, for $r_2,\, s_2 \in \R $.
If $\ \overline \mu \not= \mu$, let $X'=i(Z^\mu-\overline Z^\mu)$ and $Y'=[Z_0,X']$. Then
$(c)$ $[Y', X]+\theta [Y', X]=[Y',\theta X]+\theta [Y',\theta X]=0$.
Let $\mu$ be a root in $\Delta^+$, with $Re(\mu)=e_j$ $($non-tube case$)$ and let $Z^\mu$ be a root vector in $\g^\mu$. Let $X=Z^\mu+\overline Z^\mu $ and $Y=[Z_0,X]=i(Z^\mu-\overline Z^\mu)$. Then
$(d)$ $[Y,X]+\theta[Y,X]=tK^j$, for $t \in\R$,
$(e)$ $[Y,\theta X]+\theta[Y,\theta X]\in\m$.
\(a) By Lemma \[COMPLEXSTRUCTURE\] (b), one has $Y={1\over 2}[K^j+K^l,X]\in\g^{e_j-e_l}
\oplus \g^{\theta(e_j-e_l)}
$. Moreover $[Y,X]={1\over 2}[[\theta E^j,X], X]+{1\over 2}[[\theta E^l,X], X]\in \g^{2e_l}
\oplus \g^{ 2e_j}.$ Since such roots spaces are 1-dimensional, then $$[Y,X]+\theta[Y,X]=r_1K^j +s_1K^l,\quad \hbox{ for some $r_1~s_1\in\R$}.$$ (b) Similarly, $[Y,\theta X]={1\over 2}[[\theta E^j,X], \theta X]+{1\over 2}[[\theta E^l,X], \theta X]\in \g^{-2e_j}
\oplus \g^{-2e_l}$, and $$[Y,\theta X]+\theta [Y,\theta X]=s_1K^j+s_2K^l,\quad \hbox{for some $s_1,s_2\in\R$}.$$ (c) One has $$[Y',X]=[[Z_0,iZ^\mu-i\overline{Z^\mu}],Z^\mu+\overline{Z^\mu}]$$ $$=i [[Z_0,Z^\mu],Z^\mu]+i[[Z_0,Z^\mu],\overline{Z^\mu}]-i [[Z_0,\overline{Z^\mu}],Z^\mu] -i [[Z_0,\overline{Z^\mu}],\overline{Z^\mu}].$$ The first and the fourth terms of the above expression sum up to zero: if $\bar \mu\not= \mu$, they are both zero because otherwise there would exist a root in $\Delta^+$ with real part equal to $2e_j$ and non-zero imaginary part; if $\bar\mu=\mu$, such terms are opposite to each other. The second and the third term sum up to zero by the Jacobi identity and the fact that $[Z_0,[Z^\mu,\overline{Z^\mu}]]=0$. One has $$[Y',\theta X]=[[Z_0,iZ^\mu-i\overline{Z^\mu}],\theta Z^\mu+\theta \overline{Z^\mu}]$$ $$=i [[Z_0,Z^\mu], \theta Z^\mu]+i[[Z_0,Z^\mu],\theta\overline{Z^\mu}]-i [[Z_0,\overline{Z^\mu}],\theta Z^\mu] -i [[Z_0, \overline{Z^\mu}],\theta\overline{Z^\mu}].$$ The above expression is automatically zero, if $\bar\mu=\mu$. So let’s assume that $\bar\mu\not =\mu$. Arguing as in the previous case, one has that the sum of the first and the fourth terms is equal to zero. The second and the third terms sum up to $2{\rm Im}([[Z_0,Z^\mu],\theta\overline{Z^\mu}])$. Then $$[Y',\theta X]+\theta [Y',\theta X]=2{\rm Im}([[Z_0,Z^\mu],\theta\overline{Z^\mu}]+\theta[[Z_0,Z^\mu],\theta\overline{Z^\mu}] )=$$ $$\label{VV}=2{\rm Im}( [[Z_0,Z^\mu],\theta\overline{Z^\mu}]+[[Z_0,\theta Z^\mu], \overline{Z^\mu}] ).$$ By the Jacobi identity $$[[Z_0,\theta Z^\mu], \overline{Z^\mu}]=[[Z_0,\overline{Z^\mu}], \theta Z^\mu]+[Z_0,[\theta Z^\mu, \overline{Z^\mu}]].$$ Observe that $[Z^\mu,\theta \overline{Z^\mu}]\in \a \oplus i\s$ and therefore $[Z_0,[\theta Z^\mu, \overline{Z^\mu}]]\in\p$. It follows that the expression in (\[VV\]) reduces to $$2{\rm Im}( [[Z_0,Z^\mu],\theta\overline{Z^\mu}]+ [[Z_0,\overline{Z^\mu}], \theta Z^\mu])=0,$$ as desired.
\(d) Since $[Y, X]\in\g^{2e_j}$ $$[Y, X]+\theta[Y,X]
=tK^j, \quad \hbox{ for some $t\in\R$} .$$ (e) Since $[Y, X]\in\g^{0}$, $$[Y, X]+\theta[Y,X]\in\m.$$
Next, we recall a fact which will be used to compute the Levi form of an arbitrary smooth $\,K$-invariant function on $G/K$. For $\,X \in \k$, denote by $\,\widetilde X\,$ the vector field induced on $\,G/K\,$ by the $\,K$-action $$\label{TILDEFIELDS}\,
\textstyle\widetilde X_z:=\dds \exp sX \cdot z, \quad z\in G/K.$$ Given a smooth $\,K$-invariant function $\,f\colon G/K \to \R\,$, set $\,d^c_{I_0}\rho:=d\rho \circ {I_0}$, so that $\,2i \partial \bar \partial_{I_0}f=-dd^c_{I_0}f$. Moreover, for $\,X\in \k$, define $\,\mu^X: G/K \to \R\,$ by $\,\mu^X(z):=d^c_{I_0}f(\widetilde X_z)\,.$ Then $$\label{DIFFER}
\,d\mu^X= -\iota_{\widetilde X}dd^c_{I_0}f\,. %(CONTROLLARE).$$ The above identity was proved in [@HeSc07], Lemma 7.1. Indeed their argument needs not the plurisubharmonicity of $\,f$.
If the function $\,f\,$ is strictly plurisubharmonic, then $\,-dd^c_{I_0}f\,$ is a $\,K$-invariant Kähler form and the map $\,\mu:G/K \to \k^*$, defined by $$\mu(z)(X)=d^c_{I_0}f(\widetilde X_z)\,,$$ for $\,X\in \k$, is a moment map. It is referred to as the moment map associated with $\,f$.
We conclude the preliminaries with a lemma needed in the next section. Let $\Delta$ be the unit disc in $\C$. Consider the ($T \ltimes \mathcal S_2$)-action on $\Delta^2$, where $T= (S^1)^2$ acts by rotations and $\mathcal S_2$ is the group of coordinate permutations. Let $W_{\R^2}=(\Z_2) \ltimes \mathcal S_2$ be the group acting on $\R^2$ by signed permutations.
\[CONVESS\] Let $f:\Delta^2 \to \R$ be a smooth $T \ltimes \mathcal S_2$-invariant strictly plurisubharmonic function and let $r$, $s$ be real numbers. Consider the $W_{\R^2}$-invariant function $\tilde f: \R^2 \to \R$ given by $\tilde f(a_1,a_2)= f(\tanh a_1, \tanh a_2)$ and define $$\textstyle G_{\tilde f}(a_1,a_2):= \frac{r \sinh (2a_1) \frac{\partial \widetilde f}
{ \partial a_1}(x, y) - s\sinh (2a_2)\frac{\partial \widetilde f}
{ \partial a_2}(a_1,a_2)}{\sinh^2 a_1 -\sinh^2a_2}\,.$$
Then
- $\frac{\partial \widetilde f}
{ \partial a_1}(a_1,a_2) >0$, for every $a_1>0$, and $\frac{\partial \widetilde f}
{ \partial a_1}(a_1,a_2) <0$, for every $a_1<0$. In particular $\frac{\partial \widetilde f}
{ \partial a_1}(0, a_2) =0$, for every real $a_2$.
- $\frac{\partial \widetilde f}
{ \partial a_2}(a_1,a_2)= \frac{\partial \widetilde f}
{ \partial a_1}(a_2,a_1).$ In particular $\frac{\partial \widetilde f}
{ \partial a_2}(a_1, 0) =0$, for every real $a_1$.
- If $G_{\tilde f}$ extends continuously on $\,\R^2\,$ to a strictly positive function, then $r= s>0$. In particular $G_{\tilde f}(a_1,a_2)$ is $W_{\R^2}$-invariant as well.
\(i) For $\,a_1> 0$ let $s_1 \in (-\infty,0)$ be such that $\tanh a_1=e^{s_1}$. Since $f$ is $T$-invariant and strictly plurisubharmonic, the function $\,s_1 \to f(e^{s_1},a_2)$ is strictly convex. Moreover, the limit $\lim_{s_1 \to -\infty} f(e^{s_1},a_2)= f(0,a_2)\, $ is finite. Hence the function is strictly increasing and its derivative $ e^{s_1} \frac{\partial f}{\partial a_1}(e^{s_1},a_2)$ is strictly positive. As $$\textstyle \frac{\partial \widetilde f}
{ \partial a_1}(a_1,a_2) =\frac{1}{\cosh^2 a_1}\frac{\partial f}{\partial x}(\tanh a_1,a_2)\,,$$ and $\tanh a_1 =e^{s_1}$, statement (i) follows.
\(ii) The $\mathcal S_2$-invariance of $\tilde f$ implies that $$\textstyle \lim_{\varepsilon \to 0} \frac{{\tilde f}(a_1, a_2+ \varepsilon)}{\varepsilon}=
\lim_{\varepsilon \to 0} \frac{ {\tilde f}(a_2+ \varepsilon, a_1)}{\varepsilon}\,,$$ and (ii) follows. (iii) Let $a_1>0$. From (ii) it follows that $G_{\tilde f}(a_1,0)= \frac{r \sinh (2a_1) \frac{\partial \widetilde f}
{ \partial a_1}(a_1,0)} {\sinh^2 a_1}\,.$ Since such quantity is assumed to be strictly positive and $\frac{\partial \widetilde f}
{ \partial a_1}(a_1, 0)>0$, it follows that $r>0$. By choosing $a_1=0$ and $a_2>0$, one obtains that $s>0$.
Next we show that $r=s$. For $a_1>a_2>0$ one has $\sinh^2 a_1 -\sinh^2a_2>0$. Then the positivity of $G_{\tilde f}(a_1,a_2)$ implies that $$\textstyle \frac{r}{s}> \frac{\sinh (2a_2)\frac{\partial \widetilde f}
{ \partial a_2}(a_1, a_2)}{\sinh (2a_1) \frac{\partial \widetilde f}
{ \partial a_1}(a_1, a_2)}\,.$$ Consequently, for $a_1$ converging to a fixed $a_2>0$, statements (i) and (ii) imply $\frac{r}{s} \geq 1$. An analogous argument, with $0<a_1<a_2$, implies $\frac{r}{s} \leq 1$. As a consequence, $\frac{r}{s}=1$.
The Levi form of a $K$-invariant function {#REALLEVI}
=========================================
Let $\,G/K\,$ an irreducible non-compact Hermitian symmetric space of rank $\,r$. From the decomposition $\,G=K \exp \a\, K\,$ one obtains a bijective correspondence between $\,W$-invariant domains in $\,\a\,$ and $\,K$-invariant domains in $\,G/K\,$, namely $$\label{DA}\,\mathcal D_{\a} \to D:=K \exp \mathcal D_{\a}K/K \,.$$ In addition, every $K$-invariant function $f:D \to \R$ is uniquely determined by the $W$-invariant function $\tilde f:\mathcal D_{\a} \to \R$, given by $$\label{EFFETILDE}\tilde f(H)=f(\exp (H)K).$$
The goal of this section is express the Hermitian form $h_f(\,\cdot\,,\,\cdot\,):= -dd^c f(\,\cdot\,,I_0\,\cdot\,)$ of a smooth $\,K$-invariant function $f$ on a $K$-invariant domain $D\subset G/K$ in terms of the first and second derivatives of the corresponding $\tilde f$ on $\,\mathcal D_{\a}$. This will enable us to characterize smooth $K$-invariant strictly plurisubharmonic functions on a Stein $K$-invariant domain $D$ in $G/K$ by an appropriate differential positivity condition on the corresponding functions on $\mathcal D_\a$ (see Thm.$\,$\[BIJECTIVEK1\] and Cor.$\,$\[PSHONR\]). As $f$ is $K$-invariant, also $-dd^cf(\,\cdot\,,I_0\,\cdot\,)$ is $K$-invariant. Therefore it will be sufficient to carry out the computation along the slice $\exp \a K$, which meets all $K$-orbits.
For $\,z=aK \,$, with $\,a=\exp(H)$ and $H\in\a$, one has $$\label{TILDEVSHAT}\widetilde X_z = a_*F_aX ,\qquad
a_*X=\widetilde{F_a^{-1} X}_z,$$ where $F_a\colon \p\to\p$ is the map given by $F_a := \pi_\# \circ \Ad_{a^{-1}}|_{\p }\,,$ and $\, \pi_\# :\g \to \p \,$ is the linear projection along $\,\k $. In particular one can verify that $$\label{TILDEFIELDS2}
%\widetilde P_z=a_*F_aP ,\quad
\widetilde K_z=-a_*\sinh\alpha(H)P,$$ for $P=X^\alpha-\theta X^\alpha\in \p[\alpha]$ and $K=X^\alpha+\theta X^\alpha\in\k[\alpha]$, with $\alpha \in \Sigma^+$.
Denote by $a_1,\ldots,a_r$ the coordinates induced on $\a$ by the basis $A_1,\ldots,A_r$ of $\a$ (cf. Rem. \[BASIS2\](a)).
\[LEVI\] Let $D\subset G/K$ be a $K$-invariant domain. Let $\,f:D \to \R\,$ be a smooth $K$-invariant function. Fix $\,a=\exp H$, with $H=\sum_ja_jA_j\in \mathcal D_{\a}$. Then, in the basis of $\,\p\,$ defined in Remark \[BASIS2\], the Hermitian form $h_f$ at $z=aK\in D$ is as follows.
- The spaces $\,a_*\a$, $\,a_*I_0\a$, $\,a_* \p[e_j+ e_l]\,$, $\,a_* \p[e_j- e_l]\,$ and $a_* \p[e_j]\,$ are pairwise $\,h_f$-orthogonal.
As the form $h_f$ is $I_0$-invariant, it is determined by its restrictions to the blocks $a_*\a$, $a_*\p[e_j+e_l]$ and $a_*\p[e_j]$. The non-zero entries of $h_f$ on each of these blocks are given as follows.
- For $A_j,A_l\in\a$ one has
$\,h_f(a_*A_j,a_*A_l)=%h_f(a_*P^j,a_*P^l) =
2\coth (2a_j)\frac{\partial \widetilde f}
{ \partial a_j}(H) \delta_{jl}+\frac{\partial^2 \widetilde f}
{ \partial a_j \partial a_l}(H)\,$;
- For $P,\, P'$ as in Remark \[BASIS2\](b) one has
$$\,h_f(a_*P ,a_*P )=h_f(a_*P' ,a_*P' )=%h_f(a_*I_0P ,a_*I_0P )=h_f(a_*I_0P' ,a_*I_0P' )=$$ $$\textstyle =\frac{B(P,P)}{b}\frac{1}{\sinh (a_j+a_l)\sinh ( a_j-a_l)}
\big ( \sinh (2a_j)\frac{\partial \widetilde f}{ \partial a_j}(H)-
\sinh (2a_l)\frac{\partial \widetilde f}{ \partial a_l}(H)\big )\,,$$ where $b:=B(A_1,A_1)= \dots =B(A_r,A_r)$. In particular, with respect to the basis of $a_* \p[e_j+e_l]$ defined in Rem. \[BASIS2\]$($b$)$, the form $h_f$ is diagonal.
- $($non-tube case$)$ For $P\in \p[e_j]$, as in Remark \[BASIS2\] (c), one has
$$\,h_f(a_*P ,a_*P )= \textstyle \frac{B(P,P)}{b}\coth(a_j)
\frac{\partial \widetilde f}
{ \partial a_j}(H)\, .$$ In particular, with respect to the basis of $a_* \p[e_j]$ defined in Rem. \[BASIS2\]$($c$)$, the form $h_f$ is diagonal.
In order to exploit the relation (\[DIFFER\]), we first compute $d^cf(\widetilde X_{z})$, for $X\in \k$ and $z \in G/K$. By the $\,K$-invariance of $\,f\,$ and of $\,I_0$ one has $$\label{DCf} \, d^cf(\widetilde X_{k\cdot z})=d^cf( \widetilde {\Ad_{k^{-1}}X}_{z})\,,$$
for every $\,z \in G/K\,$ and $\,k \in K$. Thus it is sufficient to carry out the computation for $z=aK$. We first assume that $\alpha(H) \not=0$ for all $ \alpha\in
\Sigma$, and later obtain the desired result by passing to the limit for $H$ approaching the hyperplanes defined by $\{\alpha=0\}$.
Recall the decomposition of $\k$ given in (\[DECO\]). For all $M \in \m$, one has $\widetilde M_z=0$, and therefore $$\label{MENOUNO}
d^cf (\widetilde M_z)=0\,.$$
For $\,K = X^\alpha+ \theta X^\alpha$ in $\,\k[\alpha]$, with $\,\alpha \not= 2e_1, \dots, 2e_r$, set $\,P = X^\alpha- \theta X^\alpha\,$ in $\,\p[\alpha]\,$. Then $\, I_0P = Y^\beta- \theta Y^\beta$ in $\p[\beta]$, for some $\beta\not= 0$. Set $\,C= Y^\beta+ \theta Y^\beta$ in $\k[\beta]$. Then, by (\[TILDEFIELDS2\]) and the $K$-invariance of $\,f\,$, one has
$$d^c f( \widetilde {K}_z )=
-d f (I_0a_* \sinh \alpha(H) P )=-d f(a_* \sinh \alpha(H) I_0P)$$ $$\label{ZERO}
=d f \big ( \textstyle \frac {\sinh \alpha (H)}{ \sinh \beta(H)} \widetilde{C}_z \big )= 0.$$ Finally, in the case of $K^j\in \k[2e_j]$, for $\,j=1, \dots\, r$ (cf. (\[KJPJ\])), one has
$$d^c f ( \widetilde{ K^j}_z)=
-d f \big (\textstyle I_0a_* \sinh (2a_j)
P^j \big ) =-d f\big(
a_*\sinh (2a_j)A_j \big)=$$ $$\label{UNO}
=\textstyle- \dds f \big (\exp (H+ \sinh (2a_j)sA_j)K \big )= -\sinh(2a_j) \frac{\partial \widetilde f}
{ \partial a_j}(H)\,.$$
Next we prove statement [**(i)**]{}. Let $\alpha$ and $\gamma$ be distinct roots with $\alpha \in \Sigma^+$ and $\gamma \in \{0\} \cup (\Sigma^+ \setminus \{2e_1,\dots,2e_r\})$, with the convention $\p[0]:=\a$. Let $P \in \p[\alpha]$ and $Q\in \p[\gamma]$. Write $P =X^\alpha-\theta X^\alpha$, with $X^\alpha\in\g^\alpha$, and $I_0Q =Y^\beta-\theta Y^\beta$, with $Y^\beta\in\g^\beta$, for some $\beta \in \Sigma^+$. Then $\, a_*P =\textstyle -\frac{1}{\sinh \alpha(H)} \widetilde K_z \,$, with $K =X^\alpha+\theta X^\alpha \in \k[\alpha]$, and $\,a_*I_0Q =\textstyle-\frac{1}{\sinh \beta(H)} \widetilde C_k \,,$ with $C =Y^\beta+\theta Y^\beta \in \k[\beta]$. Therefore $$h_f(a_*P ,a_*Q)=-dd^c f(a_*P ,\,a_*I_0Q)=\textstyle
\frac{1}{\sinh \alpha(H) \sinh \beta(H)}
\ddt \mu^{K}( \exp tC \cdot z)=$$ $$\textstyle =\frac{1}{\sinh \alpha(H) \sinh \beta(H)}
\ddt d^c f( \widetilde K_{\exp tC \cdot z})$$ which, by (\[DCf\]), becomes $$\textstyle \frac{1}{\sinh \alpha(H) \sinh \beta(H)}
\ddt d^c f( \widetilde {\Ad_{\exp -tC }K} _{z})=$$ $$\textstyle =\frac{1}{\sinh \alpha(H) \sinh \beta(H)}
\ddt d^c f \big ( \widetilde{K}_{z} -t
\widetilde{[C ,K ]}_{z} +o(t^2) \big ) =$$ $$\textstyle =-\frac{1}{\sinh \alpha(H) \sinh \beta(H)}
d^c f( \widetilde {[C ,K ]} _{z})\,.$$ Hence $$\label{FORMULONE}
\textstyle h_f(a_*P ,a_*Q)=
-\frac{1}{\sinh \alpha(H) \sinh \beta(H)}
d^c f( \widetilde {[C ,K ]} _{z})\,.$$ The brackets $$[C ,K ]= ([Y^\beta,X^\alpha]+\theta [Y^\beta,X^\alpha])+
([Y^\beta,\theta X^\alpha]+\theta [Y^\beta,\theta X^\alpha])\,,$$ lie in $\,\k[\alpha+\beta] + \k[\alpha-\beta]$. Since $\alpha\in\Sigma^+$ and $\gamma\in \{0\} \cup (\Sigma^+ \setminus \{2e_1,\dots,2e_r\})$ are distinct, the spaces $\k[\alpha+\beta]$ and $\k[\alpha-\beta]$ have zero intersection with $\oplus_j\k[2e_j]$. Then the expression (\[FORMULONE\]) vanishes by (\[MENOUNO\]) and (\[ZERO\]), and the spaces $a_*\p[\alpha]$ and $ a_*\p[\gamma]$ are $h_f$-orthogonal, as claimed. By the $I_0$ invariance of $h_f$, this also implies that $a_*\a$ is $h_f$-orthogonal to $\oplus_{\alpha \in \Sigma^+}a_*\p[\alpha]$. This concludes the proof of (i).
Next we examine the Hermitian tensor $h_f$ on the various blocks.
Let $A_j,A_l\in\a$. Since $I_0A_l=-P^l$, one has $$h_f(a_*A_j,a_*A_l)=
-dd^cf(a_*P^l ,\,a_*A_j)=
\textstyle {1\over \sinh (2a_l)}dd^c f
((\widetilde{K^l })_z,(\widetilde{A_j})_z)=$$ $$\textstyle
- {1\over \sinh (2a_l)} \ddt \mu^{K^l }(\exp tA_j \cdot z)=
\textstyle -{1\over \sinh (2a_l)} \ddt d^cf( \widetilde K^l_{\exp tA_j \, z})$$ $$=\textstyle {1\over \sinh (2a_j)} \ddt df( I_0 ({\exp (H+tA_j) })_* \sinh 2e_j(H+tA_j))\widetilde P^l)
=$$ $$\textstyle
{1\over \sinh (2a_l)}
\ddt \sinh 2e_j(H+tA_j) \frac{\partial \widetilde f}
{ \partial a_l}(H+tA_j)
=$$ $$\textstyle {1\over \sinh (2a_l)} \big (
2 \cosh(2a_l)\frac{\partial \widetilde f}
{ \partial a_l}(H) \delta_{j,l}+ \small{ \sinh (2a_l)}\frac{
\partial^2 \widetilde f}
{ \partial a_j \partial a_l}(H) \big)\,.$$
The above expression is well defined also for those $H=\sum_ja_jA_j$ with some zero coordinate. Assume for example $a_l=0$. By the $\,W$-invariance, $\widetilde f$ is even with respect to the coordinate $a_l$ and consequently its derivative $\frac{\partial \widetilde f}
{\partial a_l}$ vanishes for $a_l=0$. Therefore $$\textstyle \lim_{a_l\to 0}= 2 \coth( 2a_l)
\frac{\partial \widetilde f}
{ \partial a_l}= \frac{\partial^2 \widetilde f}
{ \partial\, a_l^2},$$ and the above quantity smoothly extends to the hyperplane $a_l=0$, for all $l=1,\ldots,r$. This concludes the proof of (ii).
Let $P,\, Q\in\p[e_j+e_l]$ be elements of the basis of Remark \[BASIS2\](b), arising from roots $\mu,\, \nu\in \Delta^+$, respectively, with $\nu\not=\mu,\,\bar \mu$. Then $h_f(a_*P,a_*Q) =0$, due the fact that for such roots $ [Z^\mu\pm \overline Z^\mu,Z^\nu\pm \overline Z^\nu]=0,$ for all $Z^\mu\in\g^\mu$ and $Z^\nu\in\g^\nu$.
Next, let $P,\,P'\in \p[e_j+e_l] $ be elements of the basis given in Remark \[BASIS2\](b), arising from the same root $\mu\in \Delta^+$. Write $P=X-\theta X$, with $X=Z^\mu+\overline{Z^\mu}$, and $I_0P=Y-\theta Y$, with $Y=[Z_0,X]$. Likewise write $P'=X'-\theta X'$, with $X'=iZ^\mu-i\overline{Z^\mu}$, and $I_0P'=Y'-\theta Y'$, with $Y'=[Z_0,X']$.
From (\[FORMULONE\]) it follows that $$h_f(a_*P,a_*P )=
\textstyle -\frac{1}{\sinh {(a_j+a_l)} \sinh {(a_j-a_l)}}
\big (d^c f(\widetilde {[ C,K]}_z)\big ),$$ where $K=X+\theta X$ and $C=Y+\theta Y$. By Lemma \[BRACKETS\](a)(b), the above expression equals $$\textstyle -\frac{1}{\sinh {(a_j+a_l)} \sinh {(a_j-a_l)}}
\big (r d^c f(\widetilde {K^j}_{z})- sd^c f(\widetilde {K^l}_{z})\big )$$ $$\label{III}
= \textstyle \frac{1}{\sinh {(a_j+a_l)} \sinh {(a_j-a_l)}}
\big (r\sinh (2a_j) \frac{\partial \widetilde f}{ \partial a_j}(H)-
s\sinh (2a_l)\frac{\partial \widetilde f}{ \partial a_l}(H)\big )\,,\quad r,s\in\R.$$
In a similar way, one obtains $$\label{IIII} h_f(a_*P',a_*P')=h_f(a_*P ,a_*P ),
%$$= \textstyle \frac{1}{\sinh {(a_j+a_l)} \sinh {(a_j-a_l)}}
%\big (r\sinh (2a_j) \frac{\partial \widetilde f}{ \partial a_j}(H)-
% s'\sinh (2a_l)\frac{\partial \widetilde f}{ \partial a_l}(H)\big )\,,\quad r',s'\in\R.$$ and, from Lemma \[BRACKETS\](c), $$h_f(a_*P,a_*P')=0.$$ An argument similar to the one used in (ii) shows that the above expressions (\[III\]) and (\[IIII\]) smoothly extend to the hyperplane $(e_j-e_l)(H)=0$.
Moreover the quantities (\[III\]) and (\[IIII\]) are strictly positive for the strictly plurisubharmonic potential $\rho$ of the Killing metric of $G/K$ given in Proposition \[POTENTIALK\]. Then (iii) in Lemma \[CONVESS\] implies that $r=s>0$. Finally, as $h_\rho(a_*P,a_*P)= B(P,P)$, a simple computation shows that $r = B(P,P)/b$. This concludes the proof of (iii).
\(iv) [**The Hermitian form $h_f$ on $a_*\p[e_j]$**]{}.
Let $P,\, Q\in\p[e_j]$ be elements of the basis given in Remark \[BASIS2\](c), arising from roots $\mu,\, \nu\in \Delta^+$, respectively, with $\nu\not=\mu,\,\bar \mu$. Then $h_f(a_*P,a_*Q) =0$, due the fact that for such roots $ [Z^\mu\pm \overline Z^\mu,Z^\nu\pm \overline Z^\nu]=0,$ for all $Z^\mu\in\g^\mu$ and $Z^\nu\in\g^\nu$. In addition, by the $I_0$-invariance of $h_f$ one has $h_f(a_*P,a_*I_0P)=0.$
In order to compute $h_f(a_*P,a_*P)$, write $P=X-\theta X$ and $I_0P=Y-\theta Y$, with $X=Z^\mu+\overline{Z^\mu}$ and $Y=iZ^\mu-i\overline{Z^\mu}$ (Lemma \[COMPLEXSTRUCTURE\](c)). Then, from (\[FORMULONE\]) it follows that $$h_f(a_*P,a_*P)=\textstyle -\frac{1}{\sinh^2(a_j)}
d^c f(\widetilde{[C,K]}_z),$$ for $K=X+\theta X$ and $C=Y+\theta Y$. By Lemma \[BRACKETS\](d)(e), one obtains $$\textstyle
h_f(a_*P,a_*P)= -\frac{1}{\sinh^2(a_j)}d^c f({t\widetilde K^j}_{z}) = 2t\textstyle \coth (a_j)
\frac{\partial \widetilde f}
{ \partial a_j}(H),\qquad t\in\R.$$
Note that the above expression smoothly extends on $\mathcal D_\a$, since $\frac{\partial \widetilde f}
{ \partial a_j}=0$ on the hyperplane $a_j=0$. Moreover, by (i) of Lemma \[CONVESS\], for $a_j>0$ one has $\frac{\partial \widetilde f}
{ \partial a_j}> 0$.
As in the previous case, one shows that $t=B(P,P)/b$ by computing the above quantity for the strictly plurisubharmonic potential $\rho$ of the Killing metric of $G/K$ given in Proposition \[POTENTIALK\]. This completes the proof of statement (iv).
${\bf Remark.}$ The Levi form $L_f^\C$ of $f$ is given by $$L_f^\C(Z,\overline W)=2(h_f(X,Y)+ih_f(X,I_0Y)),$$ where $Z=X-I_0X$ and $W=Y-I_0Y$ are elements in $(\p^\C)^{1,0}$. One easily sees that $L_f^\C$ is (strictly) positive definite if and only if $h_f$ is (strictly) positive definite.
$K$-invariant psh functions vs. $W$-invariant logcvx functions {#PSH}
==============================================================
Let $\,G/K\,$ be an irreducible non-compact Hermitian symmetric space of rank $\,r$ and let $D\subset G/K$ be a Stein, $K$-invariant domain. The goal of this section is to prove a characterization of various classes of $K$-invariant plurisubharmonic functions on $D$ in terms of appropriate conditions of the corresponding functions on $\mathcal D_\a$ (see (\[DA\])). As an application, we reproduce the characterization of Stein $\,K$-invariant domains in $\,G/K$ (Cor.\[CHARACTER\]), outlined in [@BeDa91], Thm.$3^\prime$ and Thm.4 (see also [@FeHu93]).
In the smooth case we prove that a smooth $K$-invariant function $f$ of $D$ is strictly plurisubharmonic if and only if the associated function $\tilde f$ (see (\[EFFETILDE\])) satisfies a positivity condition arising from Proposition \[LEVI\].
Denote by $\,\Delta^r\,$ the orbit of the base point $eK\in G/K$ under the commuting $\,SL_2(\R)$’s generated by the triples defined in (\[NORMALIZ1\])-(\[KJPJ\]). It is well-known (cf. [@Wol72]) that $\Delta^r$ is biholomorphic to the unit polydisk in $\C^r$. One has $\,\Delta^r=T\exp \a K$, where $T\cong (S^1)^r$ is the $\,r$-dimensional torus in $\,K\,$ whose Lie algebra is generated by $\,K^1,\dots ,K^r$, and $$\exp (a_1,\ldots,a_r)K=(\tanh(a_1),\ldots,\tanh(a_1)),\quad\hbox{for $(a_1,\ldots,a_r)\in\a$}.$$ The polydisk $\Delta^r$ is a “thick slice" for the $K$-action in $G/K$, in the sense that $\,K \cdot\Delta^r =G/K.$ If $D$ is a $K$-invariant domain in $G/K$, then the Reinhardt domain associated to $D$ is defined as $$\,R:=D \cap \Delta^r \quad \quad {\rm and } \quad \quad D=K\cdot R\,.$$ We will show that if $D$ is Stein, then $R$ is necessarily connected. (It should be remarked that, despite its appellation, a Reinhardt “domain" is open in $\,\C^r\,$ but need not be connected).
For a Reinhardt domain $R$ in $\Delta^r$, define the set $\,\mathcal D = \{(a_1,\ldots, a_r) \in \R^r \ : \ (\tanh a_1,\ldots,\tanh a_r) \in R\},$ with the property that the image of the map $$\mathcal D \to R\, \quad \quad (a_1,\ldots, a_r) \to (\tanh a_1,\ldots,\tanh a_r)$$ coincides with $\,R \cap \R^r$. One has $R=T\cdot (R \cap \R^r),$ with $T\cong (S^1)^r$. Given a smooth $\,T$-invariant function $\,f\,$ on $\,R\,$ define $\,\tilde f: \mathcal D \to \R $ by $$\tilde f (a_1,\ldots, a_r) =f (\tanh a_1,\ldots,\tanh a_r)\,.$$ By the $T$-invariance of $f$, the function $\tilde f$ is $\,(\Z^2)^r$-invariant.
Denote by $LogConv^{\infty,+}(\mathcal D )^{(\Z_2)^r}$ the class of smooth functions on $\,\mathcal D \,$ which are even in each variable and such that the form defined in (ii) of Proposition \[LEVI\] is strictly positive definite. The next proposition characterizes $T$-invariant smooth strictly plurisubharmonic functions on $R$ by elements in $LogConv^{\infty,+}(\mathcal D )^{(\Z_2)^r}$. It is an intermediate step in the proof of the main theorem in the smooth case, but it may be of independent interest in the context of Reinhardt domains.
\[PSHPOLIDISC\] Let $\,f\,$ be a smooth $\,T$-invariant function on a Reinhardt domain $\,R\,$ in $\,\Delta^r$. Then $\,f\,$ is strictly plurisubharmonic if and only if $\,\tilde f\,$ belongs to $LogConv^{\infty,+}(\mathcal D)^{(\Z_2)^r}$.
In polar coordinates $\,(\rho_j, \theta_j)\,$, with $z_j=\rho_je^{i\theta_j}\not=0$, one has $$\,\textstyle \partial_{z_j}=
\frac{e^{-i\theta_j}}{2\rho_j}(\rho_j\partial_{\rho_j}
-i\partial_{\theta_j})\, \ \ \quad\ \ \, \partial_{\bar z_j}=
\frac{e^{i\theta_j}}{2\rho_j}(\rho_j\partial_{\rho_j}
+i\partial_{\theta_j})\,.$$ One easily sees that, for $z_jz_l\not=0$, $$\label{LEVIREIN}
\textstyle 4\frac{\partial^2 f}{\partial \bar
z_j \partial z_l} (z_1,\dots, z_r)
= \frac{1}{\rho_j}\frac{\partial f}{\partial \rho_j }(\rho_1,\dots,
\rho_r) \delta_{jl}
+e^{i(\theta_j-\theta_l)}\frac{\partial^2 f}{\partial \rho_j \partial
\rho_l}(\rho_1,\dots, \rho_r)\,.$$ The above quantity extends smoothly through the hyperplanes $z_j=0$ (and therefore to the whole domain) whenever $j=l$, while $$\textstyle 4\frac{\partial^2 f}{\partial \bar z_j \partial z_l}
(z_1,\dots, z_r)=0, \quad \hbox{ for $j\not=l$ and $z_jz_l=0$}.$$
For $\,\rho_1=\tanh a_1, \dots, \rho_r=\tanh a_r$, one has $$\label{DEREFFETILDE}
\textstyle
\frac{\partial \widetilde f}{\partial a_j }(a_1,\dots, a_r)=\frac{\partial
f}
{ \partial \rho_j } {\scriptstyle (\tanh a_1, \dots, \tanh a_r)}\frac{1}{\cosh^2
a_j}$$ $$\label{HESSIANEFFETILDE}\textstyle
\frac{\partial^2 \widetilde f}{\partial a_j \partial
a_l}(H)=\frac{\partial^2 f}
{ \partial \rho_j \partial \rho_l}{\scriptstyle (\tanh a_1, \dots, \tanh a_r)}
\frac{1}{\cosh^2 a_j \cosh^2 a_l}-\delta_{jl}
\frac{\partial f}
{ \partial \rho_j }{\scriptstyle (\tanh a_1, \dots, \tanh a_r)} \frac{2\sinh
a_j}{\cosh^3 a_j}\,,$$ and likewise $$\textstyle \frac{\partial^2 \widetilde f}{\partial a_j \partial a_l}(H)=0,\quad
\hbox{for $j\not= l$ and $a_ja_l=0$}.$$
A simple computation combining formulas (\[DEREFFETILDE\]) and (\[HESSIANEFFETILDE\]) with (\[LEVIREIN\]), shows that $\textstyle 4
\frac{\partial^2 f}{\partial \bar z_j \partial z_l}(z_1,\ldots,z_r)$ is given by $$\begin{cases}\textstyle \cosh^4 a_j \,
\big (2\coth (2a_j)\frac{\partial \widetilde f}
{ \partial a_j}(a_1,\dots,a_r) +\frac{\partial^2 \widetilde f}
{ \partial a_j^2 }(a_1,\dots,a_r) \big), ~\hbox{for $j=l,$}\\
\cosh^2 a_j\,e^{i\theta_j} \cosh^2 a_l\,e^{-i\theta_l}\, \frac{\partial^2 \widetilde f}
{ \partial a_j \partial a_l}(a_1,\dots,a_r), ~\hbox{for $j\not=l$ and
$z_jz_l\not=0,$}\\
0,~\hbox{for $j\not=l$ and $z_jz_l=0$.} \end{cases}$$
Then, for $(z_1,\ldots,z_r)\in R$, one has $$\textstyle \big (4 \frac{\partial^2 f}{\partial \bar z_j \partial
z_l}\big )_{j,l}=C \big ( \frac{\partial^2 \tilde f}{\partial a_j
\partial a_l}+\delta_{jl}\, 2\coth(2a_j)\frac{\partial \tilde f}{\partial
a_j } \big )_{j,l}\overline C,$$ where $C$ is the diagonal matrix with diagonal entries $$\textstyle c_{jj}= \begin{cases}\cosh^2(a_j)e^{i\theta_j}, ~~\hbox{for
$z_j\not=0,$}\\
\cosh^2(a_j), ~~\hbox{for $z_j=0$}.\end{cases}$$ It follows that $f$ is strictly plurisubharmonic if and only if $\,\tilde f\,$ belongs to the class $LogConv^{\infty,+}(\mathcal D)^{(\Z_2)^r}$.
Let $R$ be a Reinhardt domain in $ (\Delta^*)^r$ and let $$\,\mathcal D_{\log}:=\,\{(s_1,\, \dots , \,s_r) \in (\R^{<0})^r \ : \
(e^{s_1},\, \dots , \,e^{s_r}) \in R\,\}\,$$ be its logarithmic image. For a $T$-invariant function $f$ on $R$, define $\widehat f\colon{\mathcal D_{\log}}\to\R$ by $$\label{EFFEHAT}\widehat f(s_1,\ldots,s_r):=f(e^{s_1}, \ldots,e^{s_r}).%f\circ {\bf exp}(s_1,\ldots,s_r),$$ It is well known that if $f$ is smooth, then it is strictly plurisubharmonic if and only if $\widehat f$ has strictly positive definite Hessian. The next remarks elucidate the significance of the class $ LogConv^{\infty,+}(\mathcal D)^{(\Z_2)^r}$.
\[LOGCONV\] Let $R $ be a Reinhardt domain in $ (\Delta^*)^r$ and let $f$ be a smooth $T$-invariant function on $R$. Then $\tilde f$ belongs to $LogConv^{\infty,+}(\mathcal D)^{(\Z_2)^r}$ if and only if the smooth function $\hat f$ has everywhere strictly positive Hessian.
One has $$\label{DEREFFEHAT}\textstyle
\frac{\partial \hat f}{\partial s_j }(s_1,\dots,s_r)=\frac{\partial f}
{ \partial \rho_j }(e^{s_1},\dots, e^{s_r})e^{s_j}\,,$$ $$\label{HESSIANEFFEHAT}\textstyle
\frac{\partial^2 \hat f}{\partial s_j \partial s_l}(s_1,\dots,s_r)=\frac{\partial^2 f}
{ \partial \rho_j \partial \rho_l}(e^{s_1},\dots, e^{s_r})e^{s_j}e^{s_l} +\delta_{jl}
\frac{\partial f}
{ \partial \rho_j }(e^{s_1},\dots, e^{s_r}) e^{s_j}\,.$$ Then, by letting $\,e^{s_1}=\tanh a_1, \dots, e^{s_r}=\tanh a_r$, with $a_1,\ldots,a_r>0$, and combining formulas (\[DEREFFEHAT\]) and (\[HESSIANEFFEHAT\]) with (\[DEREFFETILDE\]) and (\[HESSIANEFFETILDE\]), one obtains $$\textstyle 2\coth (2a_j)\frac{\partial \widetilde f}
{ \partial a_j}(H) \delta_{jl}+\frac{\partial^2 \widetilde f}
{ \partial a_j \partial a_l}(H)\,=\textstyle \frac{4}{\sinh 2a_j \sinh 2a_l}
\frac{\partial^2 \hat f}{\partial s_j \partial s_l}(s_1,\dots,s_r)\,.$$ Hence $\tilde f \in LogConv^{\infty,+}(\mathcal D)^{(\Z_2)^r}$ if and only if $\,\hat f \,$ has everywhere strictly positive Hessian.
Let $R$ be an arbitrary Reinhardt domain and let $f|_{R \cap(\Delta^*)^r}$ denote the restriction of $f$ to $R \cap(\Delta^*)^r$. The strict positivity of the Hessian of $ f|_{R \cap(\Delta^*)^r}$ on $R \cap(\Delta^*)^r$ does not imply the strict plurisubharmonicity of $f$ on the coordinate hyperplanes (and therefore on the whole $R$). For instance, despite the fact that it has strictly positive Hessian on $R \cap(\Delta^*)^r$, the function $g(z)=|z|^4$ is not plurisubharmonic at $z=0$. In contrast, this fact is detected by the vanishing of the form $$\textstyle \frac{\partial^2 \widetilde g}
{ \partial a \partial a}+2\coth (2a)\frac{\partial \widetilde g}
{ \partial a}=
% 4\frac{\tanh^2 a}{\cosh^4 a}(3-2\sinh^2a) +2\coth(2a)4\tanh^3a\frac{1}{\cosh^2a}$$ 16,$$ at $a=0$, which shows that the associated function $\tilde g(a)= \tanh(a)^4$ does not belong to $LogConv^{\infty,+}(\R)^{(\Z_2)}$.
Let $R\subset \Delta^r$ be a Reinhardt domain associated to a $K$-invariant domain in $G/K$. In this case, $R$ is also invariant under coordinate permutations, which arise from the Weyl group action on $\a$. If such a Reinhardt domain is Stein, then there are two possibilities:
- $\,R\,$ intersects the coordinate hyperplanes. Then it is complete (cf. [@Car73], Thm. 2.12). In particular it contains the origin and is connected.
- $\,R\,$ does not intersects the coordinate hyperplanes, i.e. $\,R\subset (\Delta^*)^r\,$. Then $R$ is logarithmically convex.
The next proposition shows that a Stein Reinhardt domain $R$ associated to a Stein $K$-invariant domain $D\subset G/K$ is necessarily connected (even when $0\not\in R$), a fact already pointed out in [@BeDa91], Thm.$3'$.
\[MINIMUM\] Let $\,D\,$ and $R$ be as above and let $\,f: D \to \R\,$ be a smooth, $\,K$-invariant strictly plurisubharmonic exhaustion of $D$.
- If $\,R\,$ contains the origin, then $R$ is connected and $\,\tilde f\,$ has a unique minimum point at the origin of $\,\mathcal D_{\a}$.
- If $\,R\,$ does not contain the origin, then $\,\tilde f $ has a unique minimum point on the diagonal line $\,\{a_1= \dots =a_r\}\,$ of $\,\mathcal D_{\a}$. In particular $\,R\,$ is connected. In this case $\,G/K\,$ is necessarily of tube type.
The minimum set of a $K$-invariant exhaustion function $f$ of $D$ intersects $R$ in a non-empty $T$-invariant set. Since $R=T \cdot \exp \mathcal D_\a K$, a point $H\in \mathcal D_\a $ is a minimum of $\tilde f$ if and only if $\exp(H)K\in R$ is a minimum of $f|_R$, the restriction of $\,f\,$ to $\,R\,$.
\(i) We already observed that $R$ is connected. Assume that $\,\tilde f\,$ has a minimum point $\,H=(a_1,\, \dots , \,a_r)\,$, different from the origin. Then the restriction $\,f|_R\,$ of $\,f\,$ to the Reinhardt domain $\,R\,$ has a minimum point in $\,P=\exp (H)K$. For $\,\varepsilon\,$ small enough there is a holomorphic immersion $$\, \iota:\Delta_{1+\varepsilon} \to R, \ \quad \quad z \to zP\,$$ from the disc of radius $\,1+ \varepsilon\,$ to $\,R$. The pull-back $\,f \circ \iota\,$ of $\,f\,$ via $\,\iota\,$ is a smooth strictly subharmonic $\,S^1$-invariant function. Hence it has a minimum point in $\,0\,$ and, by construction, in $\,1$. It follows that $\,f \circ \iota\,$ is necessarily constant, contradicting the fact that it is strictly subharmonic. (ii) Let $\,H=(a_1,\, \dots , \,a_r)\,$ be a minimum point of $\,\tilde f\,$. In this case, all $a_j$’s are different from $\,0\,$. As a consequence $\,\textstyle 2\coth(a_j)
\frac{\partial \widetilde f}
{ \partial a_j}(H)=0\, $, for $j=1,\ldots, r$. By $\,(iv)\,$ of Proposition \[LEVI\], in the non-tube case this contradicts the strict plurisubharmonicity of $f$, implying that the space $\,G/K\,$ is necessarily of tube type. The strict plurisubharmonicity of $\,f\,$ along with $\,(iii)\,$ of Proposition \[LEVI\], also implies that $\,a_j=a_k\,$ for every $\,j,k=1, \dots, r$. Hence $H$ lies on the diagonal of $\a$. The uniqueness of the minimum follows from standard arguments as in [@AzLo93] or by the the following direct argument.
Recall that $\,\mathcal D_{\log} \,$ is convex by the Steinness of $D$. By Remark \[LOGCONV\], the associated function $\,\hat f \,$ has everywhere strictly positive definite Hessian. In particular its restriction to the diagonal $\,\mathcal D_{\log} \cap
\{s_1= \dots =s_r\}\,$ is a strictly convex exhaustion function. Hence it has a unique minimum, implying that $\,\tilde f$ has a unique minimum on $\,\mathcal D_{\a} \cap
\{a_1= \dots =a_r\}\,.$
Consider the following classes of functions:
- $C^0(\mathcal D_\a)^W$: continuous $\,W$-invariant functions on $\,\mathcal D_\a$,
- $C^{\infty}(\mathcal D_\a)^W$ : smooth $\,W$-invariant functions on $\,\mathcal D_\a$,
- $C^0(D)^K$: continuous $\,K$-invariant functions on $\,D$,
- $C^{\infty}(D)^K$: smooth $\,K$-invariant functions on $\,D$.
Since the $K$-action on $D$ is proper and every $K$-orbit intersects the slice $\exp \mathcal D_\a K$ in a $W$-orbit, it is easy to check that the map $f \to \tilde f$ is a bijection from $C^0(D)^K$ onto $C^0(\mathcal D_\a)^W$. By Theorem 4.1 in [@Fle78] (see also [@Dad82]) such a map is also a bijection from $C^\infty(D)^K$ onto $C^\infty(\mathcal D_\a)^W$. Define
- $LogConv^{\infty,+}(\mathcal D_\a)^W$: smooth, $\,W$-invariant functions on $\,\mathcal D_\a\,$ such that the form defined in (ii) of Proposition \[LEVI\] is strictly positive definite,
- $P^{{\infty},+}(D)^K$: smooth, $\,K$-invariant, strictly plurisubharmonic functions (i.e.with strictly positive definite Levi form) on $\,D$.
Our first result is the following theorem.
\[BIJECTIVEK1\] Let $\,D\,$ be a Stein $\,K$-invariant domain in an irreducible non-compact Hermitian symmetric space $\,G/K\,$ of rank $\,r$. Then $\,f\,\in P^{{\infty},+}(D)^K$ if and only if $\,\tilde f \in LogConv^{\infty,+}(\mathcal D_\a)^W$.
By (ii) of Proposition \[LEVI\], if $\,f\,$ is strictly plurisubharmonic on $D$, then $\,\tilde f \in LogConv^{\infty,+}(\mathcal D_\a)^W$.
Conversely, assume that $\,\tilde f \in LogConv^{\infty,+}(\mathcal D_\a)^W$ and $\,r>1\,$. We need to show that the terms in (iii) and (iv) of Proposition \[LEVI\] are strictly positive, the ones in (iv) occurring only in the non-tube case.
For the terms in (iii), without loss of generality, it is sufficient to consider the case $r=2$, and $H=(a_1,a_2) \in \a^+$, where $a_1\geq a_2 \geq 0$. Assume first $\,a_1> a_2 > 0$. Then $(\tanh a_1,\tanh a_2)=(e^{s_1},e^{s_2}) \in R^*$, where $R$ is the Reinhardt domain associated to $D$. Let $d_0<0$ and $t_0>0$ be real numbers defined by $\,(s_1,s_2)=(d_0+t_0,d_0-t_0)$.
Denote by SSC (smooth stably convex) those smooth functions with everywhere positive definite Hessian. The function $\,\hat f\,$, which is invariant under coordinate permutations, is SSC by Remark \[LOGCONV\]. Therefore $\,g(t):=\hat f(d_0+t,d_0-t)\,$ is even and SSC. Consequently, for $\,t_0>0$, the inequality $$g'(t_0)= \textstyle \frac{\partial f}{ \partial \rho_1}(e^{d_0+t_0},e^{d_0-t_0})e^{d_0+t_0}-
\frac{\partial f}{ \partial \rho_2}(e^{d_0+t_0},e^{d_0-t_0})e^{d_0-t_0}>0\,,$$ holds true, which combined with formulas (\[DEREFFETILDE\]) yields the desired result. $$g'(t_0)= \textstyle \frac{1}{2}
\big ( \sinh (2a_j)\frac{\partial \widetilde f}{ \partial a_j}(H)-
\sinh (2a_l)\frac{\partial \widetilde f}{ \partial a_l}(H)\big )>0\,.$$
Next we need to estimate the terms (iii) of Proposition \[LEVI\], when $H$ lies on the boundary of the Weyl chamber $ \a^+$. Consider $\,H=(a,a)\,$, with $\,a \not=0$. Set $\tanh a= e^{d_0}$ and $(\tanh a_1,\tanh a_2)=(e^{d_0+t},e^{d_0-t})$, and recall that $\,g'(0)=0$. Then the corresponding term in (iii) of Proposition \[LEVI\] is the limit $$\textstyle \frac{1}{2}\lim_{t \to 0} \frac{g'(t)}{\sinh(a_1+a_2)\sinh(a_1-a_2)}
=\frac{1}{4}\lim_{t \to 0} \frac{g'(t)}{t} \frac{s_1-s_2}{\sinh(a_1+a_2)\sinh(a_1-a_2)}$$ $$=\textstyle \frac{1}{4} \frac{g''(0)}{\sinh(2a)\cosh(2a)} \lim_{a_1-a_2 \to 0} \frac{\log \tanh a_1-\log \tanh a_2}{ a_1-a_2}=g''(0)c(a)\,,$$ which is positive since $\,c(a)\,$ is a positive real number and $g''(0)>0$ ($g$ is even and SSC).
If $H=(a_1,0) \in \mathcal D_\a$, with $\,a_1> 0$, then the Reinhardt domain $R$ associated to $D$ is necessarily complete and the term to be evaluated reduces to $$\textstyle \frac{1}{\sinh^2 a_1}
\sinh (2a_1)\frac{\partial \widetilde f}{ \partial a_1}(a_1,0)\,.$$ Moreover $$\textstyle 2\coth (2a_1)\frac{\partial \widetilde f}
{ \partial a_1}(a_1,0)+\frac{\partial^2 \widetilde f}
{ \partial a_1^2}(a_1,0)>0,$$ implying that the function $s_1 \to f(e^{s_1},0)$ is SSC (cf.Rem.\[LOGCONV\]). Since $R$ is complete, then $\lim_{s_1 \to -\infty} f(e^{s_1},0)$ is finite. As a consequence $s_1 \to f(e^{s_1},0)$ is strictly increasing and so is $a_1 \to \tilde f(a_1,0)= f(\tanh a_1,0)$. Hence $\frac{\partial \widetilde f}{ \partial a_1}(a_1,0)\,$ is positive, as wished.
Finally note that for $a_1=a_2=0$ the analytic extension of our term is given by $$\textstyle 2 \frac{\partial^2 \widetilde f}{ \partial a_1^2}(0,0)
=2 \frac{\partial^2 \widetilde f}{ \partial a_2^2}(0,0)\,,$$ which is strictly positive by assumption.
We are left to examine the terms in (iv), which only appear in the non-tube case. Our arguments are similar to the ones used above. By Proposition \[MINIMUM\], the Reinhardt domain $R$ associated to $D$ is complete. Then $\,\lim_{s_j \to -\infty } \hat f(s_1,\dots,s_j,\dots\,s_r)$ is finite. Since $\, \hat f\,$ is SSC, the function $\,s_j \to \hat f(s_1,\dots,s_j, \dots \,s_r)\,$ is strictly increasing and so is $\,a_j
\to \tilde f(a_1,\dots,a_j, \dots \,a_r)$. Hence $$\textstyle 2\coth(a_j)
\frac{\partial \widetilde f}
{ \partial a_j}(a_1,\dots a_r)>0,\quad \hbox{ for $a_j>0$}.\,$$ The limit $$\,\lim_{a_j \to 0} \textstyle 2\coth(a_j)
\frac{\partial \widetilde f}
{ \partial a_j}(a_1,\dots,a_j,\dots,a_r)=2
\frac{\partial^2 \widetilde f}
{ \partial a_j^2}(a_1,\dots,0,\dots,a_r)\,$$ is strictly positive as well, by assumption.
Consider the ($T \ltimes \mathcal S_r$)-action on $\Delta^r$, where $\mathcal S_r$ denotes the group of coordinate permutations. As a consequence of Proposition \[PSHPOLIDISC\] one has the following corollary.
\[PSHONR\] Let $\,D\,$ be a Stein $\,K$-invariant domain in an irreducible non-compact Hermitian symmetric space $\,G/K\,$ and let $\,R\,$ be the associated Reinhardt domain. The map $f\to f|_R$ is a bijection between $\,P^{{\infty},+}(D)^K\,$ and $\,P^{{\infty},+}(R)^{T \ltimes \mathcal S_r}$.
\[PSHREINAHARDT\] If $\,R\,$ does not contain the origin, then, by Remark \[LOGCONV\], the condition $\,f\,\in P^{{\infty},+}(D)^K$ is also equivalent to requiring that the smooth invariant function $\,\hat f\,$ has strictly positive definite Hessian on $\,\mathcal D_{log}$.
\[CHARACTER\] $($see [@BeDa91], Thm.$3'$ and Thm.4$)$ Let $\,D\,$ be a Stein $\,K$-invariant domain in an irreducible non-compact Hermitian symmetric space $\,G/K\,$ and let $\,R\,$ be the associated Reinhardt domain. Then
- If $G/K$ is of tube type, then $\,D\,$ is Stein if and only if $\,R\,$ is Stein and connected.
- If $G/K$ is not of tube type, then $\,D\,$ is Stein if and only if $\,R\,$ is Stein and complete. In particular $R$ contains the origin and is connected.
By Proposition \[MINIMUM\], if $\,D\,$ is Stein then the intersection $\,R =D \cap \Delta^r\,$ is Stein, connected and, in the non-tube case, complete. Conversely, let $\,R\,$ be a Stein, connected Reinhardt domain, invariant under coordinate permutations which, in the non-tube case, is also assumed to be complete. Let $f$ be a smooth, strictly plurisubharmonic exhaustion function $\,f\,$ of $\,R$. By averaging, $\,f\,$ may be assumed to be invariant with respect to $\,T\,$ and to coordinate permutations. Proposition \[PSHPOLIDISC\] implies that the function $\ \tilde f:\mathcal D_\a \to \R\,$, associated to $f$, belongs to $\,LogConv^{\infty,+}(\mathcal D_\a)^{W}$. By Theorem \[BIJECTIVEK1\], $\ \tilde f$ extends to a smooth, $K$-invariant, strictly plurisubharmonic exhaustion function of $\,D$. Hence $D$ is Stein.
\[SCHLICHT\] [ The envelope of holomorphy of a $K$-invariant domain $D$ in $G/K$ is described, without proof, in terms of the associate Reinhardt domain $R$ in Theorem 5 of [@BeDa91]: ]{}
if $G/K$ is of tube type, then $\widehat D= K\cdot \widetilde R$, where $\widetilde R$ is the smallest connected Stein, Reinhardt domain containing $R$;
if $G/K$ is not of tube type, then $\widehat D= K\cdot \widetilde R$, where $\widetilde R$ is the smallest connected and complete Stein, Reinhardt domain containing $R$.
Our next goal is to extend the characterization of smooth, $K$-invariant, strictly plurisubharmonic functions on $D$ obtained in Theorem \[BIJECTIVEK1\] to some wider classes of $K$-invariant functions. Namely:
- $P(D)^K$: plurisubharmonic, $\,K$-invariant functions on $\,D$,
- $P^{\infty}(D)^K$: smooth, plurisubharmonic, $\,K$-invariant functions on $\,D$,
- $P^+(D)^K$: functions which, on every relatively compact $K$-invariant domain $\,C\,$ in $\,D\,$, are the sum $\,g\,+\,h\,$, for some $g \in P(C)^K$ and $h \in P^{\infty,+}(C)^K$.
In order to do that we need to define the appropriate classes of functions on the associated domain $\mathcal D_\a$:
- $LogConv(\mathcal D_\a, [-\infty, \infty))^W$: limits of decreasing sequences in $LogConv^{\infty,+}(\mathcal D_\a)^W$,
- $LogConv^{\infty}(\mathcal D_\a)^W$: smooth functions in $LogConv(\mathcal D_\a, [-\infty, \infty))^W$,
- $LogConv^+(\mathcal D_\a, [-\infty, \infty))^W$: functions which, on every relatively compact $W$-invariant domain $\,\mathcal C\,$ of $\mathcal D_\a$, are the sum $\,\tilde g\,+\,\tilde h\,$ for some $\tilde g \in LogConv( \mathcal C, [-\infty, \infty))^W$ and $\tilde h \in LogConv^{\infty,+}(\mathcal C)^W$.
\[PROPCLASSES\] $(i)$ The class $\,LogConv^{\infty}(\mathcal D_\a)^W\,$ coincides with the family of smooth $\,W$-invariant functions on $\,\mathcal D_\a$ for which the form in $(ii)$ of Proposition \[LEVI\] is non-negative. One inclusion is clear. Conversely, if $\,\tilde f\,$ is smooth and the form in $(ii)$ of Proposition \[LEVI\] is non-negative, then $\,\tilde f\,$ is the limit of the sequence $\,
\tilde f_n(a_1,\dots a_r)= \tilde f(a_1,\dots a_r)+ \frac{1}{n}\sum a_j^2$. Hence $\tilde f$ belongs to $LogConv^{\infty}(\mathcal D_\a)^W$. In particular $$LogConv^{\infty,+}(\mathcal D_\a)^W
\subset LogConv^{\infty}(\mathcal D_\a)^W\,.$$
$(ii)$ The class $P^+(D)^K$ coincides with the family of functions which are locally the sum of some $\,g\,$ plurisubharmonic and $\,h\,$ smooth strictly plurisubharmonic, i.e. the strictly plurisubharmonic functions according to [@Gun90], Def.$\,$1, Sect. L, p. 118. Indeed, assume that $f$ is strictly plurisubharmonic according to such definition. Choose a $K$-invariant, smooth strictly plurisubharmonic function $\psi$ on $D$ and let $C$ be a relatively compact $K$-invariant domain of $D$. Then there exists $\varepsilon>0$ such that $f-\varepsilon \psi$ is plurisubharmonic on $C$. That is, $\,f= g +\varepsilon \psi$, with $g$ psh and $K$-invariant on $C$.
The following lemma shows that all functions in the above classes are continuous.
Let $R$ be a Reinhardt domain.
- Any $T$-invariant plurisubharmonic function $f$ on $R$ is continuous. Its pluripolar set is the union of the intersections of $R$ with some coordinate subspaces.
- The class $LogConv(\mathcal D_\a, [-\infty, \infty))^W$ is contained in $\ C^0(\mathcal D_\a, [-\infty, \infty))^W$.
\(i) First consider the case $r=1$. On $R^*=R \setminus \{0\}$ one has $f(z)=\hat f(\log |z|)$, with $\hat f$ convex. Hence the restriction of $f$ to $R^*$ is continuos. If $0 \in R$, by the upper semicontinuity of subharmonic functions, one has $f(0)= \limsup_{z \to 0}f(z)$. Assume by contradiction that $$\liminf_{z \to 0}f(z)<f(0) \,.$$ Then there exists $z_1 \in R$ close to the origin such that $f(z_1) <f(0)$. By the submean value property and the $S^1$-invariance of $f$ one has $$\textstyle f(0)\leq\frac{1}{2\pi} \int_0^{2\pi} f(e^{i\theta}z_1)d\theta =f(z_1)<f(0)\,,$$ which is a contradiction.
For $r=2$, an argument analogous to the above one shows that $f$ is continuos on $R^*= R \cap (\Delta^*)^r$. We now prove continuity on the coordinate hyperplanes $\{(z,w) \in R \ : \ zw=0\,\}$ (on each hyperplane it can be constant and equal to $-\infty$). Assume by contradiction that there exists $(z_0,0) \in R$ such that
$$\label{ASSURDO}
\limsup_{(z,w) \to (z_0,0)}f(z,w)- \liminf_{(z,w) \to (z_0,0)}f(z,w)> \varepsilon >0\,.$$
Since $ \limsup_{(z,w) \to (z_0,0)}f(z,w)=f(z_0,0)$ by plurisubharmonicity and $f$ is continuous on the hyperplane $w=0$, there exists a neighborhood $B$ of $z_0$ in $\C$ such that $$\limsup_{(z,w) \to (z_0,0)}f(z,w)- \varepsilon/2=f(z_0,0)-\varepsilon/2<f(\zeta,0)\,,$$ for every $\zeta \in B$. By (\[ASSURDO\]), we can choose $(\zeta_1,w_1)$ close to $(z_0,0)$ such that $\zeta_1 \in B$ and $$f(\zeta_1,w_1)< \liminf_{(z,w) \to (z_0,0)}f(z,w)+\varepsilon/2 <\limsup_{(z,w) \to (z_0,0)}f(z,w)- \varepsilon/2.$$ Then, by the submean value property for subharmonic functions and by the $T$-invariance of $f$ one has $$\textstyle
\limsup_{(z,w) \to (z_0,0)}f(z,w)- \varepsilon/2<f(\zeta_1,0)\leq
\frac{1}{2\pi} \int_0^{2\pi} f(\zeta_1,e^{i\theta}w_1)d\theta =$$ $$=f(\zeta_1,w_1)<\limsup_{(z,w) \to (z_0,0)}f(z,w)- \varepsilon/2\,.$$ giving a contradiction.
The above argument also shows that the pluripolar set of $f$ consists of either the origin, or the intersection of $R$ with one of the the coordinate lines $\,\{z=0\}$, $\,\{w=0\}\,$ or with both of them.
Now we can proceed inductively and obtain the statement for $\,r>2$.
\(ii) By Theorem \[BIJECTIVEK1\], to a decreasing sequence $\tilde f_n$ of functions in $LogConv^{\infty,+}(\mathcal D_\a)^W$ there corresponds a decreasing sequence $f_n$ in $P^{\infty,+}(D)^K$, whose limit $f$ necessarily belongs to $P(D)^K$. The restriction $f|_R$ of $f$ to $R$ is a plurisubharmonic $T$-invariant function. So (i) implies that $f|_R$ is continuous and consequently so is the corresponding $\tilde f$ in $LogConv(\mathcal D_\a, [-\infty, \infty))^W$, which is the limit of the $\tilde f_n$.
Summarizing, the following inclusions hold true
$$\begin{matrix}
LogConv^+(\mathcal D_\a, [-\infty,\infty))^W & \subset & LogConv(\mathcal D_\a, [-\infty,\infty))^W &
\subset &C^0(\mathcal D_\a,[-\infty,\infty))^W \cr
\cup & & \cup & & \cup \cr
LogConv^{{\infty},+}(\mathcal D_\a)^W & \subset & LogConv^{\infty}(\mathcal D_\a)^W & \subset &C^\infty(\mathcal D_\a)^W
\end{matrix}$$
and our complete result is stated in the next theorem.
\[BIJECTIVEK\] Let $D$ be a Stein $K$-invariant domain in an irreducible non-compact Hermitian symmetric space $G/K$. The map $f\to \tilde f$ is a bijection between the following classes of functions
- $\,P^{{\infty},+}(D)^K\,$ and $\ \ \,LogConv^{{\infty},+}(\mathcal D_\a)^W$,
- $\,P\,(D)^K\,\ \ \ \,$ and $\, \ \ \,LogConv(\mathcal D_\a, [-\infty,\infty))^W$,
- $\,P^{\infty}(D)^K\,\ \ $ and $\ \ \,LogConv^{\infty}(\mathcal D_\a)^W$,
- $\,P^{+}(D)^K\,\ \ $ and $\ \ \,LogConv^+(\mathcal D_\a, [-\infty,\infty))^W$.
In particular $K$-invariant plurisubharmonic functions on $D$ are continuous.
\(i) is the content of Theorem \[BIJECTIVEK1\]. By averaging over $K$, a $K$-invariant, plurisubharmonic function on $\,D\,$ is the decreasing limit of smooth $K$-invariant, strictly plurisubharmonic functions (cf.[@Gun90],Sect.K). Then the assert (ii) follows from (i). As smooth $K$-invariant functions on $D$ correspond to smooth $W$-invariant functions on $\,\mathcal D_\a$, the previous argument also proves statement (iii). Finally (iv) follows from the definitions of $LogConv^{+}(\mathcal D)^W\,$ and $P^{+}(D)^K$, by averaging all the involved functions over $K$.
Let $T \ltimes \mathcal S_r$ act on $\Delta^r$ as in Corollary \[PSHONR\]. The previous theorem can be reformulated as follows.
\[PLURIREIN\] Let $\,D\,$ be a Stein $\,K$-invariant domain in an irreducible non-compact Hermitian symmetric space $\,G/K\,$ and let $\,R\,$ be the associated Reinhardt domain. The map $f\to f|_R$ is a bijection between
- $\,P^{{\infty},+}(D)^K\,$ and $\,\ \ \ P^{{\infty},+}(R)^{T \ltimes \mathcal S_r}$,
- $\,P\,(D)^K\,\ \ \ \,$ and $\,\ \ \ P\,(R)^{T \ltimes \mathcal S_r}$,
- $\,P^{\infty}(D)^K\,\ \ $ and $\,\ \ \ P^{\infty}(R)^{T \ltimes
\mathcal S_r}$,
- $\,P^{+}(D)^K\,\ \ $ and $\,\ \ \ P^{+}(R)^{T \ltimes
\mathcal S_r}$.
Appendix: a $K$-invariant potential for the Killing metric: the logarithm of the Bergman kernel function. {#OMEGAK}
=========================================================================================================
Let $G/K$ be an irreducible non-compact Hermitian symmetric space. The Killing form $B$ of $\g$, restricted to $\p$, induces a $G$-invariant Kähler metric $h$ on $G/K$. In this section we exhibit a $K$-invariant potential $\,\rho\,$ of the Killing metric $\, h$ in a Lie theoretical fashion. As such a $K$-invariant potential is unique up to an additive constant (Rem.\[UNIQUE\]), then, up to an additive constant, it coincides with the logarithm of the Bergman kernel function (Cor.\[BERGMAN\]).
In order to define $\,\rho$, recall the decomposition $\,G=K \exp \a\, K\,$ and write an element of $\,G/K\,$ as $\,kaK$, where $\,k \in K\,$ and $\,a=\exp H\,$, with $H=\sum_ja_jA_j\in\a$.
\[POTENTIALK\] Let $\,\widehat\rho\,$ be a real valued function satisfying $\, \widehat\rho\,'(t)= \frac{\cosh t-1}{\sinh t} \, $. Then
[$(i)$]{} the $\,K$-invariant function $\,\rho:G/K \to \R\,$ defined by $$\, \textstyle \rho(ka K):= \textstyle \frac{1}{4}{\sum_{j=1}^r}
\widehat\rho(2a_j)B(A_j,\,A_j)\,,$$ is a potential of the Killing metric $\,h$;
[$(ii)$]{} the moment map $\,\mu: G/K \to \k^*\,$ associated with $\,\rho\,$ is given by $$\,\mu(kaK)(X) =\textstyle {\frac{1}{2}\sum_{j=1}^r}
\sinh (2a_j)\widehat\rho\,'(2a_j)B (\Ad_{k^{-1}} C,K^j)\,,$$ where $X\in\k$.
\(ii) Resume the notation of Proposition \[LEVI\]. It was shown in the proof Proposition \[LEVI\] that for $z=aK$ one has $$\,d^c \rho (\widetilde C_z)=0\,,$$ for all $C\in \m \oplus \bigoplus_{\alpha\in \Sigma^+\atop\alpha\not=2e_j}\k[\alpha]$. Let $K^j\in\k[2e_j]$ be defined as in (\[KJPJ\]). From (\[UNO\]) it follows that $$\label{MOME}
d^c \rho ( \widetilde{ K^j}_z)=
\textstyle -\frac{1}{2}{\sinh (2a_j)}\widehat\rho\,'(2a_j)B(A_j,\,A_j)\,.$$ As $B(A_j,A_j)=-B(K^j,K^j)$, then (\[MOME\]) and (\[DCf\]) imply (ii).
\(i) Define $h_\rho(\,\cdot\,,\,\cdot\,)=-dd^c \rho (\,\cdot\,,I_0\,\cdot\,)$. Because of the invariance of $\rho$ and of the orthogonality relations proved in Proposition \[LEVI\], it is sufficient to prove that $
h_\rho (a_* P ,a_* Q)= B(P,Q)
$ for $P$, $Q$ both in one of the blocks $a_*\a$, $a_*\p[e_j+e_l]$ and $a_*\p[e_j]$.
[**The Hermitian form $h_\rho$ on $a_*\a$.**]{} Let $A_j, A_l \in\a$, be as in (\[NORMALIZ1\]). Then, by (\[TILDEFIELDS2\]) and (\[DIFFER\]), $$h_\rho(A_j,A_l)=- dd^c\rho(a_*A_j
,\,a_*I_0A_l)=-dd^c\rho(a_*P^l,a_*A_j)$$ $$\textstyle
= {1\over \sinh
(2a_l)}dd^c\rho((\widetilde{K^l})_z,(\widetilde{A_j})_z)=
-{1\over \sinh(2a_l)} \dds \mu^{K^l }(\exp sA_j \, z),$$ where $P^l \in\p[2e_l]$ and $K^l\in \k[2e_l]$ are defined in (\[KJPJ\]). By (ii) and (\[MOME\]) such quantity vanishes if $l\not=j$. For for $j=l$, it becomes $$\textstyle
{1\over \sinh (2a_j)}{1\over 2} \dds
{\sinh (2a_j+2s)}\widehat\rho\,'(2a_j+2s))B(A_l,\,A_l)=$$ $$\textstyle
{1\over \sinh (2a_j)} \big(
{\cosh (2a_j)}\widehat \rho\,'(2a_j)+
{\sinh(2a_j)}\widehat \rho\,''(2a_j)
\big)
B(A_l,\,A_l)=B(A_l,\,A_l)\,,$$ where the last equality follows from the assumption $\, \widehat \rho'(t)= \frac {\cosh t-1}{\sinh t}$.
Recall that if $\alpha \in \Sigma^+\setminus\{2e_j\}$, then $I_0\p[\alpha]=\p[\beta]$, for some $\beta \not=0$. Let $P \in \p[\alpha]$, with $\alpha$ as above. Write $\,P=X^\alpha-\theta X^\alpha\in\p[\alpha],$ and $\,Q=I_0P =X^\beta-\theta
X^\beta \in \p[\beta]$. Define $K:=X^\alpha+\theta X^\alpha \in \k[\alpha]$ and $C
:=X^\beta+\theta X^\beta \in \k[\beta]$. As $K^j=[I_0A_j,A_j]$, by (\[FORMULONE\]) and (ii), $$h_\rho(a_*P,a_*P )= \textstyle
-\frac{1}{ \sinh \alpha(H) \sinh \beta(H)}{1\over 2} \sum_k {\sinh (2a_k)} \widehat\rho\,'(2a_k)
B \big ([C,K] ,\,[I_0A_k,\,A_k]\big)\,=$$ $$\textstyle
= - \frac{1}{ \sinh \alpha(H) \sinh \beta(H)}{1\over 2} \sum_k \sinh (2a_k) \widehat\rho\,'(2a_k)
B \big (K ,\,[[I_0A_k,A_k],\,C ] \big)\,.$$
From the Jacobi identity, one has $$B \big (K ,\,[[I_0A_k,A_k],\,C ] \big)=-B\big (K ,\,[[C
,\,I_0A_k],\,A_k]+[[A_k,\,C ],\,I_0A_k] \big)=$$ $$=B\big ([A_k,\,K] ,\,[I_0A_k,\,C]\big)-
B\big ([I_0A_k,\,K],\, [A_k,\,C] \big)=$$ $$\alpha(A_k)B\big (P,\,I_0[A_k,\,C]\big)-
\beta(A_k)B\big (I_0[A_k,\,K],\, Q \big)=$$ $$=\big (\alpha(A_k)\beta(A_k)+
\beta(A_k) \alpha(A_k)\big) B\big (P ,\, I_0Q \big)\,.$$ As $I_0Q=-P$, one obtains $$h_\rho(a_*P,a_*P)=
\textstyle \frac{1}{2 \sinh \alpha(H)\sinh \beta(H)}\sum_k
\sinh (2a_k)\widehat\rho\,'(2a_k)\big(
\alpha(A_k)\beta(A_k)+\beta(A_k)\alpha(A_k)\big) B(P ,\,P).$$
We are left to check the following cases.
[**The Hermitian form $h_\rho$ on $a_*\p[e_j+ e_l] $.**]{} Here $\, \alpha =e_j+e_l\,$ and $\, \beta=e_j-e_l$. Then for $P\in\p[e_j+e_l]$, one has $$h_\rho(a_*P,a_*P)=
\textstyle
{1\over {2\sinh (a_j+a_l) \sinh (a_j-a_l)}} \big ( \sinh (2a_j)\widehat\rho\,'(2a_j)
- \sinh (2a_l)\widehat\rho\,'(2a_l)
\big )
B( P ,\,P)=$$ $$\textstyle
\frac{ \cosh (2a_j)-\cosh(2a_l)}{2\sin(a_j+a_l) \sin (a_j-a_l)}B( P ,\, P
)=B( P ,\, P ),$$ due to the identity $\,\cosh (2a_j)-\cosh (2a_l)=2\sinh (a_j+a_l)\sinh(a_j-a_l)$.
[**The Hermitian form $h_\rho$ on $a_*\p[e_j]$.**]{} Here $\, \alpha=\beta =e_j$. Then for $P\in\p[e_j]$, one has $$h_\rho(a_*P,a_*P)=
\textstyle
{1\over {2\sinh^2( a_j)}} \sinh (2a_j)\widehat\rho\,'(2a_j)B(P ,\,P )$$ $$\textstyle= {1\over { 2\sinh^2( a_j)}}(\cosh (2 a_j)-1)B( P ,P )=B( P ,P ).$$ This concludes the proof of (i).
The following remark shows that, up to an additive constant, the $K$-invariant potential $\rho$ of the Killing metric $h$ is unique.
\[UNIQUE\] Let $\rho_1$ and $\rho_2$ be smooth $K$-invariant functions on $G/K$ such that $dd^c\rho_1=dd^c\rho_2$. Then $\rho_1-\rho_2$ is constant.
As $\rho_1-\rho_2$ is pluriharmonic and $G/K$ is contractible, there exists a unique holomorphic function $f : G/K\to \C$ such that Re$f=\rho_1-\rho_2$ (cf. [@Gun90], Sect. K). By $K$-invariance, one has Re$f\,=$Re$f\,\circ \, k $, for every $k \in K$, and $f \circ k - f \equiv \lambda(k)$, where $\lambda(k)$ is a constant in $i\R$. Moreover, given $h,\ k \in K$ and $z \in G/K$, one has $$f (z)+\lambda(hk)=f((hk) \cdot z)=f(k \cdot z)+\lambda(h)
=f(z)+\lambda(h) +\lambda(k)\,.$$ Hence the map $\lambda: K \to i\R$ is a Lie group homomorphism and it is necessarily trivial by the compactness of $K$. Thus $f$ is $K$-invariant and it is is also invariant with respect to the induced local $K^\C$-action on $G/K$. Since $K^\C$ acts locally transitively on an open subset of $G/K$ (cf. [@Wol72]), the holomorphic function $f$ is constant and so is its real part $\rho_1-\rho_2$.
Since the logarithm of the Bergman kernel function is a $K$-invariant potential of the Killing metric (see [@KoNo69], Vol.2, Exa. 6.6 p. 162 and Thm. 9.6 p. 262), one can draw the following conclusion.
\[BERGMAN\] Up to an addictive constant, the smooth $K$-invariant exhaustion function $\rho$ coincides with the logarithm of the Bergman kernel function.
\[DISC\] As an example, consider the unit disc $\Delta=G/K, $ where $G=SU(1,1)$ acts on $\,\Delta\,$ by linear fractional transformations. Fix the basis of $\g$, normalized as in $(\ref{KJPJ}):$ $$K^1= \begin{pmatrix}
i &\ 0 \cr
0 & -i
\end{pmatrix} , \quad A_1= \begin{pmatrix}
0 & 1 \cr
1 & 0
\end{pmatrix}, \quad P^1= \begin{pmatrix}
0 & -i \cr
i & \ 0
\end{pmatrix}.$$ Then $\, \exp a_1A_1 K = \tanh a_1 = |z|$. Choose $\widehat \rho(t)= -\ln \frac{1}{\cosh t +1}$, satisfying $\, \widehat\rho\,'(t)= \frac{\cosh t-1}{\sinh t} \, $. Since $B(A_1,A_1)=8$, then up to an addictive constant, the logarithm of the Bergman kernel function is given by $$\textstyle \rho(\exp a_1A_1 K)=-
\frac {1}{4}\log \frac{1}{\cosh 2a_1 +1} B(A_1,A_1)=$$ $$\textstyle
-2\log \frac {\cosh^2a_1 - \sinh^2a_1}{2\cosh^2a_1} =
-2\log (1-|z|^2)+const \,.$$
[DaSz97]{}
Bull. London Math. Soc. [**25**]{} (1993) 162–168.
J. Geom. Anal. [**1**]{} (1991) 1–17.
Tans. of the AMS [**179**]{} (1973) 415-431.
Adv. in Math. [**44**]{} (1982) 121-131.
In: “Complex analysis and geometry", Univ. Ser. Math., Plenum, New York, 1993, 223-234.
J. Funct. Anal. (1), [**30**]{} (1978) 106-146.
Math. Zeit. [**278**]{}, 3-4 (2014) 769–793.
Introduction to Holomorphic Functions of Several Variables, Vol I: Function Theory. Wadsworth & Brooks/Cole, 1990.
Math. Ann. [**337**]{} (2007), 197–232.
Foundations of Differential Geometry, Vol. 2. Interscience Tracts in Pure and Applied Mathematics, No. 15 John Wiley et Sons, Inc., New York-London-Sydney 1969.
J. Funct. Anal. [**13**]{} (1973) 324-389.
Algebraic Structures of Symmetric Domains Princeton University Press, Princeton, New Jersey, 1980.
in Boothby, William; Weiss, Guido, Symmetric spaces (Short Courses, Washington University), Pure and Applied Mathematics 8, Dekker, New York, 1972, pp. 271-357.
[^1]: [*Mathematics Subject Classification (2010):*]{} 32M15, 31C10, 32T05
[^2]: [*Key words*]{}: Hermitian symmetric spaces, Stein domains, plurisubharmonic functions
[^3]: The authors acknowledge the MIUR Excellence Department Project awarded to the Department of Mathematics, University of Rome Tor Vergata, CUP E83C18000100006
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Supercapacitors have recently gained popularity as possible energy storage systems due to their high cycling ability and increased power density. However, one of the major drawbacks of supercapacitors is that they have a low energy density, which makes them less effective than batteries. Herein, we explore different methods of increasing the supercapacitor performance of the perovskite SrCoO$_3$. We carry out first-principles calculations to systematically study how SrCoO$_3$$/$graphene interface, oxygen vacancies, and doping improve the performance of strontium cobaltite as an anion-intercalation-type supercapacitor. The results show that the SrCoO$_3$$/$graphene interface is relatively stable with a formation energy of 1.3 eV and is highly conductive, which makes it a promising material for supercapacitors. We also find that inducing oxygen vacancies in SrCoO$_3$ significantly increases the conductivity of this material. Results of doping calculations reveal that doping with Mo, V, P, and Nb all increase the stability and conductivity of SrCoO$_3$. We find that niobium is the most stable and most conductive of all four dopants. In addition, we find that vanadium is a very promising novel dopant for SrCoO$_3$ as an anion-intercalation-type supercapacitor electrode material.'
author:
- Sadhana Lolla
- Xuan Luo
bibliography:
- 'my\_ref.bib'
title: 'A Systematic Study to Improve the Performance of SrCoO$_3$ as an Anion-Intercalation-Type Electrode for Supercapacitors Through Interface, Oxygen Vacancies, and Doping'
---
Introduction
============
Global warming is wreaking destruction on habitats around the world, polluting the air and water, and putting endangered species at risk. In light of this growing environmental crisis, renewable energy sources such as wind, solar, and hydrothermal energy are becoming crucial to the growth of sustainable clean energy[@saw2019]. In addition, over the past decade, concerns regarding environmental health have propelled the movement away from traditional gasoline-powered vehicles to more environmentally friendly such as electric vehicles (EVs) and hybrid electric vehicles (HEVs). However, both clean energy sources and EVs require fast and powerful energy storage devices [@wei2011]. Supercapacitors have the potential to fill this need due to their unique properties and widespread applications [@horn2019].
Though the first patent for a supercapacitor was filed in 1957[@wang2012], supercapacitors have only recently begun to attract widespread interest because of their fast charge-discharge rate, high power density, and long cycle life [@tomar2018]. However, there are two main disadvantages facing the development of supercapacitors: they have high production costs and low energy densities (ED), which make them less effective than lithium ion batteries. The types of supercapacitors include traditional electric double layer capacitors (EDLCs) and pseudocapacitors. Of these two types, pseudocapacitors are generally preferred because they have a higher ED than EDLCs due to the reversible redox reactions between electrode materials and electrolytes [@george2018]. However, the ED of pseudocapacitors still does not match the ED of batteries. Therefore, one popular method of improving supercapacitor performance has been to develop new electrode materials for pseudocapacitors. Materials used in pseudocapacitors include transition metal oxides (TMOs) [@yang2018controllable; @yang2018; @guo2018; @xiong2013; @nagarani2018; @wang2014], which have been extensively researched for this purpose. TMOs, specifically ruthenium oxide, are widely regarded as some of the most promising materials for pseudocapacitors[@wang2012]. However, ruthenium is an extremely expensive material and therefore electrodes made of ruthenium are not desirable. Recently, a third type of supercapacitor electrode has been proposed based on anion-intercalation mechanisms. These anion-intercalated electrode materials often produce higher energy densities than either pseudocapacitors or EDLCs because electrode reactions in the intercalation-type supercapacitors are similar to those in lithium-ion batteries[@liu2016].
Perovskite oxides of the form ABO$_3$ have the potential to be extremely effective anion-intercalation-type supercapacitor electrode materials due to their natural oxygen vacancies and defects[@ozolins2013], which make them extremely conductive and give them the potential for both high energy density[@zhu2016] and high power density. Preliminary experimental studies have demonstrated high cycling ability and promising energy densities [@liu2018; @lang2017; @mo2018] of ABO$_3$ as a supercapacitor electrode.
Perovskite/graphene composites have the potential to further improve the energy density for supercapacitor applications, as graphene can enhance the conductivity of the perovskite oxide and the cluster structure of the perovskite can effectively reduce the agglomeration of graphene[@lang2017]. In addition, inducing oxygen vacancies in perovskite oxides can improve their performance as anion-intercalation-type supercapacitor electrodes. Furthermore, doping the perovskite oxide at the B-site increases the stability and conductivity of the complex, especially of cubic perovskite oxides [@aguadero2012; @zhu2016; @mo2018]. However, we are currently unaware of studies that have focused on perovskite oxides combined with graphene for supercapacitor applications. Therefore, our objective is to determine whether the performance of perovskite oxides can be improved through various methods in order to increase the energy density and/or power density of supercapacitors.
We choose SrCoO$_3$ as our perovskite oxide because of its high Oxygen Evolution Rate (OER) [@jia2017] and its behavior as an extremely active catalyst[@calle2015], which means that oxygen vacancies can be created very easily in this material. Using first-principles calculations, we explore the conductivity and stability of an SrCoO$_3$$/$graphene interface. We also induce oxygen vacancies in SrCoO$_3$ to observe their effect on the conductivity of the perovskite for supercapacitor applications. In addition, we conduct doping calculations with Mo, V, P, and Nb to determine which dopant is the most effective in terms of conductivity and stability. We calculate band structure, DOS, and formation energy for all of the compounds.
The rest of this paper is structured as follows: first, we will discuss the computational methods and material selection process and we will then discuss the results of our DFT calculations and their impact on the field of supercapacitors as a whole.
Methods
=======
Computational Methods
---------------------
All first-principle density functional theory (DFT) calculations were conducted using ABINIT [@gonze2009]. The Generalized Gradient Approximation (GGA) and the Perdew-Burke-Ernzerhof (PBE) [@perdew1996] exchange-correlation functionals were utilized to determine the electronic structure properties of various materials. The pseudopotentials were based on the Projector Augmented Wave (PAW)[@blochl1994] method and the projectors were generated using AtomPAW[@holzwarth2001].
Element Electronic Configuration (core/val) PAW radius cutoff (a.u.)
------------ ---------------------------------------------------------- --------------------------
Strontium \[Ar ${3d}^{10}$\] ${4s}^2$ ${4p}^6$ ${5s}^1$ ${4d}^{1}$ 2.2
Cobalt \[Ne\] ${3s}^2$ ${3p}^6$ ${4s}^1$ ${3d}^8$ 2.1
Oxygen \[He\] ${2s}^2$ ${2p}^4$ 1.4
Carbon \[He\] ${2s}^2$ ${2p}^2$ 1.5
Molybdenum \[Ar ${3d}^{10}$\] ${4s}^2$ ${4p}^6$ ${5s}^1$ ${4d}^{5}$ 2.2
Phosphorus \[Ne\] ${3s}^2$ ${3p}^3$ 1.9
Niobium \[Ar\] ${3d}^{10} $ ${4s}^2$ ${4p}^6$ ${5s}^1$ ${4d}^4$ 1.4
: Electronic Configuration and PAW radius cutoff used to generate PAW pseudopotentials
Convergence
-----------
The kinetic energy cutoff and Monkhorst-Pack k-point grids were converged for all materials. The self-consistent field (SCF) total energy tolerance was set as 1.0 $\times$ $10^{-10}$ Ha. Once this tolerance was reached twice consecutively, the SCF iterations were terminated. The kinetic energy cutoff and Monkhorst-Pack k-point grids were considered converged when the differences in total energies were less than 1.0 $\times$ $10^{-4}$ Ha (about 3 meV) twice consecutively. The converged values for all materials can be found in Table \[tab:ecutkpt\].
Material Energy Cutoff (Ha) k-mesh
----------- -------------------- -------------------------
SrCoO$_3$ 20 4 $\times$ 4 $\times$ 4
Graphene 19 4 $\times$ 4 $\times$ 1
Mo 13 6 $\times$ 6 $\times$ 6
P 23 8 $\times$ 8 $\times$ 8
Nb 17 6 $\times$ 6 $\times$ 6
V 13 6 $\times$ 6 $\times$ 6
: Converged values of the kinetic energy cutoff and Monkhorst-Pack k-point grids for each material.[]{data-label="tab:ecutkpt"}
Relaxation
----------
The converged values for the kinetic energy cutoff and k-mesh were used to fully relax all systems studied. All structures were relaxed using the Broyden$-$Fletcher$-$Goldfarb$-$Shanno algorithm. The SCF iterations were terminated when the Hellman-Feynman forces were less than 5.0 $\times$ ${10}^{-6}$ Ha/Bohr twice consecutively. Structures were considered fully relaxed when the maximal force was less than 5 $\times$ ${10}^{-5}$ Ha/Bohr.
Materials
---------
When used in supercapacitor applications, perovskite oxides usually follow the structure ABO$_3$. We choose SrCoO$_3$ as our perovskite because of its high OER[@jia2017]. Other theoretical studies have also predicted that strontium cobaltite has the highest OER ability among all perovskites through DFT calculations [@calle2015]. This makes SrCoO$_3$ a good candidate for supercapacitors because oxygen vacancies increase conductivity, especially in anion-intercalation supercapacitors [@alexander2019].
The cubic structure of the perovskite SrCoO$_3$ was chosen as the base material for all calculations due to its high conductivity.
### Interface
Experimental studies have speculated that graphene has the potential to increase the stability conductivity of perovskite oxides [@lang2017]. However, we are unaware of any experimental or theoretical studies to date that have focused on the effects of graphene on perovskite oxides for supercapacitor applications. In this study, we simulate an interface by combining an SrCoO$_3$ supercell with graphene. To do this, the kinetic energy cutoff was calculated for both pristine SrCoO$_3$ and pure graphene and the larger of the two was used for interface calculations. These values can be found in Table \[tab:ecutkpt\]. The relaxed lattice constant for SrCoO$_3$ was chosen for the interface in order to best analyze the effects of graphene on strontium cobaltite. To obtain the layer-by-layer distance $d$ between SrCoO$_3$ and graphene, we systematically increase $d$ in increments of 0.5 Bohr and compare the total energy. The layer-by-layer distance with the lowest total energy is used in band structure and PDOS calculations.
### Oxygen Vacancy
Numerous studies have demonstrated that oxygen vacancies improve the conductivity of compounds in supercapacitor applications [@yang2018; @zhai2014; @xiang2017]. This is especially true for perovskites of the form ABO$_3$ where B is a transition metal because oxygen vacancies change the geometric and chemical properties [@yang2018] of TMOs. To the best of our knowledge, oxygen vacancies have not been theoretically studied in perovskite oxides for supercapacitor applications. In the present study, we induce oxygen vacancies in strontium cobaltite to determine the mechanism by which vacancies improve the conductivity of this perovskite oxide. We simulated one oxygen vacancy in a 2 $\times$ 2 $\times$ 1 supercell by removing an oxygen. This structure was then relaxed and the band structure, PDOS, and formation energy were calculated.
Two oxygen vacancies were created by taking out an oxygen and its nearest neighbor. The structure was fully relaxed again and the above procedures were repeated for the complex with two vacancies.
### Doping
We choose four different dopants for SrCoO$_3$: Mo, V, P, and Nb because they have been shown to increase the OER of SrCoO$_3$ and other similar perovskites. Substitutional doping of 25% was achieved at the B-site by replacing one cobalt with a dopant in a 2 $\times$ 2 $\times$ 1 supercell. The kinetic energy cutoff was calculated for each dopant and for pristine SrCoO$_3$ and the maximum of the two was chosen for the doping calculations. These values are displayed in Table \[tab:ecutkpt\]. Previous experimental studies have explored Mo and Nb as dopants for SrCoO$_3$ as an anion-intercalation-type supercapacitor and concluded that both Mo and Nb increase the conductivity of strontium cobaltite [@tomar2018; @li2017niobium]. We use Mo and Nb as dopants in this study to examine how they improve the performance of SrCoO$_3$ as an electrode material for anion-intercalation-type supercapacitors. P-doped SrCoO$_3$ has previously been studied as a water oxidation electrocatalyst and has been shown to increase the OER of SrCoO$_3$[@zhu2016]. We use phosphorus as a dopant to determine whether phosphorus can be as effective as the other metallic dopants for SrCoO$_3$. Vanadium-doped transition metal oxides have been explored as possible electrode materials[@yang2013] for pseudocapacitors. In this study, we analyze the conductivity and stability of the V-doped SrCoO$_3$ to determine if it is also an effective dopant for perovskite oxides for anion-intercalation-type supercapacitors.
Band Structure
--------------
Using the cubic atomic structure of SrCoO$_3$ shown in Figure \[bandstructure\] (a), the band structure was calculated for the interface, vacancy, and doping calculations. To transform the lattice vectors $a_1$, $a_2$, and $a_3$ into the reciprocal space lattice vectors $b_1$, $b_2$, and $b_3$, Eqs. (1)-(3) were used. High symmetry k-points were then selected to comprehensively sample the first Brillouin zone. For doping and vacancy calculations, the high symmetry k-points for simple cubic structures were chosen and are: $\Gamma$ (0, 0, 0), X(0, 0.5, 0), M(0.5, 0.5, 0.0), and R(0.5, 0.5, 0.5). The first Brillouin zone and high symmetry k-points for simple cubic structures are shown in Figure \[bandstructure\] (b). When plotting the band structures, each band contained two electrons and four additional conduction bands were used.
Likewise, the band structure was also calculated for pure hexagonal monolayer graphene, which is shown in Figure \[bandstructure\](c). Eqs. (1)-(3) were used to transform the lattice vectors into the reciprocal space lattice vectors, and the high symmetry k-points for hexagonal structures were used: $\Gamma$ (0, 0, 0), K(1/3, 2/3, 0), and M(0, 0, 0). Figure \[bandstructure\](d) shows the first Brillouin zone and hexagonal high symmetry k-points.
We plotted the fat-band structures for the doped complex structures to analyze the contributions of dopants to the overall band structure of the compound. Similarly, we also calculate the fat-band structures for SrCoO$_3$ with oxygen vacancies to study the contributions of the Oxygen $2p$ orbital to the band structure of the complex.
$$\vec{b_1} = 2\pi\frac{\vec{a_2} \times \vec{a_3}}{\vec{a_1} \cdot (\vec{a_2} \times \vec{a_3})}
\label{eqn:recspace1}$$
$$\vec{b_2} = 2\pi\frac{\vec{a_3} \times \vec{a_1}}{\vec{a_2} \cdot (\vec{a_3} \times \vec{a_1})}
\label{eqn:recspace2}$$
$$\vec{b_3} = 2\pi\frac{\vec{a_1} \times \vec{a_2}}{\vec{a_3} \cdot (\vec{a_1} \times \vec{a_2})}
\label{eqn:recspace3}$$
Density Of States
-----------------
Total DOS (TDOS) and projected DOS (PDOS) were calculated for the SrCoO$_3$$/$graphene interface, oxygen vacancy, and doping complex systems. The TDOS was used to analyze the overall conductivity of the complex and the PDOS was analyzed to find the contributions of each orbital to the conduction process. We plot the PDOS of the $3d$ orbital for Sr, the $3d$ orbital for Co, the $2p$ orbital for Oxygen, and the $d$ orbitals for all the dopants except Phosphorus, which lacks a $d$ orbital, so the $p$ orbital was used instead.
Charge Transfer/Density
-----------------------
We calculate the charge density for structures with oxygen vacancies in order to determine where the charge accumulates after the vacancy is induced. To do this, the converged charge density of the complex structure was obtained by calculating the total energy using the converged values for the kinetic energy cutoff and Monkhorst-Pack k-point grids along with the relaxed lattice constants. This total energy calculation is then used to generate isosurface diagrams for both one vacancy and two vacancy calculations.
We also calculate the charge transfer for the SrCoO$_3$$/$graphene interface using the formalism in Eq. (\[eqn:chargetr\]). To calculate the charge transfer, the converged charge density was obtained through the total energy calculation. The total energy is calculated using the converged kinetic energy cutoff, Monkhorst-Pack k-point grids, and layer-by-layer distance $d$. The charge density of pristine SrCoO$_3$ and pure graphene are subtracted from the charge density of the interface to determine the electrons gained or lost. The red regions in the isosurface diagrams represent the charge accumulation and the blue regions represent the charge depletion.
$$Q_{transfer} = Q_{interface} - Q_{SrCoO3} - Q_{graphene}
\label{eqn:chargetr}$$
Formation energy
----------------
The formation energies for doping, oxygen vacancy, and interface complex systems were calculated. A lower formation energy indicates a more stable compound, and a negative formation energy is preferred because it indicates that an exothermic reaction has occurred. The formation energy for all complex systems were calculated using Eq. (\[eqn:formation energy\]), where $E_f$ represents the formation energy.
$$E_{f} = E_{complex} - E_{A} - E_{B}
\label{eqn:formation energy}$$
For the SrCoO$_3$$/$graphene interface, the total energy of pure SrCoO$_3$ and pure graphene were calculated using the converged values for the kinetic energy cutoff and k-point grids. In this case, $E_{complex}$ represents the total energy of the interface. $E_A$ represents the total energy of 1 $\times$ 1 $\times$ 2 SrCoO$_3$, and $E_B$ represents the total energy of 2 $\times$ 2 graphene. The formalism is displayed in Eq. (\[eqn:interface\_energy\]).
$$E_{f} = E_{complex} - 2 \times E_{SrCoO3} - 4 \times E_{graphene}
\label{eqn:interface_energy}$$
In order to determine the formation energy of strontium cobaltite with vacancies, the total energy of oxygen gas was first calculated. $E_{complex}$ represents the formation energy of the complex with either one or two $V_O$. $E_A$ is the total energy of 2 $\times$ 2 $\times$ 1 SrCoO$_3$. $E_B$ is equal to $n$ multiplied by the total energy of one oxygen atom, where $n$ is the number of vacancies. The equation for the defect formation energy of the vacancy calculations can be found in Eq. (\[eqn:vacancy\_energy\]).
$$E_{f} = E_{complex} - 4 \times E_{SrCoO_3} + n \times E_{O_2}/2
\label{eqn:vacancy_energy}$$
To obtain the formation energy of doped SrCoO$_3$, the total energies of the dopants and cobalt were calculated, along with the total energy of pure SrCoO$_3$. $E_{complex}$ represents the total energy of the doped compound, which was calculated using the converged kinetic energy cutoff, Monkhorst-Pack k-point grids, and relaxed lattice constants. As the dopant is being substituted for a cobalt atom, we add the total energy of the cobalt atom and subtract the total energy of the dopant atom. Eq. (\[eqn:dopant\_energy\]) shows the formalism for the formation energy of the doping calculations.
$$E_{f} = E_{complex} - 4 \times E_{SrCoO_3} + E_{Co} - E_{dopant}
\label{eqn:dopant_energy}$$
Results and Discussion
======================
In this section, we use DFT calculations to analyze the electronic and structural properties of an SrCoO$_3$$/$graphene interface, oxygen vacancies, and doping calculations as they apply to supercapacitor performance.
Interface
---------
Previous studies have suggested that graphene has the potential to improve supercapacitor performance. Yang et al[@yang2018], for example, has demonstrated that creating a Co$_3$O$_4$$/$graphene interface increased the energy density dramatically from 32 W h$/$kg to 43.1 W h$/$kg. Lang et al.[@lang2017] also suggested that a perovskite$/$graphene interface has the potential to further increase the conductivity and stability of this material. Therefore, our goal is to theoretically study a perovskite/graphene interface to understand why graphene improves perovskite performance for supercapacitors.
An SrCoO$_3$$/$graphene interface was formed by 1 $\times$ 1 $\times$ 2 SrCoO$_3$ and 2 $\times$ 2 graphene. The lattice parameters of the relaxed SrCoO$_3$ were used for the interface calculations. In lieu of relaxation, the optimal layer distance $d$ between the perovskite and graphene was determined by incrementally increasing $d$ by 0.5 Bohr and calculating the total energy. The layer distance with the lowest total energy was 5.5 Bohr, so this structure was used for DOS and formation energy calculations. We conclude that this value is reasonable because the graphene layer distance is around 6.6 Bohr[@affoune2001].
The lattice constants of the relaxed compounds are displayed in table \[interface\_structure\]. The percentage error between the relaxed lattice constants and the published theoretical and experimental lattice constants for SrCoO$_3$ is always less than 1%. Furthermore, the relaxed lattice constants for graphene calculations fall within the range indicated by Girit et al[@girit2009graphene], indicating that our relaxation calculations are reliable for both SrCoO$_3$ and graphene.
Compounds Calculation Method a (Bohr) b (Bohr) c (Bohr)
----------- --------------------------------- ------------- ------------- ----------
SrCoO$_3$ GGA (this study) 7.21 7.21 7.21
Experiment [@wang2012porous] 7.28 7.28 7.28
GGA [@jia2017] 7.19 7.19 7.19
Graphene GGA(this study) 4.62 4.62
Experiment [@girit2009graphene] 4.35 - 4.72 4.35 - 4.72
GGA [@giovannetti2008] 4.62 4.62
: Relaxed Lattice Constants of SrCoO$_3$ and graphene[]{data-label="interface_structure"}
To calculate the formation energy of the SrCoO$_3$$/$graphene interface, the total energy of pure SrCoO$_3$ and pure graphene were first calculated using the relaxed lattice constants given in Table \[interface\_structure\] and the converged kinetic energy cutoff and Monkhorst-Pack k-point grids given in Table \[tab:ecutkpt\]. The formalism displayed in Eq. (\[eqn:interface\_energy\]) was then applied to calculate the formation energy of the interface, which is 1.3 eV. Previous theoretical studies such as Hussain et al[@hussain2019] have studied interactions between transition metal oxides (TMOs) and graphene for supercapacitor applications and determined the formation energy of CuO$/$graphene as 0.3 eV and FeO$/$graphene as 3.7 eV. Since the formation energy of the SrCoO$_3$$/$graphene interface falls between these two values, it can be concluded that the formation energy for this interface is reasonable and that the interface is relatively stable.
The charge transfer between graphene and SrCoO$_3$ was calculated in order to determine the areas of charge accumulation and charge depletion. To calculate the charge transfer, the converged charge density of the interface, pure SrCoO$_3$, and pure graphene were first obtained using the total energy calculations. The charge density of pure SrCoO$_3$ and pure graphene were then subtracted from the charge density of the interface to obtain the charge transfer. The resulting isosurface is displayed in Figure \[isosurface\], and the isovalue used was 0.002 electrons$/$Bohr$^3$. As shown in the figure, red regions indicate charge accumulation and blue areas indicate charge depletion.
The charge transfer could be caused by the strong hybridization of the C $2p$, Sr $3d$, Co $3d$, and O $2p$ orbitals. Figure \[isosurface\] (a) clearly shows that the oxygen atoms facilitate the charge transfer through the interface. Charge is transported away from the carbon atoms towards SrCoO$_3$ through the oxygen atoms. These results are similar to those obtained by Hussain et al [@hussain2019] and Xiong et al [@xiong2013], where the oxygen exhibits weak accumulation of charge and acts as a conducting channel. Figure \[isosurface\] (b) indicates that there is also accumulation of charge on the graphene layer, confirming the conductivity of the interface as a whole. It has been suggested by Lang et al[@lang2017] that SrCoO$_3$ has the potential to prevent the agglomeration of graphene sheets through homogeneous dispersion. Due to this, it is possible that graphene not only increases the conductivity of the interface as a whole, but also can provide a buffer to accommodate the expansion and contraction of the cluster structure during charge/discharge cycles[@hussain2019].
The total DOS (TDOS) and projected DOS (PDOS) were calculated for the SrCoO$_3$$/$graphene interface to analyze the contribution of the Sr $3d$, Co $3d$, O $2p$, and C $2p$ orbitals to the conduction process. The TDOS of pure SrCoO$_3$ is displayed in Figure \[interface\_dos\] (a), and the TDOS of the SrCoO$_3$$/$graphene interface can be found in Figure \[interface\_dos\] (b). Compared to the TDOS of the pure perovskite, the interface calculation has more peaks in both the conduction and valence bands, indicating that the interface is more conductive overall than pure SrCoO$_3$. Figure \[interface\_dos\] (b) shows no band gap, which is to be expected because of the metallic nature of the SrCoO$_3$$/$graphene interface. Furthermore, the non-zero states at and around the Fermi level indicate an overlap between the valence and conduction bands and show that the states are shifting away from the valence bands towards the conduction bands.
The PDOS of the interface was calculated and is displayed in Figure \[interface\_dos\] (d), and the PDOS of the pure SrCoO$_3$ is shown in Figure \[interface\_dos\] (c) for comparison. The increased presence of non-zero localized states at and around the Fermi level indicates the increased conductivity of this material. It is clearly evident that the PDOS plot has many sharp peaks and that peaks are present for every material used in the interface, so it can be said that SrCoO$_3$ and graphene contribute equally to the conductivity of the interface. The increased number of coinciding peaks indicates the strong hybridization between C $2p$, O $2p$, Co $3d$, and Sr $3d$. The PDOS of the composite system is similar to the DOS of Sr, Co, and O as plotted by earlier studies[@van2001bulk]. Therefore, we can conclude that due to the relative stability and extremely high conductivity of this SrCoO$_3$$/$graphene interface, it has the potential to be an effective supercapacitor electrode.
Oxygen vacancies
----------------
A multitude of studies have concluded that inducing oxygen vacancies improves the conductivity of materials for supercapacitor applications[@yang2018; @xiang2017two; @wang2014]. Cheng et al[@cheng2017] specifically found that inducing oxygen vacancies in Co$_3$O$_4$ increased the specific capacitance of the complex material significantly. In the present study, we simulate one oxygen vacancy by removing one oxygen atom in a 2 $\times$ 2 $\times$ 1 supercell; similarly, we induce two oxygen vacancies by taking out one oxygen and its nearest neighboring oxygen atom. We then calculate the formation energy, band structure, and PDOS of the supercell with one V$_O$ and two V$_O$ and compare the two with pure SrCoO$_3$.
Oxygen vacancies were first induced in a pure SrCoO$_3$ supercell of size 2 $\times$ 2 $\times$ 1. After the oxygen was removed, the structure was relaxed again and the band structure and the PDOS were calculated. To simulate two oxygen vacancies, an oxygen atom and its nearest neighbor were both taken out. Table \[tab:vacancy\_lattice\] shows that with two vacancies, the relaxed lattice constant increases slightly.
Number of vacancies a (Bohr) b (Bohr) c (Bohr)
--------------------- ---------- ---------- ----------
One 14.4 14.4 7.22
Two 14.5 14.5 7.23
: Relaxed lattice constants for SrCoO$_3$ supercell with oxygen vacancies[]{data-label="tab:vacancy_lattice"}
The defect formation energy of the complex structures was calculated. To do this, the total energy of pure SrCoO$_3$ was determined using the relaxed lattice constants in Table \[interface\_structure\] and converged kinetic energy cutoff and Monkhorst-Pack k-point grids displayed in Table \[tab:ecutkpt\]. The formalism defined in Eq. (\[eqn:vacancy\_energy\]) was then used to determine the defect formation energy of the oxygen vacancies. The formation energy of the complex with one oxygen vacancy is 2.75 eV, while the formation energy of the complex with two oxygen vacancies is 7.62 eV. Since a lower formation energy is indicative of a more stable compound, the complex with one oxygen vacancy is more stable than the complex with two oxygen vacancies. This is to be expected, as inducing vacancies disrupts the crystal structure and therefore decreases the stability of the system. Compared to other studies which have analyzed oxygen vacancies with transition metal oxides, the complex with one oxygen vacancy is relatively stable. Previous studies have shown that inducing an oxygen vacancy in TiO$_2$ results in a formation energy of around 2-5 eV [@na2006]. Since the formation energy of one oxygen vacancy is similar to this range, we conclude that the formation energies of these complexes are reasonable.
The fat-band structures of the one-vacancy complex and the two-vacancy complex were calculated and are displayed in Figure \[vacancy\_bands\]. The contributions by the oxygen $2p$ orbital are plotted in red, and it is evident that inducing oxygen vacancies increases the contributions of oxygen to the bands around the Fermi level. The oxygen $2p$ orbital contributes more to the two-vacancy complex than to the one-vacancy complex, demonstrating that the contributions of the oxygen $2p$ orbital increase as the number of vacancies increases. As expected, there is little to no band gap near the Fermi level, which indicates that the materials in question are metallic.
In order to yield further insights into the conductivity of the complexes with vacancies, TDOS and PDOS were calculated. To calculate the DOS, the total energy was determined using the converged kinetic energy cutoffs and Monkhorst-Pack k-point grids, as well as the relaxed lattice vectors shown in Table \[tab:vacancy\_lattice\]. The TDOS of the complexes can be found in Figure \[vacancy\_dos\] (b) and (c). The total DOS of the pure SrCoO$_3$ is also shown in Figure \[vacancy\_dos\] (a) for comparison. It is clearly evident from \[vacancy\_dos\] (b) and \[vacancy\_dos\] (c) that inducing oxygen vacancies improves the conductivity of the complex because of the increased number of peaks in the TDOS in figures \[vacancy\_dos\](b) and (c). Figures \[vacancy\_dos\] (b) and (c) show no band gap, which is to be expected because of the metallic nature of SrCoO$_3$. In addition, the non-zero states at and around the Fermi level indicate an overlap between the valence and conduction bands and show that the states are shifting away from the valence bands towards the conduction bands.
![Total DOS for (a) pure SrCoO$_3$, (b) an SrCoO$_3$ complex with one vacancy, and (c) an SrCoO$_3$ complex with two vacancies. Since there is little to no band gap around the Fermi level, we can conclude that these materials are metallic.[]{data-label="vacancy_dos"}](vacancy_tdos.pdf){width="10cm"}
The PDOS of the complexes were also calculated in order to determine the contributions of Sr $3d$, Co $3d$, and O $2p$ with vacancies to the conduction process. The PDOS reveals that vacancies increase the conductivity of SrCoO$_3$ tremendously, and there is also a considerable increase in conductivity from one vacancy to two vacancies. Figures \[vacancy\_pdos\] (a) and (b) show that there are localized states at and around the Fermi level, indicating an increase in conductivity. In addition, peaks are present for all elements, indicating that Sr, Co, and O contribute equally to the conduction process. The non-zero states at the Fermi level show that there is an overlap between the valence and conduction band edges that reveals a shift away from the valence band (VB) toward conduction bands (CB). The increased number of states from one vacancy to two vacancies reveals the strong hybridization of the Sr $3d$, Co $3d$, and O $2p$ orbitals caused by the vacancy. Due to the increase in conductivity caused by vacancies and their relative stability, we conclude that inducing oxygen vacancies in SrCoO$_3$ increase its performance as a supercapacitor electrode.
The converged charge density was calculated for both one V$_O$ and 2 V$_O$ and the results are displayed in figure \[one\_vacancy\_iso\]. The charge density was obtained by calculating the total energy of the complex using the kinetic energy cutoffs and k-point grids of pure SrCoO$_3$ and the relaxed lattice constants of the complex. The isovalue used was 0.03 electrons$/$Bohr$^3$. Figure \[one\_vacancy\_iso\] (a) shows that the charge accumulation is centered around the vacancies, which is well in agreement with published studies [@yang2018]. Similarly, Figure \[one\_vacancy\_iso\](b) shows two concentrated areas of charge accumulation caused by the two vacancies in the structure. This accumulation is most likely caused by the increased electron delocalization of the surrounding Sr and Co atoms as a result of the oxygen vacancy. In addition, the charge accumulation could also have been caused by the strong hybridization of the Sr $3d$, Co $3d$, and O $2p$ orbitals, which supports the PDOS calculations shown in Figure \[vacancy\_pdos\].
Doping
------
We choose Mo, V, P, and Nb as dopants to investigate how they increase the OER of SrCoO$_3$. We also study whether nonmetallic dopants can achieve the same level of conductivity and stability as the metallic dopants.
To achieve 25% substitutional doping at the B-site of SrCoO$_3$, a 2 $\times$ 2 $\times$ 1 supercell was used, in which one cobalt atom was replaced with the dopant. The complex was then relaxed and the band structure, PDOS, and formation energy were calculated for each dopant and compared to determine the most effective dopant.
Once a cobalt atom was substituted for one of the dopants, the resulting structure was relaxed again. The fully relaxed structure of the complex remained cubic when the dopant was Mo, V, or Nb, which confirms experimental results, as these dopants do not alter the basic structure of SrCoO$_3$. For P-doped SrCoO$_3$, however, the structure was altered from cubic to tetragonal, which supports the experimental results found by Zhu et al [@zhu2016]. Table \[tab:dopant\_relaxed\_dim\] displays the relaxed lattice constants of each complex. The atomic structure is displayed in Figure \[dopant\_atomic\_structure\].
Compound a (Bohr) b (Bohr) c (Bohr)
-------------------- ---------- ---------- ----------
Mo-doped SrCoO$_3$ 14.64 14.64 7.32
V-doped SrCoO$_3$ 14.52 14.52 7.26
P-doped SrCoO$_3$ 14.77 14.77 7.39
Nb-doped SrCoO$_3$ 14.69 14.69 7.34
: Relaxed lattice constants of doped SrCoO$_3$[]{data-label="tab:dopant_relaxed_dim"}
The formation energy for all dopants was calculated. To do this, the total energy was calculated for the doped complex, pure SrCoO$_3$, pure cobalt, and the pure dopant using the converged kinetic energy cutoffs, Monkhorst-Pack k-point grids, and relaxed lattice vectors. The formalism displayed in Eq. (\[eqn:dopant\_energy\]) was used to calculate the formation energy of the complex, and the results are displayed in Table \[tab:dopant\_energy\]. It has been well-documented by Hussain et al[@hussain2018] and others that a lower formation energy is an indicator of stability of a structure. These DFT calculations reveal that niobium-doped SrCoO$_3$ has the lowest formation energy of -9.89 eV, indicating that it is the most stable compound of the four dopants. All of the formation energies are negative, which indicates that exothermic reactions are taking place. This further confirms that all of the complexes are stable as they do not gain heat during the doping reaction.
Compound Formation Energy (eV)
----------------- -----------------------
Mo-doped SrCoO3 -7.46
V-doped SrCoO3 -7.88
P-doped SrCoO3 -6.07
Nb-doped SrCoO3 -9.89
: Formation energy for doped SrCoO$_3$ with different dopants[]{data-label="tab:dopant_energy"}
The band structure of all four dopants reveals that there is little to no band gap around the Fermi level as expected, which reveals the metallic behavior of these systems. The contributions of the dopant to the band structure is shown in Figure \[dopant\_band\_structure\] in the form of fat-bands in order to analyze how the dopant affects the band structure. Figure \[dopant\_band\_structure\] shows that niobium, vanadium and molybdenum substantially affect the band structure of the complex, while Phosphorus does so to a lesser extent because of the lack of a $d$ orbital. This reveals that metallic dopants are more effective than non-metallic dopants for supercapacitor applications. Therefore, we can conclude that metallic dopants are more effective than nonmetallic dopants for supercapacitor electrodes.
The total DOS was analyzed for further insight into the conductive behavior of the dopants and how they affect the total conductivity of the complex. The TDOS was calculated by first calculating the total energy using the converged kinetic energy cutoffs and k-point grids, along with the relaxed lattice constants of the complexes, and is plotted in Figure \[dopant\_tdos\]. The total DOS of pure SrCoO$_3$ is also plotted in Figure \[dopant\_tdos\] (e) for comparison.
![TDOS for all of the doped compounds. Figure (a) shows the TDOS of pure SrCoO$_3$, (b) shows the TDOS for Mo-doped SrCoO$_3$, (c) shows the TDOS for V-doped SrCoO$_3$, (d) shows the TDOS for P-doped SrCoO$_3$, and (e) shows the TDOS for Nb-doped SrCoO$_3$.[]{data-label="dopant_tdos"}](doping_tdos.pdf){width="10cm"}
Analysis of the TDOS and PDOS reveals the conductive behavior of doped SrCo$O_3$. Figure \[dopant\_pdos\] reveals that all four dopants increase the conductivity substantively, as shown by the non-zero localized states around the Fermi level. However, it is clear that phosphorus is the least conductive dopant because the values of the states are considerably lower for phosphorus TDOS than for any of the other dopants. The PDOS shows that Mo and V achieve similar results as dopants, with an increased conductivity shown by the $3p$-orbital for Sr, the $d$-orbital for Cobalt, the $2p$-orbital for Oxygen, and the $d$-orbital for the dopants. It can be said that the increased number of peaks shown in the PDOS are a result of the hybridization of the Sr $3d$, Co $3d$, O $2p$ and the $3d$ orbitals of the metallic dopants. It is evident that Niobium is by far the best dopant for SrCoO$_3$, as the states are much higher and there are more states for Nb-doped SrCoO$_3$ than any other dopant. For all four dopants, it is evident from the nonzero states around the fermi level that there is an overlap between the conduction bands (CB) and valence bands (VB) and that the states shift away from the CB towards the VB. For all of the dopants, there are states around the Fermi level for all four elements (Sr, Co, O, and the dopant), showing that all of the elements contribute equally to the conduction process. Taken together, the formation energy, band structure and PDOS reveal that while all of the dopants studied increase the compound’s conductivity, niobium is the best dopant due to its higher stability (as given by the lower formation energy) and its greater conductivity (as shown by the PDOS and the band structure). Furthermore, the PDOS and fat-band structure reveal that phosphorus is not as effective as the rest of the dopants. This shows that metallic dopants should be used to improve supercapacitor performance. In addition, we have discovered that vanadium is a promising new dopant for SrCoO$_3$ as anion-intercalation-type supercapacitor electrode.
Our findings suggest the viability of a novel SrCoO$_3$$/$graphene interface for supercapacitor applications. Previous studies involving TMO/graphene interfacial structures have suggested that oxygen acts as a conducting channel between graphene and the cluster structure[@hussain2019; @xiong2013], which increases the conductivity of the interface, which supports our electronic charge density calculations. Compared to FeO$/$ graphene, which has a formation energy of 3.7 eV[@hussain2019], and MnO$_2$$/$graphene, which has a formation energy of 5.5 eV[@xiong2013], we find that our SrCoO$_3$$/$graphene interface is relatively stable with a formation energy of only 1.3 eV. Our PDOS calculations are also supported by Xiong et al[@xiong2013] and Hussain et al[@hussain2019], who report that TMO/graphene interfacial structures exhibit more peaks that coincide with each other due to the strong hybridization of the Sr $3d$, Co $3d$, C $2p$ and O $2p$ orbitals.
We also theoretically verify that inducing oxygen vacancies in SrCoO$_3$ increases the conductivity of SrCoO$_3$ due to the strong hybridization of the Sr $3d$, Co $3d$, and O $2p$ orbitals. Yang et al [@yang2018] and Cheng et al [@cheng2017], among others, have concluded that inducing oxygen vacancies in TMOs decreases the stability but increases the conductivity of the material, which we confirm through formation energy and PDOS calculations. In addition, previous studies have also found that inducing oxygen vacancies results in charge accumulation around the vacancy due to electron delocalization in the surrounding atoms[@cheng2017; @cheng2013] this is supported by our electronic charge density calculations.
Furthermore, we find that doping at the B-site with Mo, V, P, and Nb increases the stability of this material and also improves the conductivity of SrCoO$_3$. Mo has been previously studied as a dopant for SrCoO$_3$ in both supercapacitor applications[@tomar2018] and for solid oxide fuel cells[@aguadero2012] due to its OER-enhancing properties, which make this perovskite more conductive. We confirm this conductive behavior through fat-band calculations, which reveal Molybdenum’s increased contribution to the band structure of Mo-doped SrCoO$_3$. Our PDOS calculations show that Mo-doped SrCoO$_3$ results in strong hybridization of Mo $d$, Sr $d$, Co $d$, and O $p$ orbitals. Zhu et al[@zhu2016] studied the effect of P-doping on SrCoO$_3$ as an OER catalyst and concluded that phosphorus significantly changes the structure of SrCoO$_3$, which supports our relaxation calculations. Although they concluded that this resulted in a more stable compound, we find that the other dopants increase the stability of SrCoO$_3$ to a much greater extent. However, previous studies have concluded that a small amount of doping at the B-site increases the stability of SrCoO$_3$[@george2018; @nan2019]. We confirm this as doping with any of the four dopants resulted in a complex with a very low formation energy, indicating the stability. Li et al[@li2017niobium] found that doping SrCoO$_3$ with niobium resulted in a very high energy density and also prolonged cycling life, which supports our conclusion that Niobium is the best dopant of the four dopants because it significantly increases the stability of this compound and also contributes tremendously to the conductivity. To the best of our knowledge, vanadium and phosphorus have not been studied as dopants for SrCoO$_3$ for supercapacitor applications. Nevertheless, our conclusions that these compounds enhance the OER of SrCoO$_3$ and also increase the stability are supported by the findings of Zhu et al[@zhu2016] and Guo et al[@guo2018]. However, we find that phosphorus is the least effective dopant due to its lack of a $d$ orbital, which reveals that metallic dopants are more effective for anion-intercalation-type supercapacitors. Therefore, we can conclude that vanadium and to a lesser degree phosphorus are both novel, promising dopants for supercapacitor applications.
Conclusion
==========
In this study, we use Density Functional Theory to explore various methods the performance of the perovskite SrCoO$_3$ for supercapacitor applications. We systematically study electronic structure properties and formation energies.
An SrCoO$_3$$/$graphene interface was studied to analyze how graphene contributes to the performance of SrCoO$_3$ as a supercapacitor electrode. The formation energy of the interface was 1.3 eV, which is relatively stable compared to other TMO$/$graphene interface structures. Both DOS calculations and electronic charge density indicate that the interface is extremely conductive.
Oxygen vacancies were induced to determine the mechanism by which vacancies improve the conductivity of SrCoO$_3$. The defect formation energy of one oxygen vacancy is 2.5 eV, which is comparable to other studies which have induced oxygen vacancies in TMOs. Fat-band structures reveal that vacancies cause the oxygen $2p$ orbital to contribute more to the band structure, and DOS calculations support this conclusion. The PDOS calculations reveal that oxygen vacancies cause a dramatic increase in the conductivity of the material.
We also doped SrCoO$_3$ with four different dopants, Mo, V, P, and Nb, at 25% concentrations at the B-site by replacing one Co atom with one dopant atom. We evaluated the performance of each dopant by calculating the formation energy, TDOS, PDOS, and fat-band structure. All calculations show that all four dopants improve SrCoO$_3$ performance drastically. Niobium is the best dopant of the four because it had the lowest formation energy of -9.89 eV. In addition, Niobium contributed the most to the fat-band structure of any of the four dopants and is the most conductive as shown by PDOS. Vanadium is also a promising new dopant for anion-intercalation-type supercapacitor electrodes.
Therefore, we conclude that a graphene interface, oxygen vacancies, and doping can all improve the performance of SrCoO$_3$ as an anion-intercalation type supercapacitor electrode to different extents.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'SciDB is a scalable, computational database management system that uses an array model for data storage. The array data model of SciDB makes it ideally suited for storing and managing large amounts of imaging data. SciDB is designed to support advanced analytics in database, thus reducing the need for extracting data for analysis. It is designed to be massively parallel and can run on commodity hardware in a high performance computing (HPC) environment. In this paper, we present the performance of SciDB using simulated image data. The Dynamic Distributed Dimensional Data Model (D4M) software is used to implement the benchmark on a cluster running the MIT SuperCloud software stack. A peak performance of 2.2M database inserts per second was achieved on a single node of this system. We also show that SciDB and the D4M toolbox provide more efficient ways to access random sub-volumes of massive datasets compared to the traditional approaches of reading volumetric data from individual files. This work describes the D4M and SciDB tools we developed and presents the initial performance results. This performance was achieved by using parallel inserts, a in-database merging of arrays as well as supercomputing techniques, such as distributed arrays and single-program-multiple-data programming.'
author:
-
bibliography:
- 'IEEEabrv.bib'
- 'references.bib'
title: Benchmarking SciDB Data Import on HPC Systems
---
Introduction
============
SciDB is an open-source database management system that uses an array data model [@stonebreaker2011; @brown2010]. The array-based data model provides support for parallel processing, efficient sparse storage, and in-database linear algebra operations that are well suited for the storage and analysis of biomedical imaging data. SciDB is a full ACID (atomicity, consistency, isolation, durability) database management system that guarantees repeatability of results across multiple users operating on the same data. Additionally, it supports array versioning. When using array versioning, SciDB creates new versions of an array instead of modifying an existing array. One of the unique advantages of SciDB is it’s ability to perform fast range selects and joins. This capability is achieved by storing data in chunks, in the same order as in the original coordinate system. By storing data in this manner, data that are close to each other can be accessed very quickly by reducing the number of reads necessary to access a given range of data. SciDB also allows a user-settable overlap between chunks of data to speed up applications such as spatial filtering of images for which fewer data reads are necessary at the boundaries to read in the required arrays.
The array data model of SciDB makes it well suited for managing multidimensional image data. With advances in image acquisition techniques [@tomer; @leica], it is possible to generate multi-Terabytes of high-resolution 2D and 3D images of biological specimens. For example, Tomer et. al. report volumetic image data of 4.8 terabytes generated from a tissue of size 0.5x0.5x0.5 cubic-micrometers. Such large datasets necessitate the development of new approaches to data management, storage and analysis. Open sources tools such as OMERO [@omero] and bisque [@bisque] offer solutions for managing a variety of microscopy image data. These tools are designed to be extensible and can be used from programming languages such as python, MATLAB and Java. However, in our experience, using these tools to manage very large datasets that require out-of-core processing has been challenging because of their attempt to load entire datasets into memory. This can make the system unusable for processing big imaging data. SciDB stores data in it’s own native format and provides database functions that allow fast query and extraction of the data of interest. It is also designed to be deployed on massively parallel processing systems and can take advantage of multiple cores on a system.
The Dynamic Distributed Dimensional Data Model (D4M) [@kepner2012] provides a uniform framework, based on the mathematics of associative arrays [@d4m2.0], that can be applied to diverse domains such as cyber, bioinformatics, free text, and social media data. D4M can also be used to perform linear algebraic operations inside a database [@vijay2015]. In this paper, we extend D4M to encompass multidimensional arrays in SciDB using image data management as an example application.
D4M also works seamlessly with the pMatlab (http://www.ll.mit.edu/pMatlab) [@pmatlab; @kepner2009] parallel computing environment which allows high performance parallel applications to be constructed with just a few lines of code. pMatlab uses a single-program-multiple-data (SPMD) parallel programming model and sits on top of a message passing interface (MPI) communication layer. SPMD and MPI are the primary tools used in much of the parallel computing world to achieve the highest levels of performance on the world’s largest systems (see hpcchallenge.org). These tools can also be used for achieving high performance on SciDB.
Section \[sec:technology\] describes SciDB, D4M, pMatlab, and the MIT Supercloud system. Section \[sec:design\] describes the driving application behind this work and Section \[sec:benchmark\] describes the experiments conducted and the measured performance. Finally, Section \[sec:summary\] summarizes the results.
Technologies {#sec:technology}
============
A variety of technologies were used to conduct benchmarks for SciDB data ingest. Together, these technologies make up the MIT SuperCloud (see Figure \[fig:supercloud\]) and are described in the following subsections.
![MIT SuperCloud architecture consists of seven components. (1) Lustre parallel file system for high performance file I/O, (2) D4M & pMatlab ingest processes, (3) parallel databases, (4) D4M & pMatlab analytic processes, (5) database web monitor page, (6) Grid Engine scheduler for allocating processes to hardware and (7) the TX-Green supercomputer.[]{data-label="fig:supercloud"}](llsc.png){width="20pc"}
SciDB {#ssec:scidb}
-----
SciDB is a scalable, computational database management system designed for advanced analytics on multidimensional data [@scidb]. SciDB uses an array data model for storage; thus, it is ideally suited for scientific data such as images, time series data, weather data, and sensor data. Data are stored in the user-defined co-ordinate system such that data that are close to each other in the coordinate system are stored in the same chunk on disk. This storage mechanism has a significant advantage for performing operations such as selecting ranges or joining multiple arrays. Additionally, the number of files read can be minimized because of the ability to specify overlaps in the chunks used to store data on disk. By appropriately specifying the array schema, it possible to optimize the data access and query speeds in SciDB.
SciDB is built to take advantage of massively parallel processing architectures and is highly scalable on commodity hardware. It also provides the capability to run advanced analytics in database, including parallel linear algebra routines.
D4M analytics library {#ssec:d4m}
---------------------
D4M is open-source software that provides a convenient mathematical representation of the kinds of data that are routinely stored in spreadsheets and large triple-store databases. Associations between multidimensional entities (tuples) using string keys and string values can be stored in data structures called associative arrays. For example, in two dimensions, a D4M associative array entry might be:
**A**(alice , bob ) = cited or **A**(alice , bob ) = 47.0
The above tuples have a 1-to-1 correspondence with their triple-store representations:
(alice ,bob ,cited ) or (alice ,bob ,47.0)
Associative arrays can represent complex relationships in either a sparse matrix or a graph form (see Figure \[fig:d4mfig\]). Thus, associative arrays are a natural data structure for performing both matrix and graph algorithms. Such algorithms are the foundation of many complex database operations across a wide range of fields [@kepner2011]. Constructing complex composable query operations can be expressed by using simple array indexing of the associative array keys and values, which themselves return associative arrays:
![A graph describing the relationship between alice, bob, and carl (left). A sparse associative array A captures the same relationships (right). The fundamental operation of graphs is finding neighbors from a vertex (breadth-first search). The fundamental operation of linear algebra is matrix vector multiply. D4M associative arrays make these two operations identical. Thus, algorithm developers can simultaneously use both graph theory and linear algebra to exploit complex data.[]{data-label="fig:d4mfig"}](d4massoc.png){width="20pc"}
----------------------- ---------------------------
**A**(alice ,:) alice row
**A**(alice bob ,:) alice and bob rows
**A**(al\* ,:) rows beginning with al
**A**(alice : bob ,:) rows alice to bob
**A**(1:2, :) first two rows
**A** == 47.0 subarray with values 47.0
----------------------- ---------------------------
The composability of associative arrays stems from their ability to define fundamental mathematical operations whose results are also associative arrays. Given two associative arrays A and B, the results of all the following operations will also be associative arrays:
--------------- --------------- --------------- -------------------- ----------------
**A** + **B** **A** - **B** **A** & **B** **A** $\mid$ **B** **A** \* **B**
--------------- --------------- --------------- -------------------- ----------------
Measurements using D4M indicate these algorithms can be implemented with a tenfold decrease in coding effort when compared to standard approaches [@kepner2012].
pMatlab parallel computing library {#ssec:pmatlab}
----------------------------------
pMatlab is open-source software that allows a Matlab program (mathworks.com) or a GNU Octave program (octave.org) to be launched in parallel. In a pMatlab program all *Np* parallel instances of the program persist for the life of the program, have a unique process identifier (PID), and can directly communicate with all the other instances of the programs. The communication between each PID is handled by message passing. In addition, pMatlab provides scalable mechanisms for creating distributed arrays so that each PID knows exactly which part of the array it owns and where to find all the other pieces.
pMatlab implements the distributed arrays parallel programming model to achieve high performance on the largest computers in the world. This model gives the application precise control of its computations and communications when it is running on a parallel computing system.
Lustre parallel file system {#ssec:lustre}
---------------------------
The MIT SuperCloud has two forms of storage: distributed and central. Distributed storage is comprised of the local storage on each of the compute nodes and this storage is used for running Hadoop and database applications. Central storage is implemented using the open-source Lustre parallel file system (lustre.org) on a commercial storage array. Lustre provides high performance data access to all the compute nodes, while maintaining the appearance of a single filesystem to the user. The Lustre filesystem is used in most of the largest supercomputers in the world.
The MIT SuperCloud leverages both types of storage to dynamically start, stop, checkpoint, relocate, and restart (or clone) databases by storing their data in the Lustre filesystem when the databases are stopped. This dynamic database management system allows many more SciDB databases to be hosted on the system than would otherwise be possible. Groups of users can quickly create their own databases to share data among themselves without interfering with other groups. In addition, because all the database instances are running directly on the compute nodes, they can run at maximum performance.
Grid Engine scheduler {#ssec:sge}
---------------------
Supercomputers require efficient mechanisms for rapidly identifying available computing resources, allocating those resources to programs, and launching the programs on the allocated resources. The open-source Grid Engine software (gridscheduler.sourceforge.net) provides these services and is independent of programming language (C, Fortran, Java, Matlab, etc.) or parallel programming model (message passing, distributed arrays, threads, map/reduce, etc.). The Grid Engine scheduler coordinates the starting and stopping of SciDB database instances in the MIT SuperCloud. SciDB users authenticate by using a web page that shows them only the databases they are allowed to access. They can then start and stop any of these databases. When a database is started, Grid Engine determines the computing requirements of the database, finds the computing resources, allocates them to the database, copies all the database files to the appropriate computing nodes, assigns dynamic alias domain name entries to the compute nodes, and starts the database processes.
TX-Green hardware {#ssec:txgreen}
-----------------
The TX-Green supercomputer consists of 270 HP servers connected to a single, non-blocking 10 GigE Voltaire core switch. The Lustre central storage system uses a 1 Petabyte Data Direct Networks (DDN) storage array and a 5.0 Petabyte Seagate storage array that are directly connected to the core switch. This architecture provides high bandwidth to all the nodes and the central storage. Each compute node has 32 cores (x86 instruction set), 128 Gigabytes of memory, and 12 Terabytes of storage. The storage on each compute node is hot-swappable RAID5 so that each node can tolerate one drive failure.
TX-Green is housed in an HP EcoPOD mobile data center that uses ambient air cooling to maximize energy efficiency. The EcoPOD is located near a hydroelectric dam that delivers clean energy that does not contribute greenhouse gases to the environment.
The MIT SuperCloud software stack, which contains all the systems and applications software, resides on every node. Hosting the application software on each node accelerates the launch of large applications (such as SciDB) and minimizes their dependency on the central storage.
Benchmark Design {#sec:design}
================
SciDB uses a multidimensional array data model for storage. This model makes it well suited for data such as images (2D or 3D), sensor data, and time-series data. Our benchmark is motivated by a biomedical imaging application that involves 3D volumetric data. Advances in imaging technologies are enabling the acquisition of larger and higher resolution biomedical datasets. Imaging techniques such as CLARITY [@clarity2013] have the potential to significantly advance our understanding of brain function by enabling molecular and optical interrogation of brain tissue. However, a significant challenge in this area is the management and analysis of 3D volumetric data generated by using such techniques. Several commercial [@imaris; @metamorph] and open source tools [@imagej; @omero; @vaa3d] are available for the analysis and visualization of biomedical imaging data. Each tool has its advantages and disadvantages, but a common limitation in many of these systems is the inability to analyze and/or visualize datasets that are significantly larger than the total amount of memory available on the system. Additionally, many of these packages are focused on single-client visualization and may not have a programmatic way to serve image data to a variety of clients.
Traditional parallel computing approaches to the analysis of large image datasets involves the use of multiple processors to analyze subsets of images, with each parallel process reading a subset of images files in order to access the data of interest. SciDB in conjunction with pMatlab enables a new mode of parallel operation where a stack of 2D images can be treated as a true volume. SciDB gives us the ability to efficiently extract random, multi-dimensional data using appropriate queries. Consider a 3D array in SciDB with the following schema:
vol3d<val:uint8> [row=1:4096,4096,0,col=1:4096,4096,0,slice=1:1000,1,0]
This schema defines a 3D array in SciDB of size 4096x4096x1000 pixels used to store 8 bit unsigned integer values. By using the SciDB AFL query language, a sub-volume can be extracted as shown below:
between(vol3d, 100, 100, 10, 300, 500, 100);
This query will extract all values in the cube bounded by rows = 100:300, cols = 100:500 and slice = 10:100. This is a powerful capability that we have extended to D4M to extract sub-volumes from SciDB using standard MATLAB indexing syntax as shown in Listing \[matlab\].
DB = DBsetupSciDB('txg-testdb');
T = DB('vol3d<gray:uint8>row=1:4096,4096,0,col=1:4096,4096,0,slice=1:1000,1,0]');
v = T(100:300, 100:500, 10:100);
SciDB data ingest {#ssec:ingest}
-----------------
Since SciDB is a fully ACID database, only one client can ingest data to a given array at a time. In order to maximize the amount of data being ingested, we use multiple MATLAB processes running in parallel. The ACID nature of SciDB results in a serialization of the data ingest process. The solution to this serialization was to ingest data into two stages, as shown in Figure \[fig:ingest\]. We launch multiple parallel MATLAB processes using pMatlab. In the first stage, each MATLAB process uses D4M to ingest data into a new array with the appropriate schema. Once all the MATLAB processes have completed their data ingest, the MATLAB process with rank 0 issues a `merge` statement to combine all the individual arrays into the desired multidimensional array. The merging of individual arrays into a large multi-dimensional array is very fast in SciDB and does not incur any appreciable overhead.
![Data is imported into SciDB in a two step process: Step 1 is to ingest each image into a new array and Step 2 is to merge the arrays into a single multi-dimensional array.[]{data-label="fig:ingest"}](ingest.png){width="20pc"}
Data are ingested into SciDB using D4M using a simple “putTriple” command as shown in Listing \[ingest\]. In this example, slice number 15 of a 1000 slice 3D volume is ingested into an array named `vol3d`.
DB = DBsetupSciDB('txg-testdb');
T = DB('vol3d<gray:uint8>row=1:4096,4096,0,col=1:4096,4096,0,slice=1:1000,1,0]');
im = imread('test-image.tif');
[nr, nc] = size(im);
[rowids, colids] = ind2sub([nr nc], [1:nr*nc]');
slicenum = 15*ones(size(ir));
T = putTriple(T, [rowids colids slicenum], im(:));
Benchmarking data ingest {#sec:benchmark}
------------------------
For this test, we used randomly generated imaging data to simulate a volume of size 5120 x 5120 x 1000 pixels. The data were ingested into SciDB instances configured with different hardware and software specifications. For a single node instance of SciDB, configurations with 1, 4, 8, 12, and 16 SciDB worker processes were used. A two-node SciDB configuration using 2, 4, 8, and 16 SciDB worker threads per node was also used for testing. Data were imported using 2, 4, 8, and 12 processes running in parallel on the same node as the database. Network bandwidth limitations were minimized by running the ingest processes on the same node as the database. Data import used the SciDB shim interface from MATLAB and D4M. Figure \[fig:results\] shows the ingest rates achieved. When importing data into a single-node instance of SciDB, we achieved a maximum ingest rate of 2.23 million entries/second by using 8 parallel MATLAB processes as shown in Figure \[subfig:single-node\]. A maximum ingest rate of 2.876 million entries/second was observed when using 8 parallel MATLAB processes importing into a two node SciDB instance. In each of these configurations, the use of more than 8 parallel processes for importing data resulted in a degradation of performance. Similarly, SciDB configurations using more than 8 SciDB instances per node did not result in an appreciable increase in ingest rates and actually resulted in a slower ingest in the case of the two-node SciDB instance, as shown in Figure \[subfig:two-node\].
\
Summary {#sec:summary}
=======
SciDB is an open-source, array-based database management system designed for advanced analytics on multidimensional data. This paper described the performance of SciDB for ingesting image data. A peak performance of 2,867,910 inserts per second was achieved with a two-node SciDB instance. This performance was achieved using parallel clients running D4M for inserting data into unique arrays and a final merge step. D4M and SciDB offer a new approach to the storage and analysis of large biomedical imaging datasets. To our knowledge, this is the highest database insert rate ever achieved and surpassed our prior single-node record of 800,000 inserts per second achieved with the Accumulo database. By leveraging D4M, it is possible to access large volumetric imaging data stored in SciDB using high-level languages such as MATLAB. The ability to efficiently access any 3D sub-volume from such a dataset gives us a capability that is not easily available with traditional approaches to image data management. Future work in this area includes the implementation of image analysis routines as linear algebra operations that can be run directly in SciDB, thus removing the need to retrieve data from the database. Other areas of research include the development of a more efficient method to import images into SciDB by exploiting sparsity and statistical distribution of the data.
Acknowledgment
==============
The authors would like to thank Prof. Kwanghun Chung (The Picower Institute for Learning and Memory, MIT) for providing imaging data used in the development of D4M extensions and Ms. Dorothy S. Ryan (MIT Lincoln Laboratory) for assistance with proof-reading this article. We would also like to thank David Martinez (Associate Division Head, Cyber Security and Information Sciences, MIT Lincoln Laboratory) and Dr. Jeffrey Palmer (Group Leader, Bioengineering Systems & Technologies, MIT Lincoln Laboratory) for their support.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article, we assign the $D_{s3}^*(2860)$ to be a D-wave $c\bar{s}$ meson, and study the mass and decay constant of the $D_{s3}^*(2860)$ with the QCD sum rules by calculating the contributions of the vacuum condensates up to dimension-6 in the operator product expansion. The predicted mass $M_{D_{s3}^*}=(2.86\pm0.10)\,\rm{GeV}$ is in excellent agreement with the experimental value $M_{D_{s3}^*}=(2860.5\pm 2.6 \pm 2.5\pm 6.0)\,\rm{ MeV}$ from the LHCb collaboration. The present prediction supports assigning the $D_{s3}^*(2860)$ to be the D-wave $c\bar{s}$ meson.'
---
\
Zhi-Gang Wang [^1]\
Department of Physics, North China Electric Power University, Baoding 071003, P. R. China
PACS number: 14.40.Lb, 12.38.Lg
Key words: $D_{s3}^*(2860)$, QCD sum rules
Introduction
============
In 2006, the BaBar collaboration observed the $D^*_{sJ}(2860)$ meson with the mass $(2856.6 \pm 1.5 \pm 5.0)\, \rm{ MeV}$ and the width $(48 \pm 7 \pm 10)\, \rm{ MeV}$ in decays to the final states $D^0 K^+$ and $D^+K^0_S$ using $240\rm{ fb}^{-1}$ of data recorded by the BaBar detector at the PEP-II asymmetric-energy $e^+e^-$ storage rings at the Stanford Linear Accelerator Center [@BaBar2006]. In 2009, the BaBar collaboration confirmed the $D^*_{sJ}(2860)$ in the $D^*K$ channel using $470 {\rm fb}^{-1}$ of data recorded by the BaBar detector, and measured the ratio $R$ among the branching fractions [@BaBar2009], $$\begin{aligned}
R&=& \frac{{\rm Br}\left(D_{sJ}^*(2860)\to D^*K\right)}{{\rm Br}\left(D_{sJ}^*(2860)\to D K\right)}=1.10 \pm 0.15 \pm 0.19\, \, .
\end{aligned}$$ The observation of the decays $D^*_{sJ}(2860)\to D^*K$ rules out the $J^P=0^+$ assignment, the possible assignments are the $1^3{\rm D}_3$ $c\bar{s}$ meson [@Colangelo0607; @Zhang2007; @Li2007; @Zhong2008; @Chen2009; @Zhong2010; @Li0911; @Badalian2011], the $c\bar{s}-cn\bar{s}\bar{n}$ mixing state [@Vijande2009], the dynamically generated $D_1(2420)K$ bound state [@FKGuo], etc.
In 2014, the LHCb collaboration observed a structure at $2.86\,\rm{GeV}$ with significance of more than $10 \sigma$ in the $\overline{D}^0K^-$ mass spectrum in the Dalitz plot analysis of the decays $B_s^0\to \overline{D}^0K^-\pi^+$ using a data sample corresponding to an integrated luminosity of $3.0\rm{ fb}^{-1}$ of $pp$ collision data recorded by the LHCb detector, the structure contains both spin-1 ($D_{s1}^{*-}(2860)$) and spin-3 ($D_{s3}^{*-}(2860)$) components, which can be assigned to be the $J^P =1^-$ and $3^-$ members of the 1D family [@LHCb7574; @LHCb7712]. The measured masses and widths are $M_{D_{s3}^*}=(2860.5\pm 2.6 \pm 2.5\pm 6.0)\,\rm{ MeV}$, $M_{D_{s1}^*}=(2859 \pm 12 \pm 6 \pm 23)\,\rm{ MeV}$, $\Gamma_{D_{s3}^*}=(53 \pm 7 \pm 4 \pm 6)\,\rm{ MeV}$, and $\Gamma_{D_{s1}^*}=(159 \pm 23\pm 27 \pm 72)\,\rm{ MeV}$, respectively. Furthermore, the LHCb collaboration obtained the conclusion that the $D^*_{sJ}(2860)$ observed by the BaBar collaboration in the inclusive $e^+e^- \to \overline{D}^0K^{-}X$ production and by the LHCb collaboration in the $pp \to \overline{D}^0K^{-}X$ processes consists of at least two particles [@BaBar2009; @LHCb1207].
If we assign the $D^*_{sJ}(2860)$ to be the $1^3{\rm D}_3$ state or the $D^*_{s3}(2860)$, the ratio $R$ from the leading order heavy meson effective theory [@Colangelo0607], the ${}^3{\rm P}_0$ model [@Zhang2007; @Li0911; @Song2014] and the relativized quark model [@Godfrey2013] cannot reproduce the experimental value $R=1.10 \pm 0.15 \pm 0.19$ [@BaBar2009]. In Ref.[@WangEPJC2860], we assign the $D_{s3}^*(2860)$ and $D_{s1}^*(2860)$ to be the $1^3{\rm D}_3$ and $1^3{\rm D}_1$ $c\bar{s}$ states, respectively, study their strong decays with the heavy meson effective theory by including the chiral symmetry breaking corrections. We can reproduce the experimental value $R =1.10 \pm 0.15 \pm 0.19$ with suitable hadronic coupling constants if the chiral symmetry breaking corrections are large. The preferred assignment is $D^*_{sJ}(2860)=D^*_{s3}(2860)$, while the assignment $D^*_{sJ}(2860)=D^*_{s1}(2860)$ is not excluded.
According to the predictions of the potential models [@mass-PQM], the masses of the 1D $c\bar{s}$ states are about $2.9\,\rm{GeV}$. It is reasonable to assign the $D_{s1}^*(2860)$ and $D_{s3}^*(2860)$ to be the $\rm{1^3D_1}$ and $\rm{1^3D_3}$ $c\bar{s}$ states, respectively. We can obtain further support by calculating the mass of the $D_{s3}^*(2860)$ based on the QCD sum rules. The QCD sum rules is a powerful theoretical tool in studying the ground state hadrons and has given many successful descriptions of the masses, decay constants, form-factors and hadronic coupling constants, etc [@SVZ79; @Reinders85]. There have been many works on the spin-parity $J^P=0^\pm$, $1^\pm$ heavy-light mesons with the full QCD sum rules [@WangEPJC-HL; @Narison-HL] (and references therein), while the works on the $J^P=2^+$ are few [@Azizi-Tmeson; @Wang-Tmeson], the $J^P=3^-$ heavy-light mesons are only studied with the QCD sum rules combined with the heavy quark effective theory [@Zhu-D-wave]. In this article, we assign the $D_{s3}^*(2860)$ to be a D-wave $c\bar{s}$ meson, study the mass and decay constant of the $D_{s3}^*(2860)$ with the full QCD sum rules in details by calculating the contributions of the vacuum condensates up to dimension-6 in the operator product expansion.
The article is arranged as follows: we derive the QCD sum rules for the mass and decay constant of the $D_{s3}^*(2860)$ in Sect.2; in Sect.3, we present the numerical results and discussions; and Sect.4 is reserved for our conclusions.
QCD sum rules for the $D_{s3}^*(2860)$ as a D-wave meson
=========================================================
In the following, we write down the two-point correlation function $\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)$ in the QCD sum rules, $$\begin{aligned}
\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)&=&i\int d^4x e^{ip \cdot (x-y)} \langle
0|T\left\{J_{\mu\nu\rho}(x)J_{\alpha\beta\sigma}^{\dagger}(y)\right\}|0\rangle\mid_{y=0} \, ,\end{aligned}$$ where the current $$\begin{aligned}
J_{\mu\nu\rho}(x)&=&\overline{c}(x)\left( \gamma_\mu\stackrel{\leftrightarrow}{D}_\nu\stackrel{\leftrightarrow}{D}_\rho
+\gamma_\nu\stackrel{\leftrightarrow}{D}_\rho\stackrel{\leftrightarrow}{D}_\mu+\gamma_\rho\stackrel{\leftrightarrow}{D}_\mu\stackrel{\leftrightarrow}{D}_\nu \right) s(x) \, ,\end{aligned}$$ with $\stackrel{\leftrightarrow}{D}_\mu=\stackrel{\rightarrow}{\partial}_\mu-ig_sG_\mu-\stackrel{\leftarrow}{\partial}_\mu-ig_sG_\mu $ interpolates the D-wave meson $D_{s3}^*(2860)$. We can rewrite the current into two parts, $$\begin{aligned}
J_{\mu\nu\rho}(x)&=&\eta_{\mu\nu\rho}(x)+J^V_{\mu\nu\rho}(x)\, ,\end{aligned}$$ where $$\begin{aligned}
\eta_{\mu\nu\rho}(x)&=&\overline{c}(x)\left( \gamma_\mu\stackrel{\leftrightarrow}{\partial}_\nu\stackrel{\leftrightarrow}{\partial}_\rho
+\gamma_\nu\stackrel{\leftrightarrow}{\partial}_\rho\stackrel{\leftrightarrow}{\partial}_\mu+\gamma_\rho\stackrel{\leftrightarrow}{\partial}_\mu\stackrel{\leftrightarrow}{\partial}_\nu \right) s(x) \, , \\
J^V_{\mu\nu\rho}(x)&=&-2i\,\overline{c}(x)\left[ \gamma_\mu\left( g_sG_\nu\stackrel{\leftrightarrow}{\partial}_\rho
+\stackrel{\leftrightarrow}{\partial}_\nu g_sG_\rho-2ig_s^2 G_\nu G_\rho\right)\right.\nonumber\\
&&\left.+ \gamma_\nu\left( g_sG_\rho\stackrel{\leftrightarrow}{\partial}_\mu
+\stackrel{\leftrightarrow}{\partial}_\rho g_sG_\mu-2ig_s^2 G_\rho G_\mu\right)\right.\nonumber\\
&&\left.+\gamma_\rho\left( g_sG_\mu\stackrel{\leftrightarrow}{\partial}_\nu
+\stackrel{\leftrightarrow}{\partial}_\mu g_sG_\nu-2ig_s^2 G_\mu G_\nu\right)\right] s(x) \, ,\end{aligned}$$ with $\stackrel{\leftrightarrow}{\partial}_\mu=\stackrel{\rightarrow}{\partial}_\mu-\stackrel{\leftarrow}{\partial}_\mu$ and the $G_\mu$ is the gluon field.
We can choose either the partial derivative $\partial_\mu$ or the covariant derivative $D_\mu$ to construct the interpolating currents. The current $J_{\mu\nu\rho}(x)$ with the covariant derivative $D_\mu$ is gauge invariant, but blurs the physical interpretation of the $\stackrel{\leftrightarrow}{D}_\mu$ being the angular momentum. The current $\eta_{\mu\nu\rho}(x)$ with the partial derivative $\partial_\mu$ is not gauge invariant, but manifests the physical interpretation of the $\stackrel{\leftrightarrow}{\partial}_\mu$ being the angular momentum. In this article, we will present the results come from the currents with both the partial derivative and the covariant derivative.
We can insert a complete set of intermediate hadronic states with the same quantum numbers as the current operator $J_{\mu\nu\rho}(x)$ into the correlation function $\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)$ to obtain the hadronic representation [@SVZ79; @Reinders85]. The current $J_{\mu\nu\rho}(0)$ has negative parity, and couples potentially to the $J^P={3}^-$ $\bar{c}s$ meson $D_{s3}^*(2860)$. On the other hand, the current $J_{\mu\nu\rho}(0)$ also couples potentially to the $J^P={2}^+$, $1^-$, $0^+$ $\bar{c}s$ mesons, $$\begin{aligned}
\langle 0|J_{\mu\nu\rho}(0)|D_{s3}^*(p)\rangle&=&f_{D_{s3}^*}\varepsilon_{\mu\nu\rho}(p,\lambda) \, ,\\
\langle 0| J_{\mu\nu\rho}(0)|D_{s2}^*(p)\rangle &=&f_{D_{s2}^*} \left[ p_\mu \varepsilon_{\nu\rho}(p,\lambda)+p_\nu\varepsilon_{\rho\mu}(p,\lambda)+p_\rho\varepsilon_{\mu\nu}(p,\lambda)\right] \, , \nonumber\\
\langle 0| J_{\mu\nu\rho}(0)|D_{s1}^*(p)\rangle &=&f_{D_{s1}^*} \left[ p_\mu p_\nu \varepsilon_{\rho}(p,\lambda)+p_\nu p_\rho \varepsilon_{\mu}(p,\lambda)+p_\rho p_\mu\varepsilon_{\nu}(p,\lambda)\right] \, , \nonumber\\
\langle 0| J_{\mu\nu\rho}(0)|D_{s0}^*(p)\rangle &=&f_{D_{s0}^*} p_\mu p_\nu p_{\rho} \, ,\end{aligned}$$ where the $f_{D_{s3}^*}$, $f_{D_{s2}^*}$, $f_{D_{s1}^*}$ and $f_{D_{s0}^*}$ are the decay constants, the $\varepsilon_{\mu\nu\rho}(p,\lambda)$, $\varepsilon_{\mu\nu}(p,\lambda)$ and $\varepsilon_{\mu}(p,\lambda)$ are the polarization vectors of the $\bar{c}s$ mesons with the following properties [@JJZhu], $$\begin{aligned}
{\rm P}_{\mu\nu\rho\alpha\beta\sigma}&=&\sum_{\lambda} \varepsilon^*_{\mu\nu\rho}(\lambda,p)\varepsilon_{\alpha\beta\sigma}(\lambda,p) \nonumber\\
&=&\frac{1}{6}\left(\widetilde{g}_{\mu\alpha}\widetilde{g}_{\nu\beta}\widetilde{g}_{\rho\sigma}+\widetilde{g}_{\mu\alpha}\widetilde{g}_{\nu\sigma}\widetilde{g}_{\rho\beta}
+\widetilde{g}_{\mu\beta}\widetilde{g}_{\nu\alpha}\widetilde{g}_{\rho\sigma} +\widetilde{g}_{\mu\beta}\widetilde{g}_{\nu\sigma}\widetilde{g}_{\rho\alpha}
+\widetilde{g}_{\mu\sigma}\widetilde{g}_{\nu\alpha}\widetilde{g}_{\rho\beta}+\widetilde{g}_{\mu\sigma}\widetilde{g}_{\nu\beta}\widetilde{g}_{\rho\alpha}\right)\nonumber\\
&&-\frac{1}{15}\left(\widetilde{g}_{\mu\alpha}\widetilde{g}_{\nu\rho}\widetilde{g}_{\beta\sigma}+\widetilde{g}_{\mu\beta}\widetilde{g}_{\nu\rho}\widetilde{g}_{\alpha\sigma}
+\widetilde{g}_{\mu\sigma}\widetilde{g}_{\nu\rho}\widetilde{g}_{\alpha\beta} +\widetilde{g}_{\nu\alpha}\widetilde{g}_{\mu\rho}\widetilde{g}_{\beta\sigma}
+\widetilde{g}_{\nu\beta}\widetilde{g}_{\mu\rho}\widetilde{g}_{\alpha\sigma} +\widetilde{g}_{\nu\sigma}\widetilde{g}_{\mu\rho}\widetilde{g}_{\alpha\beta}\right. \nonumber\\
&&\left. +\widetilde{g}_{\rho\alpha}\widetilde{g}_{\mu\nu}\widetilde{g}_{\beta\sigma} +\widetilde{g}_{\rho\beta}\widetilde{g}_{\mu\nu}\widetilde{g}_{\alpha\sigma}
+\widetilde{g}_{\rho\sigma}\widetilde{g}_{\mu\nu}\widetilde{g}_{\alpha\beta}\right) \, , \\
{\rm P}_{\mu\nu\alpha\beta}&=& \sum_\lambda \varepsilon^*_{\mu\nu}(\lambda,p)\varepsilon_{\alpha\beta}(\lambda,p)=\frac{\widetilde{g}_{\mu\alpha}\widetilde{g}_{\nu\beta}
+\widetilde{g}_{\mu\beta}\widetilde{g}_{\nu\alpha}}{2}-\frac{\widetilde{g}_{\mu\nu}\widetilde{g}_{\alpha\beta}}{3} \, , \\
\widetilde{g}_{\mu\nu} &=&\sum_\lambda \varepsilon^*_{\mu}(\lambda,p)\varepsilon_{\nu}(\lambda,p)=-g_{\mu\nu}+\frac{p_\mu p_\nu}{p^2} \, .
\end{aligned}$$
The correlation function can be written into the following form according to Lorentz covariance, $$\begin{aligned}
\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)&=&\Pi(p^2){\rm P}_{\mu\nu\rho\alpha\beta\sigma} +\Pi_2(p^2)\left[ {\rm P}_{\nu\rho\beta\sigma}\,p_\mu p_\alpha + {\rm P}_{\nu\rho\alpha\sigma}\,p_\mu p_\beta + {\rm P}_{\nu\rho\alpha\beta}\,p_\mu p_\sigma + {\rm P}_{\mu\rho\beta\sigma}\,p_\nu p_\alpha\right.\nonumber\\
&&\left.+ {\rm P}_{\mu\rho\alpha\sigma}\,p_\nu p_\beta+ {\rm P}_{\mu\rho\alpha\beta}\,p_\nu p_\sigma+ {\rm P}_{\mu\nu\beta\sigma}\,p_\rho p_\alpha+ {\rm P}_{\mu\nu\alpha\sigma}\,p_\rho p_\beta+ {\rm P}_{\mu\nu\alpha\beta}\,p_\rho p_\sigma\right] \nonumber\\
&&+\Pi_1(p^2)\left[\widetilde{g}_{\mu \alpha}\, p_\nu p_\rho p_\beta p_\sigma+\widetilde{g}_{\mu \beta}\, p_\nu p_\rho p_\alpha p_\sigma
+\widetilde{g}_{\mu \sigma}\, p_\nu p_\rho p_\alpha p_\beta +\widetilde{g}_{\nu \alpha}\, p_\mu p_\rho p_\beta p_\sigma\right.\nonumber\\
&&\left.+\widetilde{g}_{\nu \beta} \,p_\mu p_\rho p_\alpha p_\sigma+\widetilde{g}_{\nu \sigma}\, p_\mu p_\rho p_\alpha p_\beta
+\widetilde{g}_{\rho \alpha} \, p_\mu p_\nu p_\beta p_\sigma+\widetilde{g}_{\rho \beta}\, p_\mu p_\nu p_\alpha p_\sigma
+\widetilde{g}_{\rho \sigma}\, p_\mu p_\nu p_\alpha p_\beta\right] \nonumber\\
&&+\Pi_0(p_2)\, p_\mu p_\nu p_\rho p_\alpha p_\beta p_\sigma \, ,\end{aligned}$$ the components $\Pi_2(p^2)$, $\Pi_1(p^2)$ and $\Pi_0(p^2)$ come from the contributions of the $J^P=2^+$, $1^-$ and $0^+$ $\bar{c}s$ mesons, respectively.
We isolate the ground state contribution from the $D_{s3}^*(2860)$ and get the following result, $$\begin{aligned}
\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)&=&\frac{f_{D_{s3}^*}^2}{M_{D_{s3}^*}^2-p^2}{\rm P}_{\mu\nu\rho\alpha\beta\sigma} +\cdots\, ,\nonumber\\
&=&\Pi(p^2){\rm P}_{\mu\nu\rho\alpha\beta\sigma}+\cdots\, .\end{aligned}$$ We can project out the component $\Pi(p^2)$, $$\begin{aligned}
\Pi(p^2)&=&\frac{1}{7}{\rm P}^{\mu\nu\rho\alpha\beta\sigma}\Pi_{\mu\nu\rho\alpha\beta\sigma}(p) \, ,\end{aligned}$$ according to the properties, $$\begin{aligned}
p^\mu{\rm P}_{\mu\nu\rho\alpha\beta\sigma}=p^\nu{\rm P}_{\mu\nu\rho\alpha\beta\sigma}=p^\rho{\rm P}_{\mu\nu\rho\alpha\beta\sigma}=p^\alpha{\rm P}_{\mu\nu\rho\alpha\beta\sigma}=p^\beta{\rm P}_{\mu\nu\rho\alpha\beta\sigma}=p^\sigma{\rm P}_{\mu\nu\rho\alpha\beta\sigma}=0\,.\end{aligned}$$
Now, we briefly outline the operator product expansion for the correlation function $\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)$ in perturbative QCD. We contract the quark fields in the correlation function $\Pi_{\mu\nu\rho\alpha\beta\sigma}(p)$ with Wick theorem firstly, $$\begin{aligned}
\Pi(p^2)&=&-\frac{i}{7}{\rm P}^{\mu\nu\rho\alpha\beta\sigma}\int d^4x e^{ip \cdot (x-y)} Tr\left\{\Gamma_{\mu\nu\rho}^{ik}(x)S_{k l}(x-y)\Gamma_{\alpha\beta\sigma}^{l j}(y) S^c_{ji}(y-x) \right\}\mid_{y=0}\, ,\end{aligned}$$ where $\Gamma_{\mu\nu\rho}(x)$ and $\Gamma_{\alpha\beta\sigma}(y)$ are the vertexes, $$\begin{aligned}
S_{ij}(x)&=& \frac{i\delta_{ij}\!\not\!{x}}{ 2\pi^2x^4}
-\frac{\delta_{ij}m_s}{4\pi^2x^2}-\frac{\delta_{ij}}{12}\langle\bar{s}s\rangle +\frac{i\delta_{ij}\!\not\!{x}m_s
\langle\bar{s}s\rangle}{48}-\frac{\delta_{ij}x^2\langle \bar{s}g_s\sigma Gs\rangle}{192}+\frac{i\delta_{ij}x^2\!\not\!{x} m_s\langle \bar{s}g_s\sigma
Gs\rangle }{1152}\nonumber\\
&& -\frac{ig_s G^{a}_{\alpha\beta}t^a_{ij}(\!\not\!{x}
\sigma^{\alpha\beta}+\sigma^{\alpha\beta} \!\not\!{x})}{32\pi^2x^2} -\frac{1}{8}\langle\bar{s}_j\sigma^{\mu\nu}s_i \rangle \sigma_{\mu\nu} +\cdots \, ,\end{aligned}$$ $$\begin{aligned}
S^c_{ij}(x)&=&\frac{i}{(2\pi)^4}\int d^4k e^{-ik \cdot x} \left\{
\frac{\delta_{ij}}{\!\not\!{k}-m_c}
-\frac{g_sG^n_{\alpha\beta}t^n_{ij}}{4}\frac{\sigma^{\alpha\beta}(\!\not\!{k}+m_c)+(\!\not\!{k}+m_c)
\sigma^{\alpha\beta}}{(k^2-m_c^2)^2}\right.\nonumber\\
&& -\frac{g_s^2 (t^at^b)_{ij} G^a_{\alpha\beta}G^b_{\mu\nu}(f^{\alpha\beta\mu\nu}+f^{\alpha\mu\beta\nu}+f^{\alpha\mu\nu\beta}) }{4(k^2-m_c^2)^5}\nonumber\\
&&\left.+\frac{i\langle g_s^3GGG\rangle}{48}\frac{(\!\not\!{k}+m_c)\left[\!\not\!{k}(k^2-3m_c^2)+2m_c(2k^2-m_c^2) \right](\!\not\!{k}+m_c)}{(k^2-m_c^2)^6}+\cdots\right\}\, ,\nonumber\\
f^{\alpha\beta\mu\nu}&=&(\!\not\!{k}+m_c)\gamma^\alpha(\!\not\!{k}+m_c)\gamma^\beta(\!\not\!{k}+m_c)\gamma^\mu(\!\not\!{k}+m_c)\gamma^\nu(\!\not\!{k}+m_c)\, ,\end{aligned}$$ $t^n=\frac{\lambda^n}{2}$, the $\lambda^n$ is the Gell-Mann matrix, the $i$, $j$ are color indexes; then compute the integrals both in the coordinate and momentum spaces; finally obtain the QCD spectral density through dispersion relation, $$\begin{aligned}
\Pi(p^2)&=&\frac{1}{\pi}\int_{m_c^2}^\infty \frac{{\rm Im}\Pi(s)}{s-p^2}=\int_{m_c^2}^\infty \frac{\rho_{QCD}(s)}{s-p^2}\, .\end{aligned}$$ In Eq.(17), we retain the term $\langle\bar{s}_j\sigma_{\mu\nu}s_i \rangle$ originates from the Fierz re-ordering of the $\langle s_i \bar{s}_j\rangle$ to absorb the gluons emitted from the heavy quark line to form $\langle\bar{s}_j g_s G^a_{\alpha\beta} t^a_{mn}\sigma_{\mu\nu} s_i \rangle$ to extract the mixed condensate $\langle\bar{s}g_s\sigma G s\rangle$. There are no contributions come from the terms $\langle\bar{s} s\rangle$ and $\langle\bar{s}g_s\sigma G s\rangle$ due to the projector ${\rm P}_{\mu\nu\rho\alpha\beta\sigma}$. So it is convenient to use the following $s$ quark propagator, $$\begin{aligned}
S_{ij}(x)&=&S^c_{ij}(x)|_{m_c \to m_s}\, .\end{aligned}$$ We take into account all the Feynman diagrams shown explicitly in Figs.1-2, which contribute to the gluon condensate and three-gluon condensate. In the Feynman diagrams, we use the solid and dashed lines to represent the light and heavy quark propagators, respectively. In the fixed point gauge, $G_\mu(x)=\frac{1}{2}x^\theta G_{\theta\mu}(0)+\cdots$ and $G_\alpha(y)=\frac{1}{2}y^\theta G_{\theta\alpha}(0)+\cdots=0$. So in the $\Gamma_{\alpha\beta\sigma}(y)$, we can set $G_\alpha(y)=G_\beta(y)=G_\sigma(y)=0$, there are no gluon lines associated with the vertexes at the point $y=0$ in Fig.2.
{width="14cm"}
{width="14cm"} +4mm {width="14cm"} +4mm {width="14cm"}
We take quark-hadron duality below the continuum threshold $s_0$ and perform the Borel transform with respect to the variable $P^2=-p^2$ to obtain the QCD sum rule, $$\begin{aligned}
f_{D_{s3}^*}^2 \exp\left(-\frac{M_{D_{s3}^*}^2}{T^2}\right)&=&B_{T^2}\Pi(p^2) \, ,\end{aligned}$$ where $$\begin{aligned}
B_{T^2}\Pi(p^2)&=& \frac{9}{140\pi^2} \int_{m_c^2}^{s_0} ds \frac{(s-m_c^2)^6(4s+3m_c^2)+14m_s m_c s(s-m_c^2)^5}{s^4} \exp\left(-\frac{s}{T^2}\right) \nonumber \\
&& +m_c^4\langle \frac{\alpha_sGG}{\pi}\rangle \left\{ \Pi_{GG}^\eta +\Pi_{GG}^V \right\} +\frac{\langle g_s^3 GGG\rangle}{48\pi^2}\left\{ \Pi_{GGG}^\eta +\Pi_{GGG}^V \right\} \, ,\end{aligned}$$
$$\begin{aligned}
\Pi_{GG}^\eta&=& \left[\frac{3}{5}+\frac{17m_c^2}{10T^2}+\frac{3m_c^4}{10T^4}+\frac{T^2}{5m_c^2}-\frac{4T^4}{5m_c^4}\right]\exp\left( -\frac{m_c^2}{T^2}\right)- \left[\frac{2m_c^2}{T^2}+\frac{2m_c^4}{T^4}+\frac{3m_c^6}{10T^6}\right]\Gamma\left( 0,\frac{m_c^2}{T^2}\right)\, , \nonumber\\\end{aligned}$$
$$\begin{aligned}
\Pi_{GGG}^\eta&=& \left[51+\frac{81m_c^2}{T^2}+\frac{12m_c^4}{T^4}-\frac{15T^2}{m_c^2}\right]\exp\left( -\frac{m_c^2}{T^2}\right)- \left[\frac{120m_c^2}{T^2}+\frac{93m_c^4}{T^4}+\frac{12m_c^6}{T^6}\right]\Gamma\left( 0,\frac{m_c^2}{T^2}\right) \nonumber\\
&&+\frac{9m_c^6}{8T^6}\log\frac{m_c^2}{T^2}\, ,\end{aligned}$$
$$\begin{aligned}
\Pi_{GG}^V&=& \left[-\frac{11}{10}+\frac{7m_c^2}{40T^2}+\frac{3m_c^4}{40T^4}+\frac{21T^2}{20m_c^2}-\frac{17T^4}{10m_c^4}\right]\exp\left( -\frac{m_c^2}{T^2}\right)+ \left[\frac{m_c^2}{T^2}-\frac{m_c^4}{4T^4}-\frac{3m_c^6}{40T^6}\right]\Gamma\left( 0,\frac{m_c^2}{T^2}\right) \, ,\nonumber\\\end{aligned}$$
$$\begin{aligned}
\Pi_{GGG}^V&=& \left[33+\frac{81m_c^2}{2T^2}+\frac{9m_c^4}{2T^4}-\frac{6T^2}{m_c^2}\right]\exp\left( -\frac{m_c^2}{T^2}\right)- \left[\frac{69m_c^2}{T^2}+\frac{45m_c^4}{T^4}+\frac{9m_c^6}{2T^6}\right]\Gamma\left( 0,\frac{m_c^2}{T^2}\right) \, , \nonumber\\\end{aligned}$$
$\Gamma(0,x)=\int_0^\infty dt \frac{1}{t}e^{-xt}$, the superscripts $\eta$ and $V$ denote the contributions come from the Feynman diagrams in Fig.1 and Fig.2, respectively.
We differentiate Eq.(21) with respect to $\frac{1}{T^2}$, then eliminate the decay constant $f_{D_{s3}^*}$, and obtain the QCD sum rule for the mass of the $D_{s3}^*(2860)$, $$\begin{aligned}
M_{D_{s3}^*}^2 &=& -\frac{\frac{d }{d(1/T^2)}B_{T^2}\Pi(p^2)}{B_{T^2}\Pi(p^2)} \, .
\end{aligned}$$ Once the mass $M_{D_{s3}^*}$ is obtained, we can take it as the input parameter and obtain the decay constant from the QCD sum rule in Eq.(21).
Numerical results and discussions
=================================
The values of the gluon condensate and three-gluon condensate can be taken to be the standard values (SV) $\langle \frac{\alpha_s GG}{\pi}\rangle=0.012 \,\rm{GeV}^4 $ and $\langle g_s^3 GGG\rangle=0.045\,\rm{GeV}^6$ [@SVZ79; @Reinders85; @CReview]. The value of the gluon condensate $\langle \frac{\alpha_s
GG}{\pi}\rangle $ has been updated from time to time, and changes greatly [@NarisonBook], we can choose the recently updated value $\langle \frac{\alpha_s GG}{\pi}\rangle=(0.022 \pm
0.004)\,\rm{GeV}^4 $ [@gg-conden], and take the three-gluon condensate as $\langle g_s^3 GGG\rangle=(8.8\pm5.5)\,{\rm{GeV}^2}\langle\alpha_s GG\rangle=(0.616\pm0.385)\,\rm{GeV}^6$ [@gg-conden]. The most recent value of the gluon condensate from the QCD sum rules is $\langle \frac{\alpha_s GG}{\pi}\rangle=(0.037 \pm 0.015)\,\rm{GeV}^4 $ [@Dominguez-gg], but the value of the three-gluon condensate is not updated. Thereafter the recently updated values (UV) of the gluon condensate and three-gluon condensate in Ref.[@gg-conden] will be referred to as UV. The SV and UV differ from each other greatly, there are no overlaps between the two sets of parameters, we obtain the mass and decay constant with the two sets of parameters separately, one can take the average values.
As the quark masses are concerned, we can take the $\overline{MS}$ masses $m_{c}(m_c)=(1.275\pm0.025)\,\rm{GeV}$ and $m_s(\mu=2\,\rm{GeV})=(0.095\pm0.005)\,\rm{GeV}$ from the Particle Data Group [@PDG], and take into account the energy-scale dependence of the $\overline{MS}$ masses from the renormalization group equation, $$\begin{aligned}
m_s(\mu)&=&m_s({\rm 2GeV} )\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}({\rm 2GeV})}\right]^{\frac{4}{9}} \, ,\nonumber\\
m_c(\mu)&=&m_c(m_c)\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}(m_c)}\right]^{\frac{12}{25}} \, ,\nonumber\\
\alpha_s(\mu)&=&\frac{1}{b_0t}\left[1-\frac{b_1}{b_0^2}\frac{\log t}{t} +\frac{b_1^2(\log^2{t}-\log{t}-1)+b_0b_2}{b_0^4t^2}\right]\, ,\end{aligned}$$ where $t=\log \frac{\mu^2}{\Lambda^2}$, $b_0=\frac{33-2n_f}{12\pi}$, $b_1=\frac{153-19n_f}{24\pi^2}$, $b_2=\frac{2857-\frac{5033}{9}n_f+\frac{325}{27}n_f^2}{128\pi^3}$, $\Lambda=213\,\rm{MeV}$, $296\,\rm{MeV}$ and $339\,\rm{MeV}$ for the flavors $n_f=5$, $4$ and $3$, respectively [@PDG]. In calculations, we can take $n_f=4$. We usually take the energy scale $\mu_{D}=1\,\rm{GeV}$ for the pseudoscalar $D$ meson, if we count the contributions of the additional P-wave and D-wave as $0.5\,\rm{GeV}$ and $1\,\rm{GeV}$, respectively, and assume the flavor $SU(3)$ breaking effect is about $0.1\,\rm{GeV}$ [@WangEPJC-HL], then $\mu_{D_{s3}^*}=2.1\,\rm{GeV}$ for the $D_{s3}^*(2860)$, which works well. Thereafter the $\overline{MS}$ masses $m_{c}({\rm 2.1\,GeV})$ and $m_s({\rm 2.1\,GeV})$ will be referred to as MS.
The heavy quark masses appearing in perturbative calculations are usually taken to be the pole masses. The $\overline{MS}$ mass $m_c(m_c)$ relates with the pole mass $m_c$ through the relation [@PDG], $$\begin{aligned}
m_c&=& m_c(m_c)\left[1+\frac{4 \alpha_s(m_c)}{3\pi}+\cdots\right]\, .\end{aligned}$$ We can take the approximation $m_c\approx m_c(m_c)$ without the perturbative corrections for consistency. The value listed in the Particle Data Group is $m_c(m_c)=(1.275\pm0.025) \,\rm{GeV}$ [@PDG], it is reasonable to take the pole mass $m_c=(1.275\pm0.025)\,\rm{GeV}$. Up to corrections of the order $\mathcal{O}\left({\alpha_s}^3\right)$, the $\overline{MS}$ mass $m_c(m_c)=(1.275\pm0.025) \,\rm{GeV}$ corresponds to the pole mass $m_c=(1.67\pm0.07)\,\rm{GeV}$ [@PDG], which is too large for the QCD sum rules, as $2m_c>M_{J/\psi}={3.096916\,\rm{GeV}}>M_{\eta_c}=2.9836\,\rm{GeV}$ [@PDG]. We can also take the pole masses $m_c=(1.275\pm0.025)\,\rm{GeV}$ and $m_s=m_s(1{\rm GeV})=0.13\,\rm{GeV}$ [@PDG].
Now we sum up the QCD input parameters, which are shown explicitly in Table 1. There are four sets of input parameters after taking into account the $\overline{MS}$ masses and pole masses, see Table 2 and Table 3.
Notations Values
----------- ---------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- --
SV $\langle\frac{\alpha_s GG}{\pi}\rangle=0.012 \,\rm{GeV}^4 $, $\langle g_s^3 GGG\rangle=0.045\,\rm{GeV}^6$ [@SVZ79; @Reinders85; @CReview]
UV $\langle \frac{\alpha_s GG}{\pi}\rangle=(0.022 \pm 0.004)\,\rm{GeV}^4 $, $\langle g_s^3 GGG\rangle=(0.616\pm0.385)\,\rm{GeV}^6$ [@gg-conden]
MS $m_{c}({\rm 2.1\,GeV})$, $m_s({\rm 2.1\,GeV})$ [@PDG]
PM $m_c=(1.275\pm0.025)\,\rm{GeV}$, $m_s=0.13\,\rm{GeV}$ [@PDG]
: The QCD input parameters, where the SV, UV, MS and PM denote the standard values, the updated values, the $\overline{MS}$ masses and the pole masses, respectively, the $\overline{MS}$ masses $m_{c}({\rm 2.1\,GeV})$ and $m_s({\rm 2.1\,GeV})$ are obtained according to Eq.(28).
The $D_{s3}^*(2860)$ is a conventional meson, we take it for granted that the energy gap between the ground state and the first radial excited state is about $0.5\,\rm{GeV}$. The measured mass and width are $M_{D_{s3}^*}=(2860.5\pm 2.6 \pm 2.5\pm 6.0)\,\rm{ MeV}$ and $\Gamma_{D_{s3}^*}=(53 \pm 7 \pm 4 \pm 6)\,\rm{ MeV}$, respectively [@LHCb7574; @LHCb7712]. So we take the threshold parameter $\sqrt{s_0}=2.9+(0.4-0.6)\,\rm{GeV}$ to avoid the contaminations of the high resonances and continuum states. We impose the two criteria (pole dominance and convergence of the operator product expansion) of the QCD sum rules on the $D_{s3}^*(2860)$, and search for the optimal values of the Borel parameters. The resulting Borel parameters and pole contributions are shown explicitly in Table 2. From Table 2, we can see that the pole dominance is well satisfied.
Input parameters $T^2 (\rm{GeV}^2)$ pole
------------------ -------------------- ------------- -- -- -- -- --
MS+SV $1.9-2.5$ $(46-78)\%$
MS+UV $1.7-2.3$ $(51-83)\%$
PM+SV $1.6-2.2$ $(51-85)\%$
PM+UV $1.4-2.0$ $(57-90)\%$
: The Borel parameters and pole contributions of the QCD sum rules.
In Fig.3, we plot the contributions come from different terms in the operator product expansion with variations of the Borel parameters $T^2$. From the figure we can see that if the $\overline{MS}$ masses are chosen, the contributions of the three-gluon condensate reach zero at about $T^2=0.9\,\rm{GeV}^2$, on the other hand, if the pole masses are chosen, the contributions of the three-gluon condensate reach zero at about $T^2=1.0-1.2\,\rm{GeV}^2$. From Table 2, we can see that the Borel parameters are larger than the low bound $T^2=0.9\,\rm{GeV}^2$ or $T^2=1.0-1.2\,\rm{GeV}^2$, the operator product expansion is well convergent. In calculations, we observe that it is impossible to obtain the Borel platforms at the low bound $T^2=0.9\,\rm{GeV}^2$ or $T^2=1.0-1.2\,\rm{GeV}^2$, and postpone the Borel parameters to larger values. The two criteria of the QCD sum rules are fully satisfied, so we expect to obtain reasonable predictions.
{width="7cm"} {width="7cm"} {width="7cm"} {width="7cm"}
Now we take into account the uncertainties of the input parameters and obtain the mass and decay constant of the $D_{s3}^*(2860)$, which are shown explicitly in Figs.4-5, and Table 3. From the figures, we can see that they are rather stable with variations of the Borel parameters in the Borel windows, it is reliable to extract the mass and decay constant.
The predicted mass $M_{D_{s3}^*}=(2.86\pm0.10)\,\rm{GeV}$ is in excellent agreement with the experimental value $M_{D_{s3}^*}=(2860.5\pm 2.6 \pm 2.5\pm 6.0)\,\rm{ MeV}$ from the LHCb collaboration [@LHCb7574; @LHCb7712]. The calculations based on the QCD sum rules also support assigning the $D_{s3}^*(2860)$ to be the D-wave $\bar{c}s$ meson, and the predicted decay constant $f_{D_{s3}^*}$ can be used to study the hadronic coupling constants involving the $D_{s3}^*(2860)$ with the three-point QCD sum rules or the light-cone QCD sum rules. In the four cases, MS+SV, MS+UV, PM+SV, PM+UV, the predicted mass $M_{D_{s3}^*}=(2.86\pm0.10)\,\rm{GeV}$ remains the same, but the predicted decay constant varies greatly. It is not unreasonable, as we extract the mass and decay constant from different Borel windows, different Borel windows correspond to different predictions, we choose the special Borel windows to reproduce the experimental value of the mass. The decay constant $f_{D_{s3}^*}$ cannot be extracted from the experimental data, we have to calculate it by some theoretical methods, the true value cannot be obtained. So in calculating the hadronic coupling constants or form-factors involving the $D_{s3}^*(2860)$ with the three-point QCD sum rules, we must use the value of the decay constant $f_{D_{s3}^*}$ in a consistent way. The average value is about $f_{D_{s3}^*}=5.46\pm1.02\,\rm{GeV}^4$.
The measured mass and width are $M_{D_{s3}^*}=(2860.5\pm 2.6 \pm 2.5\pm 6.0)\,\rm{ MeV}$ and $\Gamma_{D_{s3}^*}=(53 \pm 7 \pm 4 \pm 6)\,\rm{ MeV}$ [@LHCb7574; @LHCb7712], the threshold parameter $\sqrt{s_0}>M_{D_{s3}^*}+\frac{\Gamma_{D_{s3}^*}}{2}\approx 2.9\,\rm{GeV}$. Now we vary the threshold parameter at a larger interval, $\sqrt{s_0}=3.4\pm 0.3 \,\rm{GeV}$ in stead of $\sqrt{s_0}=3.4\pm0.1\,\rm{GeV}$, then the uncertainty $\delta \sqrt{s_0}$ leads to the uncertainty $\delta M_{D_{s3}^*}=\pm 0.14\,\rm{GeV}$, $\pm 0.13\,\rm{GeV}$, $\pm 0.13\,\rm{GeV}$, $\pm 0.12\,\rm{GeV}$ in stead of $\delta M_{D_{s3}^*}=\pm 0.05\,\rm{GeV}$, $0.05\,\rm{GeV}$, $0.05\,\rm{GeV}$, $0.04\,\rm{GeV}$ for the input parameters MS+SV, MS+UV, PM+SV, PM+UV, respectively. The larger uncertainty $\delta\sqrt{s_0}= \pm 0.3 \,\rm{GeV}$ leads to additional uncertainty about $\delta M_{D_{s3}^*}/M_{D_{s3}^*}\approx \pm 3\%$ compared to the uncertainty $\delta\sqrt{s_0}= \pm 0.1 \,\rm{GeV}$.
{width="7cm"} {width="7cm"} {width="7cm"} {width="7cm"}
{width="7cm"} {width="7cm"} {width="7cm"} {width="7cm"}
Now we explore which interplaiting current is preferred. In Figs.6-7, we plot the mass and decay constant in the case “MS+SV” with variations of the Borel parameter in the QCD sum rules, where the currents $J_{\mu\nu\rho}(x)$ and $\eta_{\mu\nu\rho}(x)$ are chosen or only the perturbative terms are included. From the figures, we can see that the gluon condensate and three-gluon condensate play an important in determining the Borel window even in the case small values of the vacuum condensates are chosen, see Table 1, while in the Borel window, they play a minor important role. Furthermore, we can obtain better QCD sum rules for the current $J_{\mu\nu\rho}(x)$ compared to the current $\eta_{\mu\nu\rho}(x)$, so the covariant derivative is preferred in constructing the currents.
{width="10cm"}
{width="10cm"}
Input parameters $ M_{D_{s3}^*} (\rm{GeV})$ $f_{D_{s3}^*}(\rm{GeV}^4)$
------------------ ---------------------------- ---------------------------- -- -- -- -- --
MS+SV $2.86\pm0.10$ $6.02\pm1.02$
MS+UV $2.86\pm0.10$ $5.82\pm1.01$
PM+SV $2.86\pm0.10$ $5.14\pm1.00$
PM+UV $2.86\pm0.10$ $4.87\pm1.05$
: The mass and decay constant from the QCD sum rules with different input parameters.
In Fig.8, we plot the predicted mass $M_{D_{s3}^*}$ with variations of the Borel parameter $T^2$ and $c$-quark mass $m_c$ for both the standard values and updated values of the gluon condensate and three-gluon condensate. From the figure, we can see that all the lineshapes of the predicted mass $M_{D_{s3}^*}$ cross the experimental value, at the vicinity of the crossover points, the lineshapes of the predicted mass $M_{D_{s3}^*}$ with smaller $c$-quark mass are more flat than that with larger $c$-quark mass. The $\overline{MS}$ mass $m_c(2.1\,\rm{GeV})=1.120\,\rm{GeV}$ is much smaller than the pole mass $m_c=1.275\,\rm{GeV}$, we prefer the $\overline{MS}$ mass. The $c$-quark mass $m_c=1.5\,\rm{GeV}$ is meaningless as $2m_c>M_{\eta_c}=2.9836\,\rm{GeV}$ [@PDG], the value $m_c=1.5\,\rm{GeV}$ should be discarded. In calculations, we observe that at the vicinity of the crossover point of the lineshapes $M_{D_{s3}^*}=2.86\,\rm{GeV}$ and $m_c=1.1\,\rm{GeV}$, if the same Borel parameter $T^2$ is chosen, the predicted mass $M_{D_{s3}^*}$ increases about $0.03\,\rm{GeV}$ with the increasement $\delta m_c=0.1\,\rm{GeV}$; while at the vicinity of the crossover point of the lineshapes $M_{D_{s3}^*}=2.86\,\rm{GeV}$ and $m_c=1.3\,\rm{GeV}$, if the same Borel parameter $T^2$ is chosen, the predicted mass $M_{D_{s3}^*}$ decreases about $(0.03-0.04)\,\rm{GeV}$ with the decreasement $\delta m_c=-0.1\,\rm{GeV}$.
{width="7cm"} {width="7cm"}
Conclusion
==========
In this article, we assign the $D_{s3}^*(2860)$ to be a D-wave $c\bar{s}$ meson, and study the mass and decay constant of the $D_{s3}^*(2860)$ with the QCD sum rules by calculating the contributions of the vacuum condensates up to dimension-6 in the operator product expansion. The predicted mass $M_{D_{s3}^*}=(2.86\pm0.10)\,\rm{GeV}$ is in excellent agreement with the experimental value $M_{D_{s3}^*}=(2860.5\pm 2.6 \pm 2.5\pm 6.0)\,\rm{ MeV}$ from the LHCb collaboration. The prediction supports assigning the $D_{s3}^*(2860)$ to be the D-wave $\bar{c}s$ meson. While the predicted decay constant $f_{D_{s3}^*}$ can be used to study the hadronic coupling constants involving the $D_{s3}^*(2860)$ with the three-point QCD sum rules or the light-cone QCD sum rules.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by National Natural Science Foundation, Grant Numbers 11375063, and Natural Science Foundation of Hebei province, Grant Number A2014502017.
[99]{}
B. Aubert et al, Phys. Rev. Lett. [**97**]{} (2006) 222001.
B. Aubert et al, Phys. Rev. [**D80**]{} (2009) 092003.
P. Colangelo, F. De Fazio and S. Nicotri, Phys. Lett. [**B642**]{} (2006) 48.
B. Zhang, X. Liu, W. Z. Deng and S. L. Zhu, Eur. Phys. J. [**C50**]{} (2007) 617.
D. M. Li, B. Ma and Y. H. Liu, Eur. Phys. J. [**C51**]{} (2007) 359.
X. H. Zhong and Q. Zhao, Phys. Rev. [**D78**]{} (2008) 014029.
B. Chen, D. X. Wang and A. Zhang, Phys. Rev. [**D80**]{} (2009) 071502.
X. H. Zhong and Q. Zhao, Phys. Rev. [**D81**]{} (2010) 014031.
D. M. Li and B. Ma, Phys. Rev. [**D81**]{} (2010) 014021.
A. M. Badalian and B. L. G. Bakker, Phys. Rev. [**D84**]{} (2011) 034006.
J. Vijande, A. Valcarce and F. Fernandez, Phys. Rev. [**D79**]{} (2009) 037501.
F. K. Guo and U. G. Meissner, Phys. Rev. [**D84**]{} (2011) 014013.
R. Aaij et al, Phys. Rev. Lett. [**113**]{} (2014) 162001.
R. Aaij et al, Phys. Rev. [**D90**]{} (2014) 072003.
R. Aaij et al, JHEP [**1210**]{} (2012) 151.
Q. T. Song, D. Y. Chen, X. Liu and T. Matsuki, Eur. Phys. J. [**C75**]{} (2015) 30. J. Segovia, D. R. Entem and F. Fernandez, Phys. Rev. [**D91**]{} (2015) 094020; B. Chen, X. Liu and A. Zhang, Phys. Rev. [**D92**]{} (2015) 034005.
S. Godfrey and I. T. Jardine, Phys. Rev. [**D89**]{} (2014) 074023.
Z. G. Wang, Eur. Phys. J. C75 (2015) 25.
D. Ebert, R. N. Faustov and V. O. Galkin, Eur. Phys. J. [**C66**]{} (2010) 197; M. Di Pierro and E. Eichten, Phys. Rev. [**D64**]{} (2001) 114004; S. Godfrey and K. Moats, Phys. Rev. [**D93**]{} (2016) 034035.
M. A. Shifman, A. I. Vainshtein and V. I. Zakharov, Nucl. Phys. [**B147**]{} (1979) 385; Nucl. Phys. [**B147**]{} (1979) 448.
L. J. Reinders, H. Rubinstein and S. Yazaki, Phys. Rept. [**127**]{} (1985) 1.
Z. G. Wang, Eur. Phys. J. [**C75**]{} (2015) 427.
S. Narison, Nucl. Part. Phys. Proc. [**270-272**]{} (2016) 143.
H. Sundu and K. Azizi, Eur. Phys. J. [**A48**]{} (2012) 81; K. Azizi, H. Sundu, J. Y. Sungu and N. Yinelek, Phys. Rev. [**D88**]{} (2013) 036005; K. Azizi, Y. Sarac and H. Sundu, Eur. Phys. J. [**C74**]{} (2014) 3106. H. A. Alhendi, T. M. Aliev and M. Savci, JHEP [**1604**]{} (2016) 050.
Z. G. Wang and Z. Y. Di, Eur. Phys. J. [**A50**]{} (2014) 143; Z. G. Wang, Eur. Phys. J. [**C74**]{} (2014) 3123; Z. Y. Li, Z. G. Wang and G. L. Yu, Mod. Phys. Lett. [**A31**]{} (2016) 1650036.
D. Zhou, E. L. Cui, H. X. Chen, L. S. Geng, X. Liu and S. L. Zhu, Phys. Rev. [**D90**]{} (2014) 114035.
J. J. Zhu and M. L. Yan, hep-ph/9903349; and references therein.
P. Colangelo and A. Khodjamirian, hep-ph/0010175.
S. Narison, Camb. Monogr. Part. Phys. Nucl. Phys. Cosmol. [**17**]{} (2002) 1.
S. Narison, Phys. Lett. [**B706**]{} (2012) 412; S. Narison, Phys. Lett. [**B707**]{} (2012) 259.
C. A. Dominguez, L. A. Hernandez and K. Schilcher, JHEP [**1507**]{} (2015) 110.
K. A. Olive et al, Chin. Phys. [**C38**]{} (2014) 090001.
[^1]: E-mail,[email protected].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an exquisite, 30-min cadence *Kepler* (K2) light curve of the Type Ia supernova (SN Ia) 2018oh (ASASSN-18bt), starting weeks before explosion, covering the moment of explosion and the subsequent rise, and continuing past peak brightness. These data are supplemented by multi-color Pan-STARRS1 and CTIO 4-m DECam observations obtained within hours of explosion. The K2 light curve has an unusual two-component shape, where the flux rises with a steep linear gradient for the first few days, followed by a quadratic rise as seen for typical SNe Ia. This “flux excess” relative to canonical SN Ia behavior is confirmed in our $i$-band light curve, and furthermore, SN2018oh is especially blue during the early epochs. The flux excess peaks 2.14$\pm0.04$ days after explosion, has a FWHM of 3.12$\pm0.04$ days, a blackbody temperature of $T=17,500^{+11,500}_{-9,000}$K, a peak luminosity of $4.3\pm0.2\times10^{37}\,{\rm erg\,s^{-1}}$, and a total integrated energy of $1.27\pm0.01\times10^{43}\,{\rm erg}$. We compare SN2018oh to several models that may provide additional heating at early times, including collision with a companion and a shallow concentration of radioactive nickel. While all of these models generally reproduce the early K2 light curve shape, we slightly favor a companion interaction, at a distance of $\sim$$2\times10^{12}\,{\rm cm}$ based on our early color measurements, although the exact distance depends on the uncertain viewing angle. Additional confirmation of a companion interaction in future modeling and observations of SN2018oh would provide strong support for a single-degenerate progenitor system.'
author:
- 'G. Dimitriadis'
- 'R. J. Foley'
- 'A. Rest'
- 'D. Kasen'
- 'A. L. Piro'
- 'A. Polin'
- 'D. O. Jones'
- 'A. Villar'
- 'G. Narayan'
- 'D. A. Coulter'
- 'C. D. Kilpatrick'
- 'Y. -C. Pan'
- 'C. Rojas-Bravo'
- 'O. D. Fox'
- 'S. W. Jha'
- 'P. E. Nugent'
- 'A. G. Riess'
- 'D. Scolnic'
- 'M. R. Drout'
- 'G. Barentsen'
- 'J. Dotson'
- 'M. Gully-Santiago'
- 'C. Hedges'
- 'A. M. Cody'
- 'T. Barclay'
- 'S. Howell'
- 'P. Garnavich'
- 'B. E. Tucker'
- 'E. Shaya'
- 'R. Mushotzky'
- 'R. P. Olling'
- 'S. Margheim'
- 'A. Zenteno'
- 'J. Coughlin'
- 'J. E. Van Cleve'
- 'J. Vinícius de Miranda Cardoso'
- 'K. A. Larson'
- 'K. M. McCalmont-Everton'
- 'C. A. Peterson'
- 'S. E. Ross'
- 'L. H. Reedy'
- 'D. Osborne'
- 'C. McGinn'
- 'L. Kohnert'
- 'L. Migliorini'
- 'A. Wheaton'
- 'B. Spencer'
- 'C. Labonde'
- 'G. Castillo'
- 'G. Beerman'
- 'K. Steward'
- 'M. Hanley'
- 'R. Larsen'
- 'R. Gangopadhyay'
- 'R. Kloetzel'
- 'T. Weschler'
- 'V. Nystrom'
- 'J. Moffatt'
- 'M. Redick'
- 'K. Griest'
- 'M. Packard'
- 'M. Muszynski'
- 'J. Kampmeier'
- 'R. Bjella'
- 'S. Flynn'
- 'B. Elsaesser'
- 'K. C. Chambers'
- 'H. A. Flewelling'
- 'M. E. Huber'
- 'E. A. Magnier'
- 'C. Z. Waters'
- 'A. S. B. Schultz'
- 'J. Bulger'
- 'T. B. Lowe'
- 'M. Willman'
- 'S. J. Smartt'
- 'K. W. Smith'
- 'S. Points'
- 'G. M. Strampelli'
- 'J. Brimacombe'
- 'P. Chen'
- 'J. A. Muñoz'
- 'R. L. Mutel'
- 'J. Shields'
- 'P. J. Vallely'
- 'S. Villanueva Jr.'
- 'W. Li'
- 'X. Wang'
- 'J. Zhang'
- 'H. Lin'
- 'J. Mo'
- 'X. Zhao'
- 'H. Sai'
- 'X. Zhang'
- 'K. Zhang'
- 'T. Zhang'
- 'L. Wang'
- 'J. Zhang'
- 'E. Baron'
- 'J. M. DerKacy'
- 'L. Li'
- 'Z. Chen'
- 'D. Xiang'
- 'L. Rui'
- 'L. Wang'
- 'F. Huang'
- 'X. Li'
- 'G. Hosseinzadeh'
- 'D. A. Howell'
- 'I. Arcavi'
- 'D. Hiramatsu'
- 'J. Burke'
- 'S. Valenti'
- 'J. L. Tonry'
- 'L. Denneau'
- 'A. N. Heinze'
- 'H. Weiland'
- 'B. Stalder'
- 'J. Vinkó'
- 'K. Sárneczky'
- 'A. Pál'
- 'A. Bódi'
- 'Zs. Bognár'
- 'B. Csák'
- 'B. Cseh'
- 'G. Csörnyei'
- 'O. Hanyecz'
- 'B. Ignácz'
- 'Cs. Kalup'
- 'R. Könyves-Tóth'
- 'L. Kriskovics'
- 'A. Ordasi'
- 'I. Rajmon'
- 'A. Sódor'
- 'R. Szabó'
- 'R. Szakáts'
- 'G. Zsidi'
- 'S. C. Williams'
- 'J. Nordin'
- 'R. Cartier'
- 'C. Frohmaier'
- 'L. Galbany'
- 'C. P. Gutiérrez'
- 'I. Hook'
- 'C. Inserra'
- 'M. Smith'
- 'D. J. Sand'
- 'J. E. Andrews'
- 'N. Smith'
- 'C. Bilinski'
bibliography:
- 'SN2018oh.bib'
title: 'K2 Observations of SN2018oh Reveal a Two-Component Rising Light Curve for a Type Ia Supernova'
---
,
Introduction {#sec:intro}
============
Through a combination of theoretical arguments and strong observational constraints, it has long been understood that Type Ia supernovae (SNe Ia) are the result of a thermonuclear explosion of a carbon/oxygen white dwarf (WD) [e.g., @Hoyle60; @Colgate69; @Woosley86; @Bloom12] in a binary system. Nevertheless, despite SNe Ia being used to discover the accelerating expansion of the Universe two decades ago [@Riess98AJ; @Perlmutter99] and continuing to be a powerful dark energy probe [e.g., @Scolnic18; @Jones18], we still do not know the nature of their progenitor systems, whether they come from multiple progenitor scenarios, and if so, in what proportion.
Roughly speaking, possible SN progenitor systems can be separated into two main classes (or channels): the single-degenerate (SD) channel, where the primary WD accretes material from a non-degenerate companion triggering a thermonuclear runaway near the Chandrasekhar mass ($\mathrm{M_{Ch}}$) [e.g., @Whelan73ApJ], and the double-degenerate (DD) channel where the SN is triggered by the merger of two WDs [e.g.; @Iben84ApJS]. Numerical modeling of explosions [e.g., @Hillebrandt00ARAA; @Hillebrandt13FrPhy] combined with radiative hydrodynamic modeling [e.g., @Kasen09Natur; @Woosley11ApJ; @Pakmor12ApJ; @Sim12MNRAS; @Sim13MNRAS] indicate that the basic properties of the SN Ia population can be reproduced by either scenario. Therefore, we must turn to observations to further constrain the possible progenitor systems of SNe Ia.
Thus far, the observations have been similarly limited, and are often inconsistent with a single scenario. No SN Ia progenitor system has yet been directly observed in the handful of SNe Ia with reasonably deep pre-explosion images [@Li11Natur; @Goobar14ApJ; @Kelly14ApJ] (although one has for a peculiar WD SN, the SN Iax 2012Z; @McCully14Natur). However, the images were not sufficiently deep to exclude all SD progenitor systems. On a different approach, a search for the surviving non-degenerate companion star at the central regions of SN remnants (SNRs), believed to have a SN Ia origin, also excludes WD + sub-giant or red giant (RG) systems [@Kerzendorf12ApJ...759....7K; @Schaefer12Natur.481..164S; @Kerzendorf14ApJ...782...27K]. Nonetheless, several indirect observations can reveal the nature of the companion with some scenarios having specific and distinct observational predictions.
Observing SNe Ia as close to explosion as possible can provide unique information for distinguishing between progenitor scenarios. For example, the earliest moments can be dominated by the shock cooling of the exploding WD [@Piro10ApJ], which was used in the case of SN 2011fe to constrain the explosion to be coming from a degenerate star [@Bloom12]. For SD progenitor systems containing a Roche-lobe-filling companion, signs of the SN ejecta interacting with the non-degenerate companion star are expected for some lines of sight [@Kasen10ApJ]. This produces strong X-ray and UV/optical emission that will surpass the radioactive luminosity of the SN at these early epochs. The amount of observed flux depends on the viewing angle and the distance between the exploding WD and the companion — which given the Roche-lobe overflow assumption, provides constraints on the companion star radius. Specifically, evolved red giants are expected to produce more flux than smaller stars.
While early excess emission is a robust prediction for the Roche-lobe-filling SD scenarios, other physical phenomena can possibly also produce early heating. In particular, $^{56}$Ni near the surface (i.e., with a mass fraction exceeding that of lower layers) should also introduce flux in excess of the canonical “expanding fireball” model [@Piro13ApJ]. This scenario, which can occur for both progenitor channels, can conceal or resemble interaction models. A specific explosion model that can produce such a configuration is the double-detonation explosion of a sub-$\mathrm{M_{ch}}$ WD, where the detonation of a surface helium layer will produce significant surface $^{56}$Ni [@Noebauer17MNRAS]. @Piro16ApJ also found that shallow $^{56}$Ni distributions and/or interaction with circumstellar material (CSM) expelled during a DD merger can modulate the early light curve shape.
Observations early enough and with sufficient cadence to search for these early light curve features are still relatively rare. Nearby events, such as SNe2011fe [@Nugent11; @Bloom12], 2014J [@Goobar14ApJ] and ASASSN-14lp [@Shappee2016ApJ] provide upper limits to the potential separation distance of the companion, ruling out stars more evolved than a RG, while for SN2009ig, a small blue excess is attributed to the unusual color evolution of the particular event [@Foley12ApJ]. SN Ia sample studies [@Hayden10ApJ2; @Bianco11ApJ; @Tucker11ApSS; @Ganeshalingam11MNRAS; @Brown12ApJ] exclude RGs for a fraction of the events, allowing less-evolved stars as companions.
On the other hand, two SNe, SNe2012cg and 2017cbv, have early light curves that are somewhat consistent with interaction with a companion star. @Marion16ApJ finds that interaction with a 6 $\mathrm{M_{\odot}}$ main sequence (MS) star can explain the early UV/optical excess of SN2012cg. For SN2017cbv, @Hosseinzadeh17ApJ, analyzing many possibilities, favors an interaction with a subgiant companion, within the uncertainties in the modeling. Both of these interpretations have been questioned by @Shappee18ApJ and @Sand18arXiv, respectively, where the authors disfavor a non-degenerate companion, based, in part, on non-detections of stripped hydrogen or helium (within some limits) in nebular spectra. For SN2012fr, @Contreras18ApJ...859...24C find an initial slow, nearly linear rise in luminosity, followed by a faster rising phase, and attribute it to a moderate amount of $^{56}$Ni mixing in the ejecta, while for the almost-linear rise of iPTF16abc, @Miller18ApJ argue in favor of either ejecta-mixing or pulsational delayed-detonation models. In the case of the SN2002es-like [@Ganeshalingam12ApJ] iPTF14atg [@Cao16ApJ], data are compatible with a companion at a separation of 70 $\mathrm{R_{\odot}}$, with @Kromer16MNRAS, using numerical simulations of explosion models, finding difficulties reconciling its peculiar spectral evolution with a non-degenerate companion. Finally, @Jiang17Natur show an early red flux excess for MUSSES1604D and, comparing different scenarios, favor a double detonation.
To search for such companion-shock emission, one would ideally conduct a survey with continuous, high-cadence observations to precisely constrain the explosion time and either track or constrain any possible early-time excess flux. The *Kepler* telescope [@Haas10ApJ] with its wide field of view and 30-minute cadence, continuous observations is particularly well suited to discover SNe within moments of explosion and continuously monitor those SNe [for recent transient studies with *Kepler* see @Garnavich16ApJ; @Rest18NatAs]. *Kepler* has the ability to observe thousands of galaxies at a time and therefore has the potential to discover $\sim$10 SNe a month if the observations are devoted to relatively nearby galaxies. During the main *Kepler* mission, @Olling15Natur discovered 3 likely SNe Ia with extraordinary coverage from the moments of explosion through the rise and decline of the SNe. Despite these extraordinary observations, there was no significant detection of interaction.
Nevertheless, it has been demonstrated that *Kepler* has unique capabilities for precise monitoring of the earliest phases after a SN explosion. To this end, the successor of the *Kepler* mission, K2, has dedicated a substantial number of targets during Campaign 16, lasting from December 7, 2017 to February 25, 2018, to the K2 Supernova Cosmology Experiment (K2 SCE). Significant advantages of K2 SCE over previous *Kepler* mission SN studies include (1) monitoring about 50 times as many galaxies (although for a shorter time) and (2) being “forward-facing,” where the field is pointed roughly away from the Sun, allowing for simultaneous ground-based observations of all transients discovered in the Campaign 16 field.
In this paper, we present observations of SN2018oh, a normal SN Ia whose host galaxy was monitored by the K2 SCE starting before explosion, continuing to first light, and through peak brightness. In addition to its impressive K2 light curve, SN2018oh SN was extensively monitored by many ground-based facilities. In this paper, we focus on the first week after explosion. In the data, we robustly identify, with unprecedented photometric coverage, an excess early time rise component.
This work is part of a series of papers analyzing SN2018oh: Shappee et al. (2018b) provide an alternative analysis of the K2 light curve data of the SN and Li et al. (2018) present the photometric and spectroscopic properties of the SN near and after- peak brightness.
This paper is organized as follows: In Section \[sec:disc\_early\_phot\], we present the discovery of SN2018oh and the early-time data we use in this paper, including the reduction and calibration steps. In Section \[sec:analysis\], we describe the analysis of the early-time lightcurve, while in Section \[sec:models\] we propose various physical models that explain it. Finally, in Section \[sec:conclusions\], we discuss our findings in the context of the progenitor problem of SNe Ia, and outline our conclusions.
Throughout this paper, Modified Julian Days (MJDs) are reported as observed days while phases are reported in rest-frame, unless where noted. We adopt the AB magnitude system, unless where noted, and a Hubble constant of $H_0 = 73$ km s$^{-1}$ Mpc$^{-1}$.
Discovery and Early-time Observations {#sec:disc_early_phot}
=====================================
SN2018oh was discovered by the All Sky Automated Survey for SuperNovae [ASAS-SN, @Shappee14ApJ] in images obtained on 2018 Feb 4.41 UT (all times presented are UT) [with discovery name ASASSN-18bt; @Brown18ATel] (Shappee et al. 2018b), at $V = 15.2$ mag, with the last non-detections at 2018 Jan 27.13. The supernova is located at $\alpha=09^{\rm{h}}06^{\rm{m}}39^{\rm{s}}.592$, $\delta=+19^{\rm{o}}20\arcmin17\arcsec.47$ (J2000.0) [@Cornect18ATel], $7.8\arcsec$ North and $2.0\arcsec$ East of the center of UGC 4780, a Sdm starforming galaxy, with a redshift of $z = 0.010981$ and a distance of 49.4 Mpc. The Milky Way reddening toward SN2018oh is $E(B-V) = 0.0368$ mag [@Schlafly11ApJ]. The transient was classified on 2018 Feb 5 as a relatively young ($-8$ to $-6$ days relative to peak brightness), normal SN Ia [@Leadbeater18TNS; @Zhang18ATel].
UGC 4780 was included as a Campaign 16 target through ‘The K2 ExtraGalactic Survey (KEGS) for Transients’ (PI Rest) and the ‘Multi-Observatory Monitoring of K2 Supernovae’ (PI Foley) programs as part of the K2 SCE (internal Kepler ID 228682548). After the end of Campaign 16, the data were transferred to MAST, from which we retrieved the UGC 4780 data. We produced a provisional light curve with the “quick look” routine <span style="font-variant:small-caps;">kadenza</span>[^1] [@2018ascl.soft03005B] by summing counts in a 5$\times$5 pixel aperture centered at the peak of each 30-minute image. The background was determined by estimating the median flux of the outermost pixels. Due to its unique observing strategy which requires regular thruster use to maintain pointing, K2 data suffer from a ‘sawtooth pattern’ and long-term sensitivity trends, partly due to temperature changes as the sun angle and the zodiacal light levels change during a Campaign. In order to correct for these effects, third-order polynomials were fit in both spatial dimensions to remove the ‘sawtooth.’ To account for the long-term trends, we performed a principal-component analysis that represents the common simultaneous trends seen in the light curves of all the (assumed non-varying) galaxies observed on the same chip. Through an iterative procedure, the optimal number of PCA vectors was determined to be only one. We then determined and removed the long-term trend for SN2018oh. Finally, the noise was estimated by computing the root-mean-squared variation just before the explosion and then scaling this by the square root of the galaxy flux plus the SN flux in the aperture. For a more detailed discussion on the K2 reduction steps, see @Shaya15AJ.
During Campaign 16, we actively observed the K2 field with both the Pan-STARRS1 telescope [PS1; @Chambers16arXiv; @Magnier16arXiv; @Waters16arXiv] and the CTIO 4-m Mayall telescope with DECam [@Honscheid08arXiv; @Flaugher15AJ]. The main goal was to discover and obtain multi-color light curves of transients in K2-observed galaxies. This program was successful where we discovered 9 and 8 such transients in C16 with PS1 [@2018ATel11218] and DECam [@2018ATel11345; @2018ATel11344], respectively. Unfortunately immediately after the explosion of SN2018oh, poor weather prevented observations for 7 nights. During that gap, we did not have scheduled DECam nights either.
All PS1 and DECam images were reduced using the <span style="font-variant:small-caps;">photpipe</span> imaging and photometry package [@Rest05ApJ; @Rest14ApJ], which performs standard reduction processes, including bias subtraction, cross-talk corrections, flat-fielding, astrometric calibration and image resampling. Instrumental PSF magnitudes are calculated by using <span style="font-variant:small-caps;">DoPhot</span> [@Schechter93PASP] on the difference images, and the final calibration is performed with PS1 standard-star fields. This photometric procedure is well-tested and has been applied in many transient studies [e.g., @Rest14ApJ; @Kilpatrick18MNRAS]. We present PS1 and DECam images from before and immediately after explosion, as well as images near peak brightness, in Figure \[fig:SN2018oh\_images\].
{width="95.00000%"}
Spectroscopic and photometric follow-up observations of the SN were performed immediately after its discovery, and a complete presentation of the SN properties is presented in Li et al. (2018).
Analysis {#sec:analysis}
========
In this Section, we present the early photometric observations of SN2018oh, both from ground-based facilities and K2. We then present a basic analysis of the early evolution of the SN, based on analytical models.
Ground-based Photometry {#sec:ground_based_phot}
-----------------------
SN2018oh is detected in PS1 $g$ and $i$ images, on UT 2018-01-26.56 and 26.57 (for $g$ and $i$ respectively), 8.9 days before the ASAS-SN discovery image, with AB magnitudes of $g_{\rm P1}=20.72 \pm 0.18$ and $i_{\rm P1}=20.94 \pm 0.25$, while the last non-detections were at UT 2018-01-23.38 and UT 2018-01-22.55. Moreover, from DECam $i$-band images taken one day later, SN2018oh was $i = 19.04 \pm 0.01$ and $18.96 \pm 0.01$ mag on 2018 Jan 27.25 and 2018 Jan 27.29, respectively, revealing a rise in the $i$ band of $\sim$1 mag in one day. A collection of ground-based images, showing pre-explosion, first detection and close-to-peak luminosity, in $g$ and $i$ bands, is presented in Figure \[fig:SN2018oh\_images\], and reported in Table \[tab:ground\_phot\].
After correcting for the Milky Way extinction using the @Fitzpatrick99PASP law with $R_{V} = 3.1$, we fit the $uBV\!griz$ photometry (Li et al. 2018) with the most recent version of the <span style="font-variant:small-caps;">SALT2</span> light curve fitter (SALT2.4; ) through the SNANA framework [@Kessler09PASP]. We measure a SALT2 shape parameter of $x_{1} = 0.879 \pm 0.012$ and a color parameter of $c = -0.09 \pm 0.01$. We determine that SN2018oh peaked at $B_{\rm peak} = 14.185 \pm 0.010$ mag on MJD $58163.339 \pm 0.016$.
To infer the distance, we use the distance estimator from , and references therein:
$$\mu = m_B - M_B + \alpha \times x_1 - \beta \times c + \Delta_M,
\label{eqn:salt2}$$
where $m_B$, $x_1$ and $c$ are given above. We use the values of the nuisance parameters $\alpha = 0.141$, $\beta = 3.099$ and $M_B = -19.17$ given by . Regarding the host galaxy mass step $\Delta_M$ [@Kelly10; @Lampeitl10; @Sullivan10], we use SDSS $g$ and $i$ magnitudes with the relation of @Taylor11 [their Equation 8] to derive the host galaxy mass of UGC 4780. We find the mass to be 8.81 dex, comfortably on the low-mass side of the step function, and we correct with $\Delta_M=-0.06$ mag. The final distance modulus, assuming $H_0 = 73$ km s$^{-1}$ Mpc$^{-1}$, is estimated to be $\mu = 33.61 \pm 0.05$ mag, corresponding to a distance of $52.7 \pm 1.2$ Mpc. As UGC 4780 is not in the Hubble flow and has no independent distance measurement, the distance using the SN itself is the most accurate and precise distance, and we use this distance for the remainder of the analysis.
The near-peak and post-peak photometric data of SN2018oh show that the SN is a normal SN Ia, while the only spectral peculiarity is the (relatively) long-lived carbon absorption features, seen even to about 3 weeks after the maximum light and discussed in Li et al. (2018). From all available data, we conclude that SN2018oh is a normal SN Ia.
*Kepler* Light Curve {#sec:kepler_phot}
--------------------
After the reduction of the SN2018oh [*Kepler*]{}/K2 light curve as described in Section \[sec:disc\_early\_phot\], which only provides a relative-flux light curve, we determine the true K2 flux as follows. We use the $uBV\!griz$ photometry (Li et al. 2018), which has been calibrated to the PS1 system to determine the SN2018oh flux as a function of time and wavelength. We then use the “max model" of the `SNooPy`[^2] package [@SNooPy] to determine the spectral-energy distribution (SED) of the SN as a function of time. This model first fits for the peak flux in each photometric band by scaling template light curves [@Burns11AJ] to the data, with the model $K$-corrections calculated by warping the @Hsiao07 SN Ia spectral series to match the observed colors. This approach accounts for assumptions about host-reddening and the distance to the SN by modeling the multi-band photometry before determining the K2 magnitudes. The best-fit parameters were used to normalize the mangled spectral series to the observed photometry and to generate a synthetic SED. As the @Burns11AJ method mangles the spectral series to match the SED in each observer-frame passband, there is a choice of which passband’s normalized SED to use as a model for the K2 band. We use the $V$ band as its effective wavelength is closest to that of the K2 band. After integrating over the K2 passband, recovering the ‘synthetic’ K2 light curve, we solve for the absolute zeropoint, using the background-subtracted *K2* flux light curve, interpolated over a range of $\pm$3 d around the time of [*B*]{}-band maximum light, where the supernova color evolves slowly. We estimate $\mathrm{ZP_{K2}}=25.324 \pm 0.004~\text{(statistical)}$. We find a $\pm 0.011 ~\text{(systematic)}$ mag, systematic uncertainty arising from the choice of which (observer-frame) passband normalized-SED is used to model the synthetic K2 light curve.
We present the SN2018oh K2 light curve in Figure \[fig:sn2018oh\_kepler\_lc\], normalized to the peak of the light curve, which we estimate by fitting a polynomial to the data from MJD 58160.0 to 58165.0. We find that the peak in the K2 band occurs at $\mathrm{MJD_{max}^{\mathrm{K2}}} = 58162.58$, $\sim$0.12 days prior to $B$-band maximum, with $\mathrm{K2_{max}=14.401\pm0.001}$. A portion of the light curve is presented in Table \[tab:k2\_phot\], while the complete dataset is available in the electronic edition.
{width="95.00000%"}
Basic Analysis of the K2 Light Curve {#sec:k2_phot}
------------------------------------
Assuming that the photospheric temperature of a SN Ia does not change significantly in the first few days after explosion, the luminosity of the Rayleigh-Jeans tail of the blackbody radiation will increase with time as $L\propto t^{2}$ [@Arnett82ApJ; @Riess99AJ], as the size of the photosphere increases. However, the K2 light curve of SN2018oh shows a prominent “two-component rise”: an initial flux excess, from $\sim$18 to 13 days before peak brightness, which eventually subsides and the usual “expanding fireball” rise dominates starting about 13 days before peak brightness.
We determine the onset of the K2 light curve as follows: For a given sliding time-window, we calculate the weighted-mean of the flux and we compare it with the flux of the time-window prior to it, marking as a detection when Flux$_{i} > 3\sigma_{i-1}$. By an iterative procedure, using decreasing time-window widths, we record the detection times, and we estimate their mean and standard deviation. We calculate $t_{ \rm det}^{\rm K2}=-17.99 \pm 0.04$ days from maximum light (at MJD$_{\rm det}^{\rm K2} = 58144.39$), shown as the black vertical line in Figure \[fig:sn2018oh\_kepler\_lc\]. We note that the first PS1 detections were 0.18 days (4.32 hours) after the K2 first detection, which we estimate to be 2018 Jan 26.04.
In order to determine the properties of the power-law rise (i.e., excluding the first-component rise), we attempt to estimate a time range by iteratively fitting, using <span style="font-variant:small-caps;">idl</span>’s MPFIT function, a $(t-t_{0})^2$ power law to the data in a window from a variable (shifting by steps of 0.02 days) start time beginning 20 days before peak brightness until the flux reaches 40% of the peak flux, as has been done with other [*Kepler*]{} SN Ia studies [@Olling15Natur]. Our best fit (reduced $\chi^{2} = 1.09$) is for a time window from 11.54 to 10.32 days before peak brightness, which we mark with a vertical two-headed arrow in Figure \[fig:sn2018oh\_kepler\_lc\]. From this fit, we estimate a time of first light, $t_{0}=-18.14 \pm 0.02$ days — $\sim$0.15 days ($\sim$3.6 hours) before our first K2 detection. We display the residual to the fit in the bottom panel of the inset in Figure \[fig:sn2018oh\_kepler\_lc\]. We find that $\sim$2 days after $t_{0}$, the flux excess is $\sim$3 times as luminous as the power-law rise, and represents $\sim$65% of the total flux at that time.
As it has been shown in previous rise time studies [@Riess99AJ; @Hayden10ApJ1; @Ganeshalingam11MNRAS; @GonzalezGaitan12; @Firth15MNRAS], the index of the power law can significantly vary from 2 for a particular SN. To account for this possibility, we repeat the previous procedure and, using , a Python-based application of an affine invariant Markov chain Monte Carlo (MCMC) with an ensemble sampler [@Foreman-Mackey13], we fit a $(t-t_{0})^\alpha$ power-law (thus, additionally fitting for the power-law index). Doing so, we find a similar best-fit region as before, with the new best-fit parameters $t_{0} = -17.86^{+0.24}_{-0.25}$ days before peak brightness, with $\alpha = 1.92 \pm 0.07$.
![(Top) SN2018oh K2 light curve, normalized to peak flux, with respect to the first K2 detection. Our full fit is shown as a solid black line, while the decomposition of the fit is shown as a red line, for the SN power law flux, and a blue line, for the first rise component. The red downward arrow denotes the time of first light $t_{0}$, estimated from the fit. (Middle) The early flux excess, plotted as the data minus the fitted power-law model. The result of the first component fit is shown as a blue line. (Bottom) The residual of the fit (data minus full model fit).[]{data-label="fig:sn2018oh_kepler_lc_fit"}](Figure_3_SN2018oh_k2_light_curve_fit.pdf){width="45.00000%"}
In order to quantify the rise of the excess flux component, and motivated by its shape, we consider a simple analytical model that consists of (1) a power law $L \propto (t-t_{0})^{\alpha}$ and (2) a skewed Gaussian to account for the early flux excess. We fit the light curve from 20 days before peak brightness through the time when the flux reaches 40% of the peak flux, with both a fixed power law index of 2, and with the index allowed to float. By fixing the index to $\alpha = 2$, we estimate $t_{0} = -18.00^{+0.03}_{-0.02}$. When simultaneously fitting for the power-law index, we find $t_{0} = -18.19 \pm 0.05$ and $a = 2.08 \pm 0.02$. The later fit is shown in Figure \[fig:sn2018oh\_kepler\_lc\_fit\]. These results are generally consistent with the canonical expanding fireball model, and the initial assumption that $L \propto (t-t_{0})^{2}$ seems reasonable given the data.
From the multi-component fit, we also estimate that the early excess flux peaked with a luminosity of $(4.3 \pm 0.2)\times 10^{37}$ erg s$^{-1}$ at $t_{\rm peak}^{c1} = -16.05 \pm 0.04$ days, approximately 2.2 days after $t_{0}$, and had a FWHM of 3.12 days. The total emitted energy above the power-law rise is $(1.27 \pm 0.01)\times 10^{43}$ erg.
Comparison to Other SNe {#sec:phot_comp}
-----------------------
Firstly, we compare the K2 light curve of SN2018oh with the *Kepler* SNe presented in @Olling15Natur, focusing on the discovery and rise epochs (Figure \[fig:sn2018oh\_comparison\_ksne\]). As mentioned in @Olling15Natur, KSN 2011b (blue full circles) and KSN 2012a (red full circles) occurred in red and passive galaxies at redshifts $\sim0.05$ and $\sim0.09$ (we exclude the 3rd *Kepler* SN of @Olling15Natur, KSN 2011c, due to the lower quality of data). Moreover, these SNe are fast decliners (thus, have lower absolute luminosities) while SN2018oh is a normal SN Ia. For this reason, we ‘stretch-correct’ [@Perlmutter97ApJ] the *Kepler* SN light curves to the K2 light curve of SN2018oh by determining the stretch factor that, when applied, best matches the light curves (see the insets in Figure \[fig:sn2018oh\_comparison\_ksne\]). The ‘stretch-corrected’ light curves are shown as open blue (KSN 2011b) and red (KSN 2012a) circles.
![Comparison of the K2 SN 2018oh light curve (black circles), normalized to peak, with respect to peak brightness, with KSN 2011b (upper panel) and KSN 2012a (lower panel). The *Kepler* light curves of KSN 2011b and KSN 2012a have been ‘stretch-corrected’ to match the SN 2018oh light curve. We show the original 12-h time binned data from @Olling15Natur in full circles, and the ‘stretch-corrected’ ones with open circles. In the insets, we show a zoom of the light curves at peak.[]{data-label="fig:sn2018oh_comparison_ksne"}](Figure_4_SN2018oh_k2_light_curve_comparisons.pdf){width="45.00000%"}
As it can be seen, the applied stretch correction successfully matches the SNe at the epochs around peak brightness. However, SN2018oh clearly deviates for the first few days after explosion, when the flux excess is observed. We estimate that, at the time of the peak of the flux excess, $t_{\rm peak}^{c1} = -16.05$ days, SN2018oh is 51% and 32% more luminous than the stretch-corrected KSN 2011b and KSN 2012a, respectively.
{width="95.00000%"}
Next, we compare the early SN2018oh light curves with those of two other SNe with very early data: the well-studied, extremely young SN2011fe [@Nugent11], a normal Type Ia supernova that shows no flux excess at the extremely early times, and SN2017cbv [@Hosseinzadeh17ApJ], a SN Ia with a prominent blue early flux excess. For this comparison, we need comparable filters. SN2017cbv has extensive early-time photometry in the desired $g$ and $i$ bands [@Hosseinzadeh17ApJ]. SN2011fe also has an early $g$-band light curve [@Nugent11], but lacks an early $i$-band light curve. In place of filtered photometry, we use the @Pereira13 spectrophotometric time series, from which we synthesize an $i$-band light curve.
In Figure \[fig:sn2018oh\_kepler\_early\_flux\], we simultaneously display the early SN2018oh K2 (gray), $g$ (blue), and $i$ (red) light curves. For comparison, we also show similar data for SNe2011fe and 2017cbv. In the inset, we show the first 6 days after explosion, where the SN rose $>$2 orders of magnitude in flux. We also display the full two-component fit to the SN2018oh light curve and just the power-law component.
While SN2011fe clearly lacks the flux excess of SN2018oh and rises close to $t^{2}$, SN2017cbv has a flux excess at early times and an early photometric behavior comparable to SN2018oh. At later times ($t > -10$ days), all three SNe evolve similarly. Notably, from that point on, SN2018oh looks identical to the “normal” SN2011fe.
Finally, we investigate the color evolution of SN2018oh, and in particular the $g-i$, $g-{\rm K2}$ and ${\rm K2}-i$ colors. We compare the SN2018oh colors to the synthetic colors of SNe2011fe and 2017cbv, calculated as described above (note that we also estimate the synthetic K2 magnitude). Additionally, we compute the color evolution of the @Hsiao07 template spectra. The results are shown in Figure \[fig:sn2018oh\_early\_colors\].
{width="95.00000%"}
While SNe2018oh and 2017cbv generally have similar colors for the epochs examined here, generally having bluer colors than both SN2011fe and the @Hsiao07 template, there is a distinct difference at the earliest epochs, when the prominent flux excess is observed. We note the difference in the ${\rm K2}-i$ color, at the onset of the excess-flux component where SN2018oh is bluer than all comparison SNe ($\sim$0.1 and 0.08 mag from SN2011fe and SN2017cbv, respectively). Unfortunately, we only have a single $i$ observation during this phase. Nonetheless, this single observation is critical in separating SN2018oh from SN2017cbv.
SED of the Excess Flux {#sec:sed_excess}
----------------------
Finally, we investigate the SED of the excess flux observed from 18 to 13 days before peak brightness. While we have no spectra during this phase, we have filtered photometry that can constrain the SED. In addition to the K2 photometry, we will use the PS1 $g$ and DECam $i$ observations at $t = -17.8$ and $-17.1$ days, respectively which were obtained while the flux excess was still rising (see inset of Figure \[fig:sn2018oh\_kepler\_lc\]).
We will focus on the crucial DECam $i$ observation at $t=-17.1$ days from the K2 maximum, which coincides with the rise of the flux excess (see onset of Figure \[fig:sn2018oh\_kepler\_lc\]). While there is no spectrum of SN2018oh taken at that epoch, motivated by the similar peak/post-peak photometric and spectroscopic behavior with SN2011fe, we use the Lick/KAST spectrum, presented initially by @Nugent11, taken $\sim$1.5 days after the SN2011fe explosion ($-16.33$ rest-frame days from $B-$band maximum light). We attempt to spectroscopically match this spectrum (for which no flux excess is observed) with the photometric colors of SN2018oh at the epoch in question. As mentioned above, we unfortunately don’t have $g$ observations at this epoch, therefore we assume no color evolution in $g-i$ for the first days (see left panel of Figure \[fig:sn2018oh\_early\_colors\]). We note that this assumption is somehow arbitrary: Our photometry at $-17.8$ days has large uncertainties, while the $g-i$ color is redder, compared to SN2017cbv. Nevertheless, after redshifting the spectrum to the redshift of SN2018oh, we scale it to match the SN component of the K2 flux at $-17.1$ days, as determined in Section \[sec:k2\_phot\], Figure \[fig:sn2018oh\_kepler\_lc\_fit\]. We then perform a MCMC fit of this spectrum and a blackbody spectrum, where the resulting spectrum reproduces the observed photometry, with the results shown in Figure \[fig:sn2018oh\_spec\_match\_11fe\].
![The +1.5 day from explosion SN2011fe spectrum (solid black line), redshifted and scaled to match the $t^{2.08}$ SN component of the SN2018oh K2 photometry at +1.09 days after $t_{0}$, obtained from the fit shown in Figure \[fig:sn2018oh\_kepler\_lc\_fit\]. The $g_{\rm P1}$ (blue), K2 (orange) and $i$ (red) photometry are shown as full diamonds, at each filter’s effective wavelength. The solid green line is the best-fit blackbody spectrum that reproduces the observed photometric colors at +1.09 days after $t_{0}$ (full circles), with the open circles representing the synthetic fluxes of the resulting spectrum (dashed line). In the upper panel, we show the response functions of the $g_{\rm P1}$ (blue), K2 (orange) and DECam $i$ (red) filters.[]{data-label="fig:sn2018oh_spec_match_11fe"}](Figure_7_SN2018oh_SED.pdf){width="45.00000%"}
Our best fit includes a blackbody with $T=17,500^{+11,500}_{-9,000}$ K. The main source of the large uncertainty comes from the constant color evolution assumption and the corresponding large photometric uncertainty at this extremely early epoch. However, the resulting fitted temperature is high, providing an indication of a hot blackbody component, on top of the normal SN spectrum.
Models {#sec:models}
======
We next consider three scenarios that may provide additional heating at early times to lead to the two-component rise seen in the SN2018oh light curve: the interaction between the SN and a nearby companion star, a double-detonation model with $^{56}$Ni near the surface of the star, and an additional model in which we tune that amount of surface $^{56}$Ni in an attempt to best match SN2018oh.
{width="95.00000%"}
Interaction with a companion star {#sec:interaction_kasen}
---------------------------------
One potential explanation for the early-time excess is shock-interaction between the supernova ejecta and a non-degenerate binary companion [@Kasen10ApJ]. The collision is characterized by prompt X-ray emission, followed by an optical/UV excess lasting about one week after explosion. Although the excess peaks in the UV, a measurable signature is observable in the *Kepler* bandpass if the system is configured in a favorable viewing angle [@Olling15Natur].
To test this scenario, we use a numerical model for the early light curve following the methods outlined in @Piro16ApJ (using the Chandrasekhar progenitor models from [@Martinez16]) that roughly matches the rise of SN2018oh, and then combine this with the analytic interaction model of @Kasen10ApJ. The interaction emission is mainly controlled through two parameters, the orbital separation $a$ and the characteristic ejecta velocity $v$. Since $a$ can vary by many orders of magnitude and $v$ is relatively well constrained to be $v\approx 10^9\,{\rm cm\,s^{-1}}$, this makes the interaction a powerful diagnostic for measuring $a$. In addition, there are viewing angle effects, but this is somewhat degenerate with the other parameters. Thus we focus on the the case when one is observing directly into the shocked region (when the companion is roughly between the explosion and the observer) and take the measured $a$ as a lower limit to the orbital separation.
From this procedure, we find that a collision with a companion at $a=2\times10^{12}\,{\rm cm}$ provides a reasonable match to the early rise. We plot this as a solid blue line in the right panel of Figure \[fig:sn2018oh\_kepler\_models\], and also show the constituent parts of the interaction (dashed blue line) and the SN itself (dotted blue line). An important assumption of this model is that the companion is overfilling its Roche lobe and therefore we can approximate its radius as [@Eggleton83] $$R = \frac{0.49q^{2/3}}{0.6q^{2/3}+\ln(1+q^{1/3})}a,$$ where $q$ is the ratio of the companion and WD’s mass. For a range of companion masses from $M\approx1-6\,$M$_\odot$, this results in $R\approx10-15\,$R$_\odot$, respectively. This is generally too large for a main sequence star, and thus we conclude that the companion must be a subgiant if interaction is the correct explanation for the early excess.
Double-detonation sub-Chandrasekhar explosion {#sec:abis_doubledetonations}
---------------------------------------------
Another possible mechanism for creating an early-time flux excess is the double-detonation scenario for exploding a sub-Chandrasekhar mass C/O WD with an accreted shell of helium on its surface. In this scenario, the helium shell detonates, producing on the surface some abundance of radioactive elements such as $^{56}$Ni and $^{48}$Cr, and sending a shockwave into the WD that then ignites the C/O core [@1994ApJ...423..371W]. The result produces observables generally consistent with a SN Ia, however, the amount of Fe-group elements synthesized during the He-shell detonation must be small to resemble SNe Ia near peak brightness. The photons produced by the radioactive decay of material on the surface quickly diffuse out of the ejecta, creating a flux excess relative to a typical SN Ia in the first few days after explosion [@Noebauer17MNRAS].
We test this scenario as a candidate for SN2018oh by exploring a hydrodynamic and radiative transfer numerical survey of double-detonations of sub-Chandrasekhar mass white dwarfs, the results of which are presented in @Polin18arXiv. The parameter space of the survey spans from $0.7-1.2\,\mathrm{M_{\odot}}$ WDs with helium shells from 0.01 to 0.08 $\mathrm{M_{\odot}}$ and a range of mixing mass from 0.05 to 0.3 $\mathrm{M_{\odot}}$. The best-fitting model, based on a reduced $\chi^{2}$ measurement, is a 0.98 M$_\odot$ WD with 0.05 M$_\odot$ of Helium on its surface, with the ejecta smoothed over a mixing length of 0.25 M$_\odot$. This model produces a total of 0.448 $\mathrm{M_{\odot}}$ of $^{56}$Ni, 3.65$\times10^{-3}$ $\mathrm{M_{\odot}}$ of $^{48}$Cr and 1.8$\times10^{-2}$ $\mathrm{M_{\odot}}$ of $^{52}$Fe. From these elements, the amount of each that is synthesized in the helium shell (i. e., in the outer layers of ejecta) is 1.22$\times10^{-2}$, 3.19$\times10^{-2}$ and 6.11$\times10^{-2}$ $\mathrm{M_{\odot}}$, corresponding to 2.7, 87.4 and 33.9%, respectively. The K2 synthetic light curve is shown in the left panel of Figure \[fig:sn2018oh\_kepler\_models\]. The approximate magnitudes of both the early-time excess and peak are reproduced, as is the duration of the excess and rise time to peak brightness.
A general off-center nickel distribution {#sec:offcenter_nickel}
----------------------------------------
The previous model is specifically applicable to the DD scenario, but it is possible in principle that other scenarios may mix $^{56}$Ni to the outermost layers. To explore this possibility more generally, we consider models in which we take a normal SN Ia explosion and place by hand some amount of $^{56}$Ni near the surface. As with the supernova model for the interaction scenario, we use the methods outlined in @Piro16ApJ with the progenitors generated in the work of @Martinez16. Using this we place the $^{56}$Ni in two distinct regions, a centrally concentrated region that provides the main rise and a shallow region above a mass coordinate of $1.3\,$M$_\odot$. The shallow abundance is varied to find the best fit with the K2 photometry, including smoothing with a $0.05\,$M$_\odot$ boxcar which prevents numerical issues from sharp compositional gradients.
Our best-fit model under this scenario has $0.03\,$M$_{\odot}$ of $^{56}$Ni near the surface of the WD as shown by the orange solid line in the middle panel of Figure \[fig:sn2018oh\_kepler\_models\]. The model reproduces the general evolution of the light curve, encapsulating the initial flux excess. We therefore provisionally consider this a viable model. Whether or not such a model can reproduce the full photometric and spectroscopic evolution of SN2018oh is less clear. Iron-peak elements at shallow depths can provide extensive line-blanketing that alters the colors and spectra of the SN at peak luminosity, potentially making it difficult for SN2018oh to be a spectroscopically normal SN Ia. Below we consider in further detail whether such a model can even reproduce the early color evolution of SN2018oh.
Detailed Model Comparisons {#sec:model_comparisons}
--------------------------
Having found both SD and DD models that can reproduce the K2 light curve, we must examine additional data that differentiate these scenarios. The earliest detections by PS1 and DECam are particularly powerful for this purpose.
In addition to detecting the flux excess in the K2 band, we also detect an excess in the $i$ band (see Figure \[fig:sn2018oh\_kepler\_early\_flux\]). Examining the ${\rm K2} - i$ color during the flux excess, we find that SN2018oh is bluer than SN2011fe by almost 0.2 mag. Moreover, SN2018oh is also similarly bluer than SN2017cbv at that epoch. This means that one day after explosion, SN2018oh is not only distinct from the normal SN2011fe but also from SN2017cbv, which also had an early-time flux excess [@Hosseinzadeh17ApJ]. After the onset of the canonical SN rise (rightwards of the grey-shaded region in the panels of Figure \[fig:sn2018oh\_early\_colors\]), the three SNe evolve in a similar manner (apart from the usual color dispersion seen in Type Ia supernovae).
All models examined above are able to reproduce the flux excess at early times of SN2018oh, but with two main different physical origins. These models predict very different SEDs and in particular different colors. Specifically the companion-interaction model is expected to be bluer than the surface-Ni model.
{width="95.00000%"}
We display the expected $g - i$, $g - {\rm K2}$ and ${\rm K2} - i$ colors for each model in Figure \[fig:sn2018oh\_model\_colors\]. The double-detonation model is particularly red, and it is unable to explain the blue colors of SN2018oh. Moreover, sub-Chandrasekhar double-detonations also leave spectral signatures such as a significant absorption features in the peak spectra, that are not seen in SN2018oh (Li et al. 2018). The generic off-center $^{56}$Ni model also has a color that is redder than SN2018oh by $\sim$0.1 mag. On the other hand, the companion-interaction model with a companion at $\approx2\times10^{12}$ cm generally matches both the early rise and the color evolution of SN2018oh.
Based on the color evolution of SN2018oh, we slightly favor the companion-interaction model over other models that can also reproduce the early flux excess. A direct prediction of this model is the presence of hydrogen and/or helium-rich material stripped from the companion star at the nebular phases [@Pan12ApJ; @Liu13ApJ]. To this end, detections of H or He features at late-time spectra of SN2018oh is crucial to confirm this model.
We note that, while SN2018oh has an exquisite K2 light curve, we lack the detailed color information to conclusively decide between models, particularly at bluer wavelengths. Additionally there are no spectra of SN2018oh during the flux excess, which would have provided key information for understanding how spectral features affect the color evolution. We do not see an abrupt flux excess on the first rise component, but rather a linear rise similar to other SNe Ia with sufficiently early, high-cadence photometry that also have two-component rising light curves [e.g., SN2012fr, which arguably shows no signs of interaction; @Contreras18ApJ...859...24C] argues against the interaction model. A linear rise may result from a particular interaction model (i.e., specific viewing angle, radius, and/or separation), however an exploration of how different parameter combinations affect the detailed rise behavior is beyond the scope of this paper.
However, another interpretation of the flux excess is possible. Given the growing sample of SNe Ia that show two-component early light curves, with different slopes and durations, the distinct early light curve evolution of SN2018oh, compared to the total SN Ia population, could not be due to some external heating source, but rather a reflection of varying SNe Ia properties, such as the density profile of the ejecta, different composition/metallicity of the progenitor star, asymmetries during the explosion etch [see @Stritzinger18arXiv180707576S for a relevant discussion]. Modern transient surveys, such as ASAS-SN and ZTF, and future powerful surveys, such as LSST, will discover very young SNe Ia, and with rapid follow-up, the early-light curve SN sample will increase, allowing us to investigate this possibility.
Discussion and Conclusions {#sec:conclusions}
==========================
In this paper, we analyze the early photometric observations of SN2018oh, a normal SN Ia, that occurred within the *Kepler* Campaign 16 field. The SN was observed with unprecedented cadence by the K2 SCE with complementary early Pan-STARRS1 and DECam imaging. The combination of an extremely early detection and unprecedented continuous coverage with *Kepler* make SN2018oh a spectacular reference object for early SN Ia studies, providing invaluable insights on the explosion physics and the progenitor system.
In the SN2018oh K2 and $i$-band light curves, we detect a distinct flux component in the first few days after explosion relative to other well-observed SNe Ia (e.g., SN2011fe) and the $t^{2}$ luminosity rise seen later in the evolution of SN2018oh. This flux excess lasts approximately 5 days, after which SN2018oh appears to evolve in a fashion consistent with typical SNe Ia.
Our work provides new insights on the early time evolution of SNe Ia, for which we find the following:
1. The early K2 light curve shows a distinct two-component rise evolution. Initially, the SN rises quickly, with a steep linear gradient, in flux. This flux subsides after about 5 days, when a $L \propto t^{2}$ rise dominates the SN evolution.
2. Using the $t^{2}$ component of the K2 light curve, we constrain $t_{0}$ to be $-18.19 \pm 0.09$ days before K2 peak brightness. This time is consistent with the onset of the flux excess, indicating that both components began at the same time.
3. Assuming that the $t^{2}$ component exists while the other component is bright, we find that the early flux excess peaks 2.14 days after $t_{0}$, has a FWHM of 3.12 days, a blackbody temperature of 17,500 K, a peak luminosity of 4.3$\times10^{37}$ erg/s, and a total emitted energy of 1.3$\times10^{43}$ erg.
4. We observed SN2018oh with Pan-STARRS1 and DECam only 4.1 and 20.6 hours after the first K2 detection, respectively, providing some of the earliest colors of a SN Ia ever. The flux excess is confirmed in our ground-based $i$ light curve.
5. The SN2018oh early photometric evolution is relatively similar to SN2017cbv, another SN Ia with a prominent two-component rising light curve. However, SN2018oh shows bluer ${\rm K2} - i$ colors than SN2017cbv. This is especially true during the epoch of the initial flux excess. Around peak brightness, SN2018oh is similar to both SNe2011fe and 2017cbv.
6. The early flux excess can potentially be explained by additional heating at the epoch in question. We investigate three possible sources:
1. The interaction with a non-degenerate companion at $a=2\times10^{12}\,{\rm cm}$, with a $M\approx 1-6\,$M$_\odot$ Roche-lobe-filling star.
2. The presence of a $0.05\,\mathrm{M_{\odot}}$ Helium shell on the surface of $0.98\,\mathrm{M_{\odot}}$ C/O WD, and a subsequent sub-Chandrasekhar mass explosion.
3. An off-center $^{56}$Ni distribution of 0.03 $\mathrm{M_{\odot}}$.
All of these models can, generally, reproduce the early shape of the K2 light curve. We slightly favor the interaction scenario, due to the blue colors at the epoch of the flux excess. However, another interpretation of the flux excess considers an intrinsic variation of early time behavior, due to varying SNe Ia properties, with no external heating source required.
While a SD origin for (at least a sizable fraction of) SNe Ia has been previously proposed [@Sternberg11Sci; @Foley12ApJ2; @Maguire13MNRAS], its validity has been questioned. Most SNe that have some observational evidence for the presence of a non-degenerate companion are either peculiar [e.g., SNe Ia-CSM; @Dilday12Sci; @Silverman13ApJS; @Fox15MNRAS.447..772F] or have contradictory observations [@Shappee18ApJ; @Sand18arXiv]. The general progenitor picture that has emerged over the last decade is that most SNe Ia have a DD origin .
Excluding the early-time flux excess, SN2018oh shows no signs of photometric and spectroscopic peculiarities. Therefore, SN2018oh represents a normal SN Ia with a potential SD origin, challenging the idea that all normal SNe Ia have DD progenitors. Additional SNe Ia observed at high cadence during the first few days after explosion are needed to determine the fraction of SNe Ia with SD progenitors. At the same time, these observations will grow the early light curve SN Ia sample, and investigate correlations of the light curve evolution with various SNe Ia properties.
The K2 SCE has finished and the data are currently collected and analysed. With some luck, we will soon have additional K2-observed SNe Ia with data similar in quality to that of SN2018oh.
We will continue to monitor SN2018oh. Late-time observations, after the SN has become optically thin, will be a direct test of our proposed models.
We thank the anonymous referee for helpful comments that improved the clarity and presentation of this paper.
This paper includes data collected by the K2 mission. Funding for the K2 mission is provided by the NASA Science Mission directorate.
KEGS is supported in part by NASA K2 cycle 4 and 5 grants NNX17AI64G and 80NSSC18K0302, respectively. AR and his groups is supported in part by HST grants GO-12577 and HST AR-12851.
Pan-STARRS (PS1) is supported in part by the National Aeronautics and Space Administration under Grants NNX12AT65G and NNX14AM74G. The Pan-STARRS1 Surveys (PS1) and the PS1 public science archive have been made possible through contributions by the Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project Office, the Max-Planck Society and its participating institutes, the Max Planck Institute for Astronomy, Heidelberg and the Max Planck Institute for Extraterrestrial Physics, Garching, The Johns Hopkins University, Durham University, the University of Edinburgh, the Queen’s University Belfast, the Harvard-Smithsonian Center for Astrophysics, the Las Cumbres Observatory Global Telescope Network Incorporated, the National Central University of Taiwan, the Space Telescope Science Institute, the National Aeronautics and Space Administration under Grant No. NNX08AR22G issued through the Planetary Science Division of the NASA Science Mission Directorate, the National Science Foundation Grant No. AST–1238877, the University of Maryland, Eotvos Lorand University (ELTE), the Los Alamos National Laboratory, and the Gordon and Betty Moore Foundation.
This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the DOE and NSF (USA), MISE (Spain), STFC (UK), HEFCE (UK), NCSA (UIUC), KICP (U. Chicago), CCAPP (Ohio State), MIFPA (Texas A&M), CNPQ, FAPERJ, FINEP (Brazil), MINECO (Spain), DFG (Germany) and the collaborating institutions in the Dark Energy Survey, which are Argonne Lab, UC Santa Cruz, University of Cambridge, CIEMAT-Madrid, University of Chicago, University College London, DES-Brazil Consortium, University of Edinburgh, ETH Zürich, Fermilab, University of Illinois, ICE (IEEC-CSIC), IFAE Barcelona, Lawrence Berkeley Lab, LMU München and the associated Excellence Cluster Universe, University of Michigan, NOAO, University of Nottingham, Ohio State University, OzDES Membership Consortium, University of Pennsylvania, University of Portsmouth, SLAC National Lab, Stanford University, University of Sussex, and Texas A&M University.
Based on observations at Cerro Tololo Inter-American Observatory, National Optical Astronomy Observatory (NOAO 2017B-0279; PI: A Rest, NOAO 2017B-0285; PI: A Rest), which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
The UCSC group is supported in part by NASA grants NNG17PX03C and 80NSSC18K0303, NSF grants AST-1518052 and AST-1815935, the Gordon & Betty Moore Foundation, the Heising-Simons Foundation, and by fellowships from the Alfred P.Sloan Foundation and the David and Lucile Packard Foundation to R.J.F.
We thank Chris Burns for providing crucial updates to `SNooPy`, relevant to the analysis of this work.
SJS acknowledges funding from STFC Grants ST/P000312/1 and ST/N002520/1.
This work makes use of observations from Las Cumbres Observatory. DAH, CM, and GH are supported by the US National Science Foundation grant 1313484. Support for IA was provided by NASA through the Einstein Fellowship Program, grant PF6-170148.
JV and his group at Konkoly Observatory is supported by the project “Transient Astrophysical Objects" GINOP 2.3.2-15-2016-00033 of the National Research, Development and Innovation Office (NKFIH), Hungary, funded by the European Union.
This project has been supported by the Lendület Program of the Hungarian Academy of Sciences, projects No. LP2018-7/2018 and LP2012-31, and the NKFIH grant K-115709.
ZsB acknowledges the support provided from the National Research, Development and Innovation Fund of Hungary, financed under the PD 17 funding scheme, project no. PD123910.
Research by DJS is supported by NSF grants AST-1821967, 1821987, 1813466 and 1813708. NS and JEA received support from NSF grant AST-1515559.
We acknowledge the support of the staff of the Lijiang 2.4m and Xinglong 2.16m telescope. Funding for the LJT has been provided by Chinese Academy of Sciences and the People’s Government of Yunnan Province. The LJT is jointly operated and administrated by Yunnan Observatories and Center for Astronomical Mega-Science, CAS. This work is supported by the National Natural Science Foundation of China (NSFC grants 11178003, 11325313, and 11633002), and the National Program on Key Research and Development Project (grant no. 2016YFA0400803). JJZ is supported by the National Science Foundation of China (NSFC, grants 11403096, 11773067), the Youth Innovation Promotion Association of the CAS, the Western Light Youth Project, and the Key Research Program of the CAS (Grant NO. KJZD-EW-M06). TMZ is supported by the NSFC (grants 11203034). This work was also partially Supported by the Open Project Program of the Key Laboratory of Optical Astronomy, National Astronomical Observatories, Chinese Academy of Sciences.
EB and JD acknowledge partial support from NASA grant NNX16AB5G.
CPG acknowledges support from EU/FP7-ERC grant No. \[615929\].
Parts of this research were supported by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO 3D), through project number CE170100013.
Support for this work was provided by NASA through Hubble Fellowship grant \#HST-HF2-51357.001-A, awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Incorporated, under NASA contract NAS5-26555, as well as NASA [*K2*]{} Cycle 4 Grant NNX17AE92G.
Photometry Tables
=================
[lcDDCC]{}\[t!\] 2018-01-26.24 & 58144.24 & -18.26 & -0.15 & i & 23.155$^{c}$\
2018-01-26.29 & 58144.29 & -18.21 & -0.10 & g & 23.322$^{c}$\
2018-01-26.56 & 58144.56 & -17.94 & 0.17 & g\_[P1]{} & 20.852 0.224\
2018-01-26.57 & 58144.57 & -17.93 & 0.18 & i\_[P1]{} & 21.022 0.268\
2018-01-27.25 & 58145.25 & -17.26 & 0.85 & i & 19.039 0.013\
2018-01-27.30 & 58145.30 & -17.21 & 0.90 & i & 18.957 0.014\
2018-02-03.33 & 58152.33 & -10.26 & 7.85 & i\_[P1]{} & 15.670 0.004\
2018-02-04.33 & 58153.33 & -9.27 & 8.84 & i\_[P1]{} & 15.445 0.004\
2018-02-04.49 & 58153.49 & -9.11 & 9.00 & i\_[P1]{} & 15.388 0.003\
2018-02-05.40 & 58154.40 & -8.21 & 9.90 & i\_[P1]{} & 15.264 0.010\
2018-02-08.37 & 58157.37 & -5.27 & 12.84 & g\_[P1]{} & 14.483 0.002\
2018-02-09.47 & 58158.47 & -4.18 & 13.93 & g\_[P1]{} & 14.359 0.002\
2018-02-09.48 & 58158.48 & -4.17 & 13.94 & i\_[P1]{} & 14.926 0.003\
2018-02-10.48 & 58159.48 & -3.18 & 14.93 & g\_[P1]{} & 14.307 0.002\
2018-02-11.35 & 58160.35 & -2.32 & 15.79 & g\_[P1]{} & 14.285 0.002\
2018-02-11.35 & 58160.35 & -2.32 & 15.79 & i\_[P1]{} & 14.961 0.003\
2018-02-11.49 & 58160.49 & -2.19 & 15.92 & g\_[P1]{} & 14.261 0.002\
2018-02-12.33 & 58161.33 & -1.36 & 16.76 & g\_[P1]{} & 14.241 0.002\
2018-02-13.56 & 58162.56 & -0.14 & 17.97 & g\_[P1]{} & 14.253 0.002\
2018-02-13.57 & 58162.57 & -0.13 & 17.98 & i\_[P1]{} & 15.002 0.003\
2018-02-14.53 & 58163.53 & 0.82 & 18.93 & g\_[P1]{} & 14.225 0.002\
2018-02-20.34 & 58169.34 & 6.57 & 24.68 & g\_[P1]{} & 14.418 0.002\
2018-02-21.49 & 58170.49 & 7.71 & 25.82 & g\_[P1]{} & 14.472 0.002\
2018-03-07.28 & 58184.28 & 21.35 & 39.46 & g\_[P1]{} & 15.465 0.005\
2018-03-07.29 & 58184.29 & 21.36 & 39.47 & i\_[P1]{} & 15.769 0.006\
2018-03-08.25 & 58185.25 & 22.30 & 40.42 & g\_[P1]{} & 15.543 0.004\
2018-03-08.41 & 58185.41 & 22.46 & 40.57 & g\_[P1]{} & 15.514 0.003\
2018-03-18.29 & 58195.29 & 32.24 & 50.35 & i\_[P1]{} & 15.635 0.006\
2018-03-18.29 & 58195.29 & 32.24 & 50.35 & g\_[P1]{} & 16.306 0.009\
[lcDDCC]{}\[t!\] 2017-12-07.99 & 58094.99 & -66.85 & -48.86 & 20.743$^{c}$\
2017-12-08.01 & 58095.01 & -66.83 & -48.84 & 20.744$^{c}$\
2017-12-08.05 & 58095.05 & -66.79 & -48.80 & 20.745$^{c}$\
2017-12-08.07 & 58095.07 & -66.77 & -48.78 & 20.745$^{c}$\
2017-12-08.09 & 58095.09 & -66.75 & -48.76 & 20.745$^{c}$\
2017-12-08.11 & 58095.11 & -66.73 & -48.74 & 20.745$^{c}$\
[^1]: <https://github.com/KeplerGO/kadenza>
[^2]: <https://users.obs.carnegiescience.edu/cburns/SNooPyDocs/html/>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
JLQCD collaboration: $^{,a,b,}$[^1], S. Aoki$^c$, H. Fukaya$^d$, T. Kaneko$^{a,b}$, H. Matsufuru$^a$, J. Noaki$^a$, T. Onogi$^e$, N. Yamada$^{a,b}$\
High Energy Accelerator Research Organization (KEK), Tsukuba 305-0801, Japan\
School of High Energy Accelerator Science, the Graduate University for Advanced Studies (Sokendai), Tsukuba 305-0801, Japan\
Graduate School of Pure and Applied Sciences, University of Tsukuba, Tsukuba 305-8571, Japan\
Theoretical Physics Laboratory, RIKEN, Wako 351-0198, Japan\
Yukawa Institute for Theoretical Physics, Kyoto University, Kyoto 606-8502, Japan
title: Lattice simulation of 2+1 flavors of overlap light quarks
---
Dynamical overlap fermion
=========================
The JLQCD collaboration is carrying out a large scale lattice QCD simulation using the overlap fermion formulation for sea quarks. (An overview of the project has been given at this conference by Matsufuru [@Matsufuru_lat07].) The first phase of the project was a two-flavor QCD simulation on a $16^3\times 32$ lattice at a lattice spacing $a\simeq$ 0.11–0.12 fm. The HMC simulations have been completed accumulating about 10,000 molecular dynamics trajectories for six values of sea quark mass ranging $m_s/6$–$m_s$. Preliminary reports of this project were already presented at Lattice 2006 [@Kaneko:2006pa; @Matsufuru:2006xr; @Hashimoto:2006rb; @Yamada:2006fr]; at this conference we have presented physics results for pion masses and decay constants [@Noaki:2007es], pion form factor [@Kaneko:2007nf], kaon $B$ parameter [@Yamada:2007nh], and topological susceptibility [@Chiu_lat07]. We have also performed simulations in the $\epsilon$-regime by reducing the sea quark mass down to 3 MeV. This lattice has been used for the analysis of low-lying eigenvalues of the overlap-Dirac operator [@Fukaya:2006xp; @Fukaya:2007fb; @Fukaya:2007yv] and for a calculation of meson correlators in the $\epsilon$-regime [@Fukaya_lat07]. The second phase of the project is to include strange quark as dynamical degrees of freedom: a 2+1-flavor QCD simulation with the overlap fermion. We aim at producing dynamical lattices of size $16^3\times 48$ at around the same lattice spacing $a\simeq$ 0.11–0.12 fm.
We use the Neuberger’s overlap-Dirac operator [@Neuberger:1997fp; @Neuberger:1998wv] $$\label{eq:ov}
D(m) = \left(m_0+\frac{m}{2}\right)
+ \left(m_0-\frac{m}{2}\right)\gamma_5
\mathrm{sgn}\left[H_W(-m_0)\right].$$ The choice for the kernel operator is the standard Wilson fermion with a large negative mass $m_0=1.6$. For the gauge sector we use the Iwasaki gauge action together with extra Wilson fermions and ghosts producing a factor $$\frac{\det\left[H_W(-m_0)^2\right]}{\det\left[H_W(-m_0)^2+\mu^2\right]}$$ in the partition function such that the near-zero modes of $H_W(-m_0)$ is naturally suppressed [@Fukaya:2006vs]. This term is essential for the feasibility of dynamical overlap fermion simulation, since it substantially reduces the cost of the approximation of the sign function in (\[eq:ov\]). Although it prevents us from changing the topological charge during the molecular dynamics evolutions, its systematic effect can be understood as a finite size effect and can be estimated (and even corrected) once the topological susceptibility is known [@Aoki:2007ka]. The topological susceptibility is in fact calculable on the lattice with a fixed topology as demonstrated in [@Chiu_lat07; @Aoki:2007pw].
Algorithms
==========
For the calculation of the sign function in (\[eq:ov\]) we use the rational approximation $$\label{eq:rational}
\mathrm{sgn}\left[H_W\right] = H_W
\left(p_0 + \sum_{l=1}^N \frac{p_l}{H_W^2+q_l}\right)$$ with the Zolotarev’s optimal coefficients $p_l$ and $q_l$. This is applied after projecting out a few low-lying modes of $H_W$. Typically, accuracy of order $10^{-(7-8)}$ is achieved with $N=10$. The multiple inversions for $(H_W^2+q_l)^{-1}$ can be done at once using the multi-shift conjugate gradient (CG).
The inversion of $D(m)$ is the most time-consuming part in the HMC simulation. In the two-flavor runs, we mainly used the nested CG with relaxed residual for the inner CG [@Cundy:2004pza]. In the 2+1-flavor runs, we use the five-dimensional solver as explained in the following.
By the Schur decomposition the overlap solver can be written in the form (for $N=2$ for example) [@Narayanan:2000qx; @Borici:2004pn; @Edwards:2005an] $$\label{eq:5D}
\left(
\begin{array}[c]{cccc|c}
H_W & -\sqrt{q_2} & & & 0 \\
-\sqrt{q_2} & -H_W & & & \sqrt{p_2} \\
& & H_W & -\sqrt{q_1} & 0 \\
& & -\sqrt{q_1} & -H_W & \sqrt{p_1} \\
\hline
0 & \sqrt{p_2} & 0 & \sqrt{p_1} & R\gamma_5+p_0 H_W
\end{array}
\right)
\left(
\begin{array}[c]{c}
\phi_{2+} \\
\phi_{2-} \\
\phi_{1+} \\
\phi_{1-} \\
\hline \psi_4
\end{array}
\right)
=
\left(
\begin{array}[c]{c}
0 \\
0 \\
0 \\
0 \\
\hline \chi_4
\end{array}
\right),$$ where $R=(1+m)/(1-m)$. By solving this equation we obtain a solution for $D(m)\phi_4=\chi_4$ with $D(m)$ approximated by the rational function. The matrix in (\[eq:5D\]) can be viewed as a five-dimensional (5D) matrix. An advantage of solving (\[eq:5D\]) is that one can use the even-odd preconditioning. Namely, rather than solving the 5D matrix $M$, we may solve a reduced matrix $(1-M_{ee}^{-1}M_{eo}M_{oo}^{-1}M_{oe})\psi_e=\chi_e'$, where even/odd blocks of $M$ are denoted by $M_{eo}$, $M_{ee}$, [*etc*]{}. The inversion $M_{ee}^{-1}$ (or $M_{oo}^{-1}$) can be easily calculated by the forward (or backward) substitution involving the 5D direction.
The low-mode projection can be implemented together with the 5D solver. The lower-right corner is replaced by $$R(1-P_H)\gamma_5(1-P_H)+p_0H_W+
\left(m_0+\frac{m}{2}\right)\sum_{j=1}^{N_{ev}}
\mathrm{sgn}(\lambda_j)v_j\otimes v_j^\dagger,$$ where $P_H$ is a projector onto the subspace orthogonal to the $N_{ev}$ low-lying modes: $P_H=1-\sum_{j=1}^{N_{ev}}\mathrm{sgn}(\lambda_j)v_j\otimes v_j^\dagger$. Then, the inversion of $M_{ee(oo)}$ becomes non-trivial, but can be calculated cheaply because the rank of the matrix is only $2(N_{ev}+1)$; the subspace is spanned by $x_e$, $\gamma_5x_e$, $v_{je}$, $\gamma_5 v_{je}$ ($j=1, .., N_{ev}$).
We compare the performance of the 5D solver with the relaxed CG in 4D. The elapsed time to solve the 5D equation is plotted in Figure \[fig:solver\] as a function of quark mass $m$. The lattice size is $16^3\times 48$ and the measurement is done on a half-rack (512 nodes) of the BlueGene/L supercomputer (2.7 TFlops peak performance). Data for $N=10$ is connected by lines for both 4D and 5D solvers. Evidently, the 5D solver is faster by about a factor of 3–4. Increasing the number of degree of the rational approximation requires more computational cost for both 4D and 5D. For the 5D case, the cost is naively expected to be proportional to $N$, but the actual measurement shows slower increase, which indicates some overhead due to the construction of low-mode projector [*etc.*]{}
![ Comparison of solver performance. Data for $N=10$ is connected by lines: 4D (red squares) and 5D (black circles). []{data-label="fig:solver"}](convergence.eps){width="9cm"}
Odd number of flavors
=====================
Introduction of the pseudo-fermions for dynamical quark flavors is the starting point of HMC. For the two-flavor case, this is straightforward by writing $\det D^2$ as $\int[d\phi][d\phi^\dagger]\exp[-|H^{-1}\phi|^2]$, where $H\equiv\gamma_5 D$. The same trick applied for one flavor introduces $D^{-1/2}$ in the pseudo-fermion action, which requires a method to calculate the inverse square-root of the Dirac operator. (For such algorithms, see [@Aoki:2001pt], for example.) For the overlap-Dirac operator this problem can be avoided as follows [@Bode:1999dd; @DeGrand:2006ws]. Thanks to the exact chiral symmetry of the overlap fermion, $H^2\equiv (\gamma_5 D)^2$ commutes with $\gamma_5$, and therefore can be decomposed into positive and negative chirality subspaces: $$\label{eq:H2}
H^2 = P_+ H^2 P_+ + P_- H^2 P_- \equiv Q_+ + Q_-,$$ where $P_\pm = (1\pm\gamma_5)/2$. Then, its determinant is factorized, $\det H^2 = \det Q_+ \cdot \det Q_-$. Since $Q_+$ and $Q_-$ share the eigenvalues except for those of zero-modes, $\det H^2 = (\det Q_+)^2 = (\det Q_-)^2$ up to the zero-mode contribution, which is a trivial factor for the topology fixed simulations. In order to simulate one flavor, one can just pick one chiral sector of $H^2$.
Thus, we introduce a pseudo-fermion field for the one-flavor piece as $S_{PF1} = \sum_x \phi_\sigma^\dagger(x) Q_\sigma^{-1} \phi_\sigma(x)$, where $\sigma$ can either be $+$ or $-$ representing the chiral sector. At the beginning of each HMC trajectory, we refresh $\phi_\sigma(x)$ from a gaussian distribution $\xi(x)$ as $\phi_\sigma(x)=Q_\sigma^{-1/2}\xi(x)$. This step requires a calculation of the square-root of $Q_\sigma$, which is done using the rational approximation. Calculation of the molecular-dynamics force is straightforward: one can simply project onto the chiral sector $\sigma$ in the calculation of the force from $H^2$.
Runs
====
The 2+1-flavor runs are done at $\beta=2.30$, which is the same value as our main two-flavor runs. The unit trajectory length $\tau$ is set to 1.0, twice longer than the two-flavor runs. Our choice of the sea quark mass parameters are summarized in Table \[tab:para\]. The up and down quark mass $m_{ud}$ ranges from $m_s$ down to $\sim m_s/6$ as in our two-flavor runs. For the strange quark mass we take two values aiming at interpolating to the physical strange quark mass.
$m_{ud}$ $m_s$ = 0.080 $m_s$ = 0.100
---------- --------------- ---------------
0.015 $\surd$ $\surd$
0.025 $\surd$ $\surd$
0.035 $\surd$ $\surd$
0.050 $\surd$ $\surd$
0.080 $\surd$
0.100 $\surd$
: Sea quark mass parameters[]{data-label="tab:para"}
At the time of the lattice conference, the runs proceeded to 500–1,000 HMC trajectories depending on the mass parameter. One trajectory takes about 1–2 hours on one rack (1,024 nodes) of BlueGene/L (5.7 TFlops peak performance). The acceptance rate is kept around 80–90% for each run.
Figure \[fig:ninv\] shows the number of the (two) 5D CG iteration in the calculation of the HMC Hamiltonian. As expected the calculation for the two-flavor piece is dominating the calculation.
![ Molecular dynamics time evolution of the number of CG iterations in the calculation of the HMC Hamiltonian. Data at $m_{ud}=0.025$ (left) and 0.050 (right) with $m_s=0.100$. In the plot “ov1” denotes up and down quarks, while “ov2” corresponds to strange. “PF2” stands for the inversion with the original sea quark mass, and “PF1” is for the preconditioner, whose mass is chosen to be 0.4 for $m_q\ge 0.035$ or 0.2 for $m_q\le 0.025$. []{data-label="fig:ninv"}](conv_0.025_0.100.eps "fig:"){width="7cm"} ![ Molecular dynamics time evolution of the number of CG iterations in the calculation of the HMC Hamiltonian. Data at $m_{ud}=0.025$ (left) and 0.050 (right) with $m_s=0.100$. In the plot “ov1” denotes up and down quarks, while “ov2” corresponds to strange. “PF2” stands for the inversion with the original sea quark mass, and “PF1” is for the preconditioner, whose mass is chosen to be 0.4 for $m_q\ge 0.035$ or 0.2 for $m_q\le 0.025$. []{data-label="fig:ninv"}](conv_0.050_0.100.eps "fig:"){width="7cm"}
Measurements of physical quantities are done at every 5 trajectories, so far only for the $m_s=0.100$ lattices. In order to use in the low-mode preconditioning and low-mode averaging, we are calculating 80 pairs of low-lying eigenmodes of the overlap-Dirac operator. The lattice spacing as determined through the Sommer scale $r_0$ (= 0.49 fm) is plotted in Figure \[fig:a\] for both 2- and 2+1-flavor lattices. At the same $\beta$ value (= 2.30) the lattice spacing decreases as more dynamical flavors are included.
![ Lattice spacing as a function of sea quark mass. At $\beta$ = 2.30, two-flavor data (black circles) are plotted together with a line of chiral extrapolation. 2+1-flavor data are plotted for both $m_s$ = 0.100 (blue squares) and 0.080 (blue triangles). A quenched result at the same $\beta$ value is shown by a red band. []{data-label="fig:a"}](a_vs_mq.eps){width="9cm"}
Preliminary results for pion and kaon mass squared and decay constant are shown in Figure \[fig:results\]. Data at $m_s=0.100$ are plotted as a function of sea quark mass. Although the statistics is still low ($<$ 1,000 trajectories for each sea quark mass), reasonably precise data are obtained using the low mode averaging technique. Detailed analysis with the chiral extrapolation is yet to be done after accumulating more statistics.
![ Preliminary results for pion and kaon mass squared (left) and their decay constants (right) as a function of sea quark mass. []{data-label="fig:results"}](mps2vm.eps "fig:"){width="7cm"} ![ Preliminary results for pion and kaon mass squared (left) and their decay constants (right) as a function of sea quark mass. []{data-label="fig:results"}](fpsvm.eps "fig:"){width="7cm"}
Numerical simulations are performed on Hitachi SR11000 and IBM System Blue Gene Solution at High Energy Accelerator Research Organization (KEK) under a support of its Large Scale Simulation Program (No. 07-16). This work is supported in part by the Grant-in-Aid of the Ministry of Education, Culture, Sports, Science and Technology (No. 17740171, 18034011, 18340075, 18740167, 18840045, 19540286 and 19740160).
[99]{}
H. Matsufuru, .
T. Kaneko [*et al.*]{} \[JLQCD Collaboration\], \[arXiv:hep-lat/0610036\]. H. Matsufuru [*et al.*]{} \[JLQCD Collaboration\], \[arXiv:hep-lat/0610026\]. S. Hashimoto [*et al.*]{} \[JLQCD Collaboration\], \[arXiv:hep-lat/0610011\]. N. Yamada [*et al.*]{} \[JLQCD Collaboration\], \[arXiv:hep-lat/0609073\]. J. Noaki [*et al.*]{} \[JLQCD Collaboration\], ; arXiv:0710.0929 \[hep-lat\]. T. Kaneko, H. Fukaya, S. Hashimoto, H. Matsufuru, J. Noaki, T. Onogi and N. Yamada \[JLQCD collaboration\], ; arXiv:0710.2390 \[hep-lat\]. N. Yamada [*et al.*]{} \[JLQCD Collaboration\], ; arXiv:0710.0462 \[hep-lat\]. T.W. Chiu [*et al.*]{} \[JLQCD and TWQCD Collaborations\], .
S. Aoki [*et al.*]{} \[JLQCD and TWQCD Collaborations\], arXiv:0710.1130 \[hep-lat\]. H. Fukaya [*et al.*]{} \[JLQCD Collaboration\], \[arXiv:hep-lat/0610024\]. H. Fukaya [*et al.*]{} \[JLQCD Collaboration\], Phys. Rev. Lett. [**98**]{}, 172001 (2007) \[arXiv:hep-lat/0702003\]. H. Fukaya [*et al.*]{}, Phys. Rev. D [**76**]{}, 054503 (2007) \[arXiv:0705.3322 \[hep-lat\]\]. H. Fukaya [*et al.*]{} \[JLQCD Collaboration\], .
H. Neuberger, Phys. Lett. B [**417**]{}, 141 (1998) \[arXiv:hep-lat/9707022\]. H. Neuberger, Phys. Lett. B [**427**]{}, 353 (1998) \[arXiv:hep-lat/9801031\]. H. Fukaya, S. Hashimoto, K. I. Ishikawa, T. Kaneko, H. Matsufuru, T. Onogi and N. Yamada \[JLQCD Collaboration\], Phys. Rev. D [**74**]{}, 094505 (2006) \[arXiv:hep-lat/0607020\]. S. Aoki, H. Fukaya, S. Hashimoto and T. Onogi, Phys. Rev. D [**76**]{}, 054508 (2007) \[arXiv:0707.0396 \[hep-lat\]\]. N. Cundy, J. van den Eshof, A. Frommer, S. Krieg, T. Lippert and K. Schafer, Comput. Phys. Commun. [**165**]{}, 221 (2005) \[arXiv:hep-lat/0405003\]. R. Narayanan and H. Neuberger, Phys. Rev. D [**62**]{}, 074504 (2000) \[arXiv:hep-lat/0005004\]. A. Borici, arXiv:hep-lat/0402035. R. G. Edwards, B. Joo, A. D. Kennedy, K. Orginos and U. Wenger, PoS [**LAT2005**]{}, 146 (2006) \[arXiv:hep-lat/0510086\]. S. Aoki [*et al.*]{} \[JLQCD Collaboration\], Phys. Rev. D [**65**]{}, 094507 (2002) \[arXiv:hep-lat/0112051\]. A. Bode, U. M. Heller, R. G. Edwards and R. Narayanan, arXiv:hep-lat/9912043. T. DeGrand and S. Schaefer, JHEP [**0607**]{}, 020 (2006) \[arXiv:hep-lat/0604015\].
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We are presenting new results on kinematics and structure of the Mrk 334 Seyfert galaxy. Panoramic (3D) spectroscopy is performed at the 6-m telescope of the Special Astrophysical Observatory of the Russian Academy of Sciences using the MPFS integral-field spectrograph and scanning Fabry–Pérot interferometer. The deep images have revealed that Mrk 334 is observed during the final stage of its merging with a massive companion. A possible mass ratio ranges from $1/5$ to $1/3$. The merger has triggered mass redistribution in the disk resulting in an intensification of nuclear activity and in a burst of star formation in the inner region of the galaxy. The circumnuclear starburst is so intense that its contribution to the gas ionization exceeds that contribution of the AGN. We interpret the nuclear gas outflow with velocities of $\sim200{\,\mbox{km}\,\mbox{s}^{-1}}$ as a galactic superwind that accompanies the violent star formation. This suggestion is consistent with the asymmetric X-ray brightness distribution in Mrk 334. The trajectory of the fragments of the disrupted satellite in the vicinity of the main galaxy nucleus can be traced. In the galaxy disk a cavern is found that is filled with a low-density ionized gas. We consider this region to be the place where the remnants of the companion have recently penetrated through the gaseous disk of the main galaxy.'
author:
- |
Aleksandrina Smirnova$^1$[^1] and Alexei Moiseev$^1$\
$^1$Special Astrophysical Observatory, Russian Academy of Sciences, Nizhnii Arkhyz, Karachaevo-Cherkesskaya Republic, 369167 Russia
title: '3D spectroscopy of merger Seyfert galaxy Mrk 334: nuclear starburst, superwind and the circumnuclear cavern'
---
galaxies: Seyfert – galaxies: individual: Mrk 334 – galaxies: kinematics and dynamics – galaxies: interactions – galaxies: starburst
Introduction {#sec0}
============
Numerical simulations demonstrate that galaxy interaction stimulates a concentration of gas in its central regions, thereby triggering nuclear activity and/or a burst of star formation [@BarnesHernquist1991; @Springel2005; @Bekki2006]. Many authors have tried to find a correlation between an AGN phenomenon and galaxy environment: the presence of companions or traces of interaction [@Dahari1985; @DeRobertis1998; @Schmitt2001; @Knapen2005]. However, statistically significant correlation has not been found.
A number of authors suggest that the activity may be triggered and sustained by a complex mechanism that includes several factors (see @Martini2004 and the references therein). It is clear that only a detailed analysis of the kinematics and dynamics of both the inner ($100-1000$ pc scale) and outer regions in active galaxies would make it possible to understand how in each particular case the ‘fuel’ (interstellar gas) is brought into the domain of action of the gravitational forces of the AGN ‘central engine’.
This paper continues a series of papers dedicated to a detailed study of the inner kinematics of active galaxies via methods of panoramic (3D) spectroscopy. This technique provides spectra for every spatial element (‘spaxel’) of a two-dimensional field of view. It is a powerful tool for studying non-circular motions and gas ionization properties both in the circumnuclear and external regions. Our work is aimed to investigate the relation between the gas kinematics, morphological features and nuclear activity in individual galaxies as well as the mechanisms of the central region feeding. We have already published the results concerning Mrk 315 [@Ciroi2005], NGC 6104 [@Smirnova2006], and Mrk 533 [@Smirnova2007]. In this paper we report a detailed study of Mrk 334.
Mrk 334 (VV 806, UGC 6) is a peculiar galaxy with Sy1.8 nucleus (according to the NED database). This object has been popular among the researchers due to its peculiar appearance on optical images: it has an asymmetric eastward-extending arm [@VorontsovV1977] and a bright H$\alpha$ condensation near the nucleus [@GonzalezDelgado1997]. The latter authors suggested that the object actually consisted of two merging galaxies. Mrk 334 is notable for violent star formation, resulting in high IR luminosity $L_{IR}=8.9\times10^{11}\,$L$_{\odot}$ [@Perez2001]. @RothbergJoseph2004 classify it as a luminous infrared galaxy (LIRG). @Maiolino1997 suggested that the peculiarities in the structure of the galaxy are indicative of its recent interaction with a companion, which has triggered the nuclear activity. However, when and what did the galaxy interact with? In the present paper we try to answer this question and to look for the feedback effects between the central and surrounding regions.
The paper has the following layout. Section \[sec1\] describes the observations and the data reduction; Section \[sec2\] analyses the distribution of ionized gas and stars both in the inner disk and in the outer regions of the galaxy. Section \[sec3\] analyses the ionization state of selected regions in Mrk 334, and in Section \[sec4\] the kinematics of the gas and stars are considered. Section \[sec5\] studies the peculiarities of X-ray radiation according to the *ROSAT* data, and Section \[sec6\] includes an overall discussion of the whole galaxy structure and the circumnuclear ionized gas cavern.
The adopted distance to the galaxy — 91.4 Mpc [@Maiolino1997] corresponds to a scale of 443 pc/arcsec (for redshift z=0.022 and $H_0=75{\,\mbox{km}\,\mbox{s}^{-1}}\,\mbox{Mpc}^{-1}$).
Observations and Data Reduction {#sec1}
===============================
All observations were made in the prime focus of the 6-m telescope of the Special Astrophysical Observatory of the Russian Academy of Sciences (SAO RAS). Table \[tnab\] provides the log of the observations. The central region of Mrk 334 was observed with the Multi-Pupil Fiber Spectrograph (MPFS). The large-scale kinematics and galactic environment were studied using the SCORPIO multi-mode focal reducer operating in the modes of scanning Fabry–Pérot interferometer (FPI) and broad-band imaging. The detectors used in 2006 and 2002 were a CCD EEV42-40 ($2048\times2048$ pixels) and a CCD TK1024 ($1024\times1024$ pixels), correspondingly.
MPFS integral-field spectrograph
--------------------------------
The integral-field spectrograph MPFS [@Afanasiev2001] takes simultaneous spectra of 256 spatial elements arranged in the form of $16\times16$ square lenses array with a scale of 1 arcsec per spaxel. Behind each lens an optical fibre is located whose other end is packed into the spectrograph slit. The sky background spectrum was simultaneously taken with 17 additional fibres located 4 arcmin away from the object. The wavelength interval included numerous emission lines of ionized gas and absorption features of the stellar population.
The preliminary data reduction steps were described earlier [@Moiseev2004; @Smirnova2007]. Reduction yields a ‘data cube’, where each pixel in the $16\times16$ arcsec field has a spectrum associated with it. The spectra of the spectrophotometric standard stars were used to convert counts into absolute fluxes. Observations were made successively in two spectral intervals (see Table \[tnab\]). The overlap of the two spectral domains allowed us to join them so as to operate with a single data cube covering a $\lambda\lambda3740-7220$Åspectral range.
To construct the stellar velocity field, we use the cross-correlation technique adapted for MPFS data [@Moiseev2001]. Spectra from MILES library [@Sanchez2006] smoothed to the instrumental resolution were adapted as templates for cross-correlation. We have mapped the line-of-sight velocity and brightness distribution fields for the main emission lines using the Gaussian fitting of their profiles. Underlying absorption lines were taken in account as approximation by a linear combination of smoothed and redshifted MILES templates.
------------ ------------ --------- -------------- -------- --------
Date Instrument Exp. Sp. Sp. seeing
time, s range resol. arcsec
23.11.2006 MPFS 7200 $3740-5850$Å 6.5 Å 1.4
7200 $4300-7220$Å 6.5 Å 1.4
05.09.2002 SCORPIO 6400 H$\alpha$ 2.8 Å 1.3
(FPI)
23.10.2006 SCORPIO 660 $V$ 1.4
(Images) 1020 $R$ 1.4
------------ ------------ --------- -------------- -------- --------
: Log of the observations.[]{data-label="tnab"}
![Images (top) and velocity fields (bottom) in the H$\alpha$ (left) and \[OIII\]$\lambda\lambda4959,5007$ (right) emission lines according to MPFS data. Regions ‘A’ and ‘B’ are marked.[]{data-label="f01"}](Smirnova_fig1a_color.eps "fig:"){width="4.2cm"} ![Images (top) and velocity fields (bottom) in the H$\alpha$ (left) and \[OIII\]$\lambda\lambda4959,5007$ (right) emission lines according to MPFS data. Regions ‘A’ and ‘B’ are marked.[]{data-label="f01"}](Smirnova_fig1b_color.eps "fig:"){width="4.2cm"}
![Images (top) and velocity fields (bottom) in the H$\alpha$ (left) and \[OIII\]$\lambda\lambda4959,5007$ (right) emission lines according to MPFS data. Regions ‘A’ and ‘B’ are marked.[]{data-label="f01"}](Smirnova_fig1c_color.eps "fig:"){width="4.2cm"} ![Images (top) and velocity fields (bottom) in the H$\alpha$ (left) and \[OIII\]$\lambda\lambda4959,5007$ (right) emission lines according to MPFS data. Regions ‘A’ and ‘B’ are marked.[]{data-label="f01"}](Smirnova_fig1d_color.eps "fig:"){width="4.2cm"}
SCORPIO
-------
SCORPIO universal instrument [@AfanasievMoiseev2005] allows various spectroscopic and photometric observations to be performed within $6$ arcmin field of view. Below we describe each of the modes employed in detail.
### Fabry–Pérot Interferometer
We used the scanning FPI operating in the H$\alpha$ emission line to study the kinematics of the ionized gas. During the observations we successively took 32 interference images of the object by changing the FPI plate gap. A detailed description of the technique of observations and data reduction can be found in the papers by @Moiseev2002 and @Moiseev2008. This reduction yields a data cube, where a 32-channel spectrum with a sampling step of 0.9 Å is attached to each $0.28$ arcsec pixel. The spectroscopic resolution is 2.8 Å. The velocity field of the ionized gas and images in the H$\alpha$ emission line were mapped using Gaussian fitting of the emission-line profiles. We also generated an image of the galaxy in the continuum close to the emission line.
We calibrated the emission-line flux map into absolute energy units ($\mbox{erg}\,\mbox{s}^{-1}\,\mbox{cm}^{-2}$) by comparing it with the H$\alpha$ distribution according to the MPFS data for the central region. The total H$\alpha$-luminosity of the galaxy was found to be $2.3\times10^{41}\,\mbox{erg}\,\mbox{s}^{-1}$ given the adopted extinction of $A_R=1.5^m$. The extinction estimate is based on the H$\alpha/$H$\beta$ line intensity ratio for the HII region to the west of the nucleus (hereinafter referred to as ‘Region A’). According to @Kennicutt1998ARA, such a luminosity corresponds to star-formation rate of SFR=$18\,$M$_\odot/\mbox{yr}$, if we neglect a contribution from the AGN in the H$\alpha$ flux.
### Direct Images
We took images of the galaxy in the Johnson–Cousins [*V*]{} and [*R*]{} bands with a sampling of 0.35 arcsec per pixel. Non-photometric weather conditions prevented the use of standard stars to calibrate fluxes. We performed a coarse calibration in the [*V*]{} band based on the aperture photometry data listed in the HyperLeda database . An accuracy of the zero-point is $0.1-0.2\,\mbox{mag}$. We calibrated the [*R*]{} image assuming that the average colour index in the disk of Mrk 334 corresponds to that of a normal Sb-Sc-type galaxy: [*(V-R)*]{}$\approx0.5$. The assumption of the normal $V-R$ color is justified by the known normal 2MASS infrared colours of Mrk 334. The depth of the surface-brightness measurements reaches $25.5\,\mbox{mag}\,\mbox{arcsec}^{-2}$ in the [*V*]{} and [*R*]{} bands, which is significantly deeper than the other known images of Mrk 334.
{width="8cm"} {width="8cm"}
{width="8cm"} {width="8cm"}
Morphological features of Mrk 334 {#sec2}
=================================
Line and Continuum Images
-------------------------
We use the MPFS spectra to construct the maps in various emission lines covering the $7\times7$ kpc central region of Mrk 334 (see Fig. \[f01\]). The maps show some other bright regions besides the nucleus. The brightest of them, the Region ‘A’, is located 4 arcsec to the west of the nucleus. @GonzalezDelgado1997 were the first to find it in H$\alpha$. At our maps it can be seen in other emission lines and in the continuum. Fainter Region ‘B’ is located $r=3-4$ arcsec to the south of the nucleus and shows up mostly only in the \[OIII\]$\lambda\lambda4959,5007$ doublet. The continuum image exhibits, in addition to Region ‘A’, an amorphous structure that we refer to as Region ‘C’.
Many authors (see Introduction) believed Mrk 334 to be an interacting system. Their conclusion was based mostly on the presence of an asymmetric spiral arm to the east of the nucleus resembling a tidal tail [@Hunt1997; @RothbergJoseph2004]. This feature is clearly visible on the FPI continuum image and is absent in the H$\alpha$ line map (Fig. \[f02\]). Therefore, this arm does not harbor star-forming regions.
---------------- ------------- ----------------- -----------------
Component parameter $V$ $R$
filter filter
Sersic’s bulge n 2 2
$\mu_{eff}$ $20.3$ mag $19.8$ mag
$r_0$ 32 (1.4 kpc) 32 (1.4 kpc)
Inner disk $\mu_0$ $21.5$ mag $20.9$ mag
$h$ 78 (3.5 kpc) 87 (3.9 kpc)
Outer disk $\mu_0$ $24.8$ mag $24.2$ mag
$h$ $32''$ (14 kpc) $43''$ (19 kpc)
---------------- ------------- ----------------- -----------------
: Parameters of the Photometric Decomposition.[]{data-label="tab_2"}
![The mean $R$-band brightness profile (the bold dots) and its model fit (the solid line). The dashed lines show the contribution of each of the disks. The dotted line marks the bulge. []{data-label="f04"}](Smirnova_fig4.eps){width="8cm"}
{width="5.5cm"} {width="5.5cm"} {width="5.5cm"}
Multicomponent Structure of the Disk and Outer Filaments
--------------------------------------------------------
The deep images of the galaxy show that the tidal arm noted above is the brightest part of the vast system of shells and lower surface brightness filaments (Fig. \[f03\]). The shells have sharp outer edges located about 70 arcsec northwest and about 100 arcsec southwest of the nucleus, which corresponds to 31 and 44 kpc, respectively. The [*R*]{} band surface brightness of the outer regions is about $24-25\,\mbox{mag}\,\mbox{arcsec}^{-2}$. Similar arc-like features are typical of galaxies currently interacting or having interacted in the past with a companion [@SchweizerSeitzer1988; @WehnerGallagher2005].
To study the brightness distribution in the filaments, we must remove the axisymmetric components of the galaxy – the bulge and the disk. To decompose the image into components, we use an iterative method of constructing one- and two-dimensional models [@Moiseev2004; @Ciroi2005]. The idea of the method is to determine the parameters of the exponential disk from the outer parts of the azimuthally averaged brightness profile and subtract the resulting brightness distribution from the original image. The residual image is averaged over round apertures and fitted to the Sérsic’s profile for the bulge. We then subtract the bulge model from the initial image and use the residual image to build the next iteration for the disk. When constructing the model we masked the high-contrast features like stars and tidal spirals. We found the galaxy image to be best approximated by the model consisting of a bulge and two exponential disks with different radial scales. Table \[tab\_2\] lists the parameters of the photometric components. Here $n,r_{e}$, and $\mu_{eff}$ are the exponent, effective radius, and brightness of the bulge, respectively, and $\mu_0$ and $h$ are the central surface brightness and radial disk scalelength. The position angle and the apparent ellipticity of the disks were fixed in accordance with the orientation of the ionized-gas disk (Section \[sec4\]). The results of decomposition performed in the two filters agree fairly well with each other, except that the scalelength of the outer disk is larger by a factor of 1.3 in the $R$ band.
Figure \[f04\] shows the surface-brightness profile computed by averaging the brightness over elliptical rings. It shows a well-defined break at $r=50-60$ arcsec and is dominated by the outer disk at larger galactocentric distances. Such multicomponent (two-tiered) disks have now become increasingly popular among the researchers. According to the classification proposed by @Erwin2005, Mrk 334 exhibits a typical type III (antitruncated) surface-brightness profile. Among the galaxies studied by @Erwin2005 such profile characterizes mostly post-interacting objects. A detailed study of individual galaxies also suggests that interaction events may result in the formation of multi-tiered disks. Examples include NGC 615 [@Silchenko2001], Mrk 315 [@Ciroi2005], NGC 7217 and NGC 7742 [@SilchenkoMoiseev2006].
Mrk 334 appears to represent configuration, where the debris of the companion torn apart by the tidal forces precess in the plane of the galaxy. The outer disk is being formed right now with a relatively long scalelength and a low central brightness. Indeed the parameters of the inner disk (Table \[tab\_2\]) are typical for a spiral galaxy, whereas the outer disk has a rather long radial scalelength and the $\mu_0$ that is typical for low surface brightness (LSB) galaxies. In Mrk334 we caught an LSB disk in the process of formation. Here the mean brightness profile has already become close to an exponential, despite the asymmetry of azimuthal light distribution that is still very inhomogeneous. The brightness of the outer disk should become more homogeneous after a few revolutions that is after about $0.5-1$ Gyr.
Figure \[f03\]b shows the $R$-band brightness distribution after subtracting the model consisting of two disks and a bulge. A complex system of bright loops becomes immediately visible. Three characteristic radial scales can be identified: the circumnuclear ring with a radius of about ($4-5$ kpc); inner filaments at a distance of $r=9-13$ kpc, and outer features — a loop and an arc located northwest and southwest of the nucleus, respectively, which can be observed beyond $r\sim 40$ kpc. This subdivision of tidal features is rather arbitrary, we are most probably observing a disruption of a single galaxy torn apart by tidal forces and spread along its orbit around Mrk 334. The bulk of the companion stars is concentrated in the central region, because the mean brightness of the inner and circumnuclear filaments is twice higher than that of the outer ones.
{width="8cm"} {width="8cm"}
{width="8cm"} {width="8cm"}
The total luminosity contribution of all inner and outer filaments determined after subtracting the two-dimensional model is $30\%$ in the $V$ and $25\%$ in the $R$ band. Under the assumption that all filaments (including the circumnuclear ones) were formed by stars of the satellite, we have derived the mass ratio for the galaxies before interaction ranging from $1/5$ to $1/3$ for the equal $M/L$ ratios in both galaxies. The scatter of the estimates is mostly due to the uncertain fraction of companion stars dropped into the outer disk. This ratio is close to the conventional boundary between minor and major merging. Whereas in the former case the interaction should rather regarded as a simple accretion of a low-mass companion by the primary galaxy, in the case of major merging this process also distorts appreciably the structure of the more massive companion.
Figure \[f05\] shows the central part of Mrk 334. The system of arcs and loops corresponding to the orbit of the disrupted companion shows up conspicuously on the map of residual brightness. According to the $(V-R)$ maps, individual fragments of these loops stand out because of their bluer colours, which are indicative of the presence of younger population that may have formed in the process of interaction between the galaxies. The red colour of Region ‘A’ is evidently due to the H$\alpha$ emission falls within the $R$-filter passband. Note that the tidal structures at $r=5-10$ arcsec are most likely located outside the plane of the galaxy, as follows from the analysis of the gas velocity field (Section \[sec4\]).
At the same time, the circumnuclear spiral at $r<5$ arcsec, which shows up on HST images (Fig. \[f05\]c) seems to belong to the galaxy disk plane. The authors having reported HST images of Mrk 334 [@MartiniPogge1999; @Martini2001] have also pointed out the minispiral and chaotic dust features in the circumnuclear region.
Sources of Gas Ionization. {#sec3}
==========================
Figure \[f06\] shows the integrated spectra of Regions ‘A’, ‘B’, and ‘C’ and of the nucleus extracted from the MPFS data cube. The boundaries of these regions are shown in Figure \[f05\]. The spectra of the nucleus and of the knot ‘A’ have much in common: both exhibit strong Balmer lines and weaker forbidden lines, mostly \[OIII\]$\lambda\lambda4959,5007$ and \[OII\]$\lambda3727$. Region ‘B’ shows the opposite pattern: the most conspicuous line is \[OIII\]$\lambda5007$, which is even brighter than H$\alpha$. The spectrum of Region ‘C’ exhibits, along with a very weak H$\alpha$ emission, high-contrast MgI, Fe, and Ca and Balmer-line absorptions. Such a spectrum is typical for a composite post-starburst region. We estimate the luminous-weighted age of the stellar population as 1.1-1.6 Gyr using the ULYSS[^2] program package.
Regions with different emission-line spectra must also differ in their ionization sources. We construct the diagnostic diagrams to determine the ionization mechanism for the inner regions of the galaxy. Given the line ratios for different excitation mechanisms, we can identify regions dominated by thermal (young stars), nonthermal (active nucleus), or shock ionization (hereinafter referred to as HII, AGN, and LINER). Figure \[f07\] shows the most typical diagrams. In the diagrams we adopted the boundaries separating domains corresponding to the different excitation mechanisms from @VeilleuxOsterbrock1987. We ignored the effect of internal extinction, because we use the intensity ratios of the lines with close wavelengths.
Mrk 334 is classified as a Sy 1.8-type galaxy and its Balmer-line profiles have a low-contrast broad component. We decomposed the H$\alpha$ and H$\beta$ line profiles into two Gaussian functions: one for the broad and one for the narrow component. Only the narrow component flux was used in the diagnostic diagrams.
The points in the diagrams of Fig. \[f07\] show the line ratios in each MPFS spaxel. In all diagrams the points corresponding to the nucleus lie at the HII-LINER boundary. Here the main ionization mechanisms are radiation of young stars and shocks, and not the nonthermal UV continuum as it is typical of an active nucleus. The points corresponding to the nucleus are located in the diagrams so far from the LINER-AGN boundary that the nucleus of Mrk 334 should be classified as a LINER rather than a Sy galaxy. This conclusion is consistent with the recent spectrophotometric studies of the nucleus reported by @Lumsden2001. The asterisks in Fig. \[f07\] indicate the line ratios in the nucleus as inferred from their long-slit spectrum[^3]. Figure \[f07\] demonstrates the good agreement between their data and our measurements for the nucleus except for the \[SII\]/H$\alpha$ ratio. The last discrepancy must be due to the fact that we corrected our spectra for the atmospheric absorption band which partially overlaps with the \[SII\] lines and decreases sulfur flux by $\sim0.1\,\mbox{dex}$. @Lumsden2001 say nothing about applying such a correction.
{height="7cm"} {height="7cm"}
{height="7cm"} {height="7cm"}
The emission-line ratios observed in the nucleus can be explained in terms of the following supposition. The star formation in the nucleus is so violent that the total line emission is determined mostly by the collective effect of photoionizing radiation of young stars and by the shocks produced by supernova explosions. At the same time, the emission lines of the active nucleus are barely visible against the circumnuclear starburst. The weak broad component with $FWHM\approx2500{\,\mbox{km}\,\mbox{s}^{-1}}$ in the hydrogen line profiles and FeII features in the spectrum of the nucleus are the only indications of an AGN central engine.
In all diagrams the points belonging to Knot ‘A’ lie deeply in the region corresponding to the ionization by OB stars radiation. Thus ‘A’ is indeed a region of intense ongoing star formation. It accounts for $15-20\%$ of the total H$\alpha$ luminosity, and such a fraction formally corresponds to a star formation rate of $SFR=3\,\mbox{M}_\odot/$yr. Such a high value (equivalent to the total SFR in the starburst galaxy M 82) in a relatively compact region (1 kpc in size) is indicative of a powerful starburst.
Condensation ‘B’ appears to be the most intriguing among these features: it exhibits unusually high \[OIII\]$/\mbox{H}\beta$ line intensity ratios (see Fig. \[f08\]a)that formally correspond to the ionization by an AGN. Two hypotheses can be suggested to explain this peculiarity. First, Region ‘B’ may be the active nucleus of the satellite. @Barth2008 recently observed a similar situation in the interacting galaxy NGC3341, where the spectrum of the nucleus of the disrupted companion exhibits Seyfert-type features. However, Condensation ‘B’ is barely visible on the optical continuum images of Mrk 334. Observations made at the 6 cm wavelength [@Ulvestad1986] with an angular resolution of $0.4-0.6$ arcsec also demonstrate the lack of a nonthermal radio emission typical of an active nuclei in Region ‘B’.
The second hypothesis is based on the fact that Region ‘B’ is located close to the tidal arclike structures identified in the circumnuclear region images (Fig. \[f05\]b). It would be safe to assume that here we see the intersection between the galaxy disk and the orbit of the disrupted companion remnants. It is the locus where the debris have ‘punched’ the gaseous disk of Mrk 334, thereby creating in Region ‘B’ a cavern with lower than ambient gas density. The high degree of ionization of the gas is due to a powerful shock. This idea is also supported by the electron-density estimates derived from the \[SII\]$\lambda\lambda6730/6717$ line flux ratio using the relation adopted from @Osterbrock1989 for $T_e=10\,000\,K$. The electron density is equal to $n_e=250-430\,\mbox{cm}^{-3}$ in the nucleus, $n_e=200-350\,\mbox{cm}^{-3}$ in Region ‘A’, and increases to $500\,\mbox{cm}^{-3}$ to the north of this HII region (Fig. \[f08\]b). At the same time, the sulphur line ratio is less than 0.7 at almost all points of Condensation ‘B’, which is impossible to interpret in terms of simple photoionization models. It is indicative of high temperature and low gas density ($n_e<20\,\mbox{cm}^{-3}$).
We tried to compare the line ratios observed in Region ‘B’ with the results by the modern shock ionization simulations adopted from @Allen2008. We found that the theoretical shock+precursor model predictions describe fairly well the line ratios observed in Region ‘B’ (Fig. \[f07\]) for parameters that can be reasonably expected for the interstellar medium: the density of $1\,\mbox{cm}^{-3}$, solar elemental abundances and magnetic-parameter $B/n^{1/2}=0.1-10\,\mu \mbox{G}\mbox{cm}^{3/2}$. The required shock speed is $250-350{\,\mbox{km}\,\mbox{s}^{-1}}$. The shock origin of the \[OIII\] emission is also evident from the line broadening. Indeed the width of the \[OIII\] lines in Region ‘B’, after instrumental width correction, corresponds to the $\sigma=160-200{\,\mbox{km}\,\mbox{s}^{-1}}$ instead $60-100{\,\mbox{km}\,\mbox{s}^{-1}}$ in other disk points, with the exception of the nucleus. More kinematic evidences will be presented in the next Section.
Region ‘C’ may be satellite debris that has recently punched a hole in the gaseous disk of Mrk 334. It is located close to Region ‘B’; it appears sufficiently bright in the continuum images and its spectrum contains lines of young stellar population; it appears compact compared to other nearby filaments. It may actually be the remnant of the nucleus of a companion galaxy. The data on the ionization sources in this region are rather scarce and available only for the boundary between Region ‘C’ and the nucleus of Mrk 334 because the spectrum shows mostly stellar absorptions. In the diagnostic diagrams Region ‘C’ lies in the HII and LINER sectors. The photoionization is caused here mostly by star-forming processes, and the shock ionization contributes appreciably in the points located close to the nucleus.
Kinematics of Ionized Gas and Stars. {#sec4}
====================================
The H$\alpha$ velocity field derived from the FPI data seems to be in a good agreement with the model of a regular rotating thin disk (Fig.\[f09\]). We fitted the velocity field by the “tilted-ring” model using the algorithms employed earlier to study NGC 6104 [@Smirnova2006]. The circular rotation explains the gas velocity field fairly well. Therefore we think that the inner tidal features ($r=5-10$ arcsec) are located outside the galaxy plane and do not perturb the entire gaseous disk but only cross it in some places. We found the disk inclination to be $i_0=34pm6$ deg and the line-of-nodes position angle of $PA_0=297\pm3$ deg. Fig. \[f10\] shows the rotation curve of ionized gas ($V_{rot}$) and the radial variations of the kinematic axis ($PA_{kin}$). The data points in the range of $r=12-22$ arcsec come from three outer HII regions located far from the central disk: $\mbox{PA}_{kin}$ could not be determined from these regions, and we assume it to be equal to the mean $PA_0$ of the disk.
![Map of the \[OIII\]$\lambda5007$/H$\beta$ (a) and \[SII\]$\lambda6731/6717$ (b) line ratios. The \[OIII\]$\lambda5007$ isophotes are overlapped. In the case of sulphur lines, darker colours correspond to higher electron density.[]{data-label="f08"}](Smirnova_fig8a.eps "fig:"){height="4.2cm"} ![Map of the \[OIII\]$\lambda5007$/H$\beta$ (a) and \[SII\]$\lambda6731/6717$ (b) line ratios. The \[OIII\]$\lambda5007$ isophotes are overlapped. In the case of sulphur lines, darker colours correspond to higher electron density.[]{data-label="f08"}](Smirnova_fig8b.eps "fig:"){height="4.2cm"}
The same figure shows kinematic parameters for the old stellar population. The rotation velocity measured for stars is about twice (by $\sim100{\,\mbox{km}\,\mbox{s}^{-1}}$) smaller than that found for the ionized gas. This discrepancy must be due to asymmetric drift, because the central velocity dispersion of stars reaches, according to our estimates, $170-200{\,\mbox{km}\,\mbox{s}^{-1}}$. The rotation curve of both gas and stars exhibits a characteristic peak near the effective radius of the bulge. At greater galactocentric distances, $r=4-12$ arcsec, the rotation velocity of gas is almost constant and equal to $210-220{\,\mbox{km}\,\mbox{s}^{-1}}$. The $V_{rot}$ of the external HII regions mentioned above is lower by $40-50{\,\mbox{km}\,\mbox{s}^{-1}}$, however, we have no grounds to believe that these regions may be located outside the disk of the galaxy. It seems to be more likely that the formal decrease of the rotation velocity results from a contribution of non-circular gas motions.
Small variations of $PA_{kin}$ at $r<5$ arcsec indicate the influence of the circumnuclear spiral on the kinematics of the gaseous and stellar subsystems. The $PA_{kin}$ abruptly deviates from the line of nodes at $r=8-11$ arcsec. Such a behaviour is indicative of large-scale noncircular motions at the edge of the HII disk. The residual-velocity map (Fig.\[f09\]b) shows the distribution of observed velocities after the subtraction of the model. Deviations from circular rotation are small ($15-20{\,\mbox{km}\,\mbox{s}^{-1}}$) in the regions with the brightest H$\alpha$ emission. However, to the south from the nucleus an extended region can be seen where peculiar velocities are much higher and vary smoothly from $-70$ to $+60{\,\mbox{km}\,\mbox{s}^{-1}}$ in the east–west direction. Region ‘B’ is located here, which we have identified earlier by its spectrophotometric properties, primarily by its high \[OIII\]/$H_\beta$ ratio. Now we see that this region is also distinguished by the peculiar kinematics of ionized gas. Such a velocity distribution corroborates the hypothesis — suggested above – that Region ‘B’ is the locus where the debris of the disrupted companion crossed the gaseous disk of the galaxy.
![Kinematics of ionized gas according to the H$\alpha$-line FPI data: the velocity field (a) and the residual velocities (observations minus model) with H$\alpha$-line isophotes overlapped (b).[]{data-label="f09"}](Smirnova_fig9a_color.eps "fig:"){width="8cm"} ![Kinematics of ionized gas according to the H$\alpha$-line FPI data: the velocity field (a) and the residual velocities (observations minus model) with H$\alpha$-line isophotes overlapped (b).[]{data-label="f09"}](Smirnova_fig9b_color.eps "fig:"){width="8cm"}
![The rotation curve (top) and variations of the position angle of the major axis with a radius (bottom). The black circles show the H$\alpha$-line FPI velocity-field measurements and diamond signs mark the measurements for the stellar component (MPFS). []{data-label="f10"}](Smirnova_fig10.eps){width="8cm"}
An alternative explanation of the observed gas kinematics is a jet–clouds interaction similar to that observed in Mrk 3 [@Capetti1999] or Mrk 533 [@Smirnova2007]. However, this mechanism is unlikely for Mrk 334. Firstly, as we already pointed out above, this galaxy has no extended radio structure. Secondly, if the jet acts on the interstellar medium then the expected gradient of peculiar velocities should be directed away from the nucleus, i.e., in the radial and not in the azimuthal direction as we see in Fig.\[f09\]b.
The velocity fields derived from the MPFS data (Fig.\[f01\]) have allowed us to study the kinematics of gas in lines excited by different mechanisms, albeit with lower accuracy and in coarser detail compared to the results based on the FPI H$\alpha$ data. The gas motions observed via most of the low-excitation emission lines (\[OI\], \[OII\], \[NII\], \[SII\]) agree well with the picture found in the $H_\alpha$ line. Namely, they show a circular rotation with appreciable deviations near Region ‘B’. Only the velocity distribution in the \[OIII\] line differs from the overall pattern (Fig.\[f01\], bottom). Figure \[f11\] shows the residual velocities in this line after the subtraction of the circular-rotation model derived from the H$\alpha$ data. Two features are apparent. First, the residual velocities in Region ‘B’ reach $-150{\,\mbox{km}\,\mbox{s}^{-1}}$, which is greater by amplitude than the corresponding velocities for the low-excitation lines. Second, the galaxy nucleus shows a significant excess of negative velocities (down to $-300{\,\mbox{km}\,\mbox{s}^{-1}}$ ). Similar gas outflows from AGN (excess of ‘blue’ velocities, first and foremost, in the \[OIII\] line) were found in the integral-field spectroscopy data for Seyfert galaxies studied by us and by other teams: Mrk 315 [@Ciroi2005], NGC 2273 [@Moiseev2004], NGC 2992 [@Garcia2001], NGC 6104 [@Smirnova2006] and others. A nuclear-blueshifted outflow associated with this feature is usually interpreted as a jet–clouds interaction [@Ferruit2002], or, in a more general case, as hot wind emerging from an active nucleus (@Komossa2008 and references therein). However, the situation in Mrk 334 must be different. Firstly, no radio jet can be seen in high-resolution radio maps, in any case, its size cannot exceed $0.5$ arcsec. Secondly, in the nucleus itself the contribution from the nonthermal component to the ionization of gas is smaller than that of star formation (see Section \[sec3\]). Thirdly, unlike the examples of active galaxies mentioned above, the high-velocity outflow in Mrk334 is observed only in the \[OIII\] line. This outflow is most likely associated with intense star formation in the nucleus rather than with the central engine itself. Thus, what we observe in the \[OIII\] line is a low-density gas ejected above the plane of the galaxy as a result of multiple supernova explosions – the so-called ‘superwind’. Below we analyse this possibility in more detail.
We computed the velocity field of the stellar component by correlating the spectra of the galaxy with the stellar spectra from the MILES library and selecting the spectral type of the template and wavelength interval so as to maximise the amplitude of the cross-correlation function. Figure \[f11\] shows the velocity field corresponding to the old stellar population determined by cross-correlation of the galactic spectra with the spectrum of a K0III type star in the interval of $\lambda5120-5800$Å. This velocity field was used to construct the stellar rotation curve presented in Fig. \[f10\]. However, a significant contribution of younger stellar population can be seen in the bluer part of the galaxy spectra in some regions. Thus in the wavelength interval $\lambda3750-4350$Å the coefficient of correlation with the spectrum of a F–type star exceeds the corresponding value for a K–type star in the $\lambda5100-5500$Åwavelength interval. The measurements made for the ‘old’ and ‘young’ populations yield different line-of-sight velocities. Figure \[f11\] shows the difference between the velocity fields of old and young stars. In the nucleus the difference is small and does not exceed $20{\,\mbox{km}\,\mbox{s}^{-1}}$, which is comparable to measurement errors. However, the velocity difference reaches $-80{\,\mbox{km}\,\mbox{s}^{-1}}$ in two regions to the east of the nucleus. Both regions identified in the field of stellar residual velocities coincide with the inner tidal loops in the circumnuclear region, including Region ‘C’, (Fig.\[f05\]b). These facts lead us to suggest that here we see two kinematic components along the same line of sight. The old stellar population belongs to the disk of Mrk 334 and exhibits normal circular rotation. At the same time, the tidal filaments formed in the process of the companion disruption are dominated by younger population (as a result of a relatively recent burst of star formation). The filaments are located outside the disk plane as is evident from their line-of-sight velocities. Fig.\[f11\]c also shows another region with significant negative differences of the young and old stars velocities to the west from the nucleus. It may also include stars from the companion galaxy.
An interesting pattern emerges if we cross-correlate the spectra in the NaD doublet spectral domain. Figure \[f11\] shows the line-of-sight velocities for this line measured after the subtraction of the velocity field of the old stellar component. An excess of negative residual velocities up to $-180{\,\mbox{km}\,\mbox{s}^{-1}}$ is immediately apparent in the nucleus of the galaxy. The NaD line is present not only in the spectra of late-type stars but also in the spectra of the interstellar medium. It is reasonable to associate the excess velocities in this line with the same superwind that we found in the \[OIII\]-line data for the ionized gas. We should note that outflow velocities may be underestimated because the contamination of the NaI line by absorption from the stellar population is also present.
![(a) Residual \[OIII\]-line velocities (observed velocities minus the rotation model in H$_\alpha$). (b) – Velocity field of the old stellar population. (c) – The difference between the velocities of the ‘young’ (F-type stars) and ‘old’ (K-type stars) populations. (d) – The residual velocities in the NaD line minus the rotation velocity of the old stellar population. Figures (a), (c), and (d) are shown with the \[OIII\]-line isophotes superimposed.[]{data-label="f11"}](Smirnova_fig11a_color.eps "fig:"){height="4cm"} ![(a) Residual \[OIII\]-line velocities (observed velocities minus the rotation model in H$_\alpha$). (b) – Velocity field of the old stellar population. (c) – The difference between the velocities of the ‘young’ (F-type stars) and ‘old’ (K-type stars) populations. (d) – The residual velocities in the NaD line minus the rotation velocity of the old stellar population. Figures (a), (c), and (d) are shown with the \[OIII\]-line isophotes superimposed.[]{data-label="f11"}](Smirnova_fig11b_color.eps "fig:"){height="4cm"}
![(a) Residual \[OIII\]-line velocities (observed velocities minus the rotation model in H$_\alpha$). (b) – Velocity field of the old stellar population. (c) – The difference between the velocities of the ‘young’ (F-type stars) and ‘old’ (K-type stars) populations. (d) – The residual velocities in the NaD line minus the rotation velocity of the old stellar population. Figures (a), (c), and (d) are shown with the \[OIII\]-line isophotes superimposed.[]{data-label="f11"}](Smirnova_fig11c_color.eps "fig:"){height="4cm"} ![(a) Residual \[OIII\]-line velocities (observed velocities minus the rotation model in H$_\alpha$). (b) – Velocity field of the old stellar population. (c) – The difference between the velocities of the ‘young’ (F-type stars) and ‘old’ (K-type stars) populations. (d) – The residual velocities in the NaD line minus the rotation velocity of the old stellar population. Figures (a), (c), and (d) are shown with the \[OIII\]-line isophotes superimposed.[]{data-label="f11"}](Smirnova_fig11d_color.eps "fig:"){height="4cm"}
X-Ray Radiation of Mrk 334 {#sec5}
==========================
Let us now briefly discuss the peculiarities of the distribution of X-ray flux of Mrk 334 according to the *ROSAT* data. @Zimmermann2001 report isophotes of the smoothed image in the energy interval of $0.1-2.4$ keV. The X-ray source with the luminosity of $L_X=2.6\cdot10^{42}\, \mbox{erg}\,\mbox{s}^{-1}$ is unambiguously identified with the galaxy. However, the contours of the X-ray image are appreciably offset with respect to the optical nucleus, and outer isophotes coincide with the tidal structures at $r=70-100$ arcsec northwest and southwest of the centre of the galaxy (Fig. \[f12\]). Hence the diffuse X-ray emission is associated with the merging galaxy system but not with the Seyfert nucleus. Especially striking is the almost exact coincidence of the outer X-ray isophotes with the edges of optical filaments which is surprising given the relatively low spatial resolution of the *ROSAT* data.
If the most of the X-ray radiation of Mrk 334 is unassociated with the active nucleus, it may be due either to unresolved stellar sources or to the outer hot gas. We believe the former variant to be unlikely. Excess number of X-ray point sources – close binaries, ULXS, young supernova remnants associated with a starburst – are observed in a number of interacting galaxies. In this case other signs of ongoing star formation and, in particular, HII regions, should also be apparent at the periphery of the galaxy. However, according to our data, H$\alpha$ emission is concentrated only in the central region, inside $r<12$ arcsec.
![The $R$-band image of the galaxy with *ROSAT* X-ray isophotes from @Zimmermann2001 superimposed.[]{data-label="f12"}](Smirnova_fig12.eps){width="8cm"}
If the source of the emission is hot gas then where does it come from? It cannot be the corona of the disrupted companion like those observed in giant elliptical galaxies, because the mass of the companion is not too large, and Mrk 334 is a disk galaxy. An HI corona may have formed from the gas scattered as a result of interaction, whereas most of this gas is concentrated in the disk plane. It is, however, unclear how this gas can be heated.
We could suggest only one more or less realistic scenario to explain the formation of the extended X-ray structure. The inner parts of the galaxy are marked by high star-formation rate. Our estimates yield SFR=$18\,$M$_\odot\,\mbox{yr}^{-1}$ (from the H$\alpha$ line) and $12\,$M$_\odot\,\mbox{yr}^{-1}$ (as inferred from the *IRAS* FIR using the relations from @Kennicutt1998). Such an intense star-forming activity in a compact region accompanied by supernova explosions may give rise to a superwind phenomenon that results manifests significant heating of the interstellar medium and its outflow within a wide cone perpendicular to the galaxy disk [@Heckman1990; @Veilleux2005]. Inside the cone the temperature may rise to $10^7$ K, and hence the gas should be a powerful X-ray source. The starburst starts early enough during the interaction of galaxies — it begins before the complete disruption and merger of the companion. As an example, we can mention the well-known superwind galaxy M 82 that is currently in the process of a tidal interaction with M 81, or the NGC 6285/6286 galaxy pair [@Shalyapina2004]. Thus, a hot gaseous ‘bubble’ or cone could be formed above the plane of Mrk 334 even before the final stage of the merging. The spatial distribution of the hot gas was then distorted because of the significant perturbations of the gravitational potential of the system. We observe the result of these distortions as the offset of the X-ray brightness centre in the sky-plane projection.
We have already pointed out in Section \[sec4\] that the MPFS spectra are indicative of the presence of the superwind in Mrk 334. Gas motions directed toward the observer have been found in the central region. The observed velocities ($180-300{\,\mbox{km}\,\mbox{s}^{-1}}$) are typical of superwind galaxies with intense star formation in their nuclei [@Veilleux2005].
Discussion. {#sec6}
===========
The analysis of various observational data lead us to conclude that Mrk 334 is in the process of merging with a companion that has already been almost completely disrupted by the tidal forces. Is the nuclear activity associated with such a close interaction? Let us first turn to the galaxy morphology. The inner region ($r\leq3$ kpc) hosts a well-defined spiral pattern (Fig. \[f05\]c). Nuclear spirals in Mrk 334, where sites of star formation are located, are relatively brighter than similar features in other galaxies [@Deo2006]. The luminous HII Region ‘A’ located in the western spiral arm has the size typical of giant star-forming complexes in nearby spiral galaxies. Infrared and UV observations [@MunozMarin2007; @RothbergJoseph2004] confirm that Mrk 334 is a starburst galaxy. This explains why in the ionization diagrams the part of the data points that belong to the nucleus lie in the domain corresponding to the ionization by young stars.
Such a powerful burst of star formation in a rather compact region produces a hot gas superwind. Low-density gas heated by frequent supernova explosions forces its way through the dense and cold gas of the disk to form a wide-cone outflow in the direction perpendicular to the galaxy plane. Superwind is usually most conspicuous in edge-on galaxies with the cone of hot gas fully open toward the observer. However, Mrk 334 has a less convenient orientation and the cone is seen projected against the bright disk of the galaxy and hence the conclusions about the presence of a superwind are to be based on the circumstantial evidences. Firstly, negative \[OIII\]-line velocities are observed toward the nucleus of the galaxy (the base of the outflow) suggesting outward motions of highly ionized gas with velocities of $200-300{\,\mbox{km}\,\mbox{s}^{-1}}$ perpendicular to the galactic plane. Also, the $150-180{\,\mbox{km}\,\mbox{s}^{-1}}$ motions are observed in the NaD absorption line. The velocities are larger in the high-excited gas than in neutral medium, which is typical for galactic winds [@Veilleux2005].
A second indication of the superwind is provided by the observed asymmetry of the X-ray flux distribution with respect to the nucleus.
Interaction-related processes become important as close to the nucleus as at the distances of $1-2$ kpc from it. We see their footprints as Regions ‘B’ and ‘C’ and a system of tidal arcs and envelopes extending out to galactocentric distances of $40$ kpc. Fig. \[[f13]{}\] shows schematically the inner region of the galaxy. An analysis of the kinematics of gas and stars led us to conclude that the orbits of the debris of the disrupted companion lie outside the disk of Mrk 334 and cross it at a considerable inclination. In the region of this cross-point we observe a cavern with gas density lower than that of the ambient surrounding medium (Region ‘B’) which has formed as a result of the crossing of the disk by a dense stellar condensation. The rotation velocity at this location is $200-250{\,\mbox{km}\,\mbox{s}^{-1}}$, implying that the fragments of the companion also move with the velocities of the same order relative to the gaseous disk. Also, we find evidence of a high velocity collision in the gas kinematics and ionization. The collision has strongly perturbed the velocity field in the $\sim 3$ kpc size region. The maximum amplitude of the line-of-sight velocity perturbations amounts to $70{\,\mbox{km}\,\mbox{s}^{-1}}$ in low-excitation lines and reaches $150{\,\mbox{km}\,\mbox{s}^{-1}}$ in the \[OIII\], because in this line we see low-density gas ionized by a powerful shock. The ionization state in Region ‘B’ can be described in terms of the shock+precursor model for a shock propagating at a speed of $250-350{\,\mbox{km}\,\mbox{s}^{-1}}$. The mutual agreement of all the three estimates for the collision velocity supports the adopted interpretation of the formation of Region ‘B’.
Unfortunately, in the literature we have not found any detailed simulations of the gaseous disk response to the intrusion of a self-gravitating body whose mass is small compared to that of the entire galaxy. As a close analogy, we can mention the paper by @Levy2000, that briefly analysed the response of a gaseous disk of a galaxy to the crossing by a globular cluster. Even in such a relatively small-scale collision the resulting shock propagates to at least five to six vertical disk scale heights.
We believe that the remnant of the companion galaxy now observed as Region ‘C’ is the most likely candidate object to have punched the disk of Mrk 334 and produced there a cavern of hot gas. The line-of-sight velocities of the stellar population associated with Region ‘C’ differ appreciably (by almost $100{\,\mbox{km}\,\mbox{s}^{-1}}$) from the velocities of the old stellar population in the galactic disk. How is Region ‘C’ located with respect to the observer? Residual velocities of ionized gas in the cavern (Region ‘B’) are mostly negative. This means that the body that produced the cavern crossed the disk plane moving toward us and must now be located above the galaxy disk with respect to the observer (Fig. \[[f13]{}\]). Residual velocities of the young stellar population in Region ‘C’ are also negative implying that the nucleus of the companion traversed less than a quarter of its orbit after the collision. In accordance with the rotation curve we estimate the dynamical age of the cavern in Region ‘B’ as $t\leq \cdot10^7$ yr. Therefore, we indeed deal with a recent collision and the cavern has not yet cooled down or condensed.
Our photometric analysis of the tidal filaments yields a mass ratio for the interacting galaxies ranging from $1/5$ to $1/3$. The merger must have occurred between two gas-rich galaxies. Indeed, the total luminosity of nonaxisymmetric features in the optical images is $25-30\%$ of the total luminosity of the galaxy. @Hopkins2008 suggest that the ‘excess flux’ is even greater in the K band where its contribution amounts to $45\%$. According to @Hopkins2008, such a high percentage of the excess flux can be reproduced in the model of gas-rich galaxies merging, and the process results in the formation of a LIRG galaxy, just as in the case of Mrk 334.
In Mrk 334 we appear to observe a transition from the LIRG stage to the phase of nuclear activity. The fact that the activity of the nucleus of the galaxy has started only recently is proved by weak water-ice absorptions in the infrared spectra that are indicative of a certain well-defined phase in the evolution of the object [@Spoon2002].
Thus, both the burst of star formation and nuclear activity in Mrk 334 date back to a rather recent epoch and their age is comparable to the dynamic time scale of the interaction. @Li2008 pointed out, by references to Yuan et al.(in preparation), that ULIRGs experienced the stage of ‘diffuse merger’ when the nuclei of the interacting galaxies already merge together but have not yet formed a single nucleus. Composite activity — active nucleus + star formation — intensifies abruptly during this stage, and this appears to be now the case in Mrk 334.
![Sketch of the proposed model describing the spatial structure of the inner ($r<5$ kpc) region of Mrk 334. The [*HST*]{} image from Fig. \[f05\] is projected onto the plane of the galactic disk[]{data-label="f13"}](Smirnova_fig13.eps){width="8cm"}
Conclusions. {#sec7}
============
We used 3D spectroscopic data and deep images to explore the structure and kinematics of the galaxy Mrk 334. The galaxy has a composite (AGN+starburst) nucleus and extended tidal structures in the form of loops and arcs observed at galactocentric distances ranging from 2 to 40 kpc. Extensive spectroscopic and photometric data allowed us to thoroughly analyse the structure of the inner regions of the galaxy. We consider the following points to be the most important:
1. The main galaxy-to-companion mass ratio is about $3$ to $5$. The average surface-brightness profile shows a multi-tired structure and can be decomposed into a bulge and two exponential disks. We caught the galaxy in the process of the formation of an outer low-surface-brightness disk from the debris of the disrupted companion.
2. The central region of the galaxy demonstrates a powerful starburst that must have been triggered by the galaxy merger. Circumnuclear star formation is so intense that its contribution to the total ionization of gas exceeds that of the active nucleus. As a result, the corresponding data points in the diagnostic diagrams lie in the HII/LINER domain. Such a powerful burst of star formation in a compact region gives rise to a superwind with velocities of $200-300{\,\mbox{km}\,\mbox{s}^{-1}}$. The asymmetric X-ray brightness distribution on *ROSAT* maps is consistent with this hypothesis.
3. We revealed a region $\sim2$ kpc east of the centre that is possibly the nucleus of a disrupted companion. The spectrum of this region exhibits stellar absorptions that are typical of a region that has undergone a burst of star formation about one Gyr ago. The radial velocities of the stars located in this region differ appreciably from the stellar disk of Mrk 334 onto which it is projected. In the disk of the galaxy we found a cavern filled with low-density ionized gas. We interpret this region as a site of a recent (about 10 Myr ago) crossing of the gaseous disk by the remnants of the disrupted companion. This supposition allows us to explain the unusually high \[OIII\]$/$H$\beta$ line ratio observed in this region that can be produced by a powerful shock propagating with a velocity of more than 250 ${\,\mbox{km}\,\mbox{s}^{-1}}$. The non-circular gas motions in this region agree with the crossing of the galaxy disk by the debris of the companion.
This work is based on observations made with the 6-m telescope of the Special Astrophysical Observatory of the Russian Academy of Sciences operated under the financial support of the Ministry of Science of the Russian Federation (Registration Number 01-43). This research has made use of the NASA/IPAC Extragalactic Database (NED) operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. The research is partly based on data obtained from the Multimission Archive at the Space Telescope Science Institute (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. We are grateful to Evgenii Churazov for his assistance in discussing the X-ray data and to Olga Sil’chenko, Victor Afanasiev and Natalia Sotnikova for useful comments; and also to our anonymous referee for his constructive advice that helped us to improve the paper. This work was supported by the Russian Foundation for Basic Research (project no. 09-02-00870). AM is also grateful to to the ‘Dynasty’ Fund.
Afanasiev V. L., Dodonov S. N., Moiseev A. V., 2001, in Ossipkov L. P., Nikiforov I. I., eds, Stellar Dynamics: from Classic to Modern, Saint Petersburg, p. 103
Afanasiev V. L., Moiseev A. V., 2005, Astronomy Letters, 31, 193; astro-ph/0502095
Allen M .G., Groves B .A., Dopita M .A., Sutherland R .S., Kewley L .J., 2008, ApJS, 178, 20
Barnes J. E., Hernquist L. E., 1991, ApJ, 370, L65
Barth A. J., Bentz M. C., Greene J. E., Ho L. C., 2008, ApJ, 683, L119
Bekki K., Shioya Y., Whiting M., 2006, MNRAS, 371, 805
Capetti A., Axon D. J., Macchetto F. D., Marconi A., Winge C., 1999, ApJ, 516, 187
Chilingarian I., Prugniel Ph., Sil’chenko O., Koleva M., 2007, in Vazdekis A., Peletier R. F., eds., Stellar Populations as Building Blocks of Galaxies, Proceedings of IAU Symposium, 241, Cambridge University Press, p.17
Ciroi S., Afanasiev V. L., Moiseev A. V., Botte V., Di Mille F., Dodonov S. N., Rafanelli P., Smirnova A. A., 2005, MNRAS, 360, 253
Dahari O., 1985, AJ, 90, 1772
De Robertis M. M., Yee H. K. C., Hayhoe K., 1998, ApJ, 496, 93
Deo R. P., Crenshaw D. M., Kraemer S. B., 2006, AJ, 132, 321
Erwin P., Beckman J. E., Pohlen M., 2005, ApJ, 626, L81
Ferruit P., 2002, RevMexAA (Serie de Conferecias), 13, 183
Garc[í]{}a-Lorenzo B., Arribas S., Mediavilla, E., 2001, A&A, 787, 2001
Gonzalez Delgado R. M., Perez E., Tadhunter C., Vilchez J. M., Rodriguez-Espinosa J. M., 1997, ApJS, 108, 155
Heckman T. M., Armus L., Miley G. K., 1990, ApJS, 74, 833
Hopkins P. F., Hernquist L., Cox T. J., Dutta S. N., Rothberg B., 2008, ApJ, 679, 156
Hunt L. K., Malkan M. A., Salvati M., Mandolesi N., Palazzi E., Wade R., 1997, ApJS, 108, 229
Keel W. C., 1996, AJ, 111, 696
Kennicutt Jr. R. C., 1998a, ARA&A, 36, 189
Kennicutt Jr. R. C., 1998b, ApJ, 498, 541
Knapen J. H., 2005, Ap&SS, 295, 85
Koleva M., Prugniel Ph., Bouchard A., Wu Y., 2009, A&A, in press, arXiv:0903.2979
Komossa S., Xu D., Zhou H., Storchi-Bergmann T., Binette L., 2008, ApJ, 680, 926
Levy V. V., 2000, A&A Transactions, 18, 621
Li C., Kauffmann G., Heckman T. M., White S. D. M., Jing Y. P., 2008, MNRAS, 385, 1915
Lumsden S. L., Heisler C. A., Bailey J. A., Hough J. H., Young S., 2001, MNRAS, 327, 459
Maiolino R., Ruiz M., Rieke G. H., Papadopoulos P., 1997, ApJ, 485, 552
Martini P., 2004, in Storchi-Bergmann T., Ho L. C., Schmitt H. R., eds, The Interplay Among Black Holes, Stars and ISM in Galactic Nuclei Vol. 222 of IAU Symposium, p. 235
Martini P., Pogge R. W., 1999, AJ, 118, 2646
Martini P., Pogge R. W., Ravindranath S., An J. H., 2001, ApJ, 562, 139
Moiseev A. V., 2001, Bull. Spec. Astrophys. Obs., 51, 11; astro-ph/0111219
Moiseev A. V. 2002, Bull. Spec. Astrophys. Obs., 54, 74; astro-ph/0211104
Moiseev A. V., Egorov O. V., 2008, Astrophys Bull., 63, 193; arXiv: 0805.2367 \[astro-ph\]
Moiseev A. V., Vald[é]{}s J. R., Chavushyan V. H., 2004, A&A, 421, 433
Mu[ñ]{}oz Mar[í]{}n V. M., Gonz[á]{}lez Delgado R. M., Schmitt H. R., Cid Fernande R., P[é]{}rez E., Storchi-Bergmann T. et al., 2007, AJ, 134, 648
Osterbrock D. E., 1989, Astrophysics of Gaseous Nebulae and Active Galactic Nuclei. University Science Books, Mill Valley, CA
P[é]{}rez Garc[í]{}a A. M., Rodr[í]{}guez Espinosa J. M., 2001, ApJ, 557, 39
Rothberg B., Joseph R. D., 2004, AJ, 128, 2098
S[á]{}nchez-Bl[á]{}zquez P., Peletier R. F., Jim[é]{}nez-Vicente J., Cardiel N., Cenarro A. J. et al., 2006, MNRAS, 371, 703
Schmitt H. R., 2001, AJ, 122, 2243
Schweizer F., Seitzer P., 1988, ApJ, 328, 88
Sil’chenko O. K., Moiseev A. V., 2006, AJ, 131, 1336
Sil’chenko O. K., Vlasyuk V. V., Alvarado F., 2001, AJ, 121, 2499
Shalyapina L.V., Moiseev A. V., Yakovleva V. A., Hagen-Thorn V. A., Burenkov A. N., 2004, Astronomy Letters, 30, 1
Smirnova A. A., Moiseev A. V., Afanasiev V. L., 2006, Astronomy Letters, 32, 520; astro-ph/0607163
Smirnova A. A., Gavrilovi[ć]{} N., Moiseev A. V., Popovi[ć]{} L. [Č]{}., Afanasiev V. L., Jovanovi[ć]{} P., Da[č]{}i[ć]{} M., 2007, MNRAS, 377, 480
Spoon H. W. W., Keane J. V., Tielens A. G. G. M., Lutz D., Moorwood A. F. M., Laurent O., 2002, A&A, 385, 1022
Springel V., Di Matteo T., Hernquist L., 2005, ApJ, 620, L79
Ulvestad J. S., 1986, ApJ, 310, 136
Veilleux S., Cecil G., Bland-Hawthorn J., 2005, ARA&A, 43, 769
Veilleux S., Osterbrock D. E., 1987, ApJS, 63, 295
Vorontsov-Vel’yaminov B. A., 1977, A&AS, 28, 1
Wehner E. H., Gallagher J. S., 2005, ApJ, 618, L21
Zimmermann H.-U., Boller T., D[ö]{}bereiner S., Pietsch W., 2001, A&A, 378, 30
[^1]: [email protected]
[^2]: ULYSS is an open code located at and based on the papers of @Koleva2009, @Chil2007 and other.
[^3]: The fluxes reported by @Lumsden2001 are extinction corrected, but, as we point out above, this correction is negligible for the diagrams
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Our previous analysis indicates that small-scale fluctuations in the intracluster medium (ICM) from cosmological hydrodynamic simulations follow the lognormal probability density function. In order to test the lognormal nature of the ICM directly against X-ray observations of galaxy clusters, we develop a method of extracting statistical information about the three-dimensional properties of the fluctuations from the two-dimensional X-ray surface brightness.
We first create a set of synthetic clusters with lognormal fluctuations around their mean profile given by spherical isothermal $\beta$ models, later considering polytropic temperature profiles as well. Performing mock observations of these synthetic clusters, we find that the resulting X-ray surface brightness fluctuations also follow the lognormal distribution fairly well. Systematic analysis of the synthetic clusters provides an empirical relation between the three-dimensional density fluctuations and the two-dimensional X-ray surface brightness.
We analyze [[*Chandra*]{} ]{}observations of the galaxy cluster Abell 3667, and find that its X-ray surface brightness fluctuations follow the lognormal distribution. While the lognormal model was originally motivated by cosmological hydrodynamic simulations, this is the first observational confirmation of the lognormal signature in a real cluster.
Finally we check the synthetic cluster results against clusters from cosmological hydrodynamic simulations. As a result of the complex structure exhibited by simulated clusters, the empirical relation between the two- and three-dimensional fluctuation properties calibrated with synthetic clusters when applied to simulated clusters shows large scatter. Nevertheless we are able to reproduce the true value of the fluctuation amplitude of simulated clusters within a factor of two from their two-dimensional X-ray surface brightness alone.
Our current methodology combined with existing observational data is useful in describing and inferring the statistical properties of the three dimensional inhomogeneity in galaxy clusters.
author:
- 'Hajime Kawahara, Erik D. Reese, Tetsu Kitayama, Shin Sasaki, and Yasushi Suto'
title: 'Extracting Galaxy Cluster Gas Inhomogeneity from X-ray Surface Brightness: A Statistical Approach and Application to Abell 3667'
---
Introduction {#sec:intro}
============
Galaxy clusters have been one of the most important probes of cosmology (e.g., Bartlett & Silk 1994; Eke, Cole, & Frenk 1996; Viana & Liddle 1996; Kitayama & Suto 1996, 1997; Kitayama, Sasaki, & Suto 1998; Holder et al. 2000; Haiman, Mohr, & Holder 2001; Majumdar & Mohr 2004). In the context of dark energy surveys, which attracts much of the attention of the cosmology and particle physics communities, galaxy cluster surveys are also unique in that they most directly probe the growth of structure rather than relying solely on distance measurements [e.g., @albrecht2006]. In order to capitalize on this, galaxy cluster surveys, in particular those utilizing the Sunyaev-Zel’dovich effect [for reviews see, for example, @carlstrom02; @birkinshaw99; @rephaeli95; @sunyaev80], are currently operating and many more are planned in the near future. However, one of the biggest challenges in interpreting these surveys is relating physical quantities of galaxy clusters, namely mass, to observable ones. In particular, these mass-observable relations may be sensitive to the inherent complex structure of clusters. Therefore we must better understand galaxy clusters to utilize fully the potential of galaxy cluster surveys in constraining cosmological parameters.
Recent observations of galaxy clusters have revealed a rich variety of structural complexity. Recent X-ray satellites with their improved angular resolution, collecting area, and simultaneous spectral measurement capabilities have unveiled complex temperature structure [e.g., @markevitch2000; @furusho01], shock fronts [e.g., @jones2002], cold fronts [e.g., @markevitch2000], and X-ray holes [e.g., @fabian2002]. Improved observational strategies and analysis methods of lensing observations of galaxy clusters show that the mass distribution, as opposed to just the gas, is often complicated as well [e.g., @bradac06]. Both X-ray and lensing observations of galaxy clusters reveal that clusters are frequently undergoing mergers [e.g., @briel04]. With such various and sundry structural complexities may galaxy clusters reliably be used as cosmological probes ?
The complex structure seen in galaxy clusters motivates our investigation of the intracluster medium (ICM) inhomogeneity. We note, however, that we take a statistical approach to modeling the inhomogeneity rather than directly modeling such complex phenomena as shocks, cold fronts, etc. Motivated by results from cosmological hydrodynamic simulations we explore the ramifications of a lognormal model of the inhomogeneity of the ICM. This model was first proposed (Kawahara et al. 2007, hereafter Paper I; Kawahara et al. 2008), in this context, to explain the discrepancies between emission weighted and spectroscopic temperature estimates from galaxy clusters [@mazzotta04; @rasia05; @vikhlinin06]. They found that local inhomogeneities of the ICM play an essential role in producing the systematic bias between spectroscopic and emission weighted temperatures.
Thus far, the lognormal model has been motivated by and applied only to clusters from cosmological hydrodynamic simulations. Therefore it is crucial to see if inhomogeneities in real galaxy clusters also show the lognormal signature. In reality, this is not a straightforward task since one can observe clusters in X-rays only through their projection over the line of sight. Thus we develop a method of extracting statistical information of the three-dimensional properties of fluctuations from the two-dimensional X-ray surface brightness.
The rest of the paper is organized as follows. We first summarize the log-normal model in §\[sec:model\]. We create synthetic clusters to explore the relationship between the intrinsic cluster inhomogeneity and X-ray observables in §\[sec:synthetic\]. In §\[sec:obs\] we apply our methodology to [[*Chandra*]{} ]{}observations of the galaxy cluster Abell 3667, and then attempt to quantify the nature of cluster inhomogeneity. We also compare our synthetic cluster results with cosmological hydrodynamic simulations in §\[sec:con\]. Finally, we summarize our results in §\[sec:sum\]. Throughout the paper the Hubble constant is parameterized by $h$ in the usual way, $H_0 = 100\, h$ km s$^{-1}$ Mpc$^{-1}$.
Model of the ICM Inhomogeneity {#sec:model}
==============================
Lognormal Distribution {#sec:analytic}
----------------------
In order to characterize the inhomogeneity of the ICM, we define the density and temperature fluctuations as the ratios ${\mbox{$\delta_n$}}\equiv
n({{\bf r}})/\overline{n}(r)$ and ${\mbox{$\delta_T$}}\equiv T({{\bf r}})/\overline{T}(r)$, where $n({{\bf r}})$ and $T({{\bf r}})$ are the local density and temperature at radius ${{\bf r}}$, and $\overline{n}(r)$ and $\overline{T}(r)$ are the angular average profiles defined by $$\begin{aligned}
\label{eq:averagen}
\overline{n}(r) &\equiv& \frac{1}{4 \pi}
\int n(r,\theta,\phi) \sin\theta\; d\theta\; d\phi\\
\label{eq:averageT}
\overline{T}(r) &\equiv& \frac{1}{4 \pi} \int T(r,\theta,\phi)
\sin\theta\; d\theta\; d\phi\end{aligned}$$ where $\theta$ and $\phi$ are polar and azimuthal angles, respectively. Analysis of hydrodynamical simulations (Paper I) found that ${\mbox{$\delta_n$}}$ and ${\mbox{$\delta_T$}}$ are approximately independent and follow the radially independent lognormal probability density function (PDF) given by $$p({\delta_x}; {\sigma_{\mathrm{LN}, \, x}}) \, d {\delta_x}= \frac{1}{\sqrt{2 \pi} {\sigma_{\mathrm{LN}, \, x}}}
\exp{\left[ \frac{-\left(\log{{\delta_x}}+{\sigma_{\mathrm{LN}, \, x}}^2/2 \right)^2}{2 {\sigma_{\mathrm{LN}, \, x}}^2}
\right]} \, \frac{d {\delta_x}}{{\delta_x}},
\label{eq:pdf_delta}$$ where $x$ denotes $n$ or $T$, ${\delta_x}\equiv x({{\bf r}})/\overline{x}(r)$, and ${\sigma_{\mathrm{LN}, \, x}}$ is the standard deviation of the logarithm of density or temperature.
To construct the two-dimensional surface brightness profile from the three-dimensional density and temperature distribution, we also need the properties of the power spectra of the density and temperature fluctuations. We adopt statistically isotropic fluctuations with a power-law type power spectrum for both the density fluctuations $P_{n}(k) \propto k^{{\alpha_{n}}}$ and the temperature fluctuations $P_{T}(k) \propto k^{{\alpha_{T}}}$. These assumptions are based on the results of the cosmological hydrodynamic simulations described in §\[subsec:hydro\_sim\].
We use this model to generate synthetic clusters to explore the relationship between the three-dimensional inhomogeneity in the ICM and the two-dimensional X-ray surface brightness.
Cosmological Hydrodynamic Simulated Clusters {#subsec:hydro_sim}
--------------------------------------------
When one considers the projection of galaxy clusters to two dimensions for mock X-ray observations, the power spectrum of the fluctuations is important in addition to the PDF of the inhomogeneity. Here, we once again turn to simulations to investigate the power spectrum of the fluctuations.
We extract the six massive clusters from cosmological hydrodynamic simulations of the local universe performed by @dolag05. The simulations utilize the smoothed particle hydrodynamic (SPH) method, and assume a flat $\Lambda $ CDM universe with $\Omega_m=0.3, \Omega_b=0.04,
\sigma_8=0.9$, and a dimensionless Hubble parameter $h=0.7$. The number of dark matter and SPH particles is $\sim 20$ million each within a high-resolution sphere of radius $\sim 110 $ Mpc, which is embedded in a periodic box $\sim 343$ Mpc on a side that is filled with nearly seven million low-resolution dark matter particles. The simulation is designed to reproduce the matter distribution of the local universe by adopting the initial conditions based on the [*IRAS*]{} galaxy distribution smoothed over a scale of $4.9 h^{-1} \mathrm{Mpc}$. Thus, the six massive clusters are identified as Coma, Perseus, Virgo, Centaurus, A3627, and Hydra. A cubic region with 6 $h^{-1}$ Mpc on a side centered on each simulated cluster is extracted and divided into $512^3$ cells. The density and temperature of each mesh point are calculated from SPH particles using the B-spline smoothing kernel. A detailed description of this procedure is given in Paper I. The distance between two adjacent grid points is given by $ {d_{\mathrm{grid}}}= 6 h^{-1}\mathrm{Mpc}/512
\sim 12 h^{-1}$ kpc, which is comparable to the gravitational force resolution (14 kpc) and the inter-particle separation reached by SPH particles in the dense centers of clusters. Therefore, the (maximum) resolution is ${d_{\mathrm{grid}}}/{r_{\mathrm{c}}}\approx 0.1$ assuming ${r_{\mathrm{c}}}\sim 100$ kpc. This is about one order of magnitude worse than that of both the synthetic clusters (§ \[sec:synthetic\]) and the observational data (§ \[sec:obs\]).
For each simulated cluster, we compute the radially averaged density and temperature profiles, $\overline{n}(r)$ and $\overline{T}(r)$, respectively (Eqn. \[\[eq:averagen\]\] and \[\[eq:averageT\]\]), and use them to compute the density and temperature fluctuations ${\mbox{$\delta_n$}}=
n/\overline{n}$ and ${\mbox{$\delta_T$}}= T/\overline{T}$ at each grid point. We extract $128^3$ cells of ${\mbox{$\delta_n$}}$ and ${\mbox{$\delta_T$}}$ around the center of a simulated cluster and compute the power spectrum. The distance from the center to the corner of the $128^3$ cells is $\sim 1.3\ h^{-1}$ Mpc which is approximately equal to the virial radius of the simulated clusters ($r_{\mathrm{200}} = 1.0$-$1.6\ h^{-1}$ Mpc). The virial radius, $r_{\mathrm{200}}$, is the radius within which the mean interior density is 200 times that of the critical density.
Figure \[fig:chspower\] shows the power spectra for each simulated cluster for both ${\mbox{$\delta_n$}}$ (upper panel) and ${\mbox{$\delta_T$}}$ (lower panel). In each panel a simple power law, $P(k) \propto k^{-3}$ (dotted line), is also plotted for comparison. The power spectra for both the density and temperature are relatively well approximated by a single power law. We therefore adopt a power-law spectral model for the density and temperature fluctuations for the synthetic cluster analysis.
![The power spectra of ${\mbox{$\delta_n$}}$ (upper) and ${\mbox{$\delta_T$}}$ (lower) of the six simulated clusters. Dashed lines indicate $P(k) \propto
k^{-3}$. \[fig:chspower\]](f1.ps){width="85mm"}
Synthetic Clusters {#sec:synthetic}
==================
Cosmological hydrodynamic simulations provide a useful test-bed for exploring cluster structure. Simulated clusters exhibit complex density and temperature structure akin to that of real galaxy clusters. The resolution of our current simulations, however, is limited, especially when compared to the resolution available from current generation X-ray satellites. In addition, we need to systematically survey the parameter space of ${\sigma_{\mathrm{LN}, \, n}}$ and $\alpha_n$ in order to relate the X-ray surface brightness fluctuations to the density fluctuations. Thus we create a set of synthetic clusters at higher resolution that have lognormal fluctuations around their mean profile. Analysis of mock observations of these synthetic clusters enables us to investigate the relation between the X-ray surface brightness and the statistical properties of the three-dimensional density fluctuations, namely ${\sigma_{\mathrm{LN}, \, n}}$ and $\alpha_n$.
Method \[ssec:method\]
----------------------
### Synthetic Cluster Generation {#sss:syn_cl_gen}
The three-dimensional synthetic clusters will be projected to two dimensions when considering the X-ray surface brightness. In order to incorporate a power-law type power spectrum of spatial fluctuations into the synthetic clusters, we follow a similar methodology as that of several studies of the interstellar medium [@Elmegreen02; @FD04]. First a Gaussian random field with a power-law power spectrum is constructed and that field is mapped into a lognormal field. Therefore, our assumption for the power spectrum is adopted for the Gaussian field $q$ as opposed to $\delta_n$. However, we will verify that the ensemble average of the power spectra of $q$ and ${\mbox{$\delta_n$}}$ ($ P_q(k) \propto
k^{{\alpha_q}}$ and $ P_{n}(k) \propto k^{{\alpha_{n}}}$) have almost the same power-law indices, ${\alpha_q}\sim {\alpha_{n}}$.
We generate the lognormal density fluctuation field as follows. We first generate the real random fields, $a({\bf k})$ and $b({\bf k})$, in $k$-space, whose distribution functions obey $$\label{eq:deviate1}
{p}(a)da = \frac{1}{ \sqrt{\pi f(k)} }
\exp{\left[-\frac{a^2}{f (k)}\right]} da,
\quad
{p}(b)db = \frac{1}{ \sqrt{\pi f(k)} }
\exp{\left[-\frac{b^2}{f (k)}\right]} db,$$ where $f(k) \equiv A k^{{\alpha_q}}$. Then we compute $q({\bf r})$, the Fourier transform of a complex field $\tilde{q} ({\bf k}) \equiv a({\bf
k}) + i b({\bf k})$. With the additional conditions $a({\bf k})=a(-{\bf
k})$ and $b({\bf k})=-b(-{\bf k})$, $q({\bf r})$ becomes a real Gaussian random field, and its power spectrum, $P_q(k)$, is equal to the input function $f(k) \equiv A k^{{\alpha_q}}$. The amplitude $A$ is related to the variance of the Gaussian random field: $$\sigma_g^2 \equiv
4 \pi \int_{k_{\rm min}}^{k_{\rm max}} k^2 f(k) dk,$$ where $k_{\rm min}$ and $k_{\rm max}$ denote the minimum and maximum value of the wavenumber. Finally the lognormal deviate, $\delta_{x}
({{\bf r}})$, is obtained from the Gaussian deviate, $q({{\bf r}})$, using the relation $$\delta_{x} ({{\bf r}}) =\exp{ \left( \frac{{\sigma_{\mathrm{LN}, \, x}}}{\sigma_g} q({{\bf r}}) -
\frac{{\sigma_{\mathrm{LN}, \, x}}^2}{2} \right)},
\label{eq:lognorm_deviate}$$ where ${\sigma_{\mathrm{LN}, \, x}}$ is the standard deviation of the lognormal field.
We construct synthetic clusters with average density given by the $\beta$ model and ${\mbox{$\delta_n$}}$ drawn from a lognormal distribution taking into account the power-law type power spectrum of spatial fluctuations. The $\beta$ model is given by [@cavaliere1976; @cavaliere1978] $$\overline{n}(r) = n_0 \left[ 1 + \left( \frac{r}{{r_{\mathrm{c}}}} \right)^2
\right]^{-3 \beta / 2},
\label{eq:beta_model}$$ where $n_0$ is the central electron number density, ${r_{\mathrm{c}}}$ is the core radius, and $\beta$ specifies a power-law index. For simplicity, we first adopt a fiducial value of $\beta=2/3$, and assume isothermality for the synthetic clusters. Later, we examine the effects of varying $\beta$ (§ \[sss:vary\_beta\]) and of temperature structure using a polytropic temperature profile (§ \[ss:tstruct\]).
The density at an arbitrary point is given by $$n ({{\bf r}}) = \delta_{n} \overline{n} (r) .
\label{eq:n_ijk}$$ The X-ray surface brightness profile is obtained by projecting the three-dimensional synthetic cluster down to two dimensions. For the isothermal case the projected X-ray surface brightness profile is $${S_{\mathrm{X}}}({{\bf R}}) \propto \int [n({{\bf r}})]^2 d l,
\label{eq:sx_jk}$$ where ${{\bf R}}$ indicates the position on the projected plane and $l$ is the projection of ${{\bf r}}$ onto the line of sight direction.
![The change of the power-law spectral index of the density (${\alpha_{n}}$) and density squared fields (${\alpha_{nn}}$) compared to that of the Gaussian field (${\alpha_q}$). Solid and dashed lines indicate ${\alpha_{n}}/{\alpha_q}-1$ (density) and ${\alpha_{nn}}/{\alpha_q}-1$ (density squared), respectively. Each symbol indicates a different value of ${\sigma_{\mathrm{LN}, \, n}}$ (cross, square, and triangle correspond to ${\sigma_{\mathrm{LN}, \, n}}=0.1$, $0.3$, and $0.5$,respectively.) The power-law index of the density field is very close ($\lesssim 3$%) to that of the Gaussian field used to generate the lognormal distribution and that of the square of the density is within $\sim 13$% for larger values of ${\sigma_{\mathrm{LN}, \, n}}$ and $\lesssim 5$% for smaller values (${\sigma_{\mathrm{LN}, \, n}}\lesssim
0.3$). []{data-label="fig:ipchange1"}](f2.ps){width="80mm"}
Performing the procedure described above, we set up a cubic mesh of $n({{\bf r}})$ in which our three-dimensional synthetic cluster is located with ${N_{\mathrm{grid}}}=512$ grid points along each axis. We choose the box size ${L_{\mathrm{box}}}= 10 \, {r_{\mathrm{c}}}$, which results in the distance between two adjacent grid points being ${d_{\mathrm{grid}}}=10 \, {r_{\mathrm{c}}}/ {N_{\mathrm{grid}}}\sim 0.02 \, {r_{\mathrm{c}}}$.
We fit the power spectrum of the ${\mbox{$\delta_n$}}$ field by a power-law spectrum so that $P_{n}({{\bf k}}) \propto k^{{\alpha_{n}}}$. We also fit the power spectrum of the square density field, ${\delta_{nn}}\equiv n^2/{\langle}n^2 {\rangle}= {\mbox{$\delta_n$}}^2
\exp{(-{\sigma_{\mathrm{LN}, \, n}}^2)}$ (Appendix B), by the power-law $P_{nn}(k) \propto
k^{{\alpha_{nn}}}$, relevant to X-ray surface brightness since ${S_{\mathrm{X}}}\propto
\int d\ell \; n^2$. Throughout this paper, the notation $\langle x \rangle$ is used to denote the ensemble average of quantity $x$ over many clusters.
Figure \[fig:ipchange1\] shows the change of the power-law spectral index of the density (${\alpha_{n}}$) and density squared fields (${\alpha_{nn}}$) compared to that of the Gaussian field (${\alpha_q}$). The change in the power-law index for the density and density squared distributions compared to the initial Gaussian field are small ($<$3% and $<$ 13%, respectively), and therefore, ${\alpha_q}\sim {\alpha_{n}}\sim {\alpha_{nn}}$, consistent with the results of [@FD04].
### X-ray Surface Brightness {#ss:em}
To quantify the relationship between the inhomogeneity of the density and the X-ray surface brightness, ${S_{\mathrm{X}}}$, we introduce the X-ray surface brightness fluctuation from the average radial surface brightness profile ${\overline{S}_{\mathrm{X}}}(R)$ $${\mbox{$\delta_\mathrm{Sx}$}}({{\bf R}}) \equiv \frac{{S_{\mathrm{X}}}({{\bf R}})}{{\overline{S}_{\mathrm{X}}}(R)},
\label{eq:dsx}$$ where $R \equiv |{{\bf R}}|$. We define the average profile ${\overline{S}_{\mathrm{X}}}(R)$ for an individual cluster by fitting the projected synthetic clusters to an isothermal $\beta$ model $${\overline{S}_{\mathrm{X}}}(R) = S_{\mathrm{X},0} \left[ 1 +
\left(\frac{R}{{r_{\mathrm{c,X}}}}\right)^2 \right]^{-3{\beta_{\mathrm{X}}}+1/2},
\label{eq:ave1}$$ where $S_{\mathrm{X},0}$ is the central X-ray surface brightness, ${r_{\mathrm{c,X}}}$ is the core radius, and ${\beta_{\mathrm{X}}}$ specifies the power-law index for the X-ray surface brightness distribution. These three parameters are derived from a model fit to each synthetic cluster. It is important to emphasize that the average in equation (\[eq:ave1\]) is defined for [*an individual cluster*]{}. We note that if we adopt directly the average X-ray surface brightness profile instead of a $\beta$ model fit (Eqn. \[\[eq:ave1\]\]), the results are unchanged. This is because the radial profile is well approximated by the $\beta$ model for the synthetic clusters. However, for observations of real galaxy clusters, the $\beta$ model approximation might break down and one should instead use an average of ${S_{\mathrm{X}}}({{\bf R}})$ directly in such cases. In §3.2, we will investigate the relation between the standard deviation of the X-ray surface brightness fluctuations, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$, and that of the intrinsic density fluctuations, ${\sigma_{\mathrm{LN}, \, n}}$.
Here, we consider the relation of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\sigma_{\mathrm{LN}, \, n}}$ for the [*ensemble average*]{} of clusters assuming they all obey the $\beta$ model with the same $\beta$, ${r_{\mathrm{c}}}$, $\alpha_q$ and ${\sigma_{\mathrm{LN}, \, n}}$: $$\begin{aligned}
{{\langle}S_{\mathrm{X}} {\rangle}}(R) &\equiv& {\langle}{S_{\mathrm{X}}}(|{{\bf R}}|) {\rangle}\label{eq:ens1}\\
{\langle}{S_{\mathrm{X}}}({{\bf R}}) {\rangle}&\propto& e^{{\sigma_{\mathrm{LN}, \, n}}^2} \int
\overline{n}^2 d l,
\label{eq:ens2}\end{aligned}$$ where the exponential term of the right hand side of equation (\[eq:ens2\]) comes from the second moment of the lognormal distribution (Paper I). Although the ensemble average is [*not*]{} an observable quantity, we can describe an analytical prediction of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}(R)$ assuming the isothermal $\beta$ model (Appendix \[a1:den\_sb\]). In addition, one expects that ${\overline{S}_{\mathrm{X}}}\sim {{\langle}S_{\mathrm{X}} {\rangle}}$ if there is a large enough volume compared with the size of fluctuations when calculating ${\overline{S}_{\mathrm{X}}}$. In other words, the spatial average approaches the ensemble average. For these reasons, it is useful to consider the ensemble average. Using equations (\[eq:ens1\]) and (\[eq:ens2\]), we define the ensemble average of fluctuations in the X-ray surface brightness as $${\delta_{\mathrm{Sx,ens}}}({{\bf R}}) \equiv \frac{{S_{\mathrm{X}}}({{\bf R}})}{{{\langle}S_{\mathrm{X}} {\rangle}}(R)}.
\label{eq:ensd}$$ We note that the distribution of the square of density fluctuations, which is proportional to the local emissivity in the isothermal case, is also distributed according to the lognormal function with a lognormal standard deviation of $2 {\sigma_{\mathrm{LN}, \, n}}$ if the density fluctuations follow the lognormal distribution with standard deviation ${\sigma_{\mathrm{LN}, \, n}}$ (Appendix \[sec:a2\_densquared\]).
Statistical Analysis of the Synthetic Clusters
----------------------------------------------
Here, we investigate the distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ of the synthetic clusters and relate quantities obtainable from observations, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\alpha_{\mathrm{Sx}}}$, to that of the underlying density, ${\sigma_{\mathrm{LN}, \, n}}$ and $\alpha_n$.
### Lognormal nature and the relation between ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\sigma_{\mathrm{LN}, \, n}}$ {#ss:synthetic_clusters}
![The probability distribution of the ensemble-averaged distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ illustrating the radial dependence. The distributions in shells of thickness $0.5 \, {r_{\mathrm{c}}}$ are shown. Each color indicates a different radial interval: $R<1.5 \, {r_{\mathrm{c}}}$ (red), $1.5 \, {r_{\mathrm{c}}}<
R < 3.5 \, {r_{\mathrm{c}}}$ (black), and $R>3.5 \, {r_{\mathrm{c}}}$ (blue). \[fig:shells\]](f3.ps){width="120.0mm"}
![The radial dependence of the standard deviations of the logarithm of X-ray surface brightness, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$. Two values of ${\sigma_{\mathrm{LN}, \, n}}$ are plotted, 0.1 and 0.5, as indicated in the figure. Solid and dotted lines show ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}(R)$ calculated using the average profile defined by the $\beta$ model (Eq. \[\[eq:ave1\]\]) and the ensemble average (Eq. \[\[eq:ens2\]\]), respectively. Dashed lines show the analytical prediction (Eq. \[\[eq:thickr\]\]) . Dash-dotted lines indicate the case including the temperature structure. Although we show results only for a single power-law index, ${\alpha_q}=-3.0$, similar results are obtained in other cases. \[fig:rdt\]](f4.ps){width="80mm"}
![The probability distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ for five individual synthetic clusters (solid) along with the best-fit lognormal distributions (dashed). Each color shows a different individual synthetic cluster. Each panel shows a different value of the power law index of the Gaussian field, ${\alpha_q}$, between $-2$ and $-4$ as indicated in each panel.[]{data-label="fig:ind"}](f5.ps){width="120.0mm"}
We investigate the distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ as a function of radial distance $R$ from the cluster center. We first divide the ${\mbox{$\delta_\mathrm{Sx}$}}$ field into shells of thickness $0.5 \, {r_{\mathrm{c}}}$. The distributions of ${\mbox{$\delta_\mathrm{Sx}$}}$ within each shell, $p({\mbox{$\delta_\mathrm{Sx}$}}; R)$, averaged over 256 synthetic clusters are shown in Figure \[fig:shells\] for various values of ${\alpha_q}$. We find that ${\mbox{$\delta_\mathrm{Sx}$}}$ also approximately follows the lognormal distribution. The standard deviation of the logarithm of ${\mbox{$\delta_\mathrm{Sx}$}}$ versus radius, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}(R)$, constructed from the averaged shells is displayed in Figure \[fig:rdt\]. Two values of ${\sigma_{\mathrm{LN}, \, n}}$ are plotted, 0.1 and 0.5, in addition to using the average profile defined by both the $\beta$ model (Eq. \[\[eq:ave1\]\]; solid) and that for the ensemble (Eq. \[\[eq:ens2\]\]; dotted). The analytic prediction (Eq. \[\[eq:thickr\]\]; dashed) and the case including the temperature structure (§\[ss:tstruct\]; dot-dashed) are also plotted. At large $R$, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx,ens}}}(R)$ is approximately ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}(R)$ because the spatial average tends to the ensemble average due to the large volume used for averaging. However, the agreement is poor near the center, where the ensemble average is not a good approximation. Although only one value for ${\alpha_q}$ is shown, similar results are obtained for other values.
Figures \[fig:shells\] and \[fig:rdt\] indicate that the probability density function is weakly dependent on the projected radius $R$. This radial dependence is caused mainly by two competing effects. Consider first the case where the typical nonlinear scale of fluctuations is much smaller than the size of the cluster itself (shallow spectrum). As equation (A1) indicates, the surface brightness at $R$ is given by $${S_{\mathrm{X}}}(R) \propto \int {\delta_{nn}}\left[1+ \left(\frac{l^2}{{r_{\mathrm{c}}}^2+R^2}\right)\right]^{-3\beta} dl.$$ This implies that the mean value of ${S_{\mathrm{X}}}(R)$ is effectively determined by the integration over the line of sight weighted towards the cluster center, roughly between $-\sqrt{{r_{\mathrm{c}}}^2+R^2}$ and $+\sqrt{{r_{\mathrm{c}}}^2+R^2}$. This is also true for the variance of ${S_{\mathrm{X}}}(R)$. Since the effective number of independent cells contributing to the variance of ${S_{\mathrm{X}}}(R)$ is smaller at smaller projected radii, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ slightly increases for smaller $R$. This explains the behavior of the shallow spectra results for $\alpha_q=-2$ and $-2.5$ in Figure \[fig:shells\]. On the contrary, if the typical nonlinear scale of fluctuations is comparable to or even larger than the cluster size (steep spectrum), the sampling at the central region significantly underestimates the real variance. So the ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ should increase toward the outer region. This is seen in Figure 3 for the steeper spectra, $\alpha_q=-3.5$ and $-4$.
Note the first effect is very small and the second effect becomes significant only when $\alpha_q < -3$. The cosmological hydrodynamic simulations imply that the typical value of $\alpha_q$ is $-3$. Therefore we neglect the radial dependence of the ${\mbox{$\delta_\mathrm{Sx}$}}$ field in the following analysis.
![The average of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ from the 256 synthetic cluster sample as functions of ${\alpha_q}$ (left) and ${\sigma_{\mathrm{LN}, \, n}}$ (right). The left panel also shows the standard deviation of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ from the 256 synthetic clusters and black, red, and blue represent different values of ${\sigma_{\mathrm{LN}, \, n}}$, $0.1$, $0.3$, and $0.5$, respectively. In both panels, symbols indicate values of ${\alpha_q}$ (cross, square, triangle, asterisk, and circle correspond to ${\alpha_q}= -2$, $-2.5$, $-3$, $-3.5$, and $-4$, respectively). Dashed lines show the best-fit approximately linear ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$-${\sigma_{\mathrm{LN}, \, n}}$ relation (Eq. \[eq:fitsKa\] and Eq. \[eq:Kalpha\]) for each pair of ${\sigma_{\mathrm{LN}, \, n}}$, ${\alpha_q}$. \[fig:mgsigout\] ](f6a.ps "fig:"){width="80mm"} ![The average of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ from the 256 synthetic cluster sample as functions of ${\alpha_q}$ (left) and ${\sigma_{\mathrm{LN}, \, n}}$ (right). The left panel also shows the standard deviation of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ from the 256 synthetic clusters and black, red, and blue represent different values of ${\sigma_{\mathrm{LN}, \, n}}$, $0.1$, $0.3$, and $0.5$, respectively. In both panels, symbols indicate values of ${\alpha_q}$ (cross, square, triangle, asterisk, and circle correspond to ${\alpha_q}= -2$, $-2.5$, $-3$, $-3.5$, and $-4$, respectively). Dashed lines show the best-fit approximately linear ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$-${\sigma_{\mathrm{LN}, \, n}}$ relation (Eq. \[eq:fitsKa\] and Eq. \[eq:Kalpha\]) for each pair of ${\sigma_{\mathrm{LN}, \, n}}$, ${\alpha_q}$. \[fig:mgsigout\] ](f6b.ps "fig:"){width="83mm"}
From actual observations, we obtain the ${\mbox{$\delta_\mathrm{Sx}$}}$ map for an individual cluster, not the ensemble average. Therefore, we evaluate the distributions of ${\mbox{$\delta_\mathrm{Sx}$}}$ in individual synthetic clusters. Figure \[fig:ind\] shows the PDF for five individual synthetic clusters (solid) along with the best-fit lognormal distributions (dashed). We neglect the radial dependence and use the distribution for the whole cluster within a diameter of ${L_{\mathrm{box}}}=10 \, {r_{\mathrm{c}}}$. Each color represents a different individual synthetic cluster and each panel shows a different value of the power-law index of the Gaussian field, ${\alpha_q}$, with values between -2 and -4. Even if the analysis is done for one cluster, the distribution approximately follows the lognormal distribution.
The noisy behavior for steeper spectra ($\alpha_q=-3.5$, $-4$) in Figure 5 is due to the presence of fluctuations on scales larger than that of the cluster, similar to the discussion above for Figure 3. In other words, steeper spectra ($\alpha_q<-3$) have relatively more larger scale fluctuations compared to shallower spectra ($\alpha_q>-3$). Cosmological hydrodynamic simulations suggest that $\alpha_q \approx 3$, placing galaxy clusters in the less noisy regime. We do not consider the noisy regime further in this paper.
The standard deviations of the logarithm of ${\mbox{$\delta_\mathrm{Sx}$}}$, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$, for the different sets of ${\alpha_q}$ (symbols) and ${\sigma_{\mathrm{LN}, \, n}}$ (colors) are shown in Figure \[fig:mgsigout\]. The relation between ${\sigma_{\mathrm{LN}, \, n}}$ and ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ is approximately linear (right panel) although the proportionality coefficient depends on ${\alpha_q}$. Therefore, we can write $$\begin{aligned}
\label{eq:fitsKa}
{\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}= Q({\alpha_q}) {\sigma_{\mathrm{LN}, \, n}}.\end{aligned}$$ We find that $Q({\alpha_q})$ can be approximated well by the following function $$Q({\alpha_q}) = \frac{c_1}{c_2 + |{\alpha_q}|^{-4}}.
\label{eq:Kalpha}$$ We calculate the average of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}/{\sigma_{\mathrm{LN}, \, n}}$ for each ${\alpha_q}$ over three different values of ${\sigma_{\mathrm{LN}, \, n}}$ (${\sigma_{\mathrm{LN}, \, n}}=0.1,0.3,$ and $0.5$). By fitting ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}/{\sigma_{\mathrm{LN}, \, n}}({\alpha_q})$ using equation (\[eq:Kalpha\]), we obtain $c_1 = 2.05 \times 10^{-2}$ and $c_2= 1.53 \times 10^{-2}$.
![Comparison of the X-ray surface brightness (${\alpha_{\mathrm{Sx}}}$) and the input Gaussian field (${\alpha_q}$) power-law indices. Symbols and error bars indicate the average and the standard deviation, respectively, of ${\alpha_{\mathrm{Sx}}}$ for 256 samples for different sets of ${\alpha_q}$ and ${\sigma_{\mathrm{LN}, \, n}}$. Symbols correspond to different values of ${\sigma_{\mathrm{LN}, \, n}}$, with cross, square, and triangle symbols indicating ${\sigma_{\mathrm{LN}, \, n}}=0.1$, $0.3$, and $0.5$, respectively, and the relations ${\alpha_{\mathrm{Sx}}}={\alpha_q}$ and ${\alpha_{\mathrm{Sx}}}={\alpha_q}+0.2$ are also shown (dotted and solid lines, respectively). We obtain ${\alpha_{\mathrm{Sx}}}$ for each individual synthetic cluster by fitting $P_{{S_{\mathrm{X}}}}({{\bf K}})$ of an individual cluster under the assumption of both statistical isotropy and a power-law ($ \propto K^{{\alpha_{\mathrm{Sx}}}}$). \[fig:ipchange2\]](f7.ps){width="65mm"}
![The average of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ over the 256 synthetic clusters as a function of ${\alpha_q}$ for different values of the $\beta$ model power-law index, $\beta$. Symbols correspond to different values of ${\alpha_q}$ as in Figure \[fig:mgsigout\]. Each color shows a different value of $\beta$ (black, red, and blue correspond to $\beta=1.0, 2/3, $ and $0.5$, respectively). Solid, dashed, and dotted lines are fits using equation (\[eq:fitsKa\]), corresponding to $\beta=1.0, 2/3,$ and $0.5$, respectively. The top, middle, and bottom sets of three different lines indicate ${\sigma_{\mathrm{LN}, \, n}}=0.5, 0.3$, and $0.1$, respectively, as indicated in the figure. \[fig:betabeta\]](f8.ps){width="80mm"}
### Spectral Considerations {#ss:sx_ps}
Because ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ is strongly dependent on the power-law index ${\alpha_q}$, the estimate of ${\alpha_q}$ from the ${\mbox{$\delta_\mathrm{Sx}$}}$ map is crucial for interpreting the value of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$. Because ${\alpha_q}$ is an un-observable quantity, we investigate the relationship between the power spectra of ${\mbox{$\delta_n$}}$ and ${\mbox{$\delta_\mathrm{Sx}$}}$ by fitting the power spectrum of ${\mbox{$\delta_\mathrm{Sx}$}}$ under the assumptions of both statistical isotropy and a power law so that $ P_{{S_{\mathrm{X}}}}({{\bf K}}) \propto K^{{\alpha_{\mathrm{Sx}}}}$, where ${{\bf K}}$ indicates the two-dimensional wave vector.
Figure \[fig:ipchange2\] shows the power-law index of the X-ray surface brightness, ${\alpha_{\mathrm{Sx}}}$, as a function of its counterpart Gaussian field, ${\alpha_q}$. Averages and standard deviations over 256 synthetic clusters are shown for three values of the standard deviation of the logarithm of density, ${\sigma_{\mathrm{LN}, \, n}}$, where crosses, squares, and triangles correspond to ${\sigma_{\mathrm{LN}, \, n}}$ of $0.1$, $0.3$, and $0.5$, respectively. The dotted line corresponds to the relation ${\alpha_{\mathrm{Sx}}}= {\alpha_q}$ and the solid line shows ${\alpha_{\mathrm{Sx}}}= {\alpha_q}+
0.2$. We find that ${\alpha_{\mathrm{Sx}}}\approx {\alpha_q}+ 0.2$ and since ${\alpha_q}\approx {\alpha_{n}}$, this implies ${\alpha_{\mathrm{Sx}}}\approx {\alpha_{n}}+ 0.2$. This can be understood as follows. As we have seen in § \[ssec:method\], the difference between ${\alpha_{n}}$ and ${\alpha_{nn}}$ is relatively small ($\lesssim 13$% and often $\lesssim
5$%). If one assumes ${\mbox{$\delta_\mathrm{Sx}$}}$ is the projection of ${\delta_{nn}}$ (although this is only strictly true if the average of the surface brightness is defined by the ensemble average as Eq.\[\[eq:ens1\]\]), ${\mbox{$\delta_\mathrm{Sx}$}}$ can be described as $${\mbox{$\delta_\mathrm{Sx}$}}({{\bf \Theta}}) = \int d l \, {\delta_{nn}}\, W({{\bf \Theta}},l),
\label{eq:dsx_theta}$$ where ${{\bf \Theta}}$ indicates celestial coordinates and $W({{\bf \Theta}},l)$ is the window function. If we neglect the ${{\bf \Theta}}$-dependence of the window function and set $W({{\bf \Theta}},l)=W(l)$, then $P_{{S_{\mathrm{X}}}}({{\bf K}})$ can be written as $$P_{{S_{\mathrm{X}}}}({{\bf K}}) = \frac{1}{2 \pi} \int d {k_l}\, P_{nn} ({{\bf k}}) \,
|\widetilde{W}({k_l})|^2,
\label{eq:ps_dsx}$$ where $\widetilde{W}({k_l})$ is the Fourier transform of $W(l)$. The assumption that the size of the cluster is much larger than the typical scales of the fluctuations yields $|\widetilde{W}({k_l})|^2 \sim 2 \pi
\delta({k_l})$, where $\delta({k_l})$ is the Dirac delta function, and therefore $K^{{\alpha_{\mathrm{Sx}}}} \propto k^{{\alpha_{nn}}}$. Thus, we find ${\alpha_{n}}\sim {\alpha_{nn}}\sim {\alpha_{\mathrm{Sx}}}$ ($\sim {\alpha_q}$).
In this section, we have found that, in principle, one can estimate the value of ${\sigma_{\mathrm{LN}, \, n}}$ from analysis of X-ray observations. From the observations one measures ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\alpha_{\mathrm{Sx}}}$ and uses them to infer ${\sigma_{\mathrm{LN}, \, n}}$, noting that ${\alpha_q}= {\alpha_{\mathrm{Sx}}}- 0.2$. Therefore, one can estimate the statistical nature of the intrinsic three dimensional fluctuations from two dimensional X-ray observations.
Potential Systematics {#ss:pot_sys}
---------------------
Using mock observations of isothermal $\beta$ models we found a relation between the intrinsic inhomogeneity of the three dimensional cluster gas and the fluctuations in the X-ray surface brightness. We turn our attention to the effects of departures from this idealized model.
### $\beta$ Model Power-law Index {#sss:vary_beta}
In the above description, we have fiducially assumed the $\beta$ model power-law index $\beta=2/3$. We investigate two other cases, $\beta=0.5$, and $\beta=1.0$, in Figure \[fig:betabeta\], where we show ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ as a function of ${\alpha_q}$ for different cases of $\beta$ (colors). The corresponding fits using equation (\[eq:fitsKa\]) are also shown. Although ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ tends to increase with increasing $\beta$, the change is relatively small ($<10$%).
### Temperature Structure {#ss:tstruct}
![The distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ for five individual clusters including the effects of temperature structure. Synthetic clusters (solid histogram) and best-fit lognormal model (dashed lines) are both shown for each cluster. Each color corresponds to a different individual synthetic cluster. Although we display only one example of the power-law index, ${\alpha_q}=-3.0$, similar results are also obtained in other cases. []{data-label="fig:tind"}](f9.ps){width="70.0mm"}
In the above discussion, we assumed isothermality for the ICM. However, the X-ray surface brightness also depends on the underlying cluster temperature structure, including a non-isothermal average temperature profile and local inhomogeneity. We investigate these effects for the X-ray surface brightness distribution.
We assume a polytropic profile for the temperature radial distribution expressed as $$\overline{T}(r) = T_0 \left(\frac{\overline{n}(r)}{n_0}\right)^{\gamma-1},
\label{eq:t_polytrope}$$ with polytropic index $\gamma = 1.2$ and $T_0 = 6$ keV, which is the typical set of values in simulated clusters (Paper I). The ensemble average of the power spectrum of ${\mbox{$\delta_T$}}$ is assumed to have a power-law form (${\langle}P_{T}(k) {\rangle}\sim P_q(k) \propto k^{{\alpha_{q,T}}}$). Because ${\alpha_{T}}\approx {\alpha_{q,T}}$ for the same reasons as described in § \[ssec:method\] for density fluctuations, we fiducially adopt the power-law index ${\alpha_{q,T}}=-3$ based on the results of cosmological hydrodynamic simulations (for details see § \[subsec:hydro\_sim\]).
![The effect of the PSF on ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ as a function of radius, $R/{r_{\mathrm{c}}}$, for the case of ${\alpha_q}=-3$ and $\beta=2/3$. Solid curves show ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}(R)$ without convolution of the PSF. Dashed, dash-dotted, and dotted curves correspond to ${\theta_{\mathrm{HPD}}}/{\theta_\mathrm{c}}=0.1,0.2$ and $0.5$, respectively. Two values of ${\sigma_{\mathrm{LN}, \, n}}$ are plotted, 0.1 and 0.5, as indicated in the figure. []{data-label="fig:psf"}](f10.ps){width="80mm"}
We create the lognormal distribution $\delta_{T}$ for temperature fluctuations in the same manner as for the density fluctuations described in § \[ssec:method\]. The temperature of an arbitrary point is assigned according to $$T({{\bf r}}) = \delta_{T}({{\bf r}}) \overline{T} (r).
\label{eq:t_ijk}$$ We adopt ${\sigma_{\mathrm{LN}, \, T}}=0.3$, because it is the typical value for simulated clusters (Paper I). In addition, we assume that ${\mbox{$\delta_n$}}$ and ${\mbox{$\delta_T$}}$ are distributed independently, following Paper I. The X-ray surface brightness is given by $${S_{\mathrm{X}}}({{\bf R}}) \; \propto \int [n({{\bf r}})]^2 \, \Lambda[T({{\bf r}})] \, dl,
\label{eq:sx_jk2}$$ where $\Lambda(T)$ is the X-ray cooling function. We calculate $\Lambda(T)$ in the energy range 0.5-10.0 keV using SPEX 2.0 [@1996uxsa.conf..411K] on the assumption of collisional ionization equilibrium and a constant metallicity of 30% solar abundances.
Examples of the distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ in individual clusters are shown in Figure \[fig:tind\] (solid histogram) along with the best fit lognormal distributions (dashed lines). Each color corresponds to a different individual synthetic cluster. Although only one value for the power-law index, $\alpha_q=-3$, is shown, similar results are obtained for other values. The radial dependence of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ including the effects of temperature structure is shown in Figure \[fig:rdt\] (dot-dashed).
There are only small differences between the isothermal and non-isothermal cases. The X-ray surface brightness depends on the density squared but roughly as $\sqrt{T}$ for bremsstrahlung emission. Therefore, the temperature structure effects on ${\mbox{$\delta_\mathrm{Sx}$}}$ are much less important than those of the density structure. Hereafter, we neglect the effects of temperature structure and focus only on the effects of density inhomogeneity.
### Finite Spatial Resolution {#ss:resolution}
Actual observations by X-ray satellites have finite spatial resolution, characterized by the point spread function (PSF). We assume that the PSF is a circularly symmetric Gaussian with standard deviation $\sigma$. The PSF can then be parameterized by a single parameter called the [*half power diameter*]{} (${\theta_{\mathrm{HPD}}}$) in which 50% of the X-rays are enclosed (${\theta_{\mathrm{HPD}}}/\sigma = 2 \sqrt{2 \log
2}$). We investigate three cases, ${\theta_{\mathrm{HPD}}}/{\theta_\mathrm{c}}=0.1,0.2$ and $0.5$. Figure \[fig:psf\] shows the effect of the PSF on ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ as a function of radius. In each case, the average over 256 synthetic clusters is shown. Results for no PSF correction (${\theta_{\mathrm{HPD}}}=0$, solid) and ${\theta_{\mathrm{HPD}}}/{\theta_\mathrm{c}}=0.1$ (dashed), 0.2 (dot-dashed), and 0.5 (dotted) are shown. As ${\theta_{\mathrm{HPD}}}/{\theta_\mathrm{c}}$ increases, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ near the center of the cluster decreases. This can be understood as follows. In each radial shell, fluctuations smaller than roughly the radius of the shell predominately contribute to the fluctuations, namely ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}(R)$. The PSF effectively smooths out the smaller scale fluctuations (roughly up to the size of the PSF), reducing ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$, while preserving the large scale fluctuations. Since the inner shells only contain small scale fluctuations, they are more strongly affected by the PSF. The case of ${\theta_{\mathrm{HPD}}}/{\theta_\mathrm{c}}=0.5$ best illustrates these effects. The reduction of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ from the PSF is seen at all radii. However, it is only a slight reduction at large radii, increasing as the radius decreases, with a very large effect near the cluster center.
In summary, when ${\mbox{$\delta_n$}}$ in three dimensions follows the lognormal distribution, ${\mbox{$\delta_\mathrm{Sx}$}}$ in two dimensions also approximately follows the lognormal distribution. The mean value of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ for an individual cluster is strongly dependent on both ${\sigma_{\mathrm{LN}, \, n}}$ and ${\alpha_q}$. Because ${\alpha_q}$ is approximately equal to ${\alpha_{\mathrm{Sx}}}$, in principle, one can infer ${\sigma_{\mathrm{LN}, \, n}}$ from ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ although there is still some dispersion even if ${\alpha_q}$ is known. In addition, the effect of the temperature structure is minimal.
Application to Abell 3667 {#sec:obs}
=========================
Simulations suggest that the lognormal model (Eq. \[\[eq:pdf\_delta\]\]) is a reasonable approximation of the small scale structure in galaxy clusters. We compare this model with [[*Chandra*]{} ]{}X-ray observations of the nearby galaxy cluster Abell 3667 at a redshift $z=0.056$ [@struble1999]. A3667 is a well observed nearby bright galaxy cluster that does not exhibit a cool core observed by [[*Chandra*]{} ]{}. With its complex structure, including a cold front [@vikhlinin2001] and possible merger scenario [e.g., @knopp96], A3667 will serve as a difficult test case for the lognormal model of density fluctuations.
![[[*Chandra*]{} ]{}image of the galaxy cluster Abell 3667 (left) and the corresponding ${\mbox{$\delta_\mathrm{Sx}$}}$ image (right). The counts image has been divided by the exposure map to yield X-ray surface brightness (cnt s$^{-1}$ cm$^{-2}$ arcmin$^{-2}$), including scaling for the pixel size. Point sources in the field have been masked. []{data-label="fig:a3667_image"}](f11a.ps "fig:"){height="8.5cm"} ![[[*Chandra*]{} ]{}image of the galaxy cluster Abell 3667 (left) and the corresponding ${\mbox{$\delta_\mathrm{Sx}$}}$ image (right). The counts image has been divided by the exposure map to yield X-ray surface brightness (cnt s$^{-1}$ cm$^{-2}$ arcmin$^{-2}$), including scaling for the pixel size. Point sources in the field have been masked. []{data-label="fig:a3667_image"}](f11b.ps "fig:"){height="8.5cm"}
[cccc]{} $\phn513$ & $\phn45$ & $20\ 12\ 50.30$ & $-56\ 50\ 56.99$\
$\phn889$ & $\phn51$ & $20\ 11\ 50.00$ & $-56\ 45\ 34.00$\
$5751$ & $131$ & $20\ 13\ 07.25$ & $-56\ 53\ 24.00$\
$5752$ & $\phn61$ & $20\ 13\ 07.25$ & $-56\ 53\ 24.00$\
$5753$ & $105$ & $20\ 13\ 07.25$ & $-56\ 53\ 24.00$\
$6292$ & $\phn47$ & $20\ 13\ 07.25$ & $-56\ 53\ 24.00$\
$6295$ & $\phn50$ & $20\ 13\ 07.25$ & $-56\ 53\ 24.00$\
$6296$ & $\phn50$ & $20\ 13\ 07.25$ & $-56\ 53\ 24.00$
Data Reduction {#subsec:obs_data_reduce}
--------------
[[*Chandra*]{} ]{}observations of the galaxy cluster A3667 are summarized in Table \[tab:a3667\_data\]. Listed are the observation identification numbers, exposure times, and pointing centers of each of the eight archival [[*Chandra*]{} ]{}observations of A3667 used in this analysis. The data are reduced with CIAO version 4.0 and calibration data base version 3.4.2. The data are processed starting with the level 1 events data, removing cosmic ray afterglows, correcting for charge transfer inefficiency and optical blocking filter contamination, and other standard corrections, in addition to generating a customized bad pixel file. The data are filtered for [[*ASCA*]{} ]{}grades 0, 2, 3, 4, 6 and status=0 events and the good time interval data provided with the observations are applied. Periods of high background count rate are excised using an iterative procedure involving creating light curves in background regions with 500 s bins, and excising time intervals that are in excess of 4 $\sigma$ from the median background count rate. This sigma clipping procedure is iterated until all remaining data lie within 4 $\sigma$ of the median. The final events list is limited to energies 0.7-7.0 keV to exclude the low and high energy data that are more strongly affected by calibration uncertainties. Finally, the images are binned by a factor of eight, resulting in a pixel size of 3.94. This pixel size matches the resolution of the synthetic clusters considered in §\[sec:synthetic\]. In particular, the ratio of pixel size to the cluster core radius of the [[*Chandra*]{} ]{}image is similar to the synthetic cluster grid spacing compared to the synthetic cluster core radius, namely, for $\theta_\mathrm{c} \sim 180\arcsec$ [@rb02; @knopp96], $\theta_{\mathrm{pix}} / \theta_\mathrm{c} \sim
{d_{\mathrm{grid}}}/ {r_{\mathrm{c}}}\sim 0.02$. Exposure maps are constructed for each observation at an energy of 1 keV. The binned images and exposure maps for each observation are then combined to make the single image and exposure map used for the analysis.
A wavelet based source detector is used to find and generate a list of potential point sources. The list is examined by eye, removing bogus or suspect detections, and then used as the basis for our point source mask. Figure \[fig:a3667\_image\] (left) shows the [[*Chandra*]{} ]{}merged image of A3667, the counts image divided by the exposure map, where the point source mask has been applied. Also shown is the ${\mbox{$\delta_\mathrm{Sx}$}}$ image (right), discussed below. A cold front [@vikhlinin2001] is clearly visible in the south-eastern region of the ${\mbox{$\delta_\mathrm{Sx}$}}$ image.
Analysis and Results {#subsec:obs_analysis}
--------------------
In order to determine the center of A3667, a $\beta$ model is fit to the data with fixed core radius ($180\arcsec$) and $\beta$ (2/3), using software originally developed for the combined analysis of X-ray and Sunyaev-Zel’dovich effect observations [@reese00; @reese02; @bonamente06]. Because A3667 is nearby and appears very large, [[*Chandra*]{} ]{}observations do not encompass the entire cluster but provide a wealth of information on the complexities inherent in galaxy cluster gas. By using a $\beta$ model fit to the diffuse emission of the cluster gas we obtain a better measurement of its center than simply using the brightest pixel or other simple estimates, which fail to take into account the complex structure manifest in this cluster. A circular region of radius $\sim 8\arcmin$ centered on A3667 is used in the analysis, corresponding to two and a half times the cluster’s core radius, the largest usable region from the arrangement of the combined [[*Chandra*]{} ]{}observations.
The average X-ray surface brightness is required to compute ${\mbox{$\delta_\mathrm{Sx}$}}=
{S_{\mathrm{X}}}/ {\overline{S}_{\mathrm{X}}}$. If one computes the average surface brightness, ${\overline{S}_{\mathrm{X}}}$, in annular shells, then one will tend to under (over) estimate ${\overline{S}_{\mathrm{X}}}$ toward the inner (outer) radius of each annulus. Therefore, this will lead to an over (under) estimate of ${\mbox{$\delta_\mathrm{Sx}$}}$ toward the inner (outer) radius of each annulus. To alleviate this systematic, we adopt the azimuthally averaged X-ray surface brightness as the model for ${\overline{S}_{\mathrm{X}}}$, and use cubic spline interpolation between radial bins. The X-ray surface brightness radial profile for A3667 is shown in Figure \[fig:a3667\_radprof\], along with the interpolated model (line).
![[[*Chandra*]{} ]{}radial profile of the galaxy cluster Abell 3667 (points) with the interpolated model (solid line). This model is used as the average X-ray surface brightness distribution in the calculation of ${\mbox{$\delta_\mathrm{Sx}$}}$.[]{data-label="fig:a3667_radprof"}](f12.ps){height="6.3cm"}
![Probability distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ from [[*Chandra*]{} ]{} observations of the galaxy cluster Abell 3667 (blue histogram) along with the best fit lognormal distribution (red line) with ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}=
0.30$. The lognormal distribution seems to be a reasonable description of the ICM inhomogeneity in A3667. Also shown are the best-fit Gaussian model (dashed green) and a Poisson model (dot-dashed magenta) using the average counts per pixel within the fitting region.[]{data-label="fig:a3667_hist"}](f13.ps){height="6.3cm"}
![Power spectrum of ${\mbox{$\delta_\mathrm{Sx}$}}$ (thick solid) from [[*Chandra*]{} ]{} observations of the galaxy cluster Abell 3667, normalized to one at the largest scale. Also plotted are three power-law power spectra with spectral indices of -2 (dashed), -3 (dot-dashed), and -4 (dotted) for comparison.[]{data-label="fig:a3667_ps"}](f14.ps){height="6.3cm"}
The probability distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$, $p({\mbox{$\delta_\mathrm{Sx}$}})$, is computed from the histogram of pixels calculated from the ${\mbox{$\delta_\mathrm{Sx}$}}$ image and shown in Figure \[fig:a3667\_hist\]. The lognormal distribution (Eq. \[\[eq:pdf\_delta\]\]) is fit to the $p({\mbox{$\delta_\mathrm{Sx}$}})$ of A3667, where the only free parameter is the standard deviation of the logarithm of ${\mbox{$\delta_\mathrm{Sx}$}}$, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$. The best fit value for the lognormal model is ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}= 0.30$. In addition, a Gaussian distribution is also fit to the data, with its usual two parameters, the mean and standard deviation. Figure \[fig:a3667\_hist\] shows the PDF of ${\mbox{$\delta_\mathrm{Sx}$}}$ for the [[*Chandra*]{} ]{}observations of the galaxy cluster A3667 (solid blue histogram). The best fit lognormal (solid red) and Gaussian (dashed green) models are also shown. A Poisson distribution (dot-dashed magenta) is also shown for comparison, using the average counts per pixel in the fitting region as the parameter for the Poisson distribution. Clearly, what is seen is not the result of Poisson statistics. The lognormal model seems to be a reasonable match to the observed PDF.
However, without information on the power spectrum of the ${\mbox{$\delta_\mathrm{Sx}$}}$ fluctuations, it is difficult to interpret the value of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ (§\[ss:sx\_ps\]) and relate it to the fluctuations in the density distribution (Eqs. \[\[eq:fitsKa\], \[eq:Kalpha\]\]; Fig. \[fig:mgsigout\]). Therefore, we take the Fourier transform of the ${\mbox{$\delta_\mathrm{Sx}$}}$ image and compute the average power spectrum in wavenumber annuli. The power spectrum of ${\mbox{$\delta_\mathrm{Sx}$}}$ fluctuations is shown in Figure \[fig:a3667\_ps\] (thick solid) along with three power-law spectra with spectral indices of -2 (dashed), -3 (dot-dashed), and -4 (dotted) for comparison. The power spectrum of ${\mbox{$\delta_\mathrm{Sx}$}}$ has been normalized to one at the largest scales. A simple power-law model fit to the power spectrum yields a spectral index of ${\alpha_{\mathrm{Sx}}}= -2.7$ using the entire spectrum, and a spectral index of ${\alpha_{\mathrm{Sx}}}= -3.0$ if excluding the larger wavenumbers ($\gtrsim 2$ arcmin$^{-1}$), roughly where the power spectrum changes shape.
Implications {#subsec:obs_disc}
------------
![An example of a ${\mbox{$\delta_\mathrm{Sx}$}}$ map from a cosmological hydrodynamic simulated cluster (“Centaurus”) both before (left) and after (right) removal of a quadrant with a large clump. Circles show the projected virial radius ($R_\mathrm{200}$). Although within the projected virial radius, $R_\mathrm{200}$, these structures often reside outside of the three-dimensional virial radius, $r_\mathrm{200}$.[]{data-label="fig:clump"}](f15.ps){width="95mm"}
Both the standard deviation of the logarithm of X-ray surface brightness fluctuations, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}= 0.30$, and the power spectrum power-law index ${\alpha_{\mathrm{Sx}}}\approx -3$, fall into the range expected from hydrodynamical galaxy clusters and therefore used in the synthesized cluster analysis (§\[subsec:hydro\_sim\]). By combining these pieces of information, we can relate the information obtained from the X-ray surface brightness distribution to that of the underlying density distribution, using the results of the synthesized cluster analysis. Using the synthetic cluster result that the spectral indices of the X-ray surface brightness fluctuations and that of the Gaussian field are simply related by ${\alpha_{\mathrm{Sx}}}\approx {\alpha_q}+0.2$, and the relation between ${\sigma_{\mathrm{LN}, \, n}}$, ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$, and ${\alpha_q}$ (Eqs. \[\[eq:fitsKa\], \[eq:Kalpha\]\]; Fig. \[fig:mgsigout\]), the [[*Chandra*]{} ]{}results of ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}= 0.30$ and ${\alpha_{\mathrm{Sx}}}= -2.7$ imply that the fluctuations in the underlying density distribution have ${\sigma_{\mathrm{LN}, \, n}}= 0.43$. A value of ${\alpha_{\mathrm{Sx}}}= -3.0$ implies ${\sigma_{\mathrm{LN}, \, n}}= 0.36$. The difficult test case of the A3667 X-ray surface brightness seems to follow the lognormal distribution of density fluctuations, thus enabling an estimate of the statistical properties of the underlying ICM density fluctuations.
Application to the Cosmological Hydrodynamic Simulated Clusters {#sec:con}
===============================================================
Results from cosmological hydrodynamic simulations motivated the lognormal model for ICM inhomogeneity. In §\[sec:synthetic\], we found that synthetic clusters with lognormal fluctuations show a linear relation between ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\sigma_{\mathrm{LN}, \, n}}$. We now return to clusters extracted from cosmological hydrodynamic simulations to further explore these results.
![The distribution of ${\mbox{$\delta_\mathrm{Sx}$}}$ for each of the six clusters from a cosmological hydrodynamic simulation (points and solid histogram). Each color indicates the projection along a different, orthogonal line of sight. For each line of sight, we show the number of quadrants used for the analysis. For example, “3/4” indicates that one quadrant is excluded and three remain. The best fit lognormal model for each projection is also shown (dotted lines). \[fig:chs\] ](f16.ps){width="120mm"}
For each cluster extracted from the simulations, we create X-ray surface brightness maps towards three orthogonal directions, and compute ${\mbox{$\delta_\mathrm{Sx}$}}({{\bf R}}) = S_{\mathrm X}({{\bf R}}) / {\overline{S}_{\mathrm{X}}}(R)$ in a similar manner as described for the synthetic clusters in § \[ss:em\]. The regions we consider are within the projected virial radius $R_{\mathrm
200}$. The projected virial radius, $R_{\mathrm{200}}$, is the radius within which the mean interior density is 200 times that of the critical density.
![ The density fluctuation standard deviation predicted by our model, ${\sigma_{\mathrm{LN}, \, n}}(\mbox{model}) = {\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}/ Q ({\alpha_q})$ versus that from the simulations, ${\sigma_{\mathrm{LN}, \, n}}(\mbox{sim})$. Symbols show different simulated clusters (see figure legend) and colors indicate different orthogonal lines of sight. Also plotted is the simple linear relation ${\sigma_{\mathrm{LN}, \, n}}\mbox{(model)} = {\sigma_{\mathrm{LN}, \, n}}\mbox{(sim)}$ for comparison. []{data-label="fig:chsss"}](f17.ps){width="95mm"}
Although the lognormal distribution is a good fit to the density (and temperature) of simulated galaxy clusters in three-dimensions, the projection to X-ray surface brightness suffers from the additional complexity of projection effects. If large clumps are present, the distribution of X-ray surface brightness fluctuations, ${\mbox{$\delta_\mathrm{Sx}$}}$, is not well approximated by the lognormal distribution. The large clumps artificially distort the average profile of the cluster and therefore bias the value of ${\mbox{$\delta_\mathrm{Sx}$}}$, which depends on the average profile. We also note that although these clumps fall within the projected virial radius, $R_{\mathrm 200}$, they usually fall outside of the three dimensional virial radius, $r_{\mathrm 200}$. We therefore exclude quadrants that contain large clumps, using ${\mbox{$\delta_\mathrm{Sx}$}}>10$ as the exclusion criterion. Then, we recompute ${\overline{S}_{\mathrm{X}}}(R)$ and ${\mbox{$\delta_\mathrm{Sx}$}}$. The complex structure of simulated clusters is illustrated in the ${\mbox{$\delta_\mathrm{Sx}$}}$ images shown in Figure \[fig:clump\], where examples of a simulated cluster both before and after removal of a quadrant are displayed. The circles show the projected virial radius, $R_\mathrm{200}$.
In Figure \[fig:chs\] the probability distributions of ${\mbox{$\delta_\mathrm{Sx}$}}$ for the simulated clusters (histograms) along with the best-fit lognormal model (dotted lines) are displayed. Each color indicates the projection along a different, orthogonal line of sight. Overall, the probability distributions of ${\mbox{$\delta_\mathrm{Sx}$}}$ are reasonably well approximated by the lognormal function, consistent with the results from the synthetic clusters (§ \[ss:synthetic\_clusters\]).
We now come full circle to compare our results from the synthetic clusters directly to the simulations. In order to do this, we look at the relationship between ${\sigma_{\mathrm{LN}, \, n}}\mbox(\mathrm{sim})$ measured in the simulated clusters and ${\sigma_{\mathrm{LN}, \, n}}\mbox(\mathrm{model})$ predicted from the synthetic cluster results, equations (\[eq:fitsKa\]) and (\[eq:Kalpha\]), where we adopt ${\alpha_q}= {\alpha_{\mathrm{Sx}}}- 0.2$ (see §\[ss:sx\_ps\]). The value of ${\alpha_{\mathrm{Sx}}}$ for each simulated cluster is obtained by fitting a power-law model, $P(K) \propto K^{{\alpha_{\mathrm{Sx}}}}$, to the power spectra of ${\mbox{$\delta_\mathrm{Sx}$}}$. Because the resolution of the simulations is much poorer than that of the synthetic clusters, we must recompute the coefficients $c_1$ and $c_2$ in equation \[eq:Kalpha\] from a set of lower resolution synthetic clusters. Assuming ${r_{\mathrm{c}}}\sim 100$ $h^{-1}$ kpc for the simulated clusters, we choose the resolution $\sim 0.1
{d_{\mathrm{grid}}}/{r_{\mathrm{c}}}$, noting that this value corresponds to the [*maximum*]{} resolution of the simulations. Performing the same procedure described in §\[sec:synthetic\], we obtain $c_1 = 3.99 \times 10^{-2}$ and $c_2= 3.36 \times 10^{-2}$.
We compare ${\sigma_{\mathrm{LN}, \, n}}\mbox{(model)}$ and ${\sigma_{\mathrm{LN}, \, n}}\mbox(\mathrm{sim})$ in Figure \[fig:chsss\]. Each color corresponds to a different line of sight. Although there is large scatter, these results indicate that it is possible to estimate ${\sigma_{\mathrm{LN}, \, n}}$ within a factor of two only using the information obtained from the X-ray surface brightness distribution.
Summary {#sec:sum}
=======
We have developed a method of extracting statistical information on the ICM inhomogeneity from X-ray observations of galaxy clusters. With a lognormal model for the fluctuations motivated by cosmological hydrodynamic simulations, we have created synthetic clusters, and have found that their X-ray surface brightness fluctuations retain the lognormal nature. In addition, the result that ${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\sigma_{\mathrm{LN}, \, n}}$ are linearly related implies that one can, in principle, estimate the statistical properties of the three dimensional density inhomogeneity (${\sigma_{\mathrm{LN}, \, n}}$) from X-ray observations of galaxy clusters (${\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}$ and ${\alpha_{\mathrm{Sx}}}$).
We have compared the predictions of our model to [[*Chandra*]{} ]{}X-ray observations of the galaxy cluster A3667. For the first time in a real galaxy cluster we were able to detect the lognormal signature of X-ray surface brightness fluctuations, which was originally motivated by simulations. Based on the synthetic cluster results, this enabled an estimate of the statistical properties of the inhomogeneity of the ICM of A3667. In the context of lognormally distributed inhomogeneity, we obtain ${\sigma_{\mathrm{LN}, \, n}}\approx 0.4$ for the gas density fluctuations of A3667. It is encouraging that the value of the fluctuation amplitude for Abell 3667 is in reasonable agreement with typical values from the simulated clusters.
Finally we check the validity and limitation of our method using several clusters from cosmological hydrodynamic simulations. Unlike the fairly idealized synthetic clusters, simulated clusters exhibit complex structure more akin to real galaxy clusters. As a result, the empirical relation between the two- and three-dimensional fluctuation properties calibrated with synthetic clusters when applied to simulated clusters shows large scatter. Nevertheless we are able to reproduce the true value of the fluctuation amplitude of simulated clusters within a factor of two from their two-dimensional X-ray surface brightness alone.
Our current methodology combined with existing observational data is useful in describing and inferring the statistical properties of the three dimensional inhomogeneity in galaxy clusters. The fluctuations in the ICM have several implications in properly interpreting galaxy cluster data. In particular, our current model may be useful in interpreting data from current and future galaxy cluster surveys using the Sunyaev-Zel’dovich effect, which have the potential to provide tight constraints on cosmology.
We thank Naomi Ota, Noriko Y. Yamasaki, and Kazuhisa Mitsuda for useful discussions and Klaus Dolag for providing a set of simulated clusters. HK is supported by a JSPS (Japan Society for Promotion of Science) Grant-in-Aid for science fellows. EDR gratefully acknowledges support from a JSPS Postdoctoral Fellowship for Foreign Researchers (P07030). This work is also supported by Grant-in-Aid for Scientific research from JSPS and from the Japanese Ministry of Education, Culture, Sports, Science and Technology (Nos. 20$\cdot$10466, 19$\cdot$07030, 16340053, 1874012, 20340041, and 20540235), and by the JSPS Core-to-Core Program “International Research Network for Dark Energy”.
Relation of Density and Surface Brightness Distributions Under the Thick-slice Approximation {#a1:den_sb}
============================================================================================
Modeling galaxy clusters with a spherical isothermal $\beta$ model (Eq. \[eq:ave1\]), the surface brightness at an arbitrary projected angular radius, $\theta$, is given by $$\begin{aligned}
{S_{\mathrm{X}}}(\theta) &\propto& \int_{-\infty}^{\infty} [n ({{\bf r}})]^2 dl
\label{eq:a_sx1}\nonumber \\
&=& \int_{-\infty}^{\infty} {\delta_{nn}}({{\bf r}}) {\langle}n^2 {\rangle}(r=\sqrt{l^2+{d_A^2 \theta^2}}) dl
\label{eq:a_sx2}\nonumber \\
&=& M_2 n_0^2 \left( 1 + \frac{{d_A^2 \theta^2}}{r_c^2} \right)^{-3 \beta}
\int_{-\infty}^{\infty} {\delta_{nn}}({{\bf r}}) \left[ 1 +
\left( \frac{l}{{r_{c,\mathrm{eff}}(\theta)}} \right)^2 \right]^{-3 \beta} dl,
\label{eq:a_sx3}
\label{eq:a_sx4}\end{aligned}$$ where ${r_{c,\mathrm{eff}}(\theta)}\equiv \sqrt{r_c^2 + {d_A^2 \theta^2}}$, and we assume the $\sigma_{{\rm LX},x}$ in equation \[eq:pdf\_delta\] is independent of $r$. Therefore, the second moment of $n$ ($M_2 \equiv {\langle}n^2 {\rangle}/{\langle}n
{\rangle}^2 = \exp{(-{\sigma_{\mathrm{LN}, \, n}}^2)}$) is also independent of $r$. In the above, we use ${\langle}n {\rangle}= \overline{n} (r)$.
The ensemble average of ${\langle}{S_{\mathrm{X}}}(\theta) {\rangle}$ can be expressed as $$\begin{aligned}
{\langle}{S_{\mathrm{X}}}(\theta) {\rangle}&\propto& \int_{-\infty}^{\infty} {\langle}n^2 {\rangle}(r=\sqrt{l^2+{d_A^2 \theta^2}}) dl
\label{eq:a_sx_avg2}\nonumber \\
&=& \sqrt{\pi} n_0^2 M_2 r_c \frac{\Gamma(3
\beta - 1/2)}{\Gamma(3 \beta)} \left(
1 + \frac{{d_A^2 \theta^2}}{r_c^2} \right)^{-3 \beta + 1/2}
\label{eq:a_sx_avg3}\end{aligned}$$
Combining equations (\[eq:a\_sx3\]) and (\[eq:a\_sx\_avg3\]), ${\delta_{\mathrm{Sx,ens}}}$ reduces to $$\begin{aligned}
{\delta_{\mathrm{Sx,ens}}}(\theta) &=& \kappa_\beta \int_{-\infty}^{\infty} {\delta_{nn}}({{\bf r}})
\left[ 1 + \left( \frac{l}{{r_{c,\mathrm{eff}}(\theta)}} \right)^2
\right]^{-3 \beta} d\left(\frac{l}{{r_{c,\mathrm{eff}}(\theta)}}\right)
\label{eq:a_dsx_ens} \nonumber \\
\kappa_\beta &\equiv& \pi^{-1/2} \frac{\Gamma(3 \beta)}{\Gamma (3
\beta -1/2)}.
\label{eq:a_kappa_beta}\end{aligned}$$
Now, fixing $\theta$, let us consider the three-dimensional field ${\delta_{nn}}({{\bf r}}) W_{\beta}(l)$ and its projected two-dimensional field ${\delta_{\mathrm{Sx,ens}}}|_\theta$, defined as $$\begin{aligned}
{\delta_{\mathrm{Sx,ens}}}|_\theta&=& \int_{-\infty}^{\infty} {\delta_{nn}}({{\bf r}}) W_{\beta}({{l^{\prime}}})
d{{l^{\prime}}}\label{eq:a_dsx_ens2} \\
W_\beta({{l^{\prime}}}) &\equiv& \kappa_\beta ( 1 + {{l^{\prime}}}^2 )^{-3 \beta},
\label{eq:a_window}\end{aligned}$$ where we use a dimensionless length normalized by ${r_{c,\mathrm{eff}}(\theta)}$ distinguished by prime (${{l^{\prime}}}\equiv l/{r_{c,\mathrm{eff}}(\theta)}$, ${k_l'}\equiv {k_l}{r_{c,\mathrm{eff}}(\theta)}$). Then, we can consider the variance of the ${\delta_{\mathrm{Sx,ens}}}|_\theta$-field, $${\sigma_{{\delta_{\mathrm{Sx,ens}}}|_\theta}^2}= \frac{1}{(2 \pi)^2} \int d {{\bf K'}}P_{{S_{\mathrm{X,ens}}}|_\theta}({{\bf K'}}),
\label{eq:s1}$$ where $P_{{S_{\mathrm{X,ens}}}|_\theta}({{\bf K'}})$ is the (two-dimensional) power spectrum of ${\delta_{\mathrm{Sx,ens}}}|_\theta$. The variance of the ${\delta_{nn}}$ field can also be written as $${\sigma_{{\delta_{nn}}}^2}= \frac{1}{(2 \pi)^3} \int d {{\bf k'}}P_{nn}({{\bf k'}}).
\label{eq:s2}$$
With this, the relation between $P_{{S_{\mathrm{X,ens}}}}({{\bf K'}})$ and $P_{nn}({{\bf k'}})$ is $$P_{{S_{\mathrm{X,ens}}}|_\theta}({{\bf K'}}) = \frac{1}{2 \pi} \int d {k_l'}P_{nn}({{\bf k'}}) |\widetilde{W}_\beta({k_l'})|^2.
\label{eq:s3}$$ The Fourier conjugate $\widetilde{W}_\beta({k_l'})$ is given by $$\widetilde{W}_\beta({k_l'}) = \kappa_\beta
\left(\frac{2}{{k_l'}}\right)^{-3\beta+1/2} \frac{2 \sqrt{\pi}}{\Gamma
(3 \beta)} K_{-3\beta+1/2} ({k_l'}),$$ where $K_{-3 \beta+1/2} ({k_l'})$ is modified Bessel function of the second kind .
In the case that the largest scale fluctuation is smaller than the physical scale (the thick-slice approximation, following [@FD04]), the Fourier conjugate of the window function becomes the Dirac-delta function, $|\widetilde{W}_\beta ({k_l'})|^2 \sim g(\beta)
\delta({k_l'})$. The normalization factor $g(\beta)$ is given by $$g(\beta) \equiv 2 \int_0^\infty d {k_l'}|\widetilde{W}_\beta ({k_l'})|^2 = 2 \sqrt{\pi}
\frac{\Gamma (3 \beta)^2 \Gamma (6 \beta -1/2)}{\Gamma(3 \beta
-1/2)^2 \Gamma(6 \beta)}.$$ Let us define the effective width $${\Delta_{\mathrm{eff}}}(\theta) \equiv 2 \pi {r_{c,\mathrm{eff}}(\theta)}/ g(\beta) = \sqrt{\pi} \frac{\Gamma(3 \beta -1/2)^2 \Gamma(6 \beta)}{\Gamma (3 \beta)^2
\Gamma (6 \beta -1/2)} {r_{c,\mathrm{eff}}(\theta)}.$$ [@FD04] explore the column density distribution assuming a plane parallel geometry with width $\Delta$. In the thick slice case, ${\Delta_{\mathrm{eff}}}$ corresponds to $\Delta$ although they consider the column density not the surface brightness. We assume statistical isotropy and a power law spectrum with upper and lower limit (${k_{\mathrm{max}}'}\equiv
{k_{\mathrm{max}}}{r_{c,\mathrm{eff}}(\theta)}$ and ${k_{\mathrm{min}}'}\equiv {k_{\mathrm{min}}}{r_{c,\mathrm{eff}}(\theta)}$), $$\begin{aligned}
P_{nn}({{\bf k'}})
\left\{
\begin{array}{lr}
\propto |{{\bf k'}}|^{{\alpha_{nn}}} & \mbox{${k_{\mathrm{min}}'}<|{{\bf k'}}|<{k_{\mathrm{max}}'}$} \\
= 0 & \mbox{otherwise.} \\
\end{array}
\right.\end{aligned}$$
Finally, using equation (\[eq:s1\]), (\[eq:s2\]), and (\[eq:s3\]) under the thick-slice approximation, we obtain $$\begin{aligned}
\label{eq:thickr}
{\sigma_{{\delta_{\mathrm{Sx,ens}}}|_\theta}^2}/{\sigma_{{\delta_{nn}}}^2}=
\left\{
\begin{array}{lr}
\displaystyle{ \frac{1({\alpha_{nn}}+3)(1-\zeta^{{\alpha_{nn}}+2})}{2({\alpha_{nn}}+2)(1-\zeta^{{\alpha_{nn}}+3})}
\left(\frac{{\Delta_{\mathrm{eff}}}(\theta)}{{l_\mathrm{max}}}\right)^{-1}} & \mbox{${\alpha_{nn}}\neq -3$ and ${\alpha_{nn}}\neq -2$}\\
\displaystyle{ \frac{\log \zeta}{2 (\zeta -1)} \left(\frac{{\Delta_{\mathrm{eff}}}(\theta)}{{l_\mathrm{max}}}\right)^{-1} } & \mbox{${\alpha_{nn}}= -2$} \\
\displaystyle{ \frac{1-1/\zeta}{2 \log \zeta} \left(\frac{{\Delta_{\mathrm{eff}}}(\theta)}{{l_\mathrm{max}}}\right)^{-1}} & \mbox{${\alpha_{nn}}= -3$,} \\
\end{array}
\right.\end{aligned}$$ where $\zeta \equiv {k_{\mathrm{max}}}/{k_{\mathrm{min}}}$ and ${l_\mathrm{max}}\equiv 2 \pi {k_{\mathrm{min}}}^{-1}$.
Then, although ${\sigma_{{\delta_{\mathrm{Sx,ens}}}|_\theta}^2}$ is the variance of ${\delta_{\mathrm{Sx,ens}}}|_\theta$-field, we regard it as the variance of the ensemble average of ${\delta_{\mathrm{Sx,ens}}}({{\bf R'}})$ at $\theta$. The conversion to the standard deviation of logarithm is expressed as $$\begin{aligned}
{\sigma_{\mathrm{LN}, \, \mathrm{Sx}}}= \sqrt{\log{(1+{\sigma_{{\delta_{\mathrm{Sx,ens}}}|_\theta}^2})}}\end{aligned}$$ In Figure \[fig:rdt\], we adopt $\zeta= {L_{\mathrm{box}}}/(2
{d_{\mathrm{grid}}})=f_{\mathrm{s}}/f_{\mathrm{Ny}}$, where $f_{\mathrm{s}}$ and $f_{\mathrm{Ny}}$ are the sampling frequency and the Nyquist frequency, respectively, and ${l_\mathrm{max}}={L_{\mathrm{box}}}$. We also adopt the fitted value of ${\alpha_{nn}}$ in equation (\[eq:thickr\]).
Distribution of the Density Squared {#sec:a2_densquared}
===================================
If one assumes density inhomogeneity fluctuations, ${\mbox{$\delta_n$}}= n/{\langle}n {\rangle}$, follow the lognormal distribution, $p_{\mathrm{LN}}({\mbox{$\delta_n$}};{\sigma_{\mathrm{LN}, \, n}})$, the fluctuations of the density squared, ${\delta_{nn}}\equiv n^2/{\langle}n^2 {\rangle}$ can be written as $$\begin{aligned}
{\delta_{nn}}= {\mbox{$\delta_n$}}^2 \frac{{\langle}n {\rangle}^2}{{\langle}n^2 {\rangle}} = {\mbox{$\delta_n$}}^2 \exp{(-{\sigma_{\mathrm{LN}, \, n}}^2)}.\end{aligned}$$ where ${\langle}\, {\rangle}$ indicates ensemble average.
For simplicity, we replace ${\mbox{$\delta_n$}}$ and ${\delta_{nn}}$ by $x$ and $y$, respectively, $$\begin{aligned}
x \equiv {\mbox{$\delta_n$}}; y \equiv {\delta_{nn}}.\end{aligned}$$ Therefore, the relation between $x$ and $y$ is $$\begin{aligned}
x = \sqrt{y} \exp{({\sigma_{\mathrm{LN}, \, n}}^2/2)}.\end{aligned}$$ Because $x$ follows $p_{\mathrm{LN}}(x;{\sigma_{\mathrm{LN}, \, n}})$, the distribution of $y$ is obtained by $$\begin{aligned}
p(y) = p_{\mathrm{LN}}(x;{\sigma_{\mathrm{LN}, \, n}}) \frac{d x}{d y} = p_{\mathrm{LN}}(y;2 {\sigma_{\mathrm{LN}, \, n}}).\end{aligned}$$ Therefore, ${\delta_{nn}}$ follows the lognormal distribution with lognormal standard deviation of $2 {\sigma_{\mathrm{LN}, \, n}}$.
[99]{} Albrecht, A., et al., 2006, astro-ph/0609591 Birkinshaw, M., 1999, , 310, 97 Bradač, M., Clowe, D., Gonzalez, A. H., Marshall, P., Forman, W., Jones, C., Markevitch, M., Randall, S., Schrabback, T., & Zaritsky, D., 2006, , 652, 937 Bartlett, J. G., Silk, J., 1994, , 423, 12 Bonamente et al., 2006, , 647, 25 Briel, U. G., Finoguenov, A., & Henry, J. P.2004, A&A 426,1 Carlstrom, J. E., Holder, G. P., and Reese, E. D. 2002, ARA&A, 40, 643 Cavaliere, A. & Fusco-Femiano, R., 1976, , 49, 137 Cavaliere, A. & Fusco-Femiano, R., 1976, , 70, 677 Dolag, K.,Hansen, F. K.,Roncarelli, & M.,Moscardini, L. 2005, , 363, 29 Eke, V. R., Cole, S., & Frenk, C. S. 1996, , 282, 263 Elmegreen, B., G., 2002, , 564,773 Fabian, A. C., Celotti, A., Blundell, K. M., Kassim, N. E., and Perley, R. A., 2002, , 331, 369 Fabian et al, 2006, 366,417 Fischera, J., & Dopita, M., A., 2004, , 611,919 Furusho, T., Yamasaki, N. Y., Ohashi, T., Shibata, R., & Ezawa, H., 2001, PASJ, 561, L165 Haiman, Z., Mohr, J. J., Holder, G. P., 2001, , 553, 545 Holder, G. P., Mohr, J. J., Carlstrom, J. E., Evrard, A. E., Leitch, E. M., 2000, , 544, 629 Jones, C. and Forman, W. and Vikhlinin, A. and Markevitch, M. and David, L. and Warmflash, A. and Murray, S. and Nulsen, P. E. J., , 567, L115 Kaastra, J. S., Mewe, R., & Nieuwenhuijzen, H. 1996, in UV and X-ray Spectroscopy of Astrophysical and Laboratory Plasmas, 411, 414 Kawahara, H., Suto, Y., Kitayama, T.,Sasaki, S., Shimizu, M, Rasia, E.,& Dolag, K. 2007, , 659, 257 (Paper I) Kawahara, H., Kitayama, T.,Sassaki, & Suto, Y. 2008, Kitayama, T., & Suto, Y. 1996, ApJ, 469, 480 Kitayama, T., & Suto, Y. 1997, ApJ, 490, 557 Kitayama, T., Sasaki, S., Suto, Y., 1998, , 50 1 Knopp, G. P., Henry, J. P., Briel, U. G. 1996, , 472, 125 Majumdar, S., Mohr, J. J., 2004, , 613, 41 Markevitch, M. et al. 2000 , 541, 542 Mazzotta, P., Rasia, E., Moscardini, L., & Tormen, G. 2004, , 354, 10 Rasia, E., Mazzotta, P., Borgani, S., Moscardini, L., Dolag, K., Tormen, G., Diaferio, A., & Murante, G. 2005, , 618, L1 Reese, E.D., Mohr, J.J., Carlstrom, J.E., Joy, M., Grego, L., Holder, G. P., Holzapfel, W.L., Hughes, J. P., Patel, S. K., & Donahue, M. 2000, , 533, 38 Reese, E.D., Carlstrom, J.E., Joy, M., Mohr, J.J., Grego, L., & Holzapfel, W.L. 2002, , 581, 53 Reiprich,T.H., & Böhringer, H. 2002, , 567, 716 Rephaeli, Y., 1995, , 33, 541 Struble, M. F., & Rood, J. H., 1999, , 125, 35 Sunyaev, R. A., & Zel’dovich, I. B., 1980, , 18, 537 Viana, P. T. P., & Liddle, A. R. 1996, , 281, 323 Vikhlinin, A., Markevitch, M. & [Murray]{}, S. S. , 2001 , 551, 160 Vikhlinin, A. 2006, , 640, 710
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Very recently, the Large Hadron Collider was turned on. There, the experiments are aiming to test different scenarios for elementary particles interactions from SUSY, Extra dimensions to others. In fact, SUSY was invented to kill the conceptual problems existing in the conventional Standard model *i.e.* the Hierarchy problem. However, in recent years, certain theories which was rejected in the past like the wrong sign $\phi^{4}$ theory is now well known to be a truly physically acceptable theory. Here, we analyze the renormalization group flow of the different parameters in the theory. We find that, rather than the conventional $\phi^{4}$ theory and because the theory is asymptotically free, it does not sufferer from the catastrophic blow up of the mass parameter for UV scales. This feature greatly recommend that this theory is a plausible candidate to play the role of the $SU(2)\times U(1)$ symmetry breaking in the standard model. If this picture will agree with the experimental results from LHC, then the Higgs particle was massless in the far past. Moreover, the cosmological constant problem as a Hierarchy problem may be solved too. However, when trying to calculate the metric operator, we expected that the equivalent Hermitian theory may attain non-remormalizabe terms and thus the theory is incalculable. In fact, we show that there exists a Hermitian and renormalizable equivalent theory though of indefinite metric. We reformulated the Klein-Gordon equation in such a way that the wrong sign in the propagator attains a new physical meaning that the Higgs particle is made from exotic matter.'
author:
- 'Abouzeid M. Shalaby[^1]'
title: 'Will the $\mathcal{PT}$-symmetric and Non-Hermitian $\phi^{4}$ Theory Solve the Hierarchy and Triviality Problems in the Standard Model?'
---
The origin of mass of the building blocks of our universe impresses the nowadays physicists [@mass]. The reason behind that is the existence of conceptual problems in the Standard model. Such conceptual problems occur because of our ultimate hope to unify all the existent forces. The unification of the electromagnetic, weak and strong interactions introduces the energy scale $10^{16}$ $GeV$, the scale at which the three couplings be the same (in the supersymmetry (SUSY) regime). However, we have another two interesting energy scales, the electroweak scale and the Planck scale. In fact, the scalar Higgs does possess two problems regarding the flow of the parameters either for small or high energy scales. For low energy scales, the theory is trivial while if one tried to let the dimensionful parameters to flow to a very high energy scales, one gets big values for those parameters which will violate the observation. The big values for the dimensionful parameters are well known as gauge Hierarchy problem.
There exists more than one algorithm suggested to solve the Hierarchy problem. For instance, SUSY has been introduced with the aim to have natural cancellation in the dimensionful parameters that turned those parameters protected against perturbations even for very high energy scales [@susy]. However, it introduces an upper limit to the Higgs mass by 130 $GeV$ and some of its mass spectra are of one $TeV$ which make this theory under the fire of the LHC experiments test by 2008. Another algorithm for the solution of the Hierarchy problem is to consider the Higgs particle as a composite state bound by a new set of interactions (Technicolors)[@tecni]. However, the technicolor model is strongly constrained from precision tests of electroweak theory at LEP and SLC experiments [@tecni1] . Also, this algorithm has mass spectra of about one $TeV$ and it is under the direct test of the LHC experiments. Once more, a recent algorithm is suggested for which the $SU(2)\times U(1)$ symmetry is broken via the compactification of an extra dimension. In fact, particles in this model attain masses through the expectation value of the fifth ( for instance) component of the gauge field. However, to some physicists, the digestion of extra dimension is not that easy and can be accepted to them at most as a mathematical modeling to the problem. In this letter, we analyze a modified algorithm which we think it can solve the famous triviality as well as the Hierarchy problems. In this algorithm, we revisit a previously rejected theories because they are non-Hermitian but in view of the recent discoveries of the reality of some class of non-Hermitian models ($\mathcal{PT}$-symmetric), there exists no reason to prevent them from playing in the scene. We think that this is fair and the final word about the correct scenario may come from the analysis of the experimental data from the LHC.
In 1998, Carl Bender and Stefan Boettcher have shown that a class of Non-Hermitian but $\mathcal{PT}$-symmetric Hamiltonians have real spectra [@bendr]. This discovery led us to reinvestigate the non-Hermitian quantum field models like the $(-g\phi^{4})$ and $\left( g\phi^{4}+h\phi^{6}\right)
$ scalar theories [@abophi41p1; @abophi61p1]. In the first model, we realized that, rather than the corresponding Hermitian model $(g\phi^{4})$, the vacuum energy is tiny for a wide range of energy scales. Besides, the vacuum energy is real and in fact, it is easy to show that the $PT$ symmetry is exact which verify that not only the lowest energy but the full energy spectrum is real [@aboinner]. However, what makes this field theory very impressive is that it is asymptotically free [@Symanzik; @bendf; @Frieder]. Moreover, we discovered that the Hermitian model $\left( g\phi^{4}+h\phi
^{6}\right) $ can have a transition to a phase for which the theory is physically acceptable though non-Hermitian [@abophi61p1]. In fact, we asserted that this model has a very interesting property namely, matter-antimatter asymmetry, which is the first time to be obtained in a spontaneously symmetry breaking regime. Relying on such interesting properties of $\mathcal{PT}$-symmetric and non-Hermitian models, one may ask if such new subject can help in solving the above mentioned problems in the standard model. In this letter, we answer this question by renormalization group analysis of the $(-\frac{\lambda}{4!}\phi^{4})$ scalar field model in $3+1$ dimension. In fact, the idea we relied on is that the theory is asymptotically free and conclusions drawn from this model can be generalized to the more reliable complex scalar field (doublets) that is used to break the $SU(2)\times U(1)$ symmetry in the standard model.
In the $\phi^{4}$ model in $3+1$ dimension, which is used to break the $SU(2)\times U(1)$ symmetry in the standard model, up to one loop, we realize that the mass term receives a correction of the form; $$M_{H}^{2}=M_{0}^{2}+\frac{3\lambda\Lambda^{2}}{8\pi^{2}},$$ where the $M_{H}^{2}$ and $M_{0}^{2}$ are the renormalized and the bare mass squared of the Higgs particle while $\lambda$ is the coupling constant. In fact, the appearance of the momentum cutoff $\Lambda$ is the reason behind the Hierarchy problem, which leads to the introduction of the supersymmetry where the cutoff $\Lambda$ from a Boson and a Fermion loops cancels. Instated of the SUSY additive cancellation, one can guess a multiplication softening of the Hierarchy problem. By this we mean, if we have an asymptotically free Higgs particle, when $\Lambda$ is very high, the coupling is very small and thus one may expect that the perturbative correction $\frac{3\lambda\Lambda^{2}}{8\pi^{2}}$ stays small. To test this idea, consider the renormalization group functions of the $\mathcal{PT}$-symmetric and non-Hermitian $(-\frac{\lambda
}{4!}\phi^{4})$ scalar field model in $3+1$ dimensions up to second order in the coupling; $$\begin{aligned}
\beta\left( \lambda\right) & =-\frac{3\lambda^{2}}{\left( 4\pi\right)
^{2}},\\
\gamma_{m}\left( \lambda\right) & =\frac{-\lambda}{\left( 4\pi\right)
^{2}}-\frac{5}{6}\left( \frac{\lambda}{\left( 4\pi\right) ^{2}}\right)
^{2},\end{aligned}$$ where $\beta\left( \lambda\right) =\mu\frac{d\lambda}{d\mu}$ and $\gamma
_{m}\left( \lambda\right) =\frac{\mu}{m}\frac{dm}{d\mu}$ are the well known renormalization group functions for the flow of the coupling and the mass parameters. Accordingly, the mass parameter can be given by;$$\begin{aligned}
m^{2}\left( \mu\right) & =m^{2}\left( \mu_{0}\right) \exp\left(
\int_{\lambda_{0}}^{\lambda_{\mu}}\frac{\gamma_{m}\left( \lambda\right)
}{\beta\left( \lambda\right) }d\lambda\right) \\
& =m^{2}\left( \mu_{0}\right) \exp\left( -\frac{1}{288}\frac
{-5\lambda_{\mu}-96\left( \ln\frac{\lambda_{\mu}}{\lambda_{0}}\right)
\pi^{2}+5\lambda_{0}}{\pi^{2}}\allowbreak\right) .\end{aligned}$$ In fact, because $\beta\left( \lambda\right) $ is negative it will drag the coupling to the origin as $\mu$ goes to higher values. This behavior is well known as the asymptotic freedom. Accordingly, the Higgs mass will go also to very small values at high energy scales. Thus, if this picture is the successful one in view of the coming analysis from LHC, not only was the quarks, leptons, vector Bosons were massless in the far past but also the Higgs particle was massless. To make the difference between the features of the Hermitian $\phi^{4}$ and the $\mathcal{PT}$-symmetric $\phi^{4}$ more illustrative, we plotted the behavior of the coupling constant as a function of the logarithm of the mass scale $\mu$ in Fig. \[lamdah\] and Fig.\[lamdanonh\], respectively. One can easily realize from the figures that the $\mathcal{PT}$-symmetric $\phi^{4}$ theory is non-trivial and asymptotically free while the Hermitian $\phi^{4}$ is both trivial and the coupling blows up for $UV$ scales which causes the hierarchy problem. In fact, the main message we want to spreed in this letter is that (i) there is no gauge Hierarchy problem with the non-Hermitian and $\mathcal{PT}$-symmetric Higgs mechanism, provided that the contribution to the renormalization group functions from other sectors in the standard model will not affect the asymptotic freedom property of the scalar sector (ii) the model is non-trivial (iii) the technical problem concerning the remedy of the indefinite metric of the theory in the conventional Hilbert space may be solved via a simple Bogoliubov transformation for which the new representation is Hermitian and thus the theory is unitary though of indefinite norm which we will try to attribute it to the existence of a new physical meaning. In fact, it seems that we are obligated to follow that route as the existing regimes for handling the $\mathcal{PT}$-symmetric theories introduces a metric operator of the form $\eta_{+}=\int d^{3}x e^{-Q(x)}$ which is expected to introduce non-renormalizable terms in the equivalent Hermitian theory, in case we are able to calculate the metric operator for this theory in $3+1$ dimensions [^2]. We will investigate this point later on in this work.
Now, we may speculate about if this is the correct picture, why we did not discovered the Higgs particle yet? The answer to this question may be that, in this picture the Higgs particle is a strongly interacting particle and one can not isolate a single Higgs, the same way of behavior of quarks and gluons. Then, one may instead talk about bound states of Higgs which we call it Higgs balls ( like glue balls). Thus, according to this picture, the discovery of the Higgs is not a matter of building bigger and bigger machines for the sake of higher and higher energies but a matter of our understanding of the nature of the Higgs particle.
Although of the above mentioned amazing features of the non-Hermitian and $\mathcal{PT}$-symmetric $\phi^{4}$ theory toward the solution of the genuine problems in the standard model, there exists a technical problem concerning the expected negative norm of the theory. In fact, in the Hilbert space with the Dirac sense inner product operation, the theory have a positive norm but unitarity is violated. This led Bender *et.al* to suggest the $PT$ inner product of the form [@bend2005]; $$\langle A|B\rangle_{PT}=(PT|A\rangle)^{T}|B\rangle.$$ This suggestion preserves unitarity but the Hilbert space with the $PT$ inner product has an indefinite metric. Again, this led Bender et.al. to introduce what is called the $C$ operator and the Hilbert space with the $CPT$ inner product preserves unitarity as well as having a positive definite metric. However, the calculation of the $C$ operator for the model under consideration is not that easy and will turn the computation cumbersome. Although this is a technical problem and not a conceptual one, up to the best of our knowledge, no body has obtained the $C$ operator for the non-Hermitian and $\mathcal{PT}$-symmetric $\phi^{4}$ theory. However, another (equivalent) regime to cure the indefinite metric problem has been suggested by Mostafazadeh [@zadah]. What is important to us here from the Mostafazadeh regime is that the non-Hermitian form of the Hamiltonian is nothing but another representation of an equivalent Hermitian representation and one can (in principle) switch between the two representations via a similarity transformation. However, the metric operator in the Mostafazadeh regime is hard to get in the non-real line theories especially in quantum field models. In fact, both Bender and Mostafazadeh regimes will lead to a dynamical Feynmann rules in the sense that the Feynmann rules will change from order to order because of the new operators introduced to the theory in the definition of the positive definite inner product. Accordingly, one may ask if there exist a simple similarity transformation by which one can obtain an equivalent Hermitian theory and thus having a Hilbert space with the conventional Dirac sense inner product. We will try to answer this question in the following part of the work.
The reality of the spectrum of a non-Hermitian $\mathcal{PT}$-symmetric Hamiltonian demands an exact $PT$ symmetry in the sense that all the wave functions respect the $PT$ symmetry in the same footing as the Hamiltonian itself. In the above discussions, we did not check if the states have an exact $PT$ symmetry or not. To show that, let $|n(k)\rangle$ is a state consisting of $n(k)$ identical particles with momentum $k$. Because the field is assumed to transform as a pseudo scalar under $PT$ transformations we get the transformation of the creation and annihilation operators as; $$PTa(PT)^{-1}=PT\left( i\int dx\left( \exp\left( -ikx\right) \pi
-\phi\overleftrightarrow{\left( \frac{\partial}{\partial t}\right) }\exp\left( ikx\right) \right) \right) (PT)^{-1}=-a,$$ where $\overleftrightarrow{\left( \frac{\partial}{\partial t}\right)
}\left( AB\right) =A\frac{\partial B}{\partial t}-\frac{\partial A}{\partial
t}B$.Also, $PTa^{\dagger}(PT)^{-1}=-a^{\dagger}$.Since, $$\begin{aligned}
|n(k)\rangle & =\frac{a^{\dagger n(k)}(k)|0\rangle}{\sqrt{n(k)}!},\\
PT|n(k)\rangle & =\left( -1\right) ^{n(k)}|n(k)\rangle,\end{aligned}$$ where we observe that $\mathcal{PT}$-symmetry is broken. To keep the $PT$ symmetry unbroken, we add the famous extra $i^{n}$ factor to the states in the following way; $$\begin{aligned}
|n(k)\rangle & =\frac{i^{n(k)}a^{\dagger n(k)}(k)|0\rangle}{\sqrt{n(k)}!},\\
PT|n(k)\rangle & =|n(k)\rangle.\end{aligned}$$ Also, a state consisting of many particles of different momenta, can be represented by $$\begin{aligned}
|n(k_{1})n(k_{1})n(k_{3})n(k_{4})....n(k_{m})\rangle & =\prod_{j=1}^{j=m}\frac{i^{n(k_{j})}a^{\dagger n(k_{j})}(k_{j})|0\rangle}{\sqrt{n(k_{j})}!},\\
PT|n(k)\rangle & =|n(k)\rangle.\end{aligned}$$ In this way one can build up states which are $\mathcal{PT}$-symmetric too and thus the $\mathcal{PT}$ symmetry is exact which is an essential requirement for the reality of the $\mathcal{PT}$-symmetric Hamiltonian.
Now consider the Hamiltonian model of the form;$$H=H=\frac{1}{2}\left( \left( \nabla\phi\right) ^{2}+\pi^{2}+m^{2}\phi
^{2}\right) -\frac{\lambda}{4!}\phi^{4},$$ where $\pi$ is the conjugate momentum. Making the field shift $\phi=\psi+B$; one can rewrite the Hamiltonian density in the form $$H=H_{0}+H_{1}+H_{I}+E, \label{quasi}$$ where $$\begin{aligned}
H_{0} & =\frac{1}{2}\left( \left( \nabla\psi\right) ^{2}+\Pi^{2}+M^{2}\psi^{2}\right) ,\\
H_{1} & =\left( m^{2}-\frac{\lambda}{6}B^{2}\right) B\psi,\\
H_{I} & =\frac{-\lambda}{4!}\left( \psi^{4}+4B\psi^{3}\right) ,\\
E & =\frac{1}{2}m^{2}B^{2}-\frac{\lambda}{4!}B^{4},\end{aligned}$$ where $B$ is the vacuum condensate, $\Pi=\overset{\cdot}{\psi}$ and $M^{2}=m^{2}-\frac{\lambda}{8}B^{2}$. A well known condition for the effective potential $E$ is to satisfy the equality $\frac{\partial E}{\partial B}=0$. Accordingly, the term $H_{1}$ is ought to be zero. Thus, the quasi-particle Hamiltonian in Eq.(\[quasi\]) has the form $$H=\int d^{3}x\left( \frac{1}{2}\left( \left( \nabla\psi\right) ^{2}+\Pi^{2}+M^{2}\psi^{2}\right) -i\frac{\lambda\left\vert B\right\vert }{6}\psi^{3}-\frac{\lambda}{4!}\psi^{4}\right) , \label{quasi2}$$ where we used the fact that the vacuum condensate of this theory is pure imaginary [@abophi41p1; @onep]. Accordingly, the theory is non-Hermitian but $\mathcal{PT}$-symmetric and thus is physically acceptable. However, one of the essential problems opposing this theory is that the metric operator has not been obtained so far. In fact, the form in Eq.(\[quasi2\]) enables us to apply the conventional tools to calculate its metric operator at least in a perturbative way. To show this consider consider the non-Hermitian Hamiltonian in Eq.(\[quasi2\]). Mostafazadeh generalized the requirement of real spectra for a non-Hermitian theory to the existence of a positive definite metric operator $\eta$ such that $\eta_{+}H\eta_{+}^{-1}=H^{\dagger}$ [@spect; @spect1] with an equivalent hermitian theory $h$ such that
$\rho H\rho^{-1}=$ $h$ with $\rho=\sqrt{\eta}=\exp\left( -\frac{Q}{2}\right)
$ where $\eta=\exp\left( -Q\right) $. Accordingly, we can get $$\begin{aligned}
H^{\dagger} & =\exp(-Q)H\exp(Q)=H+[-Q,H]+[-Q,[-Q,H]]\\
& +[-Q,[-Q,[-Q,H]]]+....\end{aligned}$$ Also, one has a similar expansion for the Hermitian Hamiltonian $h=\exp
(\frac{-Q}{2})H\exp(\frac{Q}{2})$, which will result in a perturpative expansion for $h$ as $$h=h_{0}+gh_{1}+g^{2}h_{2}+.....$$
Now, we have for $H^{\dagger}$ the expansion; $$\begin{aligned}
\exp(-Q)H\exp(Q) & =H_{0}+gH_{I}+[-Q,H_{0}]+[-Q,gH_{I}]+[-Q,[-Q,H_{0}]]+\\
& [-Q,[-Q,gH_{I}]]+[-Q,[-Q,[-Q,H_{0}]]+[-Q,[-Q,[-Q,gH_{I}]]...\\
& =H_{0}+gH_{I}^{\dagger},\end{aligned}$$ with $$Q=Q_{0}+gQ_{1}+g^{2}Q_{2}++g^{3}Q_{3}+......$$
Thus, we get a set of coupled equations for the operators $Q_{n}$, where the first few equations are given by $$\begin{aligned}
0 & =[-Q_{0},H_{0}]\text{ }\Rightarrow\text{ }Q_{0}=0\ \text{ is a good
choice,}\nonumber\\
H_{I}^{\dagger} & -H_{I}=-2ig\int d^{3}x\psi^{3}=-\frac{1}{2}[-Q_{1},H_{0}],\nonumber\\
0 & =\frac{1}{2}[-Q_{2},H_{0}]+\frac{1}{2}[-Q_{1},H_{I}]+\frac{1}{3!}[Q_{1},[Q_{1},H_{0}]],\nonumber\\
0 & =\frac{1}{2}[-Q_{3},H_{0}]+\frac{1}{2}[-Q_{2},H_{I}]+\frac{1}{3!}[Q_{2},[Q_{1},H_{0}]]\nonumber\\
& +\frac{1}{3!}[Q_{1},[Q_{2},H_{0}]]+\frac{1}{4!}[-Q_{1},[-Q_{1},[-Q_{1},H_{0}]]]\label{qpert1}\\
& +\frac{1}{3!}[-Q_{1},[-Q_{1},H_{I}]],\nonumber\\
0 & =\frac{1}{2}[-Q_{4},H_{0}]+\frac{1}{4}[-Q_{3},H_{I}]+\frac{1}{3!}[-Q_{2},[-Q_{2},H_{0}]]\nonumber\\
& +\frac{1}{5!}[Q_{1},[Q_{1},[Q_{1},[Q_{1},H_{0}]]]]+\frac{1}{3!}[-Q_{2},[-Q_{1},H_{I}]\nonumber\\
& +\frac{1}{3!}[-Q_{1},[-Q_{2},H_{I}]+\frac{1}{4!}[-Q_{1},[-Q_{1},[-Q_{1},H_{I}]]]]\nonumber\\
& +\frac{1}{8\times4!}[-Q_{1},[-Q_{1},[-Q_{2},H_{0}]]]]\nonumber\\
& +[-Q_{1},[-Q_{2},[-Q_{1},H_{0}]]]]+[-Q_{2},[-Q_{1},[-Q_{1},H_{0}]]]].\nonumber\end{aligned}$$
In fact, this regime has been used before to calculate the $Q$ operator for a Hamiltonian form that is similar to the effective form of the $-\frac{\lambda
}{4!}\phi^{4}$ theory in Eq.(\[quasi\]) in $0+1$ dimensions [@bendx4q]. However, one can expect that the equivalent Hermitian theory is non-renormalizable. Moreover, following the work in Ref.[@cop], one can obtain the $Q$ operator up to first order for the more simpler $i\phi^{3}$ theory which will take the form; $$Q_{1}=\int\int\int d^{3}xd^{3}yd^{3}z\left( M_{(xyz)}\Pi(x)\Pi(y)\Pi
(z)+N_{x(yz)}\psi(y)\Pi(x)\psi(z)\right) ,$$ where the functions $M_{(xyz)}$ and $N_{x(yz)}$ are defined there. However, the resulting Hermitian Hamiltonian $h$ have terms for which the coupling has a negative mass dimension and the situation will be worst in higher orders as more powers of both the fields $\Pi$ and $\psi$ are appearing in the operator $Q$ and thus in the Hermitian Hamiltonian $h$. In other words, the transformation $\rho H\rho^{-1}=$ $h$ does not respect the superficial degree of divergence [@abonon]. In fact, this was the reason that leads the authors of Ref.[@ptsym] to have a Hermitian theory which is incalculable equivalent to the calculable non-Hermitian $ix^{3}$ theory. Relying on these analysis, we expect that this is will be the case for the $-\phi^{4}$ theory and the calculation of the metric operator which depends on the field and its conjugate momentum is thus meaningless.
Now, we have a theory (the $-\lambda\phi^{4}$ theory) which shows up interesting behaviors like asymptotic freedom and it seems that it is free from the hierarchy problem. However, the theory seems to be incalculable as well. To escape from this puzzle one can seek another metric operator for which the Hamiltonian is pseudo-Hermitian and leading to an equivalent Hermitian form though with a wrong sign propagator. In fact, one may gausses the parity operator. To do that, we take into account that the quasi-particle field $\psi$ is pseudo scalar and thus $$PHP^{-1}=H^{\dag},$$ where $P$ is the parity operator. Then, one can introduce the operator $\rho$ such that $P=\rho^{2}$. In this case, we have the following transformation properties$$\rho\psi\rho^{-1}=-i\psi\text{, \ }\rho\Pi\rho^{-1}=i\Pi.\text{\ \ }$$ Note that both $P$ and $\rho$ preserve the commutation relations of the field operators $\left[ \psi(x),\Pi(y)\right] =i\delta^{3}(x-y)$. In view of this, one can obtain $$h=\rho H\rho^{-1}=H=\int d^{3}x\left( \frac{-1}{2}\left( \left( \nabla
\psi\right) ^{2}+\Pi^{2}+M^{2}\psi^{2}\right) -i\frac{\lambda\left\vert
B\right\vert }{6}\psi^{3}-\frac{\lambda}{4!}\psi^{4}\right) .\label{hermn}$$ Note that $h$ is Hermitian but the propagator has an opposite sign to the conventional $\phi^{4}$ theory. Moreover, all the Feynman diagrams calculated either with $h$ or $H$ have the same numerical value as it should be. Also, both $h$ and $H$ are normalizable theories rather than the expected Hermitian Hamiltonian obtained from $\exp(-Q)H\exp(Q)$, with $Q$ is a functional in $\psi$ and $\Pi$ fields and calculated from the coupled set in Eq.(\[qpert1\]). Now, the Hamiltonian has ghost states. However, one can attribute this to a new physical meaning. To show this, consider the propagator $\frac{-i}{p^{2}-M^{2}}$, in multiplying by $M$ we obtain a new propagator of the form $$\frac{-iM}{p^{2}-M^{2}}=\frac{-i}{\frac{p^{2}}{M}-M},$$ which can be considered as the matter probability amplitude with the new hypothesis that matter density can be negative or positive. Accordingly, the wrong sign appears in the propagator can be attributed to a particle of negative mass *i.e.* made of exotic matter, provided that the jump from the non-relativistic quantum mechanics to the relativistic quantum mechanics has done via $\frac{p^{2}}{2M}\rightarrow\frac{p^{\mu}p_{\mu}}{2M}$. Or equivalently, have the klein-Gordon equation of the form; $$\left( \frac{\nabla^{2}-\frac{\partial^{2}}{\partial t^{2}}}{2m}-\frac{m}{2}\right) \psi=0,$$ which is the same for both positive and negative values of $m$. In the presence of interactions, there exists two different Klein-Gordon equations one for positive $m$ and another one for negative $m$. However, for a negative mass particle, the quantum field Hamiltonian for the theory under consideration have the form;$$h=\rho H\rho^{-1}=H=\int d^{3}x\left( \frac{-1}{2}\left( \frac{\left(
\nabla\psi\right) ^{2}+\Pi^{2}}{m}+m\psi^{2}\right) -i\frac{\lambda
\left\vert B\right\vert }{6}\psi^{3}-\frac{\lambda}{4!}\psi^{4}\right) ,$$ and in this way the negative sign can attributed to a theory of negative mass. Note that, this form is Hermitian and thus the theory is unitary in the Dirac sense inner product.
In conclusion, we showed that the non-Hermitian and $\mathcal{PT}$-symmetric $\phi^{4}$ theory has very interesting features as an asymptotically free theory. The most important feature of the theory is that the parameters of the theory are perturbative for UV scales rather than the corresponding Hermitian theory used to break the $SU(2)\times U(1)$ symmetry in the standard model. This suggests that using the negative coupling $\phi^{4}$ theory instead, will solve many problems in the standard model. Out of this problems, is that the negative coupling $\phi^{4}$ is not trivial as it has interactions allover the energy scale because of the asymptotic freedom property. Also, it might save the standard model if the experiments in the LHC was not able to detect the Higgs particle in a direct manner. Our reasoning is that in this picture the Higgs particle is a strongly interacting particle and need an infinite amount of energy to be isolated. Thus, according to this picture, the search of the Higgs has to be twisted to go the same way we feel the gluons.
A note to be mentioned is that this work does not give a final answer about the solution of the standard model problems like the Hierarchy and triviality problems. This is because, the Higgs mass receives other corrections from the vector Boson fields coupled to the Higgs field and the top quark contribution should also be taken into account. Taking this into account yield the result [@DJOUADI] $$M_{H}^{2}=M_{0}^{2}-\frac{\Lambda^{2}}{8\pi^{2}v^{2}}[M_{H}^{2}+2M_{w}^{2}+M_{Z}^{2}-4M_{t}^{2}],$$ where $M_{0}$ is the bare mass contained in the unrenormalized Lagrangian. By the renormalization group analysis mentioned above, we made sure that the first term will be small as $\Lambda$ goes to higher values. For other terms, we know that all the masses in the standard model depend on the vacuum condensate which has been shown to have an exponential sharp decrease near $\lambda\rightarrow0^{+}$ [@onep]. Accordingly, one may claim that the Higgs mass will stay protected against perturbations even for high energy scales. However, this claim should be tested in view of the renormalization group functions for the other sectors. In fact, this will take a substantial amount of time but it naturally becomes a target of our future work. The main aim we wanted by this work to spread the message that non-Hermitian and $\mathcal{PT}$-Symmetric $\phi^{4}$ as now a physically acceptable model may help in the solution of the genuine existing problems in the standard model. These problems are well known to exist because the Hermitian Higgs mechanism used is both trivial and suffers from the gauge Hierarchy problem. Another important message that we need to mention is the that the vacuum energy of the non-Hermitian and $\mathcal{PT}$-Symmetric $\phi^{4}$ is tiny and negative in $1+1$ dimensions compared to the Hermitian one [@abophi41p1; @thesis]. In another work [@cern], we showed that in $2+1$ dimensions, the vacuum energy is tiny as well as positive for a wide range of energy scales. In fact, these features are very interesting regarding the very false result of the prediction of the cosmological constant from quantum field theory which again is a manifestation of the gauge Hierarchy problem. In fact, positiveness of the vacuum energy is essential as it describes an expanding universe ( negative pressure). We hope that using the asymptotically free non-Hermitian and $\mathcal{PT}$-symmetric $\phi^{4}$ theory in the standard model will solve such genuine problems relying on the interesting features we explored above.
Since up till now the metric operator is not known, we tried to give the negative norm a physical meaning by considering the Higgs mass as a charge which can be positive or negative. We assert that, the form $\eta=\exp(-Q)$ can be obtained perturbatively using effective field representation. However, we did not prefer this direction as the resulting Hermitian Hamiltonian will be non-Renormalizable and thus incalculable.
A Higgs particle with negative mass makes sense in understanding how a potential bounded from above can have stable states. In this case, negative mass particles have the property of maximizing the classical action rather than the conventional positive mass particles which minimizing the classical action. In other words, a negative mass particle feels the bounded from above potential as the positive mass particle feels the bounded from below potentials.
![The dependence of the renormalized coupling $\lambda(\mu)$ on the mass scale $\mu$ for the Hermitian $\phi^{4}$ theory. In this figure, the theory is shown to be trivial for $IR$ scales, while the coupling blows up for $UV$ scales which means that the mass parameter will explode non-logarithmically at $UV$ scales causing the Hierarchy problem.[]{data-label="lamdah"}](lamdah__1.eps)
![The dependence of the renormalized coupling $\lambda(\mu)$on the mass scale $\mu$ for the non-Hermitian $\phi^{4}$ theory. One can realize that the theory is non-trivial as well as asymptotically free. Accordingly, the mass is finite for ultra $UV$ scales.[]{data-label="lamdanonh"}](lamdanonh__1.eps)
[99]{}
Physics Motivations for Future CERN Accelerators,Albert De Roeck, John Ellis, Fabiola Gianotti , Report presented to CERN council in Dec. 2001, hep-ex/0112004(2004).
Carl Bender and Stefan Boettcher, Phys.Rev.Lett.80:5243-5246 (1998).
Abouzeid shalaby, Eur.Phys.J.C50:999-1006 (2007).
Abouzeid shalaby, Phys.Rev.D76:041702 (2007 ).
Abouzeid shalaby, Hermiticity Breaking and Restoration in the $(g\phi^{4}+h\phi^{6})_{1+1}$ Field Theoretic Model (submitted to the Journal of Physics A).
K. Symanzik, Commun. Math. Phys. 45, 79 (1975).
C. M. Bender, K. A. Milton, and V. M. Savage, Phys. Rev. D 62, 85001 (2000).
Frieder Kleefeld, J. Phys. A: Math. Gen. 39 L9–L15 (2006). Carl M. Bender, Jun-Hua Chen, Kimball A. Milton, J.Phys.A39, 1657 (2006). A. Mostafazadeh, J. Phys. A: Math. Gen. 38, 6557 (2005) and Erratum 38, 8185 (2005). Carl Bender, Peter N. Meisinger and Haitang Yang, Phys.Rev.D63:045001 (2001). A. Mostafazadeh, J. Math. Phys., 43, 3944 (2002).
A. Mostafazadeh, J. Math. Phys. 43, 205 (2002). Carl M. Bender , Dorje C. Brody and Hugh F. Jones, Phys.Rev.D73:025002 (2006 ). Carl M. Bender, Dorje C. Brody and Hugh F. Jones, Phys.Rev. D70, 025001 (2004). Is Non-Renormalizability a Technical or a Conceptual Problem? A Clue from Quasi-Hermitian Representations; hep-0805.1247.
Carl M. Bender, Jun-Hua Chen and Kimball A. Milton, J.Phys. A39, 1657-1668 (2006). Abdelhak DJOUADI, The Anatomy of Electro–Weak Symmetry Breaking, arXiv:hep-ph/0503172. Non-Trivial Vacuum Solutions of Low Dimensional Scalar Field Theories in the Oscillator Representation Method, Thesis by Abouzeid Shalaby,http://www.lib.ncsu.edu/theses/available/etd-08122004-160155/. Abouzeid shalaby, Toward A Strongly Interacting Scalar Higgs Particle, UAE-CERN Workshop, High Energy Physics and Applications, 26-28 November 2007, Al-Ain, United Arab Emirates.
[^1]: E-mail:amshalab@ mans.eg.edu
[^2]: Up to the best our knowledge, the metric operator for the $\mathcal{PT}$-symmetric ($-\phi^{4}$) theory has never been obtained before.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Complex systems are often inherently non-ergodic and non-Markovian for which Shannon entropy loses its applicability. In particular accelerating, path-dependent, and aging random walks offer an intuitive picture for these non-ergodic and non-Markovian systems. It was shown that the entropy of non-ergodic systems can still be derived from three of the Shannon-Khinchin axioms, and by violating the fourth – the so-called composition axiom. The corresponding entropy is of the form $S_{c,d} \sim \sum_i \Gamma(1+d,1-c\ln p_i)$ and depends on two system-specific scaling exponents, $c$ and $d$. This entropy contains many recently proposed entropy functionals as special cases, including Shannon and Tsallis entropy. It was shown that this entropy is relevant for a special class of non-Markovian random walks. In this work we generalize these walks to a much wider class of stochastic systems that can be characterized as ‘aging’ systems. These are systems whose transition rates between states are path- and time-dependent. We show that for particular aging walks $S_{c,d}$ is again the correct extensive entropy. Before the central part of the paper we review the concept of $(c,d)$-entropy in a self-contained way.'
author:
- 'Rudolf Hanel$^1$ and Stefan Thurner$^{1,2,3*}$'
title: 'Generalized (c,d)-entropy and aging random walks'
---
Introduction - mini-review of $(c,d)$-entropy
=============================================
In their seminal works, Shannon and Khinchin showed that assuming four information theoretic axioms the entropy must be of Boltzmann-Gibbs type, $S=-\sum_i p_i \log p_i$. In many physical systems one of these axioms may be violated. For non-ergodic systems the so called separation axiom (Shannon-Khinchin axiom 4) is not valid. We show that whenever this axiom is violated the entropy takes a more general form, $S_{c,d}\propto \sum_i ^W \Gamma(d+1, 1- c \log p_i)$, where $c$ and $d$ are scaling exponents and $\Gamma(a,b)$ is the incomplete gamma function. These exponents $(c,d)$ define equivalence classes for [*all*]{}!, interacting and non interacting, systems and unambiguously characterize any statistical system in its thermodynamic limit. The proof is possible because of two newly discovered scaling laws which any entropic form has to fulfill, if the first three Shannon-Khinchin axioms hold [@Hanel2011]. $(c,d)$ can be used to define equivalence classes of statistical systems. A series of known entropies can be classified in terms of these equivalence classes. We show that the corresponding distribution functions are special forms of Lambert-${\cal W}$ exponentials containing – as special cases – Boltzmann, stretched exponential, and Tsallis distributions (power-laws). We go on by showing how the dependence of phase space volume $W(N)$ of a classical system on its size $N$, uniquely determines its extensive entropy, and in particular that the requirement of extensively fixes the exponents $(c,d)$, [@Hanel2011_2]. We give a concise criterion when this entropy is not of Boltzmann-Gibbs type but has to assume a [*generalized*]{} (non-additive) form. We showed that generalized entropies can only exist when the dynamically (statistically) relevant fraction of degrees of freedom in the system vanishes in the thermodynamic limit [@Hanel2011_2]. These are systems where the bulk of the degrees of freedom is frozen and is practically statistically inactive. Systems governed by generalized entropies are therefore systems whose phase space volume effectively collapses to a lower-dimensional ’surface’. We explicitly illustrated the situation for binomial processes and argue that generalized entropies could be relevant for self organized critical systems such as sand piles, for spin systems which form meta-structures such as vortices, domains, instantons, etc., and for problems associated with anomalous diffusion [@Hanel2011_2]. In this contribution we largely follow the lines of thought presented in [@Hanel2011; @Hanel2011_2; @ebook].
Theorem number 2 in the seminal 1948 paper, [*The Mathematical Theory of Communication*]{} [@shannon], by Claude Shannon, proves the existence of the one and only form of entropy, given that three fundamental requirements hold. A few years later A.I. Khinchin remarked in his [*Mathematical Foundations of Information Theory*]{} [@kinchin_1]: “However, Shannon’s treatment is not always sufficiently complete and mathematically correct so that, besides having to free the theory from practical details, in many instances I have amplified and changed both the statement of definitions and the statement of proofs of theorems.” Khinchin adds a fourth axiom. The three fundamental requirements of Shannon, in the ‘amplified’ version of Khinchin, are known as the Shannon-Khinchin (SK) axioms. These axioms list the requirements needed for an entropy to be a reasonable measure of the ‘uncertainty’ about a finite probabilistic system. Khinchin further suggests to also use entropy as a measure of the information [*gained*]{} about a system when making an ’experiment’, i.e. by observing a realization of the probabilistic system.
$\bullet$ Khinchin’s first axiom states that for a system with $W$ potential outcomes (states) each of which is given by a probability $p_i\geq0$, with $\sum_{i=1}^W p_i=1$, the entropy $S(p_1, \cdots, p_W)$ as a measure of uncertainty about the system must take its maximum for the equi-distribution $p_i=1/W$, for all $i$.
$\bullet$ Khinchin’s second axiom (missing in [@shannon]) states that any entropy should remain invariant under adding zero-probability states to the system, i.e. $S(p_1, \cdots, p_W)=S(p_1, \cdots, p_W,0)$.
$\bullet$ Khinchin’s third axiom (separability axiom) finally makes a statement of the composition of two finite probabilistic systems $A$ and $B$. If the systems are independent of each other, entropy should be additive, meaning that the entropy of the combined system $A+B$ should be the sum of the individual systems, $S({A+B}) = S(A) + S(B)$. If the two systems are dependent on each other, the entropy of the combined system, i.e. the information given by the realization of the two finite schemes $A$ and $B$, $S(A+B)$, is equal to the information gained by a realization of system $A$, $S(A)$, plus the mathematical expectation of information gained by a realization of system $B$, after the realization of system $A$, $S({A+B}) = S(A) + S|_A(B)$.
$\bullet$ Khinchin’s fourth axiom is the requirement that entropy is a continuous function of all its arguments $p_i$ and does not depend on anything else.
Given these axioms, the [*Uniqueness theorem*]{} [@kinchin_1] states that the one and only possible entropy is $$S(p_1,\cdots , p_W) = -k \sum_{i=1}^{W}p_i\log p_i \quad ,$$ where $k$ is an arbitrary positive constant. The result is of course the same as Shannon’s. We call the combination of 4 axioms the Shannon-Khinchin (SK) axioms.
From information theory now to physics, where systems may exist that violate the separability axiom. This might especially be the case for non-ergodic, complex systems exhibiting long-range and strong interactions. Such complex systems may show extremely rich behavior in contrast to simple ones, such as gases. There exists some hope that it should be possible to understand such systems also on a thermodynamical basis, meaning that a few measurable quantities would be sufficient to understand their macroscopic phenomena. If this would be possible, through an equivalent to the second law of thermodynamics, some appropriate entropy would enter as a fundamental concept relating the number of microstates in the system to its macroscopic properties. Guided by this hope, a series of so called generalized entropies have been suggested over the past decades, see [@tsallis88; @celia; @kaniadakis; @curado; @expo_ent; @ggent] and Table 1. These entropies have been designed for different purposes and have not been related to a fundamental origin. Here we ask how generalized entropies can look like if they fulfill some of the Shannon-Khinchin axioms, but explicitly violate the separability axiom. We do this axiomatically as first presented in [@Hanel2011]. By doing so we can relate a large class of generalized entropies to a single fundamental origin.
The reason why this axiom is violated in some physical, biological or social systems is [*broken ergodicity*]{}, i.e. that not all regions in phase space are visited and many micro states are effectively ‘forbidden’. Entropy relates the number of micro states of a system to an [*extensive*]{} quantity, which plays the fundamental role in the systems thermodynamical description. Extensive means that if two initially isolated, i.e. sufficiently separated systems, $A$ and $B$, with $W_A$ and $W_B$ the respective numbers of states, are brought together, the entropy of the combined system $A+B$ is $S(W_{A+B}) = S(W_A) + S(W_B)$. $W_{A+B}$ is the number of states in the combined system $A+B$. This is not to be confused with [*additivity*]{} which is the property that $S(W_A W_B) = S(W_A) + S(W_B)$. Both, extensivity and additivity coincide if number of states in the combined system is $W_{A+B}=W_AW_B$. Clearly, for a non-interacting system Boltzmann-Gibbs-Shannon entropy, $S_{\rm BG}[p]= - \sum_i^W p_i\ln p_i$, is extensive [*and*]{} additive. By ’non-interacting’ (short-range, ergodic, sufficiently mixing, Markovian, ...) systems we mean $W_{A+B}=W_AW_B$. For interacting statistical systems the latter is in general not true; phase space is only partly visited and $W_{A+B} < W_AW_B$. In this case, an additive entropy such as Boltzmann-Gibbs-Shannon can no longer be extensive and vice versa. To ensure extensivity of entropy, an entropic form should be found for the particular interacting statistical systems at hand. These entropic forms are called [*generalized entropies*]{} and usually assume trace form [@tsallis88; @celia; @kaniadakis; @curado; @expo_ent; @ggent] $$S_g[p]=\sum_{i=1}^W g(p_i) \quad ,
\label{S_g}$$ $W$ being the number of states. Obviously not all generalized entropic forms are of this type. Rényi entropy e.g. is of the form $G(\sum_{i}^W g(p_i))$, with $G$ a monotonic function. We use trace forms Eq. (\[S\_g\]) for simplicity. Rényi forms can be studied in exactly the same way as will be shown, however at more technical cost.
Let us revisit the Shannon-Khinchin axioms in the light of generalized entropies of trace form Eq. (\[S\_g\]). Specifically axioms SK1-SK3 (now re-ordered) have implications on the functional form of $g$
- SK1: The requirement that $S$ depends continuously on $p$ implies that $g$ is a continuous function.
- SK2: The requirement that the entropy is maximal for the equi-distribution $p_i=1/W$ (for all $i$) implies that $g$ is a concave function.
- SK3: The requirement that adding a zero-probability state to a system, $W+1$ with $p_{W+1}=0$, does not change the entropy, implies that $g(0)=0$.
- SK4 (separability axiom): The entropy of a system – composed of sub-systems $A$ and $B$ – equals the entropy of $A$ plus the expectation value of the entropy of $B$, conditional on $A$. Note that this also corresponds exactly to Markovian processes.
As mentioned, if SK1 to SK4 hold, the only possible entropy is the Boltzmann-Gibbs-Shannon entropy. We are now going to derive the extensive entropy when the separability axiom SK4 is violated. Obviously this entropy will be more general and should contain BG entropy as a special case.
We now assume that axioms SK1, SK2, SK3 hold, i.e. we restrict ourselves to trace form entropies with $g$ continuous, concave and $g(0)=0$. These systems we call [*admissible*]{} systems. Admissible systems when combined with a maximum entropy principle show remarkably simple mathematical properties [@Hanel2011_b; @Hanel2012_b].
This generalized entropy for (large) admissible statistical systems (SK1-SK3 hold) is derived from two hitherto unexplored fundamental scaling laws of extensive entropies [@Hanel2011]. Both scaling laws are characterized by exponents $c$ and $d$, respectively, which allow to uniquely define equivalence classes of entropies, meaning that two entropies are equivalent in the thermodynamic limit if their exponents $(c,d)$ coincide. Each admissible system belongs to one of these equivalence classes $(c,d)$, [@Hanel2011].
In terms of the exponents $(c,d)$ we showed in [@Hanel2011] that all generalized entropies have the form $$S_{c,d}\propto\sum_i ^W \Gamma(d+1, 1- c \log p_i)
\label{gent}$$ with $\Gamma(a,b)=\int_b^\infty dt\,t^{a-1}\exp(-t)$ the incomplete Gamma-function.
![ Entropies parametrized in the $(c,d)$-plane, with their associated distribution functions. BG entropy corresponds to $(1,1)$, Tsallis entropy to $(c,0)$, and entropies for stretched exponentials to $(1,d>0)$. Entropies leading to distribution functions with compact support, belong to equivalence class $(1,0)$. Figure from [@ebook]. \[classfig\] ](classes_6.eps){width="\columnwidth"}
### Special cases of equivalence classes
Let us look at some specific equivalence classes $(c,d)$
- Boltzmann-Gibbs entropy belongs to the $(c,d)=(1,1)$ class. One gets from Eq. (\[gent\]) $$S_{1,1}[p]= \sum_i g_{1,1}(p_i)= -\sum_i p_i\ln p_i +1
\quad .$$
- Tsallis entropy belongs to the $(c,d)=(c,0)$ class. From Eq. (\[gent\]) and the choice $r=1/(1-c)$ (see below) we get $$\begin{array}{lcl}
S_{c,0}[p] = \sum_i g_{c,0}(p_i)= \frac{1-\sum_i p_i^c}{c-1} +1 \, .
\end{array}$$ Note, that although the [*pointwise*]{} limit $c\to 1$ of Tsallis entropy yields BG entropy, the asymptotic properties $(c,0)$ do [*not*]{} change continuously to $(1,1)$ in this limit! In other words the thermodynamic limit and the limit $c\to 1$ do not commute.
- The entropy related to stretched exponentials [@celia] belongs to the $(c,d)=(1,d)$ classes, see Table 1. As a specific example we compute the $(c,d)=(1,2)$ case, $$S_{1,2}[p]= 2 \left(1-\sum_i p_i \ln p_i \right) + \frac{1}{2}\sum_i p_i\left(\ln p_i \right)^2,
\label{c1d2}$$ leading to a superposition of two entropy terms, the asymptotic behavior being dominated by the second.
Other entropies which are special cases of our scheme are found in Table 1.
Inversely, for any given entropy we are now in the remarkable position to characterize [*all*]{} large SK1-SK3 systems by a pair of two exponents $(c,d)$, see Fig. \[classfig\]. For example, for $g_{\rm BG}(x)=-x\ln(x)$ we have $c=1$, and $d=1$. $S_{\rm BG}$ therefore belongs to the universality class $(c,d)=(1,1)$. For $g_{q}(x)= (x-x^q)/(1-q)$ (Tsallis entropy) and $0<q<1$ one finds $c=q$ and $d=0$, and Tsallis entropy, $S_{q}$, belongs to the universality class $(c,d)=(q,0)$. Other examples are listed in Table 1.
The universality classes $(c,d)$ are equivalence classes with the equivalence relation given by: $g_{\alpha} \equiv g_{\beta} \Leftrightarrow c_\alpha=c_\beta$ and $d_\alpha=d_\beta$. This relation partitions the space of all admissible $g$ into equivalence classes completely specified by the pair $(c,d)$.
Distribution functions
----------------------
Distribution functions associated with our $\Gamma$-entropy, Eq. (\[gent\]), can be derived from so-called generalized logarithms of the entropy. Under the maximum entropy principle (given ordinary constraints) the inverse functions of these logarithms, ${{\cal E}}={\Lambda}^{-1}$, are the distribution functions, $p(\epsilon) = {{\cal E}}_{c,d,r}(-\epsilon)$, where for example $r$ can be chosen $r=(1-c+cd)^{-1}$. One finds [@Hanel2011] $${{\cal E}}_{c,d,r}(x)= e^{ - \frac{d}{1-c} \left[ {\cal W}_k \left( B (1-x/r )^{ \frac{1}{d} } \right) - {\cal W}_k(B) \right] }\, ,
\label{gexp}$$ with the constant $B\equiv \frac{(1-c)r}{1-(1-c)r} \exp \left( \frac{(1-c)r}{1-(1-c)r} \right) $. The function ${\cal W}_k$ is the $k$’th branch of the Lambert-${\cal W}$ function which – as a solution to the equation $x={\cal W}(x)\exp({\cal W}(x))$ – has only two real solutions $W_k$, the branch $k=0$ and branch $k=-1$. Branch $k=0$ covers the classes for $d\geq 0$, branch $k=-1$ those for $d<0$.
### Special cases of distribution functions
It is easy to verify that the class $(c,d)=(1,1)$ leads to Boltzmann distributions, and the class $(c,d)=(c,0)$ yields power-laws, or more precisely, Tsallis distributions i.e. $q$-exponentials. All classes associated with $(c,d)=(1,d)$, for $d>0$ are associated with stretched exponential distributions. Expanding the $k=0$ branch of the Lambert-${\cal W}$ function $W_0(x)\sim x-x^2+\dots$ for $1\gg|x|$, the limit $c\to 1$ is shown to be a stretched exponential. It was shown that $r$ does not effect its asymptotic properties (tail of the distributions), but can be used to incorporate finite size properties of the distribution function for small $x$.
How to determine the exponents $c$ and $d$?
-------------------------------------------
In [@Hanel2011_2] we have shown that the requirement of extensivity determines uniquely both exponents $c$ and $d$. What does extensivity mean? Consider a system with $N$ elements. The number of system configurations (microstates) as a function of $N$ are denoted by $W(N)$. Starting with SK2, $p_i=1/W$ (for all $i$), we have $S_g=\sum_{i=1}^{W} g(p_i) = W g(1/W)$. As mentioned above extensivity for two subsystems $A$ and $B$ means that $$W_{A+B}g\left( 1/W_{A+B} \right) = W_{A}g\left( 1/ W_{A} \right) + W_{B}g\left( 1/ W_{B} \right) \quad .
\label{dum1}$$ Using this equation one can straight forwardly derive the formulas (for details see [@Hanel2011_2]) $$\frac{1}{1-c}= \lim_{N \to \infty} N \frac{ W'(N)}{W(N)} \quad .
\label{cformel}$$ $$d= \lim_{N \to \infty} \log W \left( \frac1N \frac{W}{ W'} +c-1 \right) \quad .
\label{dformel}$$ Here $W'$ means the derivative with respect to $N$.
A note on Rényi-type entropies
------------------------------
Rényi entropy is obtained by relaxing SK4 to the unconditional additivity condition. Following the same scaling idea for Rényi-type entropies, $S=G(\sum_{i=1}^W g(p_i))$, with $G$ and $g$ some functions, one gets $$\lim_{W\to\infty} \frac{S(\lambda W)}{S(W)}= \lim_{s\to\infty} \frac{ G \left( \lambda f_g(\lambda^{-1})s \right) }{G(s)} \quad ,$$ where $f_g(z)=\lim_{x\to 0} g(zx)/g(x)$. The expression $f_G(s)\equiv \lim_{s} G(s y)/G(s)$, provides the starting point for deeper analysis which now gets more involved. In particular, for Rényi entropy with $G(x) \equiv \ln(x)/(1-\alpha)$ and $g(x)\equiv x^{\alpha}$, the asymptotic properties yield the class $(c,d)=(1,1)$, (BG entropy) meaning that Rényi entropy is additive. However, in contrast to the trace form entropies used above, Rényi entropy can be shown to be [*not*]{} Lesche stable, as was observed before [@lesche; @Abe_2002; @Arimitsu; @Kaniadakis_2004; @HT_robustness_1]. All of the $S=\sum_i^W g(p_i)$ entropies can be shown to be Lesche stable, see [@ebook].
Aging random walks
==================
In [@Hanel2011_2] we have discussed a particular type of an accelerating random walk that requires generalized entropy. We first revisit the example of this [*auto-correlated*]{} random walk $x$ and point out that all moments of this random walk are identical to the moments of an [*accelerating*]{} random walk. This means that two processes, where the the first requires a generalized entropy and the second requires Shannon entropy, they both have the same distribution function asymptotically. We then show that auto-correlated random walks are asymptotically equivalent to [*aging*]{} random walks.
Random walks of length $N$ consist of sequences of $N$ decisions $\omega_n$ with $n=1,2,\cdots,N$. Each decision determines whether to take a step of size $\Delta x$ to the left, $\omega_n = -1$, or to the right, $\omega_n =1$, at time $t=n\Delta t$, with a probability $q_+$ and $q_-$. The path $x(N\Delta t)$ is given by $$x(N\Delta t)=\sum_{n=1}^N\omega_n\Delta x \quad .$$ In the following we set the time increment $\Delta t=1$ and the step size $\Delta x=1$.
For the usual random walk each decision $\omega_n$ has no bias for any direction, i.e. $q_{+} =q_{-} =1/2$ and the expectation value $\langle\omega_n\rangle=q_{+} -q_{-} =0$. Further, all decisions are independent, meaning that $\langle \omega_m \omega_n \rangle=\delta_{nm}$, where $\delta_{nm}$ is the Kronecker delta. The number of possible paths $W$ such a random walk can take – its phase-space volume for $N$ steps – is given by $W(N)=2^N$. Using Eq. (\[cformel\]) and Eq. (\[dformel\]) one immediately finds $(c,d)=(1,1)$. Random walks consisting of independent decisions are described by Shannon’s entropy.
Accelerating and auto-correlated random walks
---------------------------------------------
In [@Hanel2011_2] we considered a different type of random walk where again decisions have no a priori bias on the direction of the walk, i.e. $\langle\omega_n\rangle=0$. However, decisions $\omega_n$ and $\omega_m$ are not independent anymore. In particular we considered a constant $0<\alpha\leq 1$ such that $$\langle\omega_m\omega_n\rangle=1\quad {\rm if}\quad z\leq n^\alpha,\,m^\alpha<z+1
\label{autocorcond}$$ for some $z=0,1,2,\cdots$, and $\langle\omega_m\omega_n\rangle=0$ otherwise. This means that the process is correlated with its history, and that after $n$ steps the number of free decisions is given by $z \sim n^{\alpha}$. As the walk progresses it heads persistently in the same direction for approximately $\frac1\alpha n^{1-\alpha}$ steps at the $n$’th step.
![ Example for an auto-correlated random walk that persistently walks in the same direction for $\propto n^{1-\alpha}$ steps ($\alpha=0.5$). \[fig:walk\] ](walk.eps){width="0.75\columnwidth"}
Therefore, the number of possible paths $W$ grows like $W(N)=2^{N^{\alpha}}$ and the random walk has a stretched exponential growth of phase-space volume. Using Eq. (\[cformel\]) and Eq. (\[dformel\]), the universality class of the process belongs to $(c,d)=(1,\frac{1}{\alpha})$. Increasing persistence of a process over time therefore can be seen as the hallmark of processes that follow generalized extensive entropies.
Computing the moments of $x$ one finds that odd moments vanish, $\langle x^{2r+1}(N)\rangle=0$, where $r$ is a natural number, and even moments behave as $$\langle x^{2r}(N)\rangle = \frac{(2r)!}{r!2^r} N^{(2-\alpha)r} \quad .$$ The [*auto-correlated*]{} random walk therefore possesses the same moments as an [*accelerated*]{} random walk, i.e. a random walk with independent decisions $\langle\omega_m\omega_n\rangle=\delta_{nm}$, however with a time-dependent step size $\Delta x(n)=D(n)\Delta x$, that increases proportional to $n^{(1-\alpha)/2}$. Here $D(n)$ is the time-dependent [*‘diffusion constant’*]{} of the process. In particular, the second moment is given by $$\langle x^2(N)\rangle = N^{2-\alpha} \quad .$$ We conclude that observable distribution functions do not necessary tell us which entropy class the process belongs to. In this example the auto-correlated random walk of class $(c,d)=(1,1/\alpha)$ has all moments in common with the accelerated random walk, which is of class $(c,d)=(1,1)$.
Generalization to aging (path-dependent) random walks
-----------------------------------------------------
The above generating rule Eq. (\[autocorcond\]), for incorporating auto-correlations into random walks is somewhat artificial. We now show that it is possible to get a completely analogous auto-correlated behavior by considering [*aging*]{} in the decision process $\omega$. This can be done as follows. Consider a second process $\eta_n$, such that $\omega_{n}=\eta_n\omega_{n-1}$. This process indicates whether at step $n$ the random walk will proceed in the direction of the previous time step ($\eta_n=1$) or whether the walk reverses direction ($\eta_n=-1$). Let $k_+(N)$ ($k_-(N)$) be the number of times that $\eta_n=+1$ ($\eta_n=-1$) for $1\leq n\leq N$, i.e. $k=(k_+,k_-)$ is the histogram of the process $\eta$ up to time step $N$. Aging can now be incorporated by considering conditional probabilities for reversing direction or not. In particular we have $$\begin{aligned}
p(\eta_{n+1}=1|k(n))=\frac{1}{1+\alpha [1+k_+(n)]^{\alpha-1}}\quad ,\nonumber \\
p(\eta_{n+1}=-1|k(n))=\frac{\alpha [1+k_+(n)]^{\alpha-1}}{1+\alpha [1+k_+(n)]^{\alpha-1}}\quad,
\label{transprob}\end{aligned}$$ where $0<\alpha\leq 1$ takes the same numerical values as in the auto-correlated random walk. As a consequence these [*aging*]{} random walks are non-Markovian processes with memory, since the conditional probabilities for making the next decision depend on the entire history of the process. The dependence is such that the process conditions its next decisions on the histogram of decisions made in the past, not on its precise trajectory and again the decisions become increasingly persistent. To handle this type of process analytically is difficult. However, we can demonstrate numerically, that the first three even moments $\langle x^2\rangle$, $\langle x^4\rangle$, and $\langle x^6\rangle$ of the auto-correlated and the aging random walk are identical, and also the number of reversal decisions $k_-$ of both processes asymptotically behave in exactly the same way. This shows that the effective number of different paths, i.e. the phase-space volume, of both processes grows in the same way and therefore the aging random walk belongs to the equivalence class $(c,d)=(1,1/\alpha)$.
It is possible to show that one can arrive at different equivalence class by altering the expression $n^{\alpha}$ in Eq. (\[autocorcond\]). In particular by exchanging $n^{\alpha}$ with $a \log n$ (same for $m$), one arrives at the Tsallis equivalence class $(c,d)=(q,0)$.
![ Comparison of the first three even moments $\langle x^2(N)\rangle$, $\langle x^4(N)\rangle$, and $\langle x^6(N)\rangle$ and the average number of direction reversal-decisions $k_-(N)$, with $1\leq N\leq 50000$ for the auto-correlated random walk (blue lines) and aging random walks (red dashed lines) for values $\alpha=0.2$, $0.5$, and $0.8$. \[fig:comparison\] ](stretched_rw.eps){width="1.1\columnwidth"}
General classes of aging random walks
-------------------------------------
We are now in the position to generalize random aging walks to different classes $(c,d)$ of entropies. This can be done by generalizing the path dependent conditional probabilities of Eq. (\[transprob\]) in the following way: $$\begin{aligned}
p(\eta_{n+1}=1|k(n))=\frac{1}{1+g(k_+(n))}\quad, \nonumber \\
p(\eta_{n+1}=-1|k(n))=\frac{g(k_+(n))}{1+g(k_+(n))}\quad,
\label{transprob2}\end{aligned}$$ where $g(k_+)$ is a monotonically decreasing function ($\lim_{k_+\to\infty}g(k_+)=0$). In the above example $g(k_+)=\alpha [1+k_+(n)]^{\alpha-1}$ corresponds to an aging process in the entropy class $(c,d)=(1,1/\alpha)$. Different choices of the function $g$ will in general lead to different entropy classes $(c,d)$ depending on the asymptotic behavior $k_-(N)$ which corresponds to the effective number of free decisions occurring during the walk and therefore to the way phase-space grows with $N$. Again, a precise analytical analysis of how the choice of $g$ determines $(c,d)$ is complicated and goes beyond the scope of the paper. However, it is known that systems with $0<c<1$ allow only a finite effective number of free decisions, e.g. [@Hanel2011_2; @TMGMS]. This can for instance be achieved with the function $$g(k_+)=\lambda^{-(k_+^\nu)}\,,$$ with $0<\nu\leq 1$ and $\lambda>1$. By using a ‘mean field’ approach and setting $$\frac{dk_+(N)}{dN}=p(+1|k(N))\quad{\rm and}\quad\frac{dk_-(N)}{dN}=p(-1|k(N))$$ one can derive the following asymptotic expression: $$k_-=\frac{1}{\nu}(\log\lambda)^{-\frac{1}{\nu}}\gamma\left(\frac{1}{\nu},k_+^\nu\log\lambda\right)\,,$$ where $\gamma(a,b)=\int_0^b dt t^{a-1}e^{-t}$ is the lower incomplete gamma function. Consequently the effective number of free decisions in this aging walks can be estimated by $k_-(\infty)$. The behavior of $k_-(\infty)$ is shown in Fig. (\[fig:maxnum\]).
![ The maximal number direction reversal decisions in random walks in entropic classes $(c,d)$ with $0<c<1$ for the values $\lambda=1.1$, $1.2$, and $1.3$. \[fig:maxnum\] ](kmininfty.eps){width="\columnwidth"}
The fact that only a finite number, $k_-$, of direction reversal decisions happen during such a random walk leads to a peculiar cross-over phenomenon that can be observed by studying the second moment $\langle x^2(N)\rangle$ of the walk. In particular $\langle x^2(N)\rangle\sim N$ for small $N$. For large $N\gg 1$ the random walk persistently heads into one direction and $\langle x^2(N)\rangle\sim N^2$. At an intermediate range of $N$ that depends on the value of $\lambda$ the behavior $\langle x^2(N)\rangle$ crosses over from $N$ to $N^2$, see Fig. (\[fig:crossover\]). The derive the exact function that relates $\nu$ and $\lambda$ to $c$ and $d$ is beyond the scope of this paper. However, we conjecture that $c=1-\nu$ since $\nu=0$ corresponds to the usual random walk and therefore we require $c=1$ in this case.
It would be desirable to have a comprehensive classfication of aging random walks in terms of equivalence classes $(c,d)$. We conjecture that this is in fact possible by exchanging the expression $n^{\alpha}$ in Eq. (\[autocorcond\]) with more general forms $n^{\alpha} \to n^{\alpha} (\log n)^{\beta}$, where $\alpha$ and $\beta$ are directly related to $c$ and $d$.
Finally, let us remark that it is not straight forward to relate aging random walks and its class $(c,d)$ with more traditional scaling exponents such as for example the Hurst exponent. The very nature of aging walks is that their persistence changes over time.
![ In the three top panes the second moment $\langle x^2(N)\rangle$ is shown for $\nu=0.2$, $0.5$, and $0.8$, for $\lambda=1.1$ (black), $1.2$ (red), and $1.3$ (green). The blue dotted and dashed lines indicate the function $N^2$ and $N$, respectively. A cross over from $\langle x^2(N)\rangle\sim N$ to $\langle x^2(N)\rangle\sim N^2$ is clearly visible for $\nu=0.5$ and $0.8$. The three bottom panes show the average number of direction reversal-decisions $k_-(N)$. Simulations were performed in the range $1\leq N\leq 50000$. For $\nu \to1$ the crossover happens at smaller $N$, for all values of $\lambda$. \[fig:crossover\] ](power_rw.eps){width="1.1\columnwidth"}
Conclusions
===========
Based on recently discovered scaling laws for trace form entropies we can classify all statistical systems and assign the a unique system-specific (extensive) generalized entropy. For non-ergodic systems these entropies may deviate from the Shannon form. The exponents for BG systems are $(c,d)=(1,1)$, systems characterized by stretched exponentials belong to the class $(c,d)=(1,d)$, and Tsallis systems have $(c,d)=(q,0)$. A further interesting feature all admissible systems is that they are all [*Lesche stable*]{}, and that the classification scheme for generalized entropies of type $S=\sum_i g(p_i)$ can be easily extended to entropies of Rényi type, i.e. $S=G(\sum_{i} g(p_i))$. For proofs see [@ebook].
We demonstrated that the auto-correlated random walk characterized by $0<\alpha\leq1$ introduced in [@Hanel2011_2] can not be distinguished from accelerating random walks. Although the presented auto-correlated random walk is of entropy class $(c,d)=(1,1/\alpha)$ and the accelerated random walk is of class $(c,d)=(1,1)$, both processes have the same distribution function since all moments $\langle x^n\rangle$ are identical. We have shown that other classes of random walks can naturally be obtained, including those belonging to the $(c,d)=(q,0)$, or Tsallis equivalence class. Moreover, we showed numerically that the auto-correlated random walk is asymptotically equivalent to a particular [*aging*]{} random walks, where the probability of a decision to reverse the direction of the walk depends on the path the random walk has taken. This concept of aging can easily be generalized to different forms of aging and it can be expected that many of the admissible systems can be represented by a specific type of aging that is specified by the aging function $g$, Eq. (\[transprob2\])). Finally, we have seen that different equivalence classes $(c,d)$ can be realized by specifying a aging function $g$. The effective number of direction reversal decisions corresponding to the aging function remains finite and therefore the associated generalized entropy requires a class $(c,d)$ with $0<c<1$. We believe that it should be possible that the scheme of aging random walks can be naturally extended to aging processes in physical, biological, and social systems in general.
biblabel\[1\][\#1. ]{}
[1]{}
Hanel R.; Thurner S. A comprehensive classification of complex statistical systems and an axiomatic derivation of their entropy and distribution functions. [*Europhys Lett*]{} [**2011**]{} [*93*]{}, 20006.
Hanel R.; Thurner S. When do generalized entropies apply? How phase space volume determines entropy [*Europhys Lett*]{} [**2011**]{} [*96*]{}, 50003.
Thurner S.; Hanel R. What do generalized entropies look like? An axiomatic approach for complex, non-ergodic systems. In [*Recent advances in Generalized Information Measures and Statistics*]{}, Kowalski A.M.; Rossignoli R.; Curado E.M.F., Eds.; Bentham Science eBook, in production 2013.
Shannon C. E. A Mathematical Theory of Communication. [*The Bell System Technical Journal*]{} [**1948**]{} [*27*]{}, 379 and 623. Khinchin A.I. [*Mathematical foundations of information theory*]{}. Dover Publ., New York 1957.
Tsallis C. [Possible generalization of Boltzmann-Gibbs statistics]{}. [*J Stat Phys*]{} [**1988**]{} [*52*]{}, 479-487.
Anteneodo C.; Plastino A.R. Maximum entropy approach to stretched exponential probability distributions. [*J Phys A: Math Gen*]{} [**1999**]{} [*32*]{}, 1089-1097.
Kaniadakis G. [Statistical mechanics in the context of special relativity]{}. [*Phys Rev E*]{} [**2002**]{} [*66*]{}, 056125.
Curado E.M.F.; Nobre F.D. On the stability of analytic entropic forms. [*Physica A*]{} [**2004**]{} [*335*]{}, 94-106.
Tsekouras G.A.; Tsallis C. Generalized entropy arising from a distribution of $q$ indices. [*Phys Rev E*]{} [**2005**]{} [*71*]{}, 046144.
Hanel R.; Thurner S. Generalized Boltzmann factors and the maximum entropy principle: entropies for complex systems. [*Physica A*]{} [**2007**]{} [*380*]{}, 109-114.
Hanel R.; Thurner S.; Gell-Mann M. Generalized entropies and the transformation group of superstatistics. [*PNAS*]{} [**2011**]{} [*108*]{}, 6390-6394.
Hanel R.; Thurner S.; Gell-Mann M. Generalized entropies and logarithms and their duality relations. [*PNAS*]{} [**2012**]{} [*109*]{}, 19151-19154.
Tsallis C. [*Introduction to Nonextensive Statistical Mechanics*]{}. Springer, New York 2009.
Shafee F. Lambert function and a new non-extensive form of entropy. [*IMA J Appl Math*]{} [**2007**]{} [*72*]{}, 785-800.
Hanel R., Thurner S., Generalized-generalized entropies and limit distributions. [*Braz J Phys*]{} [**2009**]{} [*39*]{}, 413-416.
Lesche B. Instabilities of Rényi entropies. [*J Stat Phys*]{} [**1982**]{} [*27*]{}, 419-422.
Abe S. Stability of Tsallis entropy and instabilities of Rényi and normalized Tsallis entropies. [*Phys Rev E*]{} [**2002**]{} [*66*]{}, 046134.
Jizba P.; Arimitsu T. Observability of RényiÕs entropy. [*Phys Rev E*]{} [*2004*]{} [*69*]{}, 026128.
Kaniadakis G.; Scarfone A.M. Lesche stability of $\kappa$-entropy. [*Physica A*]{} [**2004**]{} [*340*]{}, 102-109.
Hanel R.; Thurner S.; Tsallis C. On the robustness of q-expectation values and Rényi entropy. [*Europhys Lett*]{} [**2009**]{} [*85*]{}, 20005.
Tsallis C.; Gell-Mann M.; Sato Y. Asymptotically scale-invariant occupancy of phase space makes the entropy $S_q$ extensive [*PNAS*]{} [**2005**]{} [*102*]{}, 15377-15382.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Ruotian Luo\
TTI-Chicago\
[[email protected]]{}
- |
Ning Zhang\
Vaitl Inc.\
[[email protected]]{}
- |
Bohyung Han\
Seoul National University\
[[email protected]]{}
- |
Linjie Yang\
ByteDance AI Lab\
[[email protected]]{}
bibliography:
- 'egbib.bib'
title: 'Context-Aware Zero-Shot Recognition'
---
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Robert R. Tucci\
P.O. Box 226\
Bedford, MA 01730\
[email protected]
date:
title: |
Java Application that Outputs\
Quantum Circuit for Some\
NAND Formula Evaluators
---
Abstract {#abstract .unnumbered}
========
This paper introduces QuanFruit v1.1, a Java application available for free. (Source code included in the distribution.) Recently, Farhi-Goldstone-Gutmann (FGG) wrote a paper arXiv:quant-ph/0702144 that proposes a quantum algorithm for evaluating NAND formulas. QuanFruit outputs a quantum circuit for the FFG algorithm.
Introduction
============
This paper introduces QuanFruit v1.1, a Java application available[@QuanSuite] for free. (Source code included in the distribution.) Recently, Farhi-Goldstone-Gutmann (FGG) wrote a paper[@FGG07] that proposes a quantum algorithm for evaluating NAND formulas. QuanFruit outputs a quantum circuit for the FFG algorithm.
We say a unitary operator acting on a set of qubits has been compiled if it has been expressed as a SEO (sequence of elementary operations, like CNOTs and single-qubit operations). SEO’s are often represented as quantum circuits.
There exist software (quantum compilers) like Qubiter[@Tuc99] for compiling arbitrary unitary operators (operators that have no a priori known structure). QuanFruit is a special purpose quantum compiler. It is special purpose in the sense that it can only compile unitary operators that have a very definite, special structure.
The QuanFruit application is part of a suite of Java applications called QuanSuite. QuanSuite applications are all based on a common class library called QWalk. Each QuanSuite application compiles a different kind of quantum evolution operator. The applications output a quantum circuit that equals the input evolution operator. We have introduced 6 other QuanSuite applications in 2 earlier papers. Ref.[@qtree] introduced QuanTree and QuanLin. Ref.[@qfou] introduced QuanFou, QuanGlue, QuanOracle, and QuanShi. QuanFruit calls methods from these 6 previous applications, so it may be viewed as a composite of them.
Before reading this paper, the reader should read Refs.[@qtree] and [@qfou]. Many explanations in Refs.[@qtree] and [@qfou] still apply to this paper. Rather than repeating such explanations in this paper, the reader will be frequently referred to Refs.[@qtree] and [@qfou].
The goal of all QuanSuite applications, including QuanFruit, is to compile an input evolution operator $U$. $U$ can be specified either directly (e.g. in QuanFou, QuanShi), or by giving a Hamiltonian $H$ such that $U = e^{iH}$ (e.g. in QuanGlue and QuanOracle).
The standard definition of the evolution operator in Quantum Mechanics is $U= e^{-itH}$, where $t$ is time and $H$ is a Hamiltonian. Throughout this paper, we will set $t = -1$ so $U = e^{iH}$. If $H$ is proportional to a coupling constant $g$, reference to time can be restored easily by replacing the symbol $g$ by $-tg$, and the symbol $H$ by $-tH$.
Input Evolution Operator
========================
The input evolution operator for QuanFruit is $U_{fruit}=e^{iH_{fruit}}$, where
H\_[fruit]{} = .
![Line (open string) with 8 nodes []{data-label="fig-line-graph"}](line-graph.pdf){width="4in"}
$H_{line}\in\RR^{N_{S,line}\times N_{S,line}}$ where $N_{S,line}=2^{N_{B,line}}$ for some positive integer $N_{B,line}$. $H_{line}$ is proportional to the incidence matrix for a line graph, where the edges of the graph connect states that are consecutive in a Gray order. For example, for $N_{B,line}=3$, the graph of Fig.\[fig-line-graph\] yields: H\_[line]{} = g
[c|c|c|c|c|c|c|c|c|]{} &&&&&&&&\
&0&1& & & & & &\
&1&0& &1& & & &\
& & &0&1& & &1&\
& &1&1&0& & & &\
& & & & &0&1& &\
& & & & &1&0& &1\
& & &1& & & &0&1\
& & & & & &1&1&0\
, \[eq-h-line\] where $g$ is a real number that we will call the [**coupling constant**]{}.
![Binary tree with 8 nodes []{data-label="fig-tree-graph"}](tree-graph.pdf){height="1.5in"}
$H_{tree}\in\RR^{N_{S,tree}\times N_{S,tree}}$ where $N_{S,tree}=2^{N_{B,tree}}$ for some positive integer $N_{B,tree}$. $H_{tree}$ is proportional to the incidence matrix for a balanced-binary tree graph. For example, for $N_{B,tree}=3$, the graph of Fig.\[fig-tree-graph\] yields: H\_[tree]{} = g
[c|c|c|c|c|c|c|c|c|]{} &&&&&&&&\
& & & & & & & &\
& & &1&1& & & &\
& &1& & &1&1& &\
& &1& & & & &1&1\
& & &1& & & & &\
& & &1& & & & &\
& & & &1& & & &\
& & & &1& & & &\
, \[eq-h-tr\] where $g$ is the same coupling constant as before.
$h_{glue}\in \RR^{1\times 1}$. In fact,
h\_[glue]{} = g . Here $\ket{N_{S,line}+1}$ labels the god state of the tree, the one with children but no parents.( $N_{S,line}$ labels the dud node) We will call $d\in Z_{0,N_{S,line}-1}$ the [**line door**]{}. If $d=0$, then the tree is connected to a tail of states. For $N_{B,line}=3$, from Fig.\[fig-line-graph\], if $d=2$, then the tree is connected to the midpoint of the line of states (“runway").
The number of leaves in the tree is half the number of nodes in the tree: $N_{S,lvs}=\frac{N_{S,tree}}{2}$. Also, $N_{S,lvs}=2^{N_{B,lvs}}$ for some positive integer $N_{B,lvs}$. $h_{oracle}\in \RR^{N_{S,lvs}\times N_{S,lvs}}$. In fact,
h\_[oracle]{} = g , where $x_k\in Bool$ are the inputs to the NAND formula.
The dimension of the matrix $H_{fruit}$ is not generally a power of two. To represent it as a quantum circuit, we need to extend it to $diag(H_{fruit},0)\in \RR^{\ns\times\ns}$, where = 2\^, = {N: N\_[S,line]{} + N\_[S,tree]{}2\^N} . \[eq-nb\]
Define H\_[glue]{} = h\_[glue]{} + h\_[glue]{}\^, H\_[ora]{} = h\_[ora]{} + h\_[ora]{}\^. \[eq-htoH\] (This last equation is fine as an operator statement, but as a matrix statement, $h_{glue}$ and $h_{glue}^\dagger$ must be “padded" with zeros to make the equation true. By “padding a matrix with zeros”, we mean embedding it in a larger matrix, the new entries being zeros.)
One can split $H_{fruit}$ into two parts, which we call the [**bulk Hamiltonian**]{} $H_{bulk}$ and the [**boundary corrections Hamiltonian**]{} $H_{corr}$:
H\_[fruit]{} = H\_[bulk]{} + H\_[corr]{} , where
H\_[bulk]{} = H\_[line]{} + H\_[tree]{} , H\_[corr]{} = H\_[glue]{} + H\_[ora]{} . (Again, this last equation requires zero padding if considered a matrix equation.) Note that $[H_{line},H_{tree}]=0$ and $[H_{glue},H_{ora}]=0$.
For $r=1,2,3,\ldots$, if $U = L_r(g) + \calo(g^{r+1})$, we say $L_r(g)$ [**approximates (or is an approximant) of order**]{} $r$ for $U$.
Given an approximant $L_r(g) + \calo(g^{r+1})$ of $U$, and some $\nt= 1, 2, 3\,\ldots$, one can approximate $U$ by $\left(L_r(\frac{g}{\nt})\right)^\nt
+ \calo(\frac{g^{r+1}}{\nt^r})$. We will refer to this as [**Trotter’s trick**]{}, and to $\nt$ as the [**number of trots**]{}.
For $N_{T,line}=1$, QuanFruit approximates $e^{iH_{line}}$ with a Suzuki approximant of order $r_{line}=2, 4, 6, \ldots$ that is derived in Ref.[@Theory]. QuanFruit also applies the Trotter trick with $N_{T,line}>1$ trots to the $N_{T,line}=1$ approximant of $e^{iH_{line}}$.
For $N_{T,tree}=1$, QuanFruit always approximates $e^{iH_{tree}}$ with an approximant of order 3, that is derived in Ref.[@Theory]. QuanFruit also applies the Trotter trick with $N_{T,tree}>1$ trots to the $N_{T,tree}=1$ approximant of $e^{iH_{tree}}$.
Ref.[@Theory] gives exact (to numerical precision) compilations of the glue and oracle parts of $U_{fruit}$. QuanFruit uses these compilations, so the Order of the Suzuki (or other) Approximant and the Number of Trots do not arise in QuanFruit, for either the glue or the oracle.
For $N_{T, meta}=1$, QuanFruit also approximates $e^{iH_{fruit}}$ with a Suzuki approximant of order $r_{meta}=2, 4, 6, \ldots$. Recall that $S_2(t) = e^{A\frac{t}{2}}
e^{Bt}e^{A\frac{t}{2}}$ for $t\in\RR$ is the second order Suzuki approximant, and higher order ones are defined recursively from this one. Thus, all Suzuki approximants are specified by giving two functions of $t$, $e^{At}$ and $e^{Bt}$. To get a “meta" Suzuki approximant, we set $e^{At} = e^{i(H_{bulk})_{g\rarrow t}}$ and $e^{Bt} = e^{i(H_{corr})_{g\rarrow t}}$. QuanFruit also applies the Trotter trick with $N_{T,meta}>1$ trots to the $N_{T,meta}=1$ approximant of $e^{iH_{fruit}}$.
The Control Panel
=================
Fig.\[fig-qfruit-main\] shows the [**Control Panel**]{} for QuanFruit. This is the main and only window of the application. This window is open if and only if the application is running.
![Control Panel of QuanFruit[]{data-label="fig-qfruit-main"}](qfruit-main.pdf){height="7in"}
The Control Panel allows you to enter the following inputs:
File Prefix:
: Prefix to the 3 output files that are written when you press the [**Write Files**]{} button. For example, if you insert [test]{} in this text field, the following 3 files will be written:
- [test\_qfru\_log.txt]{}
- [test\_qfru\_eng.txt]{}
- [test\_qfru\_pic.txt]{}
Line: Number of Qubits:
: The parameter $N_{B,line}$ defined above.
Tree: Number of Qubits:
: The parameter $N_{B,tree}$ defined above.
Coupling Constant:
: The parameter $g\in \RR$ defined above.
Line Door:
: The parameter $d\in Z_{0,N_{S,line}-1}$ defined above.
Bands:
: You must enter here an even number of integers separated by any non-integer, non-white space symbols. Say you enter $a_1,b_1,a_2,b_2,\ldots ,a_n,b_n$. If $x_k\in Bool$ for $k\in Z_{0,\nlvs-1}$ are as defined above, then $x_k=1$ iff $k\in
Z_{a_1,b_1}\cup Z_{a_2,b_2} \ldots \cup Z_{a_n,b_n}$. Each set $Z_{a_k,b_k}$ is a “band". If $a_k=b_k$, the band has a single element. QuanFruit checks that $0\leq a_0$, $b_n\leq(\nlvs-1)$, and $b_k-a_k\geq 0$ for all $k$. It also checks that $a_{k+1}-b_k\geq 2$. (If $a_{k+1}-b_k= 1$, bands $k+1$ and $k$ can be merged. If $a_{k+1}-b_k= 0, -1, -2, \ldots$, bands $k+1$ and $k$ overlap.)
Line: Number of Trots:
: The parameter $N_{T,line}$ defined above.
Line: Order of Approximant:
: The parameter $r_{line}$ defined above.
Tree: Number of Trots:
: The parameter $N_{T,tree}$ defined above.
Tree: Order of Approximant:
: This parameter is always 3.
Meta: Number of Trots:
: The parameter $N_{T,meta}$ defined above.
Meta: Order of Approximant:
: The parameter $r_{meta}$ defined above.
The Control Panel displays the following outputs:
Number of Qubits:
: The parameter $\nb$ defined by Eq.(\[eq-nb\]).
Number of Elementary Operations:
: The number of elementary operations in the output quantum circuit. If there are no LOOPs, this is the number of lines in the English File, which equals the number of lines in the Picture File. When there are LOOPs, the “[LOOP k REPS:$\nt$]{}" and “[NEXT k]{}" lines are not counted, whereas the lines between “[LOOP k REPS:$\nt$]{}" and “[NEXT k]{}" are counted $\nt$ times.
Error:
: The distance in the Frobenius norm between the input evolution operator and the output quantum circuit (i.e., the SEO given in the English File). For a nice review of matrix norms, see Ref.[@Golub]. For any matrix $A\in\CC^{n\times n}$, its Frobenius norm is defined as $\|A\|_F = \sqrt{\sum_{j,k} A_{j,k}A^*_{j,k}}
$. Another common matrix norm is the 2-norm. The 2-norm $\|A\|_2$ of $A$ equals the largest singular value of $A$. The Frobenius and 2-norm of $A$ are related by[@Golub]: $
\|A\|_2 \leq \|A\|_F \leq \sqrt{2}\|A\|_2
$.
Message:
: A message appears in this text field if you press [**Write Files**]{} with a bad input. The message tries to explain the mistake in the input.
Output Files
============
Pressing the [**Write Files**]{} button of the Control Panel of QuanFruit generates 3 files (Log, English, Picture). These files are analogous to their namesakes for QuanTree, QuanLin and other QuanSuite applications. Ref.[@qtree] explains how to interpret them.
Behind the Scenes:\
Code Innovations in QuanSuite, QWalk
====================================
The QuanSuite applications, based on the QWalk class library, exhibit some code innovations that you will find very helpful. Hopefully, these innovations will become commonplace in future quantum computer software.
- [**QWalk class library does most of the work in all QuanSuite applications:**]{} Look in the source folder for any of the QuanSuite applications. You’ll find that it contains only 3 or 4 classes. Most of the classes are in the source folder for QWalk. That’s because most of the work is done by the QWalk class library, which is independent of the QuanSuite application.
- [**Reusability of SEO writers:**]{} Look at the class [FruitSEO\_writer]{} in the source folder for QuanFruit. You’ll find that [FruitSEO\_writer]{} utilizes the methods [GlueSEO\_writer()]{}, [OracleSEO\_writer()]{}, [TreeSEO\_writer()]{}, [LineSEO\_writer()]{}, and [ShiftSEO\_writer()]{}. Thus, [FruitSEO\_writer]{} delegates its SEO writing to methods from the QuanSuite applications: QuanGlue, QuanOracle, QuanTree, QuanLin and QuanShi. In fact, QuanFruit can be viewed as a composite of these simpler QuanSuite applications. This reusability of SEO writers is made possible by the novel technique described in Appendix \[app-pad\].
- [**Nested Loops:**]{} The English and Picture files of QuanSuite applications can have LOOPs within LOOPs. This makes the English and Picture files shorter, without loss of information. However, if you want to multiply out all the operations in an English file (this is what the class [SEO\_reader]{} in QWalk does), then having nested loops makes this task more difficult. [SEO\_reader]{} of QWalk is sophisticated enough to understand nested loops.
- [**Painless object oriented implementation of Suzuki approximants and Trotter’s trick:**]{} Higher order Suzuki approximants can be implemented painlessly by using the classes: [QWalk/src/SuzFunctions]{} and [QWalk/src/SuzWriter]{}. See the class [QuanLin/src/LineSEO\_writer]{} for an example of how it’s done. Essentially, all you have to do is to override the two abstract methods in [QWalk/src/SuzFunctions]{}.
Trotter’s trick can also be easily implemented in a QuanSuite application, by using LOOP and NEXT lines in the English file. See the [write()]{} method of [QuanLin/src/LineSEO\_writer]{} for an example.
Appendix: $P_0$ Padding and State Shifting {#app-pad}
==========================================
Suppose we know how to compile $e^{iH}$. Is it possible to use this compilation to compile $e^{i\;\;diag(Z, H, Z')}$, where $Z$ and $Z'$ are square matrices of zeros? The answer is yes, as we show next.
Suppose $\hat{N}_S > \ns$ where $\ns = 2^\nb$ and $\hat{N}_S = 2^{\hat{N}_B}$, for some positive integers $\nb$ and $\hat{N}_B$. Given a Hamiltonian $H$, define a zero padded version of it called $\hat{H}$:
\_[\_S\_S]{} &=&\
&=& (\_B-1) (\_B-2) (+1) () H\_ . As usual, $\nbar() = P_0()$. We will say that $H$ has been padded with $P_0$’s to obtain $\hat{H}$. Now let $U_{shift}^{(k)}$ be the unitary operation that shifts state $\ket{x}$ to $\ket{(x+k)\mod
\hat{N}_S}$, with $x,k\in Z_{0,\hat{N}_S-1}$. The application QuanShi gives a compilation of $U_{shift}^{(k)}$. Using $U_{shift}^{(k)}$, one can define a matrix $\hat{H}^{(k)}$ from $\hat{H}$ as follows:
(\^[(k)]{})\_[\_S\_S]{} &=&\
&=& (U\_[shift]{}\^[(k)]{})\^ \_[\_S\_S]{} U\_[shift]{}\^[(k)]{} .
It is now readily apparent that a SEO for $e^{i\;\;diag(Z, H, Z')}$ can be obtained from a SEO for $e^{iH}$ by $P_0$ padding $H$ and then state shifting it with $U_{shift}^{(k)}$.
The compilations of $e^{iH_{line}}$ (given in QuanLin), $e^{iH_{tree}}$ (given in QuanTree) and $e^{iH_{ora}}$ (given in QuanOracle), are all utilized by QuanFruit via this $P_0$ padding/state-shifting method.
[99]{}
QuanFruit software available at [www.ar-tiste.com/QuanSuite.html]{}
E. Farhi, J. Goldstone, S. Gutmann, “A Quantum Algorithm for the Hamiltonian NAND Tree", arXiv:quant-ph/0702144
R.R. Tucci, “A Rudimentary Quantum Compiler(2cnd Ed.)", arXiv:quant-ph/9902062 . Qubiter software available at [www.ar-tiste.com/qubiter.html]{}
R.R. Tucci, “QuanTree and QuanLin, Two Special Purpose Quantum Compilers", arXiv:0712.3887
R.R. Tucci, “QuanFou, QuanGlue, QuanOracle and QuanShi, Four Special Purpose Quantum Compilers", arXiv:0802.2367
R.R.Tucci, “How to Compile Some NAND Formula Evaluators", arXiv:0706.0479
G.H. Golub and C.F. Van Loan, [*Matrix Computations, Third Edition*]{} (John Hopkins Univ. Press, 1996).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ alloys evolve from local moment magnetism (x = 0) to heavy fermion superconductivity (x =1). Magnetic order is observed over a broad range of $x$. For a substantial range of $x$ (0.83 $\leq $ x $\leq $ 0.95) in the temperature - composition phase diagram we find that superconductivity may coexist with spin - density wave magnetic order at the Fermi surface. We show that a delicate balance betwen superconducting and magnetic instabilities can be reversibly tuned by both the Ce/Nd ratio and magnetic field, offering a new and unique model electronic system.'
author:
- 'Rongwei Hu,$^{1,2}$ Y. Lee,$^{3}$ J. Hudis,$^{4}$ V. F. Mitrovi[c,]{}$^{2}$ and C. Petrovic$^{1}$'
title: 'Composition and field tuned magnetism and superconductivity in Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$'
---
Introduction
============
In contrast to simple metals such as Pb, phase diagrams of unconventional superconductors often show a multitude of electronic phases of matter. In particular, magnetic order is ubiquitous in cuprate oxides and heavy fermion superconductor (HFSC) phase diagrams alike, as well as in some ruthenates and cobaltates. The proximity, competition, or coexistence of two distinct types of electronic order at the Fermi surface raised speculations that their driving mechanisms could be closely related.[@Fisk1]$^{,}$ [Demler]{} Exotic superconductivity in heavy fermion materials [@Fisk2] usually appears near the quantum critical point [@Senthil] where the magnetic ordering temperature is tuned to T =0 by a variety of external parameters. For example annealing, composition or magnetic field.[Trovarelli]{} However, the most frequently used tuning parameter in HFSC is pressure.[@Yuan] This is in variance with the cuprate family where the interplay of superconductivity and magnetism is balanced by adjusting the in plane charge density of the CuO$_{2}$ layers. The discovery of pressure - induced superconductivity in CeIn$_{3}$[@Mathur] stimulated exploratory synthesis of AuCu$_{3}$ superstructures by Fisk, where the magnetic entropy might be further suppressed by crystallographic arguments.[@Fisk3] Indeed, superconductivity in CeRhIn$_{5}$,[@Rh] and CeIrIn$_{5}$,[Ir]{} was soon discovered, as well as coexistence of heavy fermion superconductivity and magnetism in CeRh$_{1-x}$Ir$_{x}$In$_{5}$.[Pagliuso]{}$^{,}$[@Christianson] CeCoIn$_{5}$,[@Co] the *primo* compound in the 115 family of HFSC, has also been recently hole doped to an antiferromagnetic ground state by Cd substitution on the In (1) site.[Pham]{} Despite a few examples of structurally tuned superconductivity in HFSC, the lack of a predictive theory suggests that there is still no clear understanding of how the delicate interplay of various degrees of freedom in these materials stabilizes superconducting, or a magnetic ground state.
The large quasiparticle mass enhancement in CeCoIn$_{5}$ is reflected in two large nearly cylindrical pieces of the Fermi surface and smaller 3D hole pockets.[@Settai] Due to the pronounced $\overrightarrow{k}$ - space inhomogeneity, hybridization of the $3d$ electrons of Co with the $5p$ electrons of In results in a small density of states at the Fermi energy, implying partially quasi - 2D electronic structure.[@Settai2] By exploring a well defined non - hybridizing local moment such as Nd, we were able to continuously tune the coupling in the lattice and consequently quasiparticle mass enhancement between 4f ions and conduction electrons. As a result, we have obtained a rich phase diagram. The magnetic ground state in Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ smoothly evolves from local moment magnetism (LMM) on the Nd - rich side to HFSC on the Ce - rich side. Superconductivity coexists with other forms of electronic order, most likely magnetic in nature, in samples for Nd concentrations between $x=0.78$ and $x=0.98$. We demonstrate that the delicate balance between coexisting ordered states can be smoothly tuned by magnetic fields near the magnetic - superconducting boundary.
Methods
=======
Single crystals of Nd$_{1-x}$CexCoIn$_{5}$ were grown from an excess In flux. Magnetic susceptibility, specific heat and resistivity measurements were performed in a Quantum Design MPMS XL 5 and PPMS - 9 instruments respectively. Single crystals were thoroughly ground to a fine powder for structural measurements. High resolution synchrotron powder X - ray patterns were taken at beamline X7A of National Synchrotron Light Source at Brookhaven National Laboratory. Monochromatic synchrotron x-ray and gas-proportional position-sensitive detector were used to measure the powder diffraction data. Rietveld refinements were performed using GSAS.[@GSAS] The samples were manually aligned to measure the magnetization, heat capacity, or resistivity for fields applied along the appropriate axis. M/H polycrystalline averages were calculated as $\chi (T)=[2\chi _{ab}(T)+\chi
_{c}(T)/3]$ and were used to obtain the high temperature effective moments. Temperature dependent magnetic susceptibility was used to estimate the relative ratio of Ce/Nd ions. At high temperatures ranging from 150 K to 350 K, H/M curves were fit with Curie - Weiss (CW) law, $\chi (T)=C/(T-\Theta ),
$ where C is the Curie constant and $\Theta $ is the Weiss temperature. Electrical contacts for in - plane resistivity measurements were made with Epotek-H20E silver epoxy on thin plate-like crystals whose dimensions were measured by optical microscope with 10$\mu $m resolution. The samples were previously etched in diluted HCl for several hours and thoroughly rinsed in ethanol in order to remove excess In.
Structural Characterization and Phase Purity
============================================
Since both NdCoIn$_{5}$ and CeCoIn$_{5}$ are grown using identical temperature profiles and an identical ratio of starting materials Nd (Ce), Co and In,[@Co]$^{,}$[@Jacob] one would expect a smooth change of the lattice parameters in the Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ alloy series.
{height="2.2in"}
Indeed, this is confirmed by high resolution structural measurements which showed that the samples crystallized in tetragonal HoCoGa$_{5}$ structure without any additional peaks introduced by Ce alloying (Fig. 1). Selected regions of the powder X - ray spectra showed monotonic evolution and uniform sharpness independent of $x$ for both \[111\] and \[003\] peaks. This implies that Ce uniformly substitutes Nd with the increase of x and that our samples are indeed alloys rather than a mixture of intergrown compounds NdCoIn$_{5}$ and CeCoIn$_{5}$. The lattice parameters increase smoothly with Ce substitution in accordance with Vegard’s law (Fig. 2b).
![Selected regions of the synchrotron powder X - ray diffraction data for Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$. Both \[111\] and \[003\] peaks shift uniformly with Ce substitution. (b) Tetragonal lattice parameters $\widehat{\mathit{a}}
$ (red symbols) and $\widehat{\mathit{c}}$ (blue symbols)](Fig2.eps){height="4in"}
Magnetic susceptibility measurements did not detect magnetically ordered second phases, such as NdIn$_{3}$ or CeIn$_{3}$ (Fig. 3). The cubic compounds CeIn$_{3}$ and NdIn$_{3}$ order antiferromagnetically at 10 K and 6.3 K respectively.[@Mathur]$^{,}$[@Amara] Magnetic ordering in NdCoIn$_{5}$ is depressed smoothly with the increase of Ce, and we were able to follow the characteristic peak in M/H (signature of the onset of the magnetic order) down to lowest temperature of our magnetic measurement, T = 1.8 K for x = 0.6. Curie - Weiss analysis of polycrystalline magnetic susceptibility average at high temperatures showed at most a 4% deviation from the nominal ratio of Ce$^{3+}$ and Nd$^{3+}$ moments (Table 1). As expected, the highest uncertainty in the nominal concentration of Ce (x) is in the middle of the alloy series, for the highest chemical disorder. Combined together, these results demonstrate that Ce uniformly substitutes Nd in the entire doping range of Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$, with a maximum $\Delta x$ = 0.04.
x Measured $\mu _{eff}$($\mu _{B}$) Expected$\mu _{eff}$($\mu _{B}$) Error (%)
------ ----------------------------------- ---------------------------------- -----------
1 2.59(1) 2.54 2
0.85 2.67(6) 2.69 0.8
0.83 2.73(1) 2.71 0.8
0.8 2.72(1) 2.75 1
0.6 2.83(1) 2.96 4
0.5 2.96(1) 3.07 3
0.2 3.37(2) 3.40 0.8
1 3.7(1) 3.62 2
Results
=======
M/H data of Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ alloy series show substantial anisotropy (Fig. 3) at low temperatures and signature of magnetic order at T$_{N}$ = 9 K. For field applied along the $\widehat{a}$-axis, $H||\widehat{a}, $ the magnetic transition is rather broad. On the other hand for $H||\widehat{c},$ the transition is relatively sharp implying complex magnetic order.
{height="4in"}
The thermodynamic properties of Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ are shown in Fig. 4. The magnetic ordering temperature in NdCoIn$_{5}$ is smoothly depressed from T$_{N}$ = 9 K with increased Ce concentration to 2.0 K by $x=0.6$ where (C-C$_{latt}$)/T shows a broad peak (Fig. 4a). Interestingly, the magnetic ordering transition becomes sharper as $x$ is further tuned towards the superconducting boundary for $x=0.7$. At $x\geq 0.78$ superconductivity emerges. With increasing $x$, the T$_{C}$ increases to the bulk T$_{C}$ = 2.3 K of CeCoIn$_{5}$. The magnetic entropy released upon emergence of LMM (Fig. 4a inset) scales with Nd concentration up to $x=0.4$, implying that Ce ions do not play a direct part in the formation of the LMM ground state. In the HFSC state for $0.78\leq x\leq 1$, the magnetic entropy at 5K is essentially invariant to changes of Ce/Nd ratio. However, a hallmark of heavy fermion magnetism is observed for $x\geq 0.5$. This can be seen on the example of $x=0.6$ where only approximately $\sim 0.2$ Rln2 is released below the magnetic ordering transition at 2 K. In the superconducting region at the Ce -rich side (Fig. 4b) additional thermodynamic anomalies $A$, $B$ and $C$ emerge in (C-C$_{latt}$)/T below the superconducting transition T$_{C}$ for $x=0.95$, $x=0.9$, but above T$_{C}$ for $x=0.83$.
{height="4in"}
By examining electronic transport we obtain further evidence for the presence of distinct types of electronic matter in the phase diagram. Temperature dependent electrical resistivities normalized to their value at 300K (Fig. 5 a,b) show rather small loss of the spin disorder scattering below Neel temperature in NdCoIn$_{5}$. This is consistent with the LMM - type of order in rare earth intermetallic compounds where local moments do not become part of Fermi surface upon cooling. By $x=0.5$ the electronic scattering strongly increases due to a single ion Kondo - type interaction and a logarithmic contribution of hybridizing Ce$^{3+}$ ions submerged in the Fermi sea. Eventually for $x\geq 0.5$ a broad coherence peak develops in the lattice. This marks the emergence of an additional Kondo energy scale arising from the coherence in the heavy fermion lattice. As expected, the coherence temperature T$_{COH}$ increases with the increase of Ce$^{3+}$ ions. Looking from the Ce - rich side, the onset of the superconducting transition in CeCoIn$_{5}$ is depressed to 0.9 K and magnetic scattering increases by $x=0.83$ (Fig. 4b).
![Results of resistivity measurements on Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ with current parallel to the \[100\]. (a): Resistivity curves normalized to their value at 300 K. (b): Magnetic contribution to the electrical resistivity ($\protect\rho $ - $\protect\rho _{lattice}$). The lattice resistivity was estimated by taking values of electrical resistivity for LaCoIn$_{5}$.](Fig5.eps){height="4in"}
Discussion
==========
Heat capacity in the HFSC state for $x\leq 0.95$ reveals two thermodynamic anomalies. Only one (C-C$_{latt}$)/T discontinuity, though, corresponds to the superconducting transition (Fig. 4b). One possible explanation for this would involve sample inhomogeneity and a distribution of the doping concentration. This however is very unlikely due to the low residual resistivity, the clean high resolution synchrotron powder X - ray diffraction pattern with uniform sharpness of peaks for the whole range of $x $ and the compositional dependence of lattice parameters $\widehat{a}$ and $\widehat{b}$ in accordance with Vegard’s law (Fig.1). Furthermore, one would expect the highest degree of metallurgical disorder around the middle of the doping range, for $x=0.5$, and not near the Ce end. Finally, in superconducting materials with *metallic* type of bonding, metallurgical inhomogeneity would shift the bulk T$_{C}$ of the *whole* sample, not only a fraction of the sample. For example, in CeCu$_{2.2}$Si$_{2}$ this is indeed seen in high resolution studies of the local structure.[@Louca] Moreover, as seen in the heat capacity data for $x=0.1$, the two thermodynamic anomalies are comparable in size. Thus secondary phases would have been easily detected in the analysis of powder X - ray spectra. Furthermore, as will be evident in Fig. 7, the smooth evolution of $T_{C}(x)$ and $T_{M}(x)$ argues against real space inhomogeneity. Our results imply that magnetic and superconducting phases coexist in the phase diagram, similarly to the situation in Cd - doping of CeCoIn$_{5}$.[@Pham]
{height="4in"}
An alternative explanation involves two superconducting energy scales on different parts of the Fermi surface and negligible interband scattering. This could explain low temperature anomalies $A,B$ for $x=0.95$ and $x=0.9$ since zero resistivity is achieved at the high temperature transition and is maintained through the low temperature transition. However, this scenario cannot explain thermodynamic anomaly $C$ above the superconducting transition in Nd$_{0.17}$Ce$_{0.83}$CoIn$_{5}$ (Fig. 4b). We speculate that low temperature discontinuities $A$ and $B$, high temperature discontinuity $C$ as well as sharp transition in \[C-C$_{latt}$\]/T (Fig. 4(a)) for x = 0.7 may be connected with some form of Fermi surface instability appearing concurrently on a different part of the Fermi surface. In what follows we show that this instability as well as the boundary between superconductivity and magnetism can be tuned by a magnetic field, in addition to composition $x$. Application of magnetic field reverses remarkable behavior seen with progressive decrease in the Ce/Nd ratio, tuning the boundary between the two ordered states.
{height="2.8in"}
The magnetic field depresses both $T_{C}$ and $B$ in Nd$_{0.9}$Ce$_{0.1}$CoIn$_{5}$ (Fig 6). The suppression is rather anisotropic. For a field applied along the $\widehat{c}$- axis superconducting anomaly in heat capacity and increasingly broad resistivity transition are suppressed to 0.5 K in 40 kOe. The anomaly $B$ is observed only below 10 kOe, merging into a single thermodynamic transition for higher fields above 0.4 K, the lowest temperature of our measurement. On the other hand, for a field applied along $\widehat{a}$- axis both thermodynamic anomalies show quite different field dependence. Superconducting T$_{C}$ at 30 kOe - defined by the simultaneous onset of zero resistivity and the start of the heat capacity anomaly - is suppressed to 1.4 K. The anomaly $B$ is observed below the T$_{C}$, just as we observe in H = 50 kOe. In contrast to H = 0 and H = 30 kOe, zero resistivity in H = 50 kOe corresponds to midpoint, rather than the onset of the heat capacity anomaly. In H = 70 kOe, Nd$_{0.1}$Ce$_{0.9}$CoIn$_{5}$ electronic matter becomes equivalent to the $x=0.83$ sample at H = 0 kOe:the main heat capacity transition is now above superconductivity. Electronic scattering for field applied in tetragonal plane (Fig. 3b) first increases and then decreases in the normal state as the magnetic field is tuned through H$_{C}$ = 50 kOe. The resistivity transition in H $\geq $ H$_{C}$ is sharper implying a magnetic field induced phase transition for H $\uparrow
\uparrow $ $\widehat{a}$ -axis. The size of the superconducting anomaly in (C-C$_{latt}$) decreases in field relative to the size of additional thermodynamic anomaly B below T$_{C}$ at H = 0 for field applied along both crystalline axes. There is no loss of spin disorder scattering at the main heat capacity transition T$_{M}$ for both x = 0.1 sample in H = 70 kOe (Fig. 6b) and for x = 0.83 sample in H = 0 (Fig. 4b). Instead, a very small upturn is observed in Fig. 4(b) (x = 0.83, H = 0) and in Fig. 6(b) (x = 0.1, H = 70 kOe) in resistivity (denoted by orange diamonds), a signature of the partial gapping of the Fermi surface. This is reminiscent of an itinerant spin-density-wave type - transition based on Fermi-surface nesting in a heavy-electron band observed in Ce(Ru$_{0.85}$Rh$_{0.15}$)$_{2}$Si$_{2}$.[@Murayama]
It appears that Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ can support simultaneously multiple electronic ordering states. Coexistence of magnetism in superconductivity in Ce 115 superconductors has been induced by pressure in CeRhIn$_{5}$ or by composition in CeRh$_{1-x}$Ir$_{x}$In$_{5}$, CeCo(In$_{1-x}$Cd$_{x}$)$_{5}$ and CeRh$_{1-x}$Co$_{x}$In$_{5}$.[@Rh]$^{,}$[Pagliuso]{}$^{,}$[@Pham]$^{,}$[@Zapf] The low temperature specific heat and resistivity data taken in Nd$_{0.1}$Ce$_{0.9}$CoIn$_{5}$ at H = 0 and in Nd$_{0.17}$Ce$_{0.83}$CoIn$_{5}$ at H = 70 kOe imply that T$_{C}$ and T$_{M}$ involve instabilities at different parts of the Fermi surface. The balance between these states is tuned only by degree of hybridization in the Kondo Lattice (Ce/Nd ratio) and by magnetic field. Though the presence of additional magnetic transition made it difficult to estimate accurate values for the electronic heat capacity coefficient $\gamma $, by taking $\gamma $ = \[C-C$_{latt}$(T)\]/T above T$_{C}$, we observe that the large jump at ambient pressure in the specific heat of CeCoIn$_{5}$ $\Delta $C/C(T$_{C}$) = 4.35 is reduced as more Nd enters into the matrix. For x = 0.98 we observe $\Delta $C/C(T$_{C}$) = 3.79 and by x = 0.1 this ratio decreases to 1.76. Nevertheless, our results may indicate general trend that increased Nd concentration decreases electron - boson coupling strength.[@Bang]
In the standard paradigm, the suppression of heavy fermion antiferromagnetic order leads to superconductivity around the quantum critical point.[Millis]{}$^{,}$[@Steglich2] One aspect of the reduced Ce/Nd ratio in the lattice is a negative pressure. Using the bulk modulus of CeCoIn$_{5}$ B = 76 GPa,[@Normile] we estimate that a rigid shift of the lattice parameters for $x=0.83$ and $x=0.9$ corresponds to 0.03 GPa of applied pressure, Given that $\Delta $T$_{C}$ = 0.9 K (50%) between these two concentrations, this result implies that chemical pressure effects are less relevant than electronic tuning through increased hybridization.[Sidorov]{} This result applies not only at the magnetic - superconducting boundary where the contribution is only up to 2%, but also for the whole range of $x$ since changes of the ground state in Nd$_{1-x}$Ce$_{x}$CoIn$_{5} $ are far more dramatic than what was observed in pressure - induced changes of the ground state in CeCoIn$_{5}$.[@Sidorov]
It would be instructive to compare our results with the reported coexistence of superconductivity and magnetism in CeRhIn$_{5}$ under high pressure and to field-induced magnetic order in the superconducting state of La$_{1.9}$Sr$_{0.1}$CuO$_{4}$.[@Tuson]$^{,}$[@Knebel]$^{,}$[@Lake] In both materials, as well as in CeCoIn$_{5}$, there are nodes in the superconducting gap. Delicate balance between antiferromagnetic and superconducting coupling near a quantum phase transition is smoothly tuned by magnetic field which generates field-induced vortices that suppress superconductivity and enhance magnetic correlations.[@Demler]$^{,}$[Zhang]{} In addition to this, in Nd$_{1-x}$Ce$_{x}$CoIn$_{5}$ suppression of d-wave superconductivity is certainly influenced by disorder, as seen in large increase in residual resistivity (Fig.5).[@Balatsky]
The evolution of superconducting T$_{C}$(x) with the increase of Nd is similar to Ce$_{1-x}$La$_{x}$CoIn$_{5}$.[@Petrovic]$^{,}$[@Nakatsuji] This is consistent with dominant potential scattering and insensitivity of T$_{C}$ to magnetic configuration of the rare earth ion, as observed in Ce$_{1-x}$R$_{x}$CoIn$_{5}$.[@Johnpierre]
Conclusion
==========
In conclusion, we have reported on Nd substitution in CeCoIn$_{5}$. Nd substitution results in the rich phase diagram, controlled apparently only by Ce/Nd ratio, i.e. by the level of 4f - conduction electron coupling. Ground state of this alloy series evolves from LMM to HFSC state via magnetically ordered heavy fermion ground state. Small concentration of Nd (5%) induces magnetic order deep in the superconducting state. We have demonstrated magnetic field tuning of the delicate balance between superconducting and magnetic ground state at the magnetic - superconducting boundary. We invite further investigation into possible microscopic coexistence of magnetic and superconducting order parameters by NMR and neutron scattering measurements. Possible existence of short range order (spin glass) between local moment magnetism and heavy fermion superconductor near the middle of alloy series would offer model electronic system with rich interplay of superconductivity and magnetism comparable to high-T$_{C}$ cuprate oxide superconductors.
We thank R. Prozorov, S. L. Bud’ko, P. C. Canfield, Johnpierre Paglione and Myron Strongin for useful communication. This work was carried out at the Brookhaven National Laboratory, which is operated for the U.S. Department of Energy by Brookhaven Science Associates (DE-Ac02-98CH10886). This work was supported by the Office of Basic Energy Sciences of the U.S. Department of Energy.
[99]{} Z. Fisk and D. Pines, Nature 394, **22**, (1998)
E. Demler, W. Hanke and S. C. Zhang, Rev. Mod. Phys. **76**, 909 (2004)
Z. Fisk, D. W. Hess, C. J. Pethick, D. Pines, J. L. Smith, J. D. Thompson, and J. O. Willis, Science **239**, 33 (1988)
T. Senthil, A. Vishwanath, L. Balents, S. Sachdev, and M. P. A. Fisher, Science **303**, 1490 (2004)
O. Trovarelli, M. Weiden, R. Müller-Reisener, M. Gomez-Berisso, P. Gegenwart, M. Deppe, C. Geibel, J. G. Sereni and F. Steglich, Phys. Rev. B **56**, 678 (1997)
H. Q. Yuan, F. M. Grosche, M. Deppe, C. Geibel, G. Sparn, and F. Steglich, Science **302**, 2104
N. D. Mathur, F. M. Grosche, S. R. Julian, I. R. Walker, D. M. Freye, R. K. W. Haselwimmer and G. G. Lonzarich, Nature (London), **394**, 39 (1988)
Z. Fisk and G. Aeppli, Science **260**, 38 (1993)
H. Hegger, C. Petrovic, E. G. Moshopoulou, M. F. Hundley, J. L. Sarrao, Z. Fisk and J. D. Thompson, Phys. Rev. Lett. **84**, 4986 (2000)
C. Petrovic, R. Movshovich, M. Jaime, P. G. Pagliuso, M. F. Hundley, J. L. Sarrao, Z. Fisk and J. D. Thompson, Europhys. Lett. **53**, 354 (2001)
P. G. Pagliuso, C. Petrovic, R. Movshovich, D. Hall, M. F. Hundley, J. L. Sarrao, J. D. Thompson, and Z. Fisk, Phys. Rev. B **64**, 100503(R) (2001)
A. Llobet, A. D. Christianson, Wei Bao, J. S. Gardner, I. P. Swainson, J. W. Lynn, J.-M. Mignot, K. Prokes, P. G. Pagliuso, N. O. Moreno, J. L. Sarrao, J. D. Thompson, and A. H. Lacerda, Phys. Rev. Lett. **95**, 217002 (2005)
C. Petrovic, P G Pagliuso, M F Hundley, R Movshovich, J L Sarrao, J D Thompson, Z Fisk and P Monthoux, J. Phys. Condens. Matter **13,** L337 (2001)
L. D. Pham, T. Park, S. Maquilon, J. D. Thompson and Z. Fisk, Phys. Rev. Lett. **97**, 056404 (2006)
R. Settai, H Shishido, S Ikeda, Y Murakawa, M Nakashima, D Aoki, Y Haga, H Harima and Y Onuki, J. Phys. Condens. Matter **13**, L627 (2001)
R. Settai, T. Takeuchi and Y. Ōnuki, J. Phys. Soc. Jpn **76**, 051003 (2007)
A. C. Larson and R. B. VonDreele, Report No. LAUR 86-748, Los Alamos National Laboratory, New Mexico, 1986
J. Hudis, Rongwei Hu, C.L. Broholm, V.F. Mitrović and C. Petrovic, J. Magn. Magn. Mater. **307**, 301 (2006)
M. Amara, R. M. Galéra, P. Morin, T. Veres and P. Burlet, J. Magn. Magn. Mater **130**, 127 (1994)
D.Louca, J. D. Thompson, J. M. Lawrence, R. Movshovich, C. Petrovic, J. L. Sarrao and G. H. Kwei, Phys. Rev B **61**, R14940 (2000)
S. Murayama, C. Sekine, A. Yokoyanagi, K. Hoshi and Y. Onuki, Phys. Rev B **56** 11092 (1997)
V. S. Zapf, E. J. Freeman, E. D. Bauer, J. Petricka, C. Sirvent, N. A. Frederick, R. P. Dickey, and M. B. Maple, Phys. Rev. B **65**, 014506 (2001)
Y. Bang and A. V. Balatsky, Rev. B **69**, 212504 (2004)
A. J. Millis, Phys. Rev. B **48**, 7183 (1993)
F. Steglich, B. Buschinger, P. Gegenwart, M. Lohmann, R. Helfrich, C. Langhammer, P. Hellmann, L. Donnevert, S. Thomas, A. Link, C. Geibel, M. Lang, G. Sparn and W Assmus, J. Phys. Condens. Matter **9,** 9909 (1996)
P. S. Normile, S. Heathman, M. Idiri, P. Boulet, J. Rebizant, F. Wastin, G. H. Lander, T. Le Bihan and A. Lindbaum, Phys. Rev. B **72**, 184508 (2005)
V. A. Sidorov, M. Nicklas, P. G. Pagliuso, J. L. Sarrao, Y. Bang, A. V. Balatsky, and J. D. Thompson, Phys. Rev. Lett. **89**, 157004 (2002)
Tuson Park, F. Ronning, H. Q. Yuan, M. B. Salomon, R. Movshovich, J. L. Sarrao and J. D. Thompson, Nature 440, **65** (2006)
G. Knebel, D. Aoki, D. Braithwaite, B. Salce and J. Flouquet, Phys. Rev. B **74**, 020501 (2006)
B. Lake, H. M. Ronnow, N. B. Christensen, G. Aeppli, K. Lefmann, D. F. McMorrow, P. Vorderwisch, P. Smeibidl, N. Mangkorntong, T. Sasagawa, M. Nohara, H. Takagi and T. E. Mason, Nature **415**, 299 (2002)
E. Demler, S. Sachdev and Y. Zhang, Phys. Rev. Lett. **87**, 067202 (2001)
Y. Zhang, E. Demler and S. Sachdev, Phys. Rev. B **66**, 094501 (2002)
A. V. Balatsky, I. Vekhter, Jian-Xin Zhu, Rev. Mod. Phys. **78**, 373 (2006)
C. Petrovic, S. L. Bud’ko, V. G. Kogan and P. C. Canfield, Phys. Rev. B **66**, 054534 (2002)
S. Nakatsuji, S. Yeo, L. Balicas, Z. Fisk, P. Schlottmann, P. G.Pagliuso, N. O. Moreno, J. L. Sarrao and J. D. Thompson, Phys. Rev. Lett. **89**, 106402 (2002)
Johnpierre Paglione, T. A. Sayles, P.-C. Ho, J. R. Jeffries and M. B. Maple, Nature Physics **3**, 703 (2007)
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.